text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Formalising Confluence in PVS
Mauricio Ayala-Rincón
Departamentos de Ciência da Computação e Matemática
Universidade de Brası́lia, Brazil
[email protected]
Confluence is a critical property of computational systems which is related with determinism and
non ambiguity and thus with other relevant computational attributes of functional specifications and
rewriting system as termination and completion. Several criteria have been explored that guarantee
confluence and their formalisations provide further interesting information. This work discusses topics and presents personal positions and views related with the formalisation of confluence properties
in the Prototype Verification System PVS developed at our research group.
1
Introduction
Syntactic criteria such as avoiding overlapping of rules as well as linearity of rules have been used as a
discipline of functional programming which avoids ambiguity. In the context of term rewriting systems
(TRSs for short), well-known results such as Newman’s Lemma [9], Rosen’s Confluence of Orthogonal
term rewriting systems [11] as well as the famous Knut-Bendix(-Huet) Critical Pair Lemma [8, 7] are of
great theoretical and practical relevance. The first one, guarantees confluence of Noetherian and locally
confluent abstract reduction systems; the second one, assures confluence of orthogonal term rewriting
systems, that are systems that avoid ambiguities generated by overlapping of their rules and whose rules
do not allow repetitions of variables in their left-hand side (i.e., left-linear); and, the third one provides
local confluence of term rewriting systems whose critical pairs are joinable.
Formalisations in PVS of these confluence criteria provide valuable and precise data about the theory
of rewriting (cf. [6], [5], [10]). All mentioned specifications and formalisations are available either
at the local site http://trs.cic.unb.br or, as part of the NASA PVS libraries, in the theories for
abstract reduction systems ars and term rewriting systems trs that belong to the TRS development, at
http://shemesh.larc.nasa.gov/fm/ftp/larc/PVS-library/pvslib.html.
2
Background
It is assumed the reader is familiar with rewriting notations and notions as given in [3] and [4].
2.1 Abstract reduction systems
In the PVS development for TRSs, specifically in the theory ars, an abstract reduction system (for
short, ARS) is specified as a binary relation R over an uninterpreted type T , R VAR : PRED[[T,T]].
This choice facilitates the definition of associated necessary relations through the use of PVS operations
for relations such as reversal, subset, union, composition and an operator for iterative applications of
compositions. For instance,
• the inverse of the relation is specified as converse(R);
C.A. Muñoz and J. A. Pérez (Eds.) :
Developments in Computational Models
EPTCS 204, 2016, pp. 11–17, doi:10.4204/EPTCS.204.2
Formalising Confluence in PVS
12
• the symmetric closure, SC(R), as union(R, converse(R);
• the reflexive closure, RC(R), as union(R, =);
• the reflexive transitive closure, RTC(R), as IUnion(LAMBDA n:
iterate(R, n));
• the equivalence closure, EC(R), as RTC(SC(R)) etc.
Properties of ARSs are then specified in a very natural manner from these relational basis. For
instance, using PVS properties for relations such as well-foundedness, the property of noetherianity,
noetherian?(R) is specified as the predicate well founded?(converse(R)). Also, the property of
confluence, confluence?(R), is specified as
FORALL x, y, z: RTC(R)(x,y) & RTC(R)(x,z) => joinable?(R)(y,z)
where the predicate joinable? is specified as
joinable?(R)(x,y): bool = EXISTS z: RTC(R)(x,z) & RTC(R)(y, z).
More synthetic specifications might be possible; for instance, the elegant set-theoretical definition of
confluence, written in the usual rewriting notation as (∗ ← ◦ →∗ ) ⊆ (→∗ ◦ ∗ ←) can be specified straightforwardly as subset?(RTC(converse(R)) o RTC(R), RTC(R) o RTC(converse(R))), using relation composition, o, and the subset predicate, subset?.
ARS results were formalised using standard proof techniques as noetherian induction. Among other
results on confluence of ARSs, a description of the formalisation of Newman Lemma, specified below,
is available in [5]
Newman: LEMMA noetherian?(R) => (confluent?(R) <=> locally confluent?(R))
2.2 Term Rewriting Systems
Terms are specified as a data type built from variables over a nonempty uninterpreted type and a signature
of function symbols with their respective arities. The arguments of a functional term, headed by a
function symbol of the signature, are specified as a finite sequence of terms, of length equal to the arity
of the function symbol. Positions of a term t, written posOF(t), are finite sequences of naturals specified
recursively as in the standard way in the theory of rewriting. Thus, the necessary operations on positions
as their concatenation resumes to concatenation of finite sequences of naturals and so, predicates such as
disjunct or parallel positions, given by the predicate parallel? or for short ||, are specified as (NOT
p <= q) & (NOT q <= p), where <= is the sequence prefix relation built as <=(p, q): bool =
(EXISTS (p1: position): q = p o p1).
Using these types for terms and positions, it is possible to build the required algebraic properties
for terms, positions, subterms and replacement of terms. Namely, the subterm of a term t at a valid
position p, usually written as t p , is specified recursively navigating through the structure of the term
according to the naturals in the position sequence and the arguments of the functional terms inside
t: stOF(t: term, (p: positions?(t)). Also, the replacement of the subterm at a valid position p of a term s by another term t is built recursively: replaceTerm(s: term, t: term, (p:
positions?(s))), that will be abbreviated as s[t] p. These design decisions give rise to algebraic
properties that are easily formalised by inductive proofs on these data structures. Among other properties,
one has formalisations for:
• Preservation of positions after replacement of subterms either positions of replacement:
posOF(s)(p) => posOF(s[t] p)(p)
or parallel positions to the position of replacement:
posOF(s)(p) & posOF(s)(q) & p||q => posOF(s[t] p)(q)
M. Ayala-Rincón
13
• Extension of possible new valid positions at the position of replacement after a replacement (where
below o stands for the concatenation of sequences):
posOF(t)(q) & posOF(s)(p) => posOF(s[t] p)(p o q)
• Preservation of replaced terms:
posOF(s)(p) & posOF(t)(q) => stOF(s[t] p,p o q) = stOF(t, q)
• Associativity of replacement:
posOF(s)(p) & posOF(t)(q) => (s[t] p)[r] (p o q) = s[t[r] q] p
• Commutativity of replacement at parallel positions:
posOF(s)(p) & posOF(s)(q) & p||q => (s[t] p)[r] q = (s[r] q)[t] p
Rewriting rules are specified as pairs of terms (l, r) satisfying the usual conditions on rules, that is
the left-hand side (lhs) cannot be a variable and the variables occurring in the right-hand side (rhs) should
belong to the lhs of the rule:
rewrite rule?(l,r):
bool = (NOT vars?(l)) & subset?(Vars(r), Vars(l))
After that, it is possible to define the type of rewriting rules as
rewrite rule :
TYPE rewrite rule?
and then, a TRS is given as a set of rewriting rules set[rewrite rule].
Substitutions are built as objects of type [V -> term], where V is a countably infinite set of variables and such that their domain is finite, that is Sub?(sig): bool = is finite(Dom(sig)), where
Dom(sig): set[(V)] = {x: (V) | sig(x) /= x} . The type of substitutions is given as Sub:
TYPE = (Sub?). From this point, renaming, variants, composition of substitutions and homomorphic
extensions of substitutions, ext(sigma), are easily built as well as a series of necessary substitution
properties formalised.
With these elements of formal design it is possible to define the reduction relation from a set of
rewriting rules say E:
reduction?(E)(s,t): bool =
EXISTS ( (e | member(e, E)), sigma, (p: positions?(s))) :
stOF(s, p) = ext(sigma)(lhs(e)) &
t = s[ext(sigma)(rhs(e))]_p
Immediately, it is possible to prove that this relation is closed under substitutions and compatible
with contexts.
After that, a predicate for critical pairs of a TRS E is built, CP?(E), and then the most famous result
on confluence of TRSs, that is the Critical Pair Lemma is formalised as described in [6].
CP: THEOREM FORALL E: locally_confluent?(reduction?(E)) <=>
(FORALL s, t) : CP?(E)(s, t) => joinable?(reduction?(E))(s,t)
Since the reduction relation built from a set of rewriting rules inherits by parameterisation, all properties of ARSs in the ars development, it is possible to apply Newman’s Lemma in order to formally
infer confluence of a terminating TRS all whose critical pairs are joinable.
The design decisions taken in the specification of ARSs and TRSs were satisfactory to accomplish
one of our main objectives in this formalisation, that is indeed maintaining formal proofs as close as
14
Formalising Confluence in PVS
possible to the analytical proofs presented in textbooks. In fact, diagrammatic didactical artefacts (used
in papers and textbook) representing rewriting properties used in the proofs as commutation diagrams
for confluence, local-confluence etc, and those ones used for representing peaks valleys and overlap
situations in the analysis of the Critical Pair criterion can be also conducted when reasoning about our
formalisations.
In particular, the formalisation of the Critical Pair Lemma follows the textbook proof organisation
which is based on the analysis of the possible peaks when trying to obtain local confluence. These peaks
can be originated from simultaneous reductions at parallel positions, which are trivially joinable, or from
reductions at nested positions, which give rise to either critical or non-critical overlaps.
A peak, from a critical overlap can be easily verified to join, by proving that it corresponds to an
instance of a critical pair and using the assumption that critical pairs are joinable.
The non-critical overlaps are the interesting ones. A such peak is originated by application of rules
l → r and g → d with substitution σ , assuming these rules have not common variables which is possible
by renaming of rules. Supposing l σ occurs in the dominating position of the overlap, one can focus on
the analysis of the joinability of the peak rσ ← l σ → l σ [d σ ] p◦q , where p is a variable position in l, say
l p = x, and q the position in xσ in which gσ occurs.
Thus, all that is solved by the careful construction of a new substitution, σ ′ such that it modifies
σ mapping x 7→ xσ [d σ ]q and maintains the images of all other variables in the domain of σ as those
mapped by σ .
In the sequel, by a like “uniform reduction sequence” one has that rσ →∗ rσ ′ and l σ [d σ ] p◦q →∗ l σ ′ ;
the former is done as rσ → rσ [d σ ]q1 ◦q → · · · → rσ [d σ ]q1 ◦q · · · [d σ ]qn ◦q = rσ ′ , where {q1 , . . . , qn } is the
set of positions of r in which x occurs, and the latter is done as l σ [d σ ] p◦q → l σ [d σ ] p◦q [d σ ] p1 ◦q → · · · →
l σ [d σ ] p◦q [d σ ] p1 ◦q · · · [d σ ] pm ◦q = l σ ′ , where {p} ∪ {p1 , . . . , pn } is the set of positions of l in which x
occurs.
So joinability is concluded by the application of rule l → r with substitution σ ′ . See for instance the
proof in Chapters 6 or 2 of respectively [3] or [4].
The formalisation, under the design choices previously mentioned, requires the construction of elements that guarantee specialised properties, such as the instantiation of a critical pair, built from the
rewriting rules, that corresponds to a critical overlap as well as the substitution σ ′ for a non-critical overlap. For the latter, it is necessary an inductive proof (using auxiliary lemmas) on the cardinality of the
sets of positions {q1 , . . . , qn } and {p1 , . . . , pm } for concluding that l σ and rσ rewrite into l σ ′ and rσ ′ ,
respectively.
3
Formalising the algebra of parallel rewriting and orthogonality
Rosen’s confluence of orthogonal TRSs [11] is a challenging formalisation. The classical proof is based
on the Parallel Moves lemma: essentially, what is necessary is to prove that under the hypothesis of
orthogonality, the associated parallel reduction relation holds the diamond property.
Intuitively, the analytical proof requires only the comprehension of properties for the notion of the
parallel reduction relation, but the intuition of parallel rewriting is usually explained through the like
“uniform reduction” as in the analysis of the Critical Pair criterion, which in fact refers to sequential
rewriting. So, any formalisation following the classical approach would require an explicit construction of such that parallel relation as well as the specialised description and formalisation of its specific
algebraic properties.
The notion of parallel reduction depends on sets of triplets of valid positions, rules and substitutions,
M. Ayala-Rincón
15
positions in which sequential replacements are simultaneously applied according to the instantiation
of the rules with the associated substitutions. Because of this dependence, two design approaches are
possible: either using sets of triplets of positions, rules and substitutions or sets of (finite) and coordinated
sequences of positions Π, rules Γ and substitutions Σ. We opted by the last design alternative since we
believe it is closer to implementations in programming languages and also because PVS offers libraries
with translations (and their necessary formalised properties) between data structures such as sets, lists
and finite sequences.
The parallel rewriting reduction relation built from a set of rewriting rules E, that in classical notation
is written as s ⇒ t, is specified as the relation parallel reduction?(E)(s,t) below, using a parallel
replacement operator, replace par pos, that is recursively specified from the sequential replaceTerm
operator, and through an auxiliary relation parallel reduction fix?(E).
parallel_reduction_fix?(E)(s,t, (fsp: SPP(s))): bool =
EXISTS ((fse | member(fse, E)), fss) :
fsp‘length = fse‘length AND fsp‘length = fss‘length
AND subtermsOF(s,fsp) = sigma_lhs(fss, fse)
AND t = replace_par_pos(s, fsp, sigma_rhs(fss, fse))
parallel_reduction?(E)(s,t): bool =
EXISTS (fsp: SPP(s)): parallel_reduction_fix?(E)(s,t,fsp)
fsp, fse and fss correspond to the sequences Π = [p1 , . . . , pn ], Γ = [(l1 , r1 ), . . . , (ln , rn )] and Σ =
[σ1 , . . . , σn ] of positions, rewriting rules and substitutions used in the parallel rewriting. fsp is a sequence of parallel positions of s obtained by its type dependency, that is SSP(s). fse is a sequence
of equations in the rewriting system given by member(fse, E), and fss is a sequence of substitutions. The required coordination of triplets of associated positions, rules and substitutions is directly
obtained by using the corresponding indexation in the respective sequences, that is the same index. The
condition subtermsOF(s,fsp) = sigma lhs(fss, fse) equals the condition that for all valid index
i of these sequences, s pi = li σi and the condition t= replace par pos(s, fsp, sigma rhs(fss,
fse)) equals t to the desired parallel contractum, that is s replacing the li σi ’s by the ri σi ’s.
In a parallel peak, say t ⇔ s ⇒ u, using positions, equations and substitutions say (Πk , Γk , Σk ) for
k = 1, 2, as in the case of the Critical Pair Lemma, the interesting cases are those of non-critical overlaps.
Without loss of generality, at some position q ∈ Π1 one has all positions p1 , ..., pn in Π2 below q. On
the one side, sq = l σ and tq = rσ , where (l, r) is the rewriting rule in Γ1 and σ the substitution in Γ1
associated with the position q in Π1 . On the other side, one has sq ⇒ uq by parallel reduction at positions
p1 , . . . pn below q accordingly to the same equations and substitutions in the triplet (Π2 , Γ2 , Σ2 ). Let
(Π′ , Γ′ , Σ′ ) denote the triplet of this last parallel reduction step, then the parallel peak rσ ⇔ l σ ⇒ uq is
an instance of the Parallel Moves Lemma. See details in Chapter 4.3 of [4] or 6.4 of [3], for instance.
Since in such a parallel peak rσ ⇔ l σ ⇒ t all overlaps are non critical, one should prove that parallel
reducing each σ -instance of a variable in l, say x at position p one has l σ p = xσ ⇒ xσ ′ = t p , where the
substitution σ ′ is built by reducing in parallel all occurrences of σ -instantiated variables in l σ uniformly,
which is possible by left-linearity assumption. Thus, rσ ⇒ rσ ′ and t ⇒ l σ ′ , that allows concluding the
joinability of the peak.
Despite in the theory the adaptation of sequential properties for parallel replacement might be intuitively clear, in the PVS development the necessary specialised algebraic properties should be formalised.
Let s[T] Π denote the parallel replacement of terms in the sequence of terms T at valid parallel positions
Π. A few of those properties are included below.
Formalising Confluence in PVS
16
• Preservation of positions after replacement of subterms either positions of replacement:
posOF(s)(Π) => posOF(s[T] Π)(Π)
or parallel positions to the position of replacement:
posOF(s)(Π) & posOF(s)(Π′) & Π′ ||Π => posOF(s[t] p)(Π′ )
• Invariance under composition of parallel replacement at parallel sequences of positions:
posOF(s)(Π1) & posOF(s)(Π2) & Π1 ||Π2 =>
(s[T 1] Π1 )[T 2] Π2 = (s[T 1 o T 2] Π1 ◦ Π2
• Commutativity of parallel replacement at parallel sequences of positions:
posOF(s)(Π1) & posOF(s)(Π2) & Π1 ||Π2 =>
(s[T 1] Π1 )[T 2] Π2 = (s[T 2] Π2 )[T 1] Π1
The formalisation of confluence of orthogonal TRS proceeds by inductive proof techniques taking
care of the specificities of the algebra of parallel positions, replacement and rewriting.
4
Conclusions and future work
The development of these formalisations on confluence of TRSs brought out several lessons.
From the theoretical point of view, the main observation is that despite the intuitive notion of ”uniform reduction”, that is used to provide intuition about the proof of the Parallel Moves Lemma, induces
to believe that the extension is obvious, a specialised development of the theory of parallel reduction
and its algebraic properties is necessary. And extending sequential to parallel rewriting formalisations is
not trivial. A preliminary thoughtful analysis would be always necessary in order to estimate accurately
the real complexity and the necessary effort of formalisations in any context, mostly when the proposed
development appears to be a simple extension of those yet available. The second lesson has to do with
the investment of enough time for fine tuning design decisions since they influence the effort required in
proofs.
The main lesson is that through this kind of exercise, our comprehension of the theory becomes more
refined, providing a better support for the formal analysis of further related developments, as well as a
more realistic insight about the possible adaptation or reuse of previous specifications and proofs. Developments in progress include use of such techniques in other contexts as the nominal syntax approach
of rewriting (cf. [1] [2]).
References
[1] M. Ayala-Rincón, M. Fernández, M. J. Gabbay & A. C. Rocha Oliveira (2015): Checking
Available at
Overlaps of Nominal Rewriting Rules.
In: Pre-proc. LSFA, pp. 199–2014.
https://www.mat.ufrn.br/~LSFA2015/preproceedings.pdf.
[2] M. Ayala-Rincón, M. Fernández & A. C. Rocha Oliveira (2015):
Completeness in PVS
Pre-proc. LSFA, pp. 19–34.
Available at
of a Nominal Unification Algorithm.
In:
https://www.mat.ufrn.br/~LSFA2015/preproceedings.pdf.
[3] F. Baader & T. Nipkow (1998):
doi:10.1017/CBO9781139172752.
Term Rewriting and All That.
Cambridge University Press,
[4] M. Bezem, J.W. Klop & R. de Vrijer, editors (2003): Term Rewriting Systems by TeReSe. Cambridge Tracts
in Theoretical Computer Science 55, Cambridge University Press.
M. Ayala-Rincón
17
[5] A. L. Galdino & M. Ayala-Rincón (2008): A Formalization of Newman’s and Yokouchi Lemmas in a HigherOrder Language. Journal of Formalized Reasoning 1(1), pp. 39–50, doi:10.6092/issn.1972-5787/1347.
[6] A. L. Galdino & M. Ayala-Rincón (2010): A Formalization of the Knuth-Bendix(-Huet) Critical Pair Theorem. J. of Automated Reasoning 45(3), pp. 301–325, doi:10.1007/s10817-010-9165-2.
[7] G. Huet (1981): A complete proof of correctness of the Knuth-Bendix completion algorithm. Journal of
Computer and Systems Sciences 23(1), pp. 11–21, doi:10.1016/0022-0000(81)90002-7.
[8] D. E. Knuth & P. B. Bendix (1970): Computational Problems in Abstract Algebra, chapter Simple
Words Problems in Universal Algebras, pp. 263–297. J. Leech, ed. Pergamon Press, Oxford, U. K.,
doi:10.1016/B978-0-08-012975-4.50028-X.
[9] M. H. A. Newman (1942): On theories with a combinatorial definition of “equivalence”. Ann. of Math.
43(2), pp. 223–243, doi:10.2307/1968867.
[10] A. C. Rocha Oliveira & M. Ayala-Rincón (2013): Formalizing the Confluence of Orthogonal Rewriting Systems.
CoRR abs/1303.7335, doi:10.4204/EPTCS.113.14.
Available at
http://arxiv.org/abs/1303.7335.
[11] B. K. Rosen (1973): Tree-Manipulating Systems and Church-Rosser Theorems. J. of the ACM 20(1), pp.
160–187, doi:10.1145/321738.321750.
| 6cs.PL
|
Distance-based Self-Attention Network for Natural Language Inference
Jinbae Im and Sungzoon Cho
Department of Industrial Engineering
Seoul National University
Seoul, South Korea
[email protected], [email protected]
Abstract
arXiv:1712.02047v1 [cs.CL] 6 Dec 2017
Attention mechanism has been used as an
ancillary means to help RNN or CNN.
However, the Transformer (Vaswani et al.,
2017) recently recorded the state-of-theart performance in machine translation
with a dramatic reduction in training
time by solely using attention. Motivated by the Transformer, Directional Self
Attention Network (Shen et al., 2017),
a fully attention-based sentence encoder,
was proposed. It showed good performance with various data by using forward
and backward directional information in
a sentence. But in their study, not considered at all was the distance between
words, an important feature when learning the local dependency to help understand the context of input text. We propose Distance-based Self-Attention Network, which considers the word distance
by using a simple distance mask in order to
model the local dependency without losing the ability of modeling global dependency which attention has inherent. Our
model shows good performance with NLI
data, and it records the new state-of-the-art
result with SNLI data. Additionally, we
show that our model has a strength in long
sentences or documents.
1
Introduction
Sequence modeling has been employing Recurrent
Neural Networks (RNN) or Convolutional Neural
Networks (CNN) mostly. More recently, models
incorporating attention mechanisms have shown
good performance in machine translation (Bahdanau et al., 2014; Sutskever et al., 2014), Natural
Language Inference (NLI) (Liu et al., 2016), and
Question Answering (QA) (Hermann et al., 2015;
Sukhbaatar et al., 2015) etc. Attention mechanisms used to be exploited in conjunction with
RNN or CNN as an ancillary means to help improve performance. Lately, Vaswani et al. (2017)
presented the first fully attention-based model,
which recorded the state-of-the-art result in machine translation. As a fully attention-based model
can consider all words in a sentence at once, parallelization leads to great reduction in training time.
Motivated by Vaswani et al. (2017), Shen et al.
(2017) proposed the first fully attention-based
sentence encoder. Shen et al. (2017) recorded
good performance in a variety of tasks. In particular, they recorded the state-of-the-art result
with Stanford Natural Language Inference (SNLI)
dataset (Bowman et al., 2015) which is a representative dataset of NLI. The NLI task aims to
classify the relationship between two sentences
as entailment, contradiction, or neutral. One
of the approaches to solving the NLI task is
to use sentence-encoding based models.∗ Shen
et al. (2017) presented a sentence-encoding based
model reflecting directional information in a sentence. However, the distance between words was
not considered at all in their model, and the directional information simply involved words before and after the reference word. Altogether, positional information of words was not fully taken
into account. As a result, the difference of importance between the distant words and the nearby
words was not appropriately reflected. Hence lo∗
The NLI task can be solved through two different approaches: sentence encoding-based models and joint models.
The former separately encode each sentence, whereas the latter take into account the direct relationship between two sentences. Between them, sentence-encoding based models focus on training sentence encoder that can represent sentences
in vector form well. We focus on the former approach, since
the objective of our work is to develop an advanced sentenceencoding model.
cal dependency was not properly modeled, which
in turn failed to capture the context information in
long sentences.
To tackle this limitation, we propose Distancebased Self-Attention Network which introduces a
distance mask which models the relative distance
between words. In conjunction with a directional
mask, the distance mask allows us to incorporate
complete positional information of words in our
model. Our Distance-based Self-Attention Network achieved good performance with NLI data,
and recorded the state-of-the-art result with SNLI.
Our model worked exceptionally well with long
sentences, in particular. We also visualized the effect of the distance mask to show that our model
can grasp both local dependency and global dependency.
2
Related Works
NLI tasks have been studied through models of
various structures. Most of all, models combining
attention with Long Short-Term Memory (LSTM)
have performed well. Liu et al. (2016) improved
the performance by adding the mean pooling vector to the conventional attention model in which
attention is applied to hidden states of LSTM.
Chen et al. (2017) used the input gates of the
LSTM as attention weights to simplify the model
structure. In Chen et al. (2017) and Ni and Bansal
(2017), short-cut connections in stacked LSTM,
in combination with max-pooling originally suggested by Conneau et al. (2017), were proven effective in improving performance, recording the
state-of-the-art performance in MultiNLI. And
Munkhdalai and Yu (2016a) used the memory for
sentence encoding motivated by Neural Turing
Machine (Graves et al., 2014).
Vaswani et al. (2017) was the first study to construct an end-to-end model with attention alone,
and recorded the state-of-the-art performance in
machine translation tasks. Vaswani et al. (2017)’s
encoder-decoder framework consists of a multihead attention and a position-wise feed forward
network as a basic building block which is deeply
stacked combined with residual connection. The
multi-head attention projects the input sentences
to multiple subspaces and then computes the
scaled dot-product attention in each subspace. The
results in each subspace are then concatenated and
projected again. Position-wise feed forward network adds non-linearity to vector representations
of each position. In this way, the fully attentionbased model was constructed without using RNN
or CNN, and the training cost was greatly reduced.
Shen et al. (2017), a very recent work, constructed a fully attention-based sentence encoder
motivated by Vaswani et al. (2017). They proposed a multi-dimensional attention mechanism
that computes the attention by each dimension
through modification of additive attention. In addition, their model exploits directional attention
as well as fusion gate motivated by bi-directional
LSTM. Directional information was reflected by
introducing a simple directional mask. By adding
a directional mask to the logit of attention, words
in a specific direction in the sentence were masked
to avoid attention. The extent to which attention
results are ultimately reflected was determined
through fusion gate. In our study, we construct
our model based on Vaswani et al. (2017)’s basic
building block, as well as Shen et al. (2017)’s key
model structures. In order to model the distance
between words, which was not considered in their
works, we transform the multi-head attention in
Vaswani et al. (2017), in particular, to fit our objective. Details can be found in section 4.
3
Background
In Vaswani et al. (2017), the attention function
is defined as follows by introducing the concept
of query, key, and value. “An attention function
can be described as mapping a query and a set
of key-value pairs to an output, where the query,
keys, values, and output are all vectors. The output is computed as a weighted sum of the values,
where the weight assigned to each value is computed by a compatibility function of the query with
the corresponding key (Vaswani et al., 2017).” The
two most commonly used attentions are additive
attention (Bahdanau et al., 2014; Shang et al.,
2015) and dot-product attention (Kim et al., 2016;
Sukhbaatar et al., 2015; Vaswani et al., 2017).
3.1
Additive Attention
Let query, ith key, and ith value be q, ki , and vi
respectively. (q ∈ Rdk , ki ∈ Rdk , and vi ∈ Rdv )
Compatibility function of the query with the ith
key is represented by the following equation 1.
f (q, ki ) = li = uT σ(q + ki ),
(1)
where u ∈ Rdk , and σ(·) is an activation function
usually chosen as tanh.
And attention weight assigned to each ith value
is computed by applying the softmax function to
li and final output is weighted sum of value as following equations.
exp(li )
j=1 exp(lj )
wi = P
Output =
X
wi v i
(2)
(3)
i=1
3.2
Dot-product Attention
Dot-product attention is the same as additive attention except for compatibility function. In dotproduct attention, compatibility function is computed by the following equation 4 in place of the
equation 1.
f (q, ki ) = li = h q, ki i
(4)
On implementation, dot-product attention is
much faster and more space-efficient than additive
attention due to optimized matrix multiplication.
In practice, however, additive attention outperforms dot product attention for large values of dk .
So Vaswani et al. (2017) used scaled dot-product
attention instead of normal dot-product attention
to prevent performance loss in large dimension as
following equation 5.
h q, ki i
f (q, ki ) = li = √
dk
4
4.1
Figure 1: Overall architecture
4.2
Sentence Encoder
The sentence encoder structure proposed in this
paper is shown in Figure 2. The term “Norm”
in Figure 2 stands for layer normalization. The
sentence encoder of Figure 2 encodes the premise
and hypothesis in a vector form. We describe each
component of our sentence encoder in detail in the
following subsections.
(5)
Proposed Model
Overall Architecture
Our model’s overall architecture is shown in Figure 1. We follow the conventional architecture for
training NLI data. First, the two input sentences,
premise and hypothesis, are encoded as vectors, u
and v respectively, through identical sentence encoders. For the encoded vectors u and v, the representation of relation between the two vectors is
generated by the concatenation of u, v, |u − v|,
and u ∗ v. Thereafter, a probability for each of the
3-class is generated through the 300D ReLU layer
and the 3-way softmax output layer. We configured the model with the setting of 1layer 300D as
in Shen et al. (2017) to focus on the performance
evaluation of the sentence encoder itself. Layer
normalization (Ba et al., 2016) and dropout are applied to 300D ReLU layer.
Figure 2: Sentence encoder
4.2.1
Word Embedding Layer
Let an input sentence be a sequence of discrete
words x = [x1 , x2 , · · ·, xn ], where xi ∈ RN is
a one-hot representation of the word i, and N is
the vocabulary size. These one-hot representations
are transformed into dense representations by us-
ing the pre-trained word embedding.
Let We ∈ Rde ×N be a pre-trained word embedding matrix. Then a sequence of dense word
representations can be written as w = We x =
[w1 , w2 , · · ·, wn ], where wi ∈ Rde is dense representation of the word i.
(a) Forward mask
4.2.2 Masked Multi-Head Attention
The masked multi-head attention is a variation
of the multi-head attention employed by Vaswani
et al. (2017). The scaled dot-product attention of
Vaswani et al. (2017) is expressed as following:
QK T
Attention(Q, K, V ) = softmax( √ )V
dk
(6)
where Q, K, V are matrices composed of a set of
queries, keys, and values, respectively.
We transform equation 6 and express the
masked attention as following:
Masked(Q, K, V )
(7)
QK T
= softmax( √
+ Mdir + αMdis )V
dk
Here, Mdir ∈ Rn×n is the directional mask
as proposed in Shen et al. (2017), while Mdis ∈
Rn×n is the distance mask proposed in this model.
Hyper parameter α is the distance-alpha tuned
through validation data.
Mdir consists of the forward mask and backward mask as explained in Figure 3. In the Forward Masked Multi-Head Attention phase, the
forward mask is selected, and in the Backward
Masked Multi-Head Attention phase, the backward mask. The forward masks prevent words that
appear after a given word from being considered in
the attention process, while backward masks prevent words that appear before from consideration
by adding −∞ to the logits before taking the softmax at the attention phase. The diagonal component of Mdir is also set to −∞ so that each token
does not consider itself to attention, and the information of each token is later transmitted through
the fusion gate of section 4.2.3
Mdis is shown in the Figure 4. The (i, j) component of the distance mask is −|i − j|, representing the distance between (i + 1)th word and
(j + 1)th word multiplied by −1. By multiplying
this value by α and adding it to logit, the attention weight becomes smaller as distance increases.
(b) Backward mask
Figure 3: Directional mask
That is, the distance mask serves to concentrate on
the local words around the reference word. Such
a structure may appear similar to a CNN filter extracting a local feature. Yet, the big difference is
that CNN only uses information in the window
size, whereas our model considers all words in a
sentence at once, concentrating on the local words
by taking account of the relative distance between
words.
By using the distance mask, the distance between words, not considered through the directional mask of Shen et al. (2017), was considered
additionally, so the complete positional information of words was taken into consideration.∗
Figure 4: Distance mask
The masked multi-head attention can be expressed as following:
Masked Multi-Head(Q, K, V )
= concat(head1 , · · · , headh )W O
(8)
where headi = Masked(QWiQ , KWiK , V WiV ),
with h as the number of heads, WiQ , WiK , WiV ∈
∗
In Vaswani et al. (2017), the positional information of
the word was used through positional encoding. By adding
the positional encoding vector to the word embedding vector,
the embedding changed according to the absolute position of
the word in the sentence. However, in sentence modeling, the
relative position with respect to the other words is important,
not the absolute position of the word. In other words, what
words are placed in order before and after the word is important, not the absolute position of the word in a sentence.
Therefore, we take the relative position directly into account
in our model through the distance mask instead of the positional encoding which considers the relative position indirectly.
Rde ×de /h , and W O ∈ Rde ×de . Q, K, V ∈ Rn×de
are matrices created from n word embedding vectors of sentences and expressed as equation 9.
− w1 −
− w2 −
Q=K=V =
..
.
− wn −
(9)
The masked multi-head attention first projects
Q, K, V into h subspaces, respectively, and performs masked attention of equation 7 for each
Q, K, V projection combination. The h attention
result is concatenated before projection.∗
4.2.3 Fusion Gate
At the fusion gate, raw word embedding S ∈
Rn×de and the result of masked multi-head attention H ∈ Rn×de in equation 10 are used as input.
− w1 −
− h1 −
− w2 −
− h2 −
S=
H =
..
..
.
.
− wn −
− hn −
(10)
First, we generate S F , H F by projecting S, H
using W S , W H ∈ Rde ×de . Mathematically:
S F = SW S
(11)
H F = HW H
Then create gate F as shown in equation 12
where bF ∈ Rde .
4.2.4 Position-wise Feed Forward Networks
We used position-wise feed forward network
structure of Vaswani et al. (2017) as it is. The
position-wise feed forward network employs the
same fully connected network to each position of
sentence, in which the fully connected layer consists of two linear transformations, with the ReLU
activation in between. Mathematically:
FFN(x) = max(0, xW1P + bP1 )W2P + bP2 (13)
where x ∈ R1×de , W1P ∈ Rde ×df f , W2P ∈
Rdf f ×de , bP1 ∈ Rdf f , and bP2 ∈ Rde .
The FFN function of the above equation 13 is
applied to each position of the result of the fusion
gate. Note that position-wise feed forward network is combined with the residual connection as
shown in Figure 2. That is, FFN learns the residuals. In our model, df f was set to 4de .
4.2.5 Pooling Layer
The vector representation of input sentence is generated through the pooling layer after the concatenation of the results of forward directional
self attention and backward directional self attention. That is, the input of pooling layer is U =
[U f w ; U bw ] ∈ Rn×2de where each directional self
attention output is U f w ∈ Rn×de , U bw ∈ Rn×de .
We use the multi-dimensional source2token
self-attention of Shen et al. (2017) for our multidimensional self-attention.
For ith row vector of U , ui , logit l(ui ) is computed as following:
M
M
l(ui ) = ELU (ui W1M + bM
1 )W2 + b2
Gate(S, H) = F
S F + (1 − F )
F
F
HF
F
where F = sigmoid(S + H + b )
(12)
Finally, we obtain the gated sum by using F . It
is common in many papers including Shen et al.
(2017) to use raw S and H in gated sum. We,
however, use the gated sum of S F and H F which
resulted in a significant increase in accuracy.
∗
Multi-head attention (Vaswani et al., 2017) is fast and
efficient because it is based on dot-product attention. However, multi-dimensional attention (Shen et al., 2017) has a disadvantage in that it consumes a lot of gpu memory because
it requires several 4-dimensional tensors on implementation.
So, in our model, the multi-head attention was used as a base
structure instead of the multi-dimensional attention. In addition, the performance of the actual implementation was also
better with multi-head attention.
(14)
where ui = Ui∗ ∈ R1×2de , W1M , W2M ∈
M
2de .
R2de ×2de , and bM
1 , b2 ∈ R
The calculations of logit consist of two linear transformations, with the Exponential Linear
Units (ELU) activation function (Clevert et al.,
2015) in between. Multi-dimensional attention
differs from general attention in that the logit for
an input vector is not a scalar but a vector with dimensions equal to the dimensions of the input vector. This allows each dimension of the input vector
to have a scalar logit, and we can perform attention
to n word tokens in each dimension, as illustrated
below by equation 15, 16. Note that softmax is
performed on the row dimension of L, not the column dimension.
M = softmax(L) U
− l(u1 ) −
− l(u2 ) −
where L =
..
.
− l(un ) −
(15)
We set h = 5, α = 1.5 in the masked multi-head
attention, and the dropout probability was set to
0.1. Batch size was 64, and the model was trained
with Adam (Kingma et al., 2014) optimizer, with
a learning rate of 0.001. All models were implemented via Tensorflow on single Nvidia Geforce
GTX 1080Ti GPU.
5.3
Multi-dimensional(U ) =
n
X
Mi∗
(16)
i=1
The 2de -dimensional output vector of multidimensional attention and the 2de -dimensional
vector obtained by applying max pooling to U
are concatenated to encode the input sentence as
a 4de -dimensional vector.
5
Experiments and Results
5.1
Dataset
The dataset used in the experiments are SNLI
(Bowman et al., 2015) and MultiNLI (Williams
et al., 2017) datasets. The SNLI dataset consists of 549,367 / 9,842 / 9,824 (train / valid
/ test) premise and hypothesis pairs; and the
MultiNLI dataset, 392,702 / 9,815 / 9,832 / 9,796
/ 9,847 (train / valid matched / valid mismatched
/ test matched / test mismatched) sentence pairs.
The two datasets have the same format, but sentences in the MultiNLI dataset are much longer
than those in SNLI dataset. In addition, MultiNLI
dataset consists of various genre information. If
genres included in the train data are also found
in valid (test) data, then the dataset is called
“matched”; if valid (test) data includes genres that
are not in the train data, then the dataset is called
“mismatched”.
5.2
Experimental results of SNLI data compared with
the existing models on the SNLI leader-board2
are shown in Table 1. Compared with the existing state-of-the-art model (Shen et al., 2017), the
number of parameters and the training time increased, but our results show the new state-of-theart record. We also looked at the model with distance mask removed to verify the effect of the distance mask proposed in this paper. Results show
that the addition of the distance mask improved
the performance without significantly affecting the
training time or increasing the number of parameters.
Figure 5: SNLI average sentence length
Training Details
We used the Glove 840B 300D1 (de = 300) for
the pre-trained word embedding without any finetuning. This is to train the more universally usable
sentence encoder.
Layer normalization (Ba et al., 2016) was applied to all linear projections of masked multihead attention, fusion gate, and multi-dimensional
attention. We applied residual dropout as used in
Vaswani et al. (2017), with dropout to the output
of masked multi-head attention and S F +H F +bF
of fusion gate.
1
SNLI Results
https://nlp.stanford.edu/projects/glove/
Figure 6: With distance mask vs. Without distance mask. Change of test accuracy on SNLI
data w.r.t average length of sentence pair.
2
https://nlp.stanford.edu/projects/snli/
Model Name
Feature-based models
Unlexicalized features (Bowman et al., 2015)
+Unigram and bigram features (Bowman et al., 2015)
Sentence encoding-based models
100D LSTM encoders (Bowman et al., 2015)
300D LSTM encoders (Bowman et al., 2016)
1024D GRU encoders (Vendrov et al., 2015)
300D Tree-based CNN encoders (Mou et al., 2015)
300D SPINN-PI encoders (Bowman et al., 2016)
600D Bi-LSTM encoders (Liu et al., 2016)
300D NTI-SLSTM-LSTM encoders (Munkhdalai and Yu, 2016b)
600D Bi-LSTM encoders+intra-attention (Liu et al., 2016)
300D NSE encoders (Munkhdalai and Yu, 2016a)
600D Deep Gated Attn. BiLSTM encoders (Chen et al., 2017)
600D Directional Self-Attention Network (Shen et al., 2017)
Our self-attention network (without distance mask)
Our Distance-based Self-Attention Network
| θ|
220k
3.0m
15m
3.5m
3.7m
2.0m
4.0m
2.8m
3.0m
11.6m
2.4m
4.7m
4.7m
T(s)/epoch
587
687
693
Train Acc(%)
Test Acc(%)
49.4
99.7
50.4
78.2
84.8
83.9
98.8
83.3
89.2
86.4
82.5
84.5
86.2
90.5
91.1
88.1
89.6
77.6
80.6
81.4
82.1
83.2
83.3
83.4
84.2
84.6
85.5
85.6
86.0
86.3
Table 1: Experimental results of different models on SNLI data. | θ| : number of parameters (excluding word embedding part). T(s)/epoch : average training time (second) per epoch.
The improvement of the test accuracy by introducing the distance mask is only by 0.3% point,
potentially because SNLI data mostly consist of
short sentences. Hence, we additionally examined
how the effect of the distance mask changes as the
average length of the two sentences of premise and
hypothesis pair changes. The distribution of the
average length of the two sentences of the SNLI
test data is shown in Figure 5, and the effect of
the distance mask according to the average length
change can be seen from Figure 6. Figure 6 shows
that the accuracy is similar until the average length
is less than 25, yet the test accuracy of the model
without the distance mask deteriorates drastically
for data of an average length exceeding 25. This
demonstrates that the distance mask has an advantage with long sentences or documents.
5.4
MultiNLI Results
The results of applying SNLI best model to
MultiNLI dataset without additional parameter
tuning are presented in Table 2. Note that
matched-test accuracy and mismatched-test accuracy were obtained by submitting our test results
to Kaggle open evaluation platforms: MultiNLI
Matched Open Evaluation3 and MultiNLI Mismatched Open Evaluation4 . First, the average
test accuracy difference is greater than 2% when
compared to the Directional Self-Attention Net3
4
https://www.kaggle.com/c/
multinli-matched-open-evaluation
https://www.kaggle.com/c/
multinli-mismatched-open-evaluation
work (Shen et al., 2017). This once again confirms
our model’s advantage in long sentences, given
that the sentence is much longer in MultiNLI.
Compared with the result of RepEVAL
2017 (Nangia et al., 2017), we can see that the
Distance-based Self-Attention Network performs
well. When compared with the model of Chen
et al. (2017), our model showed similar average
test accuracy with much lower number of parameters. Also, considering that the model of
Chen et al. (2017) is a complex LSTM model, our
model has an advantage in training time as a fully
attention-based model.
Ni and Bansal (2017) showed the best performance with 74.5% accuracy in Matched Test.
However, it is a very deep structured LSTM model
with 140.2m parameters. In our model, the inference layer is simply composed of 1 layer of
300D in order to focus on the training of sentence encoder. Both in Chen et al. (2017) and
Ni and Bansal (2017) models, the inference layer
was set very complex in order to improve the
MultiNLI accuracy. Taking this into consideration, it can be seen that our Distance-based SelfAttention Network performs competitively given
its simpler structure.
5.5
Case Study
A case study was conducted to investigate the
role of each structure of the Distance-based SelfAttention Network. For this, a sentence “A lady
stands outside of a Mexican market.” is picked
Model Name
Baseline
CBOW (Williams et al., 2017)
BiLSTM (Williams et al., 2017)
RepEval 2017 (Nangia et al., 2017)
Cha-level Intra-attention BiLSTM encoders (Yang et al., 2017)
BiLSTM + enhanced embedding + max pooling (Vu et al., 2017)
BiLSTM + Inner-attention (Balazs et al., 2017)
Deep Gated Attn. BiLSTM encoders (Chen et al., 2017)
Shortcut-Stacked BiLSTM encoders (Ni and Bansal, 2017)
Fully attention-based models
Directional Self-Attention Network (Shen et al., 2017)
Our Distance-based Self-Attention Network
SNLI Mix
| θ|
Matched Test Acc(%)
Mismatched Test Acc(%)
O
O
66.2
67.5
64.6
67.1
O
X
O
X
O
11.6m
140.2m
67.9
70.7
72.1
73.5
74.5
68.2
70.8
72.1
73.6
73.5
X
X
2.4m
4.7m
71.0
74.1
71.4
72.9
Table 2: Experimental results of different models on MultiNLI data. SNLI Mix : use of SNLI
training dataset. | θ| : number of parameters (excluding word embedding part).
among the premise sentences of SNLI test data.
We focused on training encoders that can represent
each sentence in a vector form well. Therefore, a
case study was conducted on a single sentence, not
a sentence pair.
Masked Multi-Head Attention We first look at
the attention weights in masked multi-head attention. Attention weights represent a n by n matrix
T
√
corresponding to softmax( QK
+ Mdir + αMdis )
dk
of equation 7, which is different for each head.
Here we look at the average attention weights obtained by averaging the attention weights of each
head. The attention weights for each head can be
found in Appendix.
(a) Forward
(b) Backward
(c) Forward
(d) Backward
Figure 8: Masked multi-head average attention
weights : without/with distance mask. (a), (b)
: without distance mask. (c), (d) : with distance
mask
(a) Forward
(b) Backward
Figure 7: Masked multi-head average attention
weights
The row of the matrix of Figure 7 represents
each word of the sentence, and the column represents the attention weights for each word at each
row. It can be seen that the attention weights are
heavier to the nearby words as compared to those
distant from the reference word. At the same time,
‘outside’ in the forward mask and ‘Mexican’ in
the backward mask have high attention weights for
several words. From this, it can be seen that important word is considered in the attention process.
Distance Mask We compared the masked
multi-head average attention weights for the
longest sentence example in the SNLI test data,
with length of 57 words to further verify the effect
of the distance mask. Panels (a) and (b) of Figure 8 show results without considering distance,
while (c) and (d) show the results with the distance
mask. In panels (a) and (b), very distant words are
considered in the attention and the overall attention weights were reduced. This implies that each
word does not focus on the important words in the
attention process, but rather takes into account almost every word, resulting in noisier figures.
However, in panels (c) and (d), the neighboring words are seen more intensively, which im-
plies that the local dependency has been well captured by our model. In addition, as shown in panel
(c), even if the word is far apart, it is still considered in the attention process if it is important.
This demonstrates the effectiveness of the distance
mask to identify local dependencies without losing
the ability to grasp the global dependency.
Fusion Gate We visualize the role of the fusion
gate F ∈ Rn×de at forward directional self attention. Figure 9 represents the average gate value
that averages de -dimensional gate value for each
word. If look at the results of both extremes, keyword ‘Mexican’ has a low gate value, resulting in
an output that greatly reflects the multi-head attention result. In contrast, ‘of’, ‘.’, the words of
little importance, have large gate values, which indicates that the original word embedding is greatly
reflected, not the multi-head attention result.
Figure 9: Fusion gate (forward)
Position-wise FFN For the FFN function of
equation 13, Figure 10(a) represents the deactivation ratio in the first hidden layer of position-wise
ffn.
As shown in Figure 2, position-wise ffn is used
in conjunction with a residual connection. That
is, the final output of position-wise ffn for input
x is the de -dimensional vector of LayerNorm(x +
FFN(x)). Figure 10(b) visualizes the maximum
value of this final output vector.
In Figure 10, keywords with a high deactivation
ratio is shown in panel (a) and a high final max
value in panel (b). In case of a word corresponding to a keyword, deactivation occurs frequently
in (a), and residual learning is hardly achieved in
the position-wise ffn, so that the output of the fusion gate is almost maintained. On the other hand,
in case of non-important words, residual learning
is performed in position-wise ffn because there is
less deactivation in (a), so that the max value of
final output becomes smaller in (b). This results
in preventing non-important words from consideration in the subsequent pooling layer. In summary, position-wise ffn plays a key role in ensuring that non-critical words are paid less attention
to in pooling layers.
(a) First hidden layer deactivation ratio
(b) Final output max value (+residual connection)
Figure 10: Position-wise ffn (forward)
Pooling Layer For the multi-dimensional attention corresponding to Figure 11(a), we visualized the attention weights averaged for each
word, where attention weights correspond to
softmax(L) ∈ Rn×2de in equation 15.
In max pooling, the max value is selected for
each column of U ∈ Rn×2de . Thus, in Figure
11(b), we visualize the percentage at which each
word is selected in the max pooling operation for
the 2de dimension.
It can be seen that panels (a) and (b) of Figure
11 are similar on the whole. In other words, both
multi-dimensional attention and max pooling utilize information about key words intensively. A
similar result can be expected by using only one
of the pooling layers. However, experiment results
show that using both multi-dimensional attention
and max pooling layer gives better performance.
(a) Multi-dimensional attention average weight
(b) Max pooling ratio (%)
Figure 11: Pooling layer
6
Conclusion
In this paper, we propose the Distance-based SelfAttention Network reflecting the distance between
words. By reflecting the word distance information, our model learns the local dependency
without losing the ability to capture the global
dependency. This was achieved through a simple distance mask, so that the performance of the
NLI task could be improved while maintaining the
number of parameters and training time. In particular, we recorded the new state-of-the-art performance for SNLI data. The introduction of the distance mask improves the performance with longer
sentences.
As the research on universal sentence encoders
using NLI data was proposed by Conneau et al.
(2017), we plan to carry out research on fully
attention-based networks for universal sentence
embedding as future work. We will also study
the fully attention-based network in image data
and speech data. Especially, regarding image data,
capsule network (Sabour et al., 2017) recently proposed, and as research on new structure to replace
CNN is going on, our future work will move in
similar directions.
In Proceedings of EMNLP 2015, pages 632–642.
Association for Computational Linguistics.
Samuel R. Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D. Manning, and
Christopher Potts. 2016. A fast unified model for
parsing and sentence understanding. In Proceedings of ACL 2016, pages 1466–1477. Association for
Computational Linguistics.
Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui
Jiang, and Diana Inkpen. 2017. Recurrent neural
network-based sentence encoder with gated attention for natural language inference. In Proceedings
of RepEval 2017: The Second Workshop on Evaluating Vector Space Representations for NLP. Association for Computational Linguistics.
Djork-Arn Clevert, Thomas Unterthiner, and Sepp
Hochreiter. 2015. Fast and accurate deep network
learning by exponential linear units (elus). In Proceedings of ICLR 2016.
Acknowledgments
We would like to thank Hyejin Lee, Hyunjoong
Kim, Taewook Kim, Jinwon An, Inbeom Park,
Minki Chung, and many others in SNUDM center, for critical feedback and discussions.
This work was supported by the BK21 Plus Program(Center for Sustainable and Innovative Industrial Systems, Department of Industrial Engineering & Institute for Industrial Systems Innovation,
Seoul National University) funded by the Ministry of Education, Korea (No. 21A20130012638),
the National Research Foundation (NRF) grant
funded by the Korea government (MSIP) (No.
2011- 0030814), and the Institute for Industrial
Systems Innovation of SNU.
Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic
Barrault, and Antoine Bordes. 2017. Supervised
learning of universal sentence representations from
natural language inference data. In Proceedings
of EMNLP 2017, pages 681–691. Association for
Computational Linguistics.
Alex Graves, Greg Wayne, and Ivo Danihelka.
2014. Neural turing machines. arXiv preprint
arXiv:1410.5401.
Karl Moritz Hermann, Tom Koisk, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman,
and Phil Blunsom. 2015. Teaching machines to read
and comprehend. In Proceedings of NIPS 2015,
pages 1693–1701.
References
Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. 2016. Character-aware neural language models. In Proceedings of AAAI 2016, pages
2741–2749.
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer normalization. In NIPS 2016 Deep
Learning Symposium.
Kingma, Diederik, and Jimmy Ba. 2014. Adam: A
method for stochastic optimization. In Proceedings
of ICLR 2015.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly
learning to align and translate. In Proceedings of
ICLR 2015.
Jorge A. Balazs, Edison Marrese-Taylor, Pablo Loyola, and Yutaka Matsuo. 2017. Refining raw sentence representations for textual entailment recognition via attention. In Proceedings of RepEval 2017:
The Second Workshop on Evaluating Vector Space
Representations for NLP. Association for Computational Linguistics.
Samuel R. Bowman, Gabor Angeli, Christopher Potts,
and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference.
Yang Liu, Chengjie Sun, Lei Lin, and Xiaolong Wang.
2016. Learning natural language inference using
bidirectional lstm model and inner-attention. arXiv
preprint arXiv:1605.09090.
Lili Mou, Rui Men, Ge Li, Yan Xu, Lu Zhang, Rui Yan,
and Zhi Jin. 2015. Natural language inference by
tree-based convolution and heuristic matching. In
Proceedings of ACL 2016, page 130. Association for
Computational Linguistics.
Tsendsuren Munkhdalai and Hong Yu. 2016a. Neural
semantic encoders. In Proceedings of EACL 2017,
pages 397–407. Association for Computational Linguistics.
Tsendsuren Munkhdalai and Hong Yu. 2016b. Neural
tree indexers for text understanding. In Proceedings
of EACL 2017, pages 11–21. Association for Computational Linguistics.
Nikita Nangia, Adina Williams, Angeliki Lazaridou,
and Samuel R. Bowman. 2017. The repeval 2017
shared task: Multi-genre natural language inference
with sentence representations. In Proceedings of
RepEval 2017: The Second Workshop on Evaluating
Vector Space Representations for NLP. Association
for Computational Linguistics.
Yixin Ni and Mohit Bansal. 2017. Shortcut-stacked
sentence encoders for multi-domain inference. In
Proceedings of RepEval 2017: The Second Workshop on Evaluating Vector Space Representations
for NLP. Association for Computational Linguistics.
Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton.
2017. Dynamic routing between capsules. arXiv
preprint arXiv:1710.09829.
Lifeng Shang, Zhengdong Lu, and Hang Li. 2015.
Neural responding machine for short-text conversation. In Proceedings of ACL 2015, pages 1577–
1586. Association for Computational Linguistics.
Tao Shen, Tianyi Zhou, Guodong Long, Jing Jiang,
Shirui Pan, and Chengqi Zhang. 2017. Disan:
Directional self-attention network for rnn/cnnfree language understanding.
arXiv preprint
arXiv:1709.04696.
Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston,
and Rob Fergus. 2015. End-to-end memory networks. In Proceedings of NIPS 2015, pages 2440–
2448.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Sequence to sequence learning with neural networks. In Proceedings of NIPS 2014, pages 3104–
3112.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. arXiv preprint arXiv:1706.03762.
Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel
Urtasun. 2015. Order-embeddings of images and
language. In Proceedings of ICLR 2016.
Hoa Trong Vu, Thuong-Hai Pham, Xiaoyu Bai Marc
Tanti, Lonneke van der Plas, and Albert Gatt. 2017.
Lct-maltas submission to repeval 2017 shared task.
In Proceedings of RepEval 2017: The Second Workshop on Evaluating Vector Space Representations
for NLP. Association for Computational Linguistics.
Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2017. A broad-coverage challenge corpus for
sentence understanding through inference. arXiv
preprint arXiv:1704.05426.
Han Yang, Marta R, Costa-juss, and Jos A. R. Fonollosa. 2017. Character-level intra attention network
for natural language inference. In Proceedings of
RepEval 2017: The Second Workshop on Evaluating
Vector Space Representations for NLP. Association
for Computational Linguistics.
Appendix
Masked Multi-Head Attention The attention
weights for each head in the masked multi-head
attention are shown in Figures 12 and 13. Figure
12 shows the result of using a forward directional
mask, and Figure 13 is the result of using a backward directional mask. It can be seen that the attention weights are different for each head. This
allows our model to capture various dependencies
between words in a sentence.
(a) head0
(a) head0
(b) head1
(c) head2
(d) head3
(b) head1
(e) head4
(c) head2
(d) head3
(e) head4
Figure 12: Forward masked multi-head attention weights
Figure 13: Backward masked multi-head attention weights
| 2cs.AI
|
arXiv:1404.4944v1 [math.OC] 19 Apr 2014
Unit commitment with valve-point
loading effect
João Pedro Pedroso
INESC Porto and Faculdade de Ciências, Universidade do Porto, Portugal
[email protected]
Mikio Kubo
Tokyo University of Marine Science and Technology, Japan
[email protected]
Ana Viana
INESC Porto and Instituto Superior de Engenharia, Instituto Politécnico do Porto, Portugal
[email protected]
Technical Report Series: DCC-2014-05
Departamento de Ciência de Computadores
Faculdade de Ciências da Universidade do Porto
Rua do Campo Alegre, 1021/1055,
4169-007 PORTO,
PORTUGAL
Tel: 220 402 900 Fax: 220 402 950
http://www.dcc.fc.up.pt/Pubs/
Unit commitment with valve-point loading effect
João Pedro Pedroso
INESC Porto and Faculdade de Ciências, Universidade do Porto, Portugal
Mikio Kubo
Tokyo University of Marine Science and Technology, Japan
Ana Viana
INESC Porto and
Instituto Superior de Engenharia, Instituto Politécnico do Porto, Portugal
Abstract
Valve-point loading affects the input-output characteristics of generating units, bringing the fuel costs
nonlinear and nonsmooth. This has been considered in the solution of load dispatch problems, but
not in the planning phase of unit commitment.
This paper presents a mathematical optimization model for the thermal unit commitment problem
considering valve-point loading. The formulation is based on a careful linearization of the fuel cost
function, which is modeled with great detail on power regions being used in the current solution, and
roughly on other regions.
A set of benchmark instances for this problem is used for analyzing the method, with recourse to
a general-purpose mixed-integer optimization solver.
Keywords: Unit Commitment, Load Dispatch, Combinatorial Optimization, Mixed-integer
Programming
1. Introduction
The unit commitment problem (UCP) consists of deciding which power generating units must be
committed/decommitted over a planning horizon, usually lasting from 1 day to 2 weeks and split
into periods of one hour. The production levels at which units must operate (pre-dispatch) must
also be determined to optimize a cost function which includes both fixed and variable costs. The
committed units must satisfy the forecasted system load and reserve requirements, as well as a large
set of technological constraints. This problem has great practical relevance, due to the savings that
can be achieved with an optimized schedule.
Simple approaches for this problem model the cost as a linear function, which provides a rather
rough approximation; accuracy can be improved by switching to a quadratic function. For these
models general-purpose mixed-integer convex optimization solvers can be used to find optimum
solutions for realistic, large instances. Nonetheless, much better approximations of the true costs
can be obtained if the valve-point loading effect is taken into account; however models become much
more difficult to optimize.
Valve-point loading affects the input-output characteristics of generating units, making the fuel
costs nonlinear and nonsmooth. This has been considered in the heuristic solution of load dispatch
problems — see, e.g., [1, 2, 3, 4], but there are no exact methods available in the literature.
Email addresses: [email protected] (João Pedro Pedroso), [email protected] (Mikio Kubo),
[email protected] (Ana Viana)
2
Concerning the planning phase of unit commitment, even for convex models most of the research
on the solution of the UCP focuses in heuristic methods. Recently, however, improvements in
the capabilities of mixed-integer programming solvers has encouraged the thorough exploitation of
their capabilities [5]. Extensive surveys of different optimization techniques and modeling issues are
provided in [6, 7, 8], but the valve-point loading effect is not considered in any of the exact methods
proposed. A mixed integer quadratically constrained model to solve unit commitment on real-life,
large-scale power systems is presented in [9]; it details some features of the units, but not the valvepoint loading effect. Another approach, where a mixed integer linear formulation is used for modeling
nonlinear output of generators and applied to unit commitment, is presented in [10]. A process with
some similarities to the method we propose, also involving the assessment of lower and upper bounds
to a nonlinear function and iteratively converging to the optimum, has been presented in [11] and
applied to a related problem: that of optimizing short-term hydro scheduling.
The main contribution of this paper is a method based on a careful linearization of the fuel cost
function, which is modeled with great detail on power regions being used in the current solution
and roughly on other regions; this formulation is used in an iterative process that converges to the
optimum. The method may be used both for load dispatch (by setting the number of planning periods
to one) and for unit commitment. When the solution process stops, e.g. due to a limit on CPU usage,
the method provides both an upper and a lower bound to the optimum; this information is usually
very important in practice.
2. Mathematical model
Cost
A concise, complete description of the unit commitment problem as a mathematical optimization
model has been provided in [5]. When valve-point loading is taken into account that model can be
used as is, except for the shape of the objective function: instead of a quadratic function, as considered
in [12], it is now a nonconvex, nonsmooth function, as has been proposed for load dispatching in [13]
(cited by [14]).
P min
Power
P max
Figure 1: Typical shape of the fuel cost considering the valve-point loading effect.
3
2.1. Objective
A commonly used function of the fuel costs in a generating unit in terms of the power produced
p, taking into account the valve-point loading effect, is the following [13]:
F (p) = a + bp + cp2 + e sin f (P min − p) ,
(1)
where a, b, c, e, f are function’s parameters (see Figure 4), as are the minimum and maximum operating powers P min and P max . The first three terms describe a quadratic function, as usually taken into
account when valve-point loading is not considered; when it is, there is a periodic factor, as described
in the rightmost term.
This function being nonlinear, nonconvex, and nonsmooth, methods for optimizing it under
constraints on load limit and demand satisfaction — i.e., in an economic load dispatch setting
— usually involve specifically designed heuristic approaches, many of them based on evolutionary
algorithms [4, 15].
There is, however, a way of tackling this problem exactly (i.e., with arbitrarily low error). Let us
first notice that the valve points — where there is a discontinuity — limit areas where the function
is concave. Thus, if we evaluate the function into several points and use them as breakpoints for
joining the two valve points, we obtain a linear function that is never larger than the true function
(see Figure 2). Given a power production level this function provides, thus, a lower bound to the
objective value; on the other hand, the evaluation of the true function provides a trivial upper bound.
For any solution satisfying all the problem’s constraints, a lower and an upper bound to possible values
of the objective can therefore be obtained by the piecewise-linear lower approximation and the true
function, respectively.
UB
Cost
Cost
LB
exact function
approximation
Power
Power
p
Figure 2: Fuel cost and its representation with a piecewise-linear function (left). Upper bound (UB, exact function
evaluation) and lower bound (LB, piecewise-linear approximation) at a given production level p (right).
For obtaining increased precision, a large number of breakpoints may be required; this is likely
to be the limiting step for large instances. In order to obviate this problem as much as possible, we
may have the cost function represented with greater detail in power regions of potential operation,
whereas on other areas the representation may be more rough; in the limit, there may be a single
straight line joining two valve points (see Figure 3). Later in this section we propose an algorithm
for determining where to expand the number of breakpoints. The final shape of the piecewise-linear
approximation may resemble that of Figure 3.
Considering a set of units U and a set of periods T , for a given unit u ∈ U and period t ∈ T , once
the coordinates (Xutk , Yutk ) for the set of breakpoints k = 0, . . . , K to be considered is known, the
lower approximation of the objective function can easily be included in a mixed-integer programming
(MIP) linear model by means of a convex combination of these points:
4
Cost
Cost
exact function
approximation
exact function
approximation
P min
Power
Power
P max
Figure 3: Piecewise-linear fine approximation in important areas, rough on the other areas (left). Fuel cost and its
possible final representation with a piecewise-linear function (right).
K
X
put =
Xutk zutk , ∀u ∈ U, ∀t ∈ T ,
(2)
k=0
Lut =
K
X
Yutk zutk ,
∀u ∈ U, ∀t ∈ T ,
(3)
∀u ∈ U, ∀t ∈ T .
(4)
k=0
K
X
zutk ≤ yut ,
k=0
PK
In a standard convex combination model the last equation is written as k=0 zutk = 1. However,
in the case a unit is not committed in a given period, both the production level put and the linearized
fuel cost Lut must be zero. For taking this into account, in the right hand size of equation (4)
instead of 1 there must be variable yut , which indicates whether unit u was committed to produce in
period t (yut = 1) or not (yut = 0). Additionally, for the model to be correct, there must be points
corresponding to minimum and maximum operating powers, i.e., Xut0 = Pumin and XutK = Pumax , for
each unit u and period t.
For the minimization of non-convex functions represented by piecewise-linear segments, as in the
present case, it is necessary to introduce binary variables, each corresponding to a zutk , limiting
which may be non-zero; also an additional constraint for forcing the convex combination to take
on two consecutive points is necessary (i.e., there can be zutk > 0 and zu,t,k+1 > 0 only for one
value of k ∈ {0, . . . , K − 1}). Alternatively, we may use a convenience of most modern MIP solvers
that assures the same thing by declaring that, for each u and t, the set of variables zutk , ∀k forms a
so-called special ordered set constraint of type II (SOS2) [16].
The objective is to minimize the total production costs,
minimize
XX
(Lut + Sut ) ,
(5)
t∈T u∈U
which include start-up costs Sut ; these are modeled as
hot
cold cold
Sut = ahot
u sut + au sut ,
(6)
(7)
hot
where ahot
u is the hot start up cost and variable sut = 1 if there was a hot start for unit u in period
t, 0 otherwise. Equivalently for cold start up, with acold
and scold
u
ut . (See also constraints (15) to (18).)
5
2.2. Constraints
In this problem the following constraints will be considered: system power balance and generation
limits (i.e., load dispatch), system reserve requirements, unit initial conditions, and unit minimum
up and down times; this MIP model was introduced in [5].
Demand satisfaction is modeled by constraint (8), and power reserve requirements by constraint (9).
X
put = Dt ,
∀t ∈ T ,
(8)
Pumax yut ≥ Dt + Rt , ∀t ∈ T .
(9)
u∈U
X
u∈U
Power production levels of thermal power units are within the range defined by the technical
minimum and maximum production levels in (10).
Pumin yut ≤ put ≤ Pumax yut , ∀u ∈ U, ∀t ∈ T .
(10)
When a unit u is switched on, it must remain on for at least Tuon consecutive periods; similarly,
it must be kept off for at least Tuoff after being switched off. Constraints (11) and (12) model this
aspect for the initial state, while constraints (13) and (14) do the same for the remaining planning
off
prev
off
horizon. In (11) θuon represents max(0, Tuon − tprev
u ), and θu in (12) stands for max(0, Tu − tu ); the
prev
previous state of unit u is the parameter yu , which is 1 if the unit was on, 0 if it was off.
yut = 1, ∀u ∈ U : yuprev = 1, for t = 0, . . . , θuon ,
yut = 0, ∀u ∈ U :
yuprev
= 0, for t =
0, . . . , θuoff .
(11)
(12)
off
Constraints (13) and (14) determine if a unit u is started/switched off in period t (xon
ut = 1, xut = 1,
off
on
off
on
respectively, or 0 otherwise); variables τut and τut stand for max(t−Tu +1, 1) and max(t−Tu +1, 1),
respectively.
t
X
xon
ui ≤ yut ,
∀u ∈ U, ∀t ∈ T ,
(13)
xoff
ui ≤ 1 − yut , ∀u ∈ U, ∀t ∈ T .
(14)
on
i=τut
t
X
off
i=τut
Constraints (15) state that every time a unit is switched on, a start-up cost will be incurred.
cold
on
shot
ut + sut = xut ,
∀u ∈ U, ∀t ∈ T .
(15)
Constraints (16) determine the start-up type of each unit, i.e., decide whether it is a cold or a
hot start type. It will be a cold start if the unit remained off for more than tcold
periods of time, and
u
a hot start otherwise.
t−1
X
yut −
yui ≤ scold
ut , ∀u ∈ U, ∀t ∈ T .
(16)
i=t−tcold
u −1
Constraints (17) determine each unit’s switch-on variables, and (18) determine the switch-off
variables.
yut − yu,t−1 ≤ xon
ut ,
xoff
ut
=
xon
ut
∀u ∈ U, ∀t ∈ T ,
(17)
+ yu,t−1 − yut , ∀u ∈ U, ∀t ∈ T .
(18)
6
2.3. Solution approach
There is not, to the best of our knowledge, exact solution method for the load dispatching problem
when the valve-point loading effect is taken into account, let alone the full unit commitment problem.
The solution approach that we propose achieves this, in the sense that it converges to a solution with
an arbitrary degree of precision.
The method is based in the replacement of the function defined by Equation (1) by a piecewiselinear function; a careful selection of the breakpoints allows the construction of a model that provides
a lower bound to the exact function. Upon a feasible solution obtained by this linear model, we
can evaluate the true function; this will provide an upper bound to the value of the objective. As
described in Algorithm 1, if the deviation between the upper and lower bounds are within a userspecified tolerance, then the method will stop. Otherwise, the number of breakpoints is increased
(though only in the required intervals), thus leading to a better approximation, and the optimization
problem is resolved. This process is repeated until the relative deviation between upper and lower
bounds is small enough (or the allowed CPU time is exceeded). Notice that the solution of an iteration
is feasible to the problem of the next, the only changes are in its evaluation. We can, therefore, supply
it as a starting point to the MIP solver in line (4) of the algorithm.
Algorithm 1: Solution procedure.
Input: UCP instance, CPU limit, tolerance
Output: A solution to the input instance
(1)
foreach unit u ∈ U , period t ∈ T :
(2)
But := set of breakpoints containing only the valve points
(3)
while true
(4)
solve problem with piecewise-linear objective, breakpoints B
(5)
if U B − LB < tolerance or CPU limit exceeded:
(6)
return solution
(7)
foreach u, t with coarse objective representation in the solution:
(8)
divide the inter-valve interval containing put into K segments
(9)
add corresponding points to But
(10)
if there were no u, t with coarse objective in put :
(11)
K := 2K
(12)
for each u, t recompute But based on K segments in the inter-valve interval containing
put
The value K used in this algorithm should ideally be setup in such a way that the initial
approximation is just good enough to solve the instance with sufficient precision. If K is set too
high, there will be too many intervals in the piecewise-linear representation of the objective function,
leading to great precision but requiring unnecessary CPU for solving the problem; if is it set too low,
additional iterations will be necessary for providing the required precision (each iteration doubling
the value of K), and again unnecessary CPU will be used.
The most important properties of this algorithm are summarized next. Let us denominate P1 the
problem of maximizing the true objective,
XX
minimize
[Sut + yut Fut (put )]
(19)
t∈T u∈U
where
min
Fut (p) = au + bu p + cu p2 + eu sin fu (Put
− p) ,
under constraints (6) to (18), and P2 its linearized counterpart, i.e., the problem of maximizing (5)
under the same constraints and additionally (2), (3) and (4).
Property 1. Feasible solutions for P2 are feasible for P1 .
Proof. P2 includes all the variables of P1 , and contraints of P2 are a superset of constraints of P1 .
Property 2. The evaluation of the objective function (19) at a feasible solution obtained for P2 is
an upper bound to the optimum of P1 .
7
wi′
wi+1
wi+1
Cost
wi
Cost
wi
wK
wK
wK−1
wK−1
Power
Power
Figure 4: Linear approximations in consecutive iterations.
Proof. Follows from Property 1 and from the notion of optimum: the objective at any feasible solution
of a minimization problem provides an upper bound to its optimum.
Property 3. An optimum of P2 is a lower bound to the optimum of P1 .
Proof. Between two consecutive valve points the function F (p) defined by Equation 1 is concave as
long as c ≤ ef 2 /2 (see Appendix B); therefore, if this condition is observed, any piecewise-linear
interpolation between consecutive valve points is not larger than F . In the construction of the
piecewise-linear approximation we are considering all the valve points, as well as the extreme point
of the domain of F , as interpolation points; hence, the linear interpolation is not larger than F
in its domain. We can thus conclude that the minimum of P2 — where the objective is a linear
interpolation in the previous conditions — is not larger than the minimum of P1 , where the objective
is function F .
Property 4. With tolerance ǫ = 0, when the number of iterations increases the solution provided by
Algorithm 1 converges to the optimum solution of P1 .
Proof. Let us denote the linear approximation of F at an iteration i by Gi , and define G = limi→∞ Gi .
Note that Gj (p) ≥ Gi (p), for any j > i, and for all p in the domain of F . As functions Gi are nondecreasing with increasing i and are limited above by F , the sequence Gi converges.
It remains to prove that the optimum solution with Gi as the objective function, when i tends to
infinity, is an optimum solution with F as the objective function. We will assume that this is not true,
i.e., that the optimum F ∗ is strictly greater than G∗ , and show that this leads to a contradiction.
Notice that F ∗ > G∗ is equivalent to F ∗ − G∗ = ǫ∗ , with ǫ∗ > 0, which implies that there must be
some u, t such that Fut (p∗ ) − Gut (p∗ ) = ǫ′ > 0. However, lines (11) and (12) of Algorithm 1 imply
that the number of segments in each inter-valve interval containing p∗ut tends to infinity when the
number of iterations tends to infinity. This means that Gut (p∗ ) becomes arbitrarily close to Fut (p∗ ),
and hence ǫ′ cannot be positive.
3. Computational results
In order to run economic load dispatch instances with the current model we must set up an
instance with only one period; generators must have the previous status as operating, and must be
forced to keep operating by adjusting the minimum number of on periods. We have thus prepared
data for the most widely used benchmark instances for load dispatching with valve-point loading
8
179300
55900
upper bound
lower bound
upper bound
lower bound
179200
55850
179100
179000
55800
Total cost
Total cost
178900
55750
55700
178800
178700
178600
55650
178500
178400
55600
178300
178200
55550
0
5
10
15
20
0
25
5
10
15
20
25
Iteration
Iteration
Figure 5: Bounds as a function of the iteration number, for instance ucp40 with one period (optimal solution was
found) and three periods (final solution was not optimal).
effect: eld13 (first described in [13]) and eld40 (first described in [17]); we have used the tables
available in [14]1 .
The setup used was the following: a computer with an Intel Xeon processor at 3.0 GHz and
running Linux version 2.6.32, using Gurobi version 5.0.1 [16]. Only one thread was assigned to
this experiment. Models were written in the Python language and the default Gurobi parameters
were used until reaching the user-defined tolerance; then, parameters were changed to the maximum
precision supported by Gurobi. The aim of this procedure is to avoid spending too much CPU in the
initial iterations, where the representation of the objective function is still rather course.
Instance
eld13
eld13
eld40
Load
1800
2520
10500
Lower bound
17963.83
24169.92
121412.53
Upper bound
17963.83
24169.92
121412.54
Error
<1.e-7
<1.e-7
<1.e-7
CPU time
19.1s
1.8s
7.6s
N
34
19
27
Table 1: Solution to economic load dispatch instances: bounds to the objective value, maximum relative error, CPU
time employed, and number of iterations performed.
The results presented in Table 1 are optima, and are approximately equal to the best solutions
found by the best methods available in the literature [4].2
As for the unit commitment problem, to the best of our knowledge there are no test instances
taking into account the valve-point effect. We have therefore prepared a set of benchmarks, based on
the instances for unit commitment provided in [12], and on the instances for economic load dispatch
mentioned above. Instance ucp10 corresponds to the standard instance described in [12]; ucp5 is
based on that instance, by selecting only half of the units and adjusting demand. Instances ucp13
and ucp40 have costs taken from eld13 and eld40, respectively, and data concerning multi-period
operation based on [12]. Notice that the difference between instances eld13 and ucp13 with one
period (besides possible differences in the demand value) is that in the former the commitment
decision has already been made, and all units considered must be operating; the same for eld40 and
ucp40. These results are presented in Table 2. Figure 5 plots the evolution, for a particular case,
of the lower and upper bounds with respect to the iteration number; as expected, the lower bound
is non-decreasing, whereas the upper bound may increase (when the linear approximation at the
solution of the previous iteration was poor).
1 All
the
data
and
programs
used
are
available
in
this
paper’s
Internet
page,
http://www.dcc.fc.up.pt/~ jpp/code/valve
2 Results presented in [15] are better than the optimum for eld13, indicating that there is likely an error in that
paper. Possibly this is due to a rounding error; if we calculate the cost for the solutions provided in the paper we
obtain values greater than those reported.
9
Instance
ucp5
ucp5
ucp5
ucp5
ucp5
ucp10
ucp10
ucp10
ucp10
ucp10
ucp13
ucp13
ucp13
ucp13
ucp13
ucp40
ucp40
ucp40
ucp40
ucp40
Periods
1
3
6
12
24
1
3
6
12
24
1
3
6
12
24
1
3
6
12
24
Lower bound
7352.46
24506.12
59127.38
153575.68
308482.03
13826.85
45929.36
109919.97
285808.54
572943.80
11701.28
38849.84
91405.97
231587.47
464053.01
55644.79
178395.51
416108.42
1112370.94
2235970.56
Upper bound
7352.46
24506.12
59127.39
153575.69
309152.47
13826.85
45929.36
109919.98
286067.44
574554.02
11701.28
38849.84
91783.98
232537.37
466186.98
55644.79
178547.12
416605.93
1113801.09
2238504.45
Error
<1.e-7
<1.e-7
<1.e-7
<1.e-7
0.22%
<1.e-7
<1.e-7
<1.e-7
0.091%
0.28%
<1.e-7
<1.e-7
0.41%
0.41%
0.46%
<1.e-7
0.085%
0.12%
0.13%
0.11%
CPU time
0.12
0.29
1.2
54.
3600.
0.17
1.3
31.
3600.
3600.
0.15
3.3
3600.
3600.
3600.
3.9
3600.
3600.
3600.
3600.
N
5
5
6
7
3
6
7
9
5
3
6
11
10
4
3
25
30
15
11
7
Table 2: Solution to unit commitment instances: bounds to the objective value, maximum relative error, CPU time
employed, and number of iterations performed.
4. Conclusions
The main contribution of this paper is a model and an algorithm for solving the unit commitment
problem taking into consideration the valve-point loading effect in fuel cost computation. This model
can be used for planning which units to commit in each period so as to match demand, meeting
reserve constraints, and under a set of technological constraints. The algorithm iteratively calls a
general-purpose mixed-integer programming solver for tackling optimization subproblems, and uses
the solution to improve the accuracy of fuel cost approximation until reaching an error below a given
tolerance, or reaching a given limit on CPU usage.
The model can also be used for load dispatching, when only one period is considered and the subset
of units that will operate is already selected; this is known as the economic load dispatch problem
with valve-point effect; the method proposed was able to solve well known benchmark instances of
this problem for which the optimum was not known.
To the best of our knowledge, there are no standard benchmarks for unit commitment taking
into account the valve-point loading effect. We present a set of them, based on data from load
dispatch and unit commitment. Using our algorithm for their solution, the optimum was obtained
for a subset of instances, and very good approximations, with a relative error below 1%, were found
for the remaining.
Acknowledgements
This work was partly funded by project “NORTE-07-0124-FEDER-000057”, under the North
Portugal Regional Operational Programme (ON.2 O Novo Norte) and the National Strategic Reference Framework through the European Regional Development Fund, and by national funds through
Fundação para a Ciência e a Tecnologia (FCT), and by FCT (under Project PTDC/EGE-GES/099120/2008)
through “Programa Operacional Temático Factores de Competitividade (COMPETE)” of “Quadro
Comunitário de Apoio III”, partially funded by FEDER.
10
References
[1] Al-Sumait J, Al-Othman A, Sykulski J. Application of pattern search method to power system
valve-point economic load dispatch. International Journal of Electrical Power & Energy Systems
2007;29(10):720 –30.
[2] Hemamalini S, Simon SP. Dynamic economic dispatch using artificial immune system for
units with valve-point effect. International Journal of Electrical Power & Energy Systems
2011;33(4):868 –74.
[3] Roy P, Roy P, Chakrabarti A. Modified shuffled frog leaping algorithm with genetic algorithm
crossover for solving economic load dispatch problem with valve-point effect. Applied Soft
Computing 2013;13(11):4244 –52.
[4] Reddy AS, Vaisakh K. Shuffled differential evolution for economic dispatch with valve point
loading effects. International Journal of Electrical Power & Energy Systems 2013;46(0):342 –52.
[5] Viana A, Pedroso JP. A new MILP-based approach for unit commitment in power production
planning. International Journal of Electrical Power and Energy Systems 2013;44:997–1005.
[6] Sen S, Kothari D. Optimal thermal generating unit commitment: a review. International Journal
of Electrical Power & Energy Systems 1998;20(7):443–51.
[7] Padhy NP. Unit commitment – a bibliographical survey. IEEE Transactions in Power Systems
2004;19(2):1196–205.
[8] Yamin HY. Review on methods of generation scheduling in electric power systems. Electric
Power Systems Research 2004;69:227–48.
[9] López JÁ, Ceciliano-Meza JL, Moya IG, Gómez RN. A MIQCP formulation to solve the unit
commitment problem for large-scale power systems. International Journal of Electrical Power &
Energy Systems 2012;36(1):68 – 75.
[10] Simoglou CK, Biskas PN, Bakirtzis AG. Optimal self-scheduling of a thermal producer in shortterm electricity markets by milp. IEEE Transactions on Power Systems 2010;25(4):1965 –77.
[11] Lima RM, Marcovecchio MG, Novais AQ, Grossmann IE. On the computational studies
of deterministic global optimization of head dependent short-term hydro scheduling. IEEE
Transactions on Power Systems 2013;28(4):4336 –47.
[12] Kazarlis SA, Bakirtzis AG, Petridis V. A genetic algorithm solution to the unit commitment
problem. IEEE Transactions on Power Systems 1996;11:83–92.
[13] Wood A, Wollenberg B. Power Generation, Operation, and Control. New York, NY: John Wiley
& Sons; 1984.
[14] dos Santos Coelho L, Mariani V. Combining of chaotic differential evolution and quadratic
programming for economic dispatch optimization with valve-point effect. IEEE Transactions on
Power Systems 2006;21(2):989–96. (See also a correction to this paper).
[15] dos Santos Coelho L, Mariani VC. An efficient cultural self-organizing migrating strategy for
economic dispatch optimization with valve-point effect. Energy Conversion and Management
2010;51(12):2580 –7.
[16] Gurobi Optimization, Inc. .
http://www.gurobi.com; 2012.
Gurobi Optimizer
Reference Manual,
Version 5.0.
[17] Sinha N, Chakrabarti R, Chattopadhyay PK. Evolutionary programming techniques for
economic load dispatch. IEEE Transactions on Evolutionary Computation 2003;7(1):83–94.
11
Appendix A. Notation
Constants
• T – length of the planning horizon.
• T = {1, . . . , T } – set of planning periods.
• U – set of units.
• Pumin , Pumax – minimum and maximum production levels for unit u.
• Tuon , Tuoff – minimum number of periods that unit u must be kept switched on/off.
• Dt – system load requirements in period t.
• Rt – spinning reserve requirements in period t.
• au , bu , cu , eu , fu – fuel cost parameters for unit u.
cold
• ahot
– hot and cold start up costs for unit u.
u , au
• tcold
– number of periods after which the start up of unit u is evaluated as cold.
u
• yuprev – previous state of unit u (1 if on, 0 if off).
– number of periods unit u has been on or off prior to the first period of the planning
• tprev
u
horizon.
Decision variables
• yut – 1 if unit u is on in period t, 0 otherwise.
• put – production level of unit u, in period t.
Auxiliary variables
off
• xon
ut , xut – 1 if unit u is started/switched off in period t, 0 otherwise;
• shot
ut – 1 if unit u has a hot start in period t, 0 otherwise;
• scold
ut – 1 if unit u has a cold start in period t, 0 otherwise;
Production costs
• Fut – fuel cost for unit u in period t.
• Sut – start up cost for unit u in period t.
Appendix B. Conditions for concavity of the fuel costs
In this appendix westate the conditions for concavity of the fuel cost function F (p) = a + bp +
cp2 + e sin f (P min − p) (Equation 1) between two valve points. Recall that a differentiable function
is concave in a given region if its second derivative is less than or equal to zero in all points of that
region.
If the term within the absolute value,
e sin f (P min − p) , is positive, the first derivative of F (p) is
F ′ (p) = 2cp + b − ef cos f (P min − p) , and the second derivative is F ′′ (p) = 2c − ef 2 sin(f (P min − p)).
2
For the function F to be concave there must be F ′′ (p) ≤ 0, i.e., c ≤ ef2 sin f (P min − p) . As
sin(x) ≤ 1, ∀x, F is concave for c ≤ ef 2 /2.
Similarly, if the term within
the first derivative of F (p) is F ′ (p) =
the absolute value is negative,
min
′′
2cp + b + ef cos f (P
− p) , and the second derivative is F (p) = 2c + ef 2 sin f (P min − p) . Again,
12
2
for F to be concave there must be c ≤ − ef2 sin f (P min − p) . As sin(x) ≥ −1, ∀x, this condition is
observed for c ≤ ef 2 /2.
Therefore, for
c ≤ ef 2 /2
the function F (p) is concave between two valve points. This condition is valid for all the instances
analyzed in this paper.
13
| 5cs.CE
|
Tunable Efficient Unitary Neural Networks (EUNN) and their application to
RNNs
Li Jing * 1 Yichen Shen * 1 Tena Dubcek 1 John Peurifoy 1 Scott Skirlo 1 Yann LeCun 2 Max Tegmark 1
Marin Soljačić 1
arXiv:1612.05231v3 [cs.LG] 3 Apr 2017
Abstract
Using unitary (instead of general) matrices in
artificial neural networks (ANNs) is a promising way to solve the gradient explosion/vanishing
problem, as well as to enable ANNs to learn
long-term correlations in the data. This approach appears particularly promising for Recurrent Neural Networks (RNNs). In this work, we
present a new architecture for implementing an
Efficient Unitary Neural Network (EUNNs); its
main advantages can be summarized as follows.
Firstly, the representation capacity of the unitary space in an EUNN is fully tunable, ranging from a subspace of SU(N) to the entire unitary space. Secondly, the computational complexity for training an EUNN is merely O(1) per
parameter. Finally, we test the performance of
EUNNs on the standard copying task, the pixelpermuted MNIST digit recognition benchmark
as well as the Speech Prediction Test (TIMIT).
We find that our architecture significantly outperforms both other state-of-the-art unitary RNNs
and the LSTM architecture, in terms of the final performance and/or the wall-clock training
speed. EUNNs are thus promising alternatives
to RNNs and LSTMs for a wide variety of applications.
1. Introduction
Deep Neural Networks (LeCun et al., 2015) have been successful on numerous difficult machine learning tasks, including image recognition(Krizhevsky et al., 2012; Donahue et al., 2015), speech recognition(Hinton et al., 2012)
and natural language processing(Collobert et al., 2011;
Bahdanau et al., 2014; Sutskever et al., 2014). However,
deep neural networks can suffer from vanishing and ex*
Equal contribution 1 Massachusetts Institute of Technology
New York University, Facebook AI Research. Correspondence
to: Li Jing <[email protected]>, Yichen Shen <[email protected]>.
2
ploding gradient problems(Hochreiter, 1991; Bengio et al.,
1994), which are known to be caused by matrix eigenvalues far from unity being raised to large powers. Because
the severity of these problems grows with the the depth of
a neural network, they are particularly grave for Recurrent
Neural Networks (RNNs), whose recurrence can be equivalent to thousands or millions of equivalent hidden layers.
Several solutions have been proposed to solve these problems for RNNs. Long Short Term Memory (LSTM) networks (Hochreiter & Schmidhuber, 1997), which help
RNNs contain information inside hidden layers with gates,
remains one of the the most popular RNN implementations.
Other recently proposed methods such as GRUs(Cho et al.,
2014) and Bidirectional RNNs (Berglund et al., 2015) also
perform well in numerous applications. However, none of
these approaches has fundamentally solved the vanishing
and exploding gradient problems, and gradient clipping is
often required to keep gradients in a reasonable range.
A recently proposed solution strategy is using orthogonal hidden weight matrices or their complex generalization
(unitary matrices) (Saxe et al., 2013; Le et al., 2015; Arjovsky et al., 2015; Henaff et al., 2016), because all their
eigenvalues will then have absolute values of unity, and can
safely be raised to large powers. This has been shown to
help both when weight matrices are initialized to be unitary (Saxe et al., 2013; Le et al., 2015) and when they are
kept unitary during training, either by restricting them to a
more tractable matrix subspace (Arjovsky et al., 2015) or
by alternating gradient-descent steps with projections onto
the unitary subspace (Wisdom et al., 2016).
In this paper, we will first present an Efficient Unitary Neural Network (EUNN) architecture that parametrizes the entire space of unitary matrices in a complete and computationally efficient way, thereby eliminating the need for
time-consuming unitary subspace-projections. Our architecture has a wide range of capacity-tunability to represent
subspace unitary models by fixing some of our parameters;
the above-mentioned unitary subspace models correspond
to special cases of our architecture. We also implemented
an EUNN with an earlier introduced FFT-like architecture
which efficiently approximates the unitary space with min-
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
imum number of required parameters(Mathieu & LeCun,
2014b).
We then benchmark EUNN’s performance on both simulated and real tasks: the standard copying task, the pixelpermuted MNIST task, and speech prediction with the
TIMIT dataset (Garofolo et al., 1993). We show that our
EUNN algorithm with an O(N ) hidden layer size can
compute up to the entire N × N gradient matrix using
O(1) computational steps and memory access per parameter. This is superior to the O(N ) computational complexity
of the existing training method for a full-space unitary network (Wisdom et al., 2016) and O(log N ) more efficient
than the subspace Unitary RNN(Arjovsky et al., 2015).
2. Background
2.1. Basic Recurrent Neural Networks
A recurrent neural network takes an input sequence and
uses the current hidden state to generate a new hidden state
during each step, memorizing past information in the hidden layer. We first review the basic RNN architecture.
Consider an RNN updated at regular time intervals t =
1, 2, ... whose input is the sequence of vectors x(t) whose
hidden layer h(t) is updated according to the following
rule:
h(t) = σ(Ux(t) + W h(t−1) ),
(1)
where σ is the nonlinear activation function. The output is
generated by
y(t) = Wh(t) + b,
(2)
where b is the bias vector for the hidden-to-output layer.
For t = 0, the hidden layer h(0) can be initialized to some
special vector or set as a trainable variable. For convenience of notation, we define z(t) = Ux(t) + Wh(t−1)
so that h(t) = σ(z(t) ).
2.2. The Vanishing and Exploding Gradient Problems
When training the neural network to minimize a cost function C that depends on a parameter vector a, the gradient
descent method updates this vector to a − λ ∂C
∂a , where λ
∂C
is a fixed learning rate and ∂a ≡ ∇C. For an RNN, the
vanishing or exploding gradient problem is most significant during back propagation from hidden to hidden layers, so we will only focus on the gradient for hidden layers.
Training the input-to-hidden and hidden-to-output matrices
is relatively trivial once the hidden-to-hidden matrix has
been successfully optimized.
In order to evaluate
∂C
∂Wij ,
one first computes the derivative
∂C
∂h(t)
using the chain rule:
∂C
∂h(t)
=
=
=
∂C ∂h(T )
∂h(T ) ∂h(t)
T −1
∂C Y ∂h(k+1)
∂h(T )
k=t
∂h(k)
T −1
∂C Y (k)
D W,
∂h(T ) k=t
(3)
(4)
(5)
where D(k) = diag{σ 0 (Ux(k) + Wh(k−1) )} is the Jacobian matrix ofQthe pointwise nonlinearity. For large times
T , the term
W plays a significant role. As long as
the eigenvalues of D(k) are of order unity, then if W has
eigenvalues λi 1, they will cause gradient explosion
| ∂h∂C
(T ) | → ∞, while if W has eigenvalues λi 1, they
can cause gradient vanishing, | ∂h∂C
(T ) | → 0. Either situation
prevents the RNN from working efficiently.
3. Unitary RNNs
3.1. Partial Space Unitary RNNs
In a breakthrough paper, Arjovsky, Shah & Bengio (Arjovsky et al., 2015) showed that unitary RNNs can overcome the exploding and vanishing gradient problems and
perform well on long term memory tasks if the hiddento-hidden matrix in parametrized in the following unitary
form:
W = D3 T2 F −1 D2 ΠT1 FD1 .
(6)
Here D1,2,3 are diagonal matrices with each element
eiωj , j = 1, 2, · · · , n. T1,2 are reflection matrices, and
bv
b†
v
b
T = I − 2 ||b
v||2 , where v is a vector with each of its entries as a parameter to be trained. Π is a fixed permutation
matrix. F and F −1 are Fourier and inverse Fourier transform matrices respectively. Since each factor matrix here
is unitary, the product W is also a unitary matrix.
This model uses O(N ) parameters, which spans merely
a part of the whole O(N 2 )-dimensional space of unitary
N × N matrices to enable computational efficiency. Several subsequent papers have tried to expand the space to
O(N 2 ) in order to achieve better performance, as summarized below.
3.2. Full Space Unitary RNNs
In order to maximize the power of Unitary RNNs, it is
preferable to have the option to optimize the weight matrix W over the full space of unitary matrices rather than
a subspace as above. A straightforward method for implementing this is by simply updating W with standard backpropagation and then projecting the resulting matrix (which
will typically no longer be unitary) back onto to the space
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
∂C
of unitary matrices. Defining Gij ≡ ∂W
as the gradient
ij
with respect to W, this can be implemented by the procedure defined by (Wisdom et al., 2016):
A(t)
W(t+1)
†
†
(7)
≡ G(t) W(t) − W(t) G(k) ,
−1
λ
λ
≡
I + A(t)
I − A(t) W(t) .(8)
2
2
RN j for j = N − 1, · · · , 1. Once all elements of the last
row except the one on the diagonal are zero, this row will
not be affected by later transformations. Since all transformations are unitary, the last column will then also contain
only zeros except on the diagonal:
WN −1
0
WN RN,N −1 RN,N −2 · ·RN,1 =
(10)
0
eiwN
This method shows that full space unitary networks are superior on many RNN tasks (Wisdom et al., 2016). A key
limitation is that the back-propation in this method cannot avoid N -dimensional matrix multiplication, incurring
O(N 3 ) computational cost.
)
W
4. Efficient Unitary Neural Network (EUNN)
Architectures
In the following, we first describe a general parametrization
method able to represent arbitrary unitary matrices with up
to N 2 degrees of freedom. We then present an efficient
algorithm for this parametrization scheme, requiring only
O(1) computational and memory access steps to obtain the
gradient for each parameter. Finally, we show that our
scheme performs significantly better than the above mentioned methods on a few well-known benchmarks.
4.1. Unitary Matrix Parametrization
Any N × N unitary matrix WN can be represented as a
product of rotation matrices {Rij } and a diagonal matrix
QN Qi−1
D, such that WN = D i=2 j=1 Rij , where Rij is
defined as the N -dimensional identity matrix with the elements Rii , Rij , Rji and Rjj replaced as follows (Reck
et al., 1994; Clements et al., 2016):
iφ
Rii Rij
e ij cos θij −eiφij sin θij
=
. (9)
Rji Rjj
sin θij
cos θij
where θij and φij are unique parameters corresponding
to Rij . Each of these matrices performs a U (2) unitary
transformation on a two-dimensional subspace of the Ndimensional Hilbert space, leaving an (N −2)-dimensional
subspace unchanged. In other words, a series of U (2) rotations can be used to successively make all off-diagonal
elements of the given N × N unitary matrix zero. This
generalizes the familiar factorization of a 3D rotation matrix into 2D rotations parametrized by the three Euler angles. To provide intuition for how this works, let us briefly
describe a simple way of doing this that is similar to Gaussian elimination by finishing one column at a time. There
are infinitely many alternative decomposition schemes as
well; Fig. 1 shows two that are particularly convenient to
implement in software (and even in neuromorphic hardware (Shen et al., 2016)). The unitary matrix WN is multiplied from the right by a succession of unitary matrices
)
)
R
Figure 1. Unitary matrix decomposition: An arbitrary unitary
matrix W can be decomposed (a) with the square decomposition method of Clements et al. (Clements et al., 2016) discussed
in section 4.2; or approximated (b) by the Fast Fourier Transformation(FFT) style decomposition method (Mathieu & LeCun,
2014b) discussed in section 4.3. Each junction in the a) and b)
graphs above represent the U(2) matrix as shown in c).
The effective dimensionality of the the matrix W is thus
reduced to N −1. The same procedure can then be repeated
N − 1 times until the effective dimension of W is reduced
to 1, leaving us with a diagonal matrix:1
WN RN,N −1 RN,N −2 · · · Ri,j Ri,j−1 · · · R3,1 R2,1 = D,
(11)
where D is a diagonal matrix whose diagonal elements are
eiwj , from which we can write the direct representation of
WN as
WN
−1
−1
−1
= DR−1
2,1 R3,1 . . . RN,N −2 RN,N −1
= DR02,1 R03,1 . . . R0N,N −2 R0N,N −1 .
(12)
where
R0ij = R(−θij , −φij ) = R(θij , φij )−1 = R−1
ij
(13)
1
Note that Gaussian Elimination would make merely the upper triangle of a matrix vanish, requiring a subsequent series of
rotations (complete Gauss-Jordan Elimination) to zero the lower
triangle. We need no such subsequent series because since W is
unitary: it is easy to show that if a unitary matrix is triangular, it
must be diagonal.
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
This parametrization thus involves N (N − 1)/2 different
θij -values, N (N − 1)/2 different φij -values and N different wi -values, combining to N 2 parameters in total and
spans the entire unitary space. Note we can always fix a
portion of our parameters, to span only a subset of unitary
space – indeed, our benchmark test below will show that
for certain tasks, full unitary space parametrization is not
necessary. 2
4.2. Tunable space implementation
The representation in Eq. 12 can be made more compact
by reordering and grouping specific rotational matrices, as
was shown in the optical community (Reck et al., 1994;
Clements et al., 2016) in the context of universal multiport
interferometers. For example (Clements et al., 2016), a unitary matrix can be decomposed as
(1) (1)
(1)
WN = D R1,2 R3,4 . . . RN/2−1,N/2
(2) (2)
(2)
× R2,3 R4,5 . . . RN/2−2,N/2−1
×...
(1)
(2)
where p = 2Ni , k ∈ {0, ..., 2i−1 } and j ∈ {1, ..., p}.
This requires only log(N ) matrices, so there are a total
of N log(N )/2 rotational pairs. This is also the minimal
number of rotations that can have all input coordinates interacting with each other, providing an approximation of
arbitrary unitary matrices.
4.4. Efficient implementation of rotation matrices
To implement this decomposition efficiently in an RNN,
we apply vector element-wise multiplications and permutations: we evaluate the product Fx as
Fx = v1 ∗ x + v2 ∗ permute(x)
where ∗ represents element-wise multiplication, F refers to
general rotational matrices such as FA/B in Eq. 14 and Fi
in Eq. 16. For the case of the tunable-space implementa(l)
tion, if we want to implement FA in Eq. 14, we define v
and the permutation as follows:
(l)
(14)
(l)
(l)
(l)
(l)
(l)
(l)
(l)
(l)
v2 = (−eiφ1 sin θ1 , sin θ1 , −eiφ2 sin θ2 , sin θ2 , · · · )
permute(x) = (x2 , x1 , x4 , x3 , x6 , x5 , · · · ).
where every
(l)
(l)
(l)
(l)
FA = R1,2 R3,4 . . . RN/2−1,N/2
is a block diagonal matrix, with N angle parameters in total, and
(l)
FB
(l)
v1 = (eiφ1 cos θ1 , cos θ1 , eiφ2 cos θ2 , cos θ2 , · · · )
(l)
(L)
= DFA FB . . . FB ,
(18)
=
(l)
(l)
R2,3 R4,5
(l)
. . . RN/2−2,N/2−1
Following this physics-inspired scheme, we decompose our
unitary hidden-to-hidden layer matrix W as
(1)
(2)
(3)
(4)
(L)
(15)
4.3. FFT-style approximation
Inspired by (Mathieu & LeCun, 2014a), an alternative way
to organize the rotation matrices is implementing an FFTstyle architecture. Instead of using adjacent rotation matrices, each F here performs a certain distance pairwise rotations as shown in Fig. 1b:
W = DF1 F2 F3 F4 · · · Flog(N ) .
(l)
(l)
(l)
(l)
(l)
(l)
(l)
(l)
v1 = (eiφ1 cos θ1 , eiφ2 cos θ2 , · · · , cos θ1 , cos θ2 , · · · )
(l)
with N −1 parameters, as is schematically shown in Fig. 1a.
By choosing different values for L , WN will span a different subspace of the unitary space. Specifically,when
L = N , WN will span the entire unitary space.
W = DFA FB FA FB · · · FB .
For the FFT-style approach, if we want to implement F1 in
Eq 16, we define v and the permutation as follows:
(l)
(l)
v2 = (−eiφ1 sin θ1 , −eiφ2 sin θ2 , · · · , sin θ1 , sin θ2 , · · · )
permute(x) = (x n2 +1 , x n2 +2 · · · xn , x1 , x2 · · · ).
In general, the pseudocode for implementing operation F
is as follows:
Algorithm 1 Efficient implementation for F with parameter θi and φi .
Input: input x, size N ; parameters θ and φ, size N/2;
constant permuatation index list ind1 and ind2 .
Output: output y, size N .
v1 ← concatenate(cos θ, cos θ * exp(iφ))
v2 ← concatenate(sin θ, - sin θ * exp(iφ))
v1 ← permute(v1 , ind1 )
v2 ← permute(v2 , ind1 )
y ← v1 ∗ x + v2 ∗ permute(x, ind2 )
(16)
Note that ind1 and ind2 are different for different F.
The rotation matrices in Fi are performed between pairs of
coordinates
(2pk + j, p(2k + 1) + j)
(17)
2
Our preliminary experimental tests even suggest that a fullcapacity unitary RNN is even undesirable for some tasks.
From a computational complexity viewpoint, since the operations ∗ and permute take O(N ) computational steps,
evaluating Fx only requires O(N ) steps. The product Dx
is trivial, consisting of an element-wise vector multiplication. Therefore, the product Wx with the total unitary
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
matrix W can be computed in only O(N L) steps, and
only requires O(N L) memory access (for full-space implementation L = N , for FFT-style approximation gives
L = log N ). A detailed comparison on computational
complexity of the existing unitary RNN architectures is
given in Table 1.
4.5. Nonlinearity
We use the same nonlinearity as (Arjovsky et al., 2015):
(modReLU(z, b))i =
zi
∗ ReLU(|zi | + bi )
|zi |
(19)
where the bias vector b is a shared trainable parameter, and
|zi | is the norm of the complex number zi .
For real number input, modReLU can be simplified to:
(modReLU(z, b))i = sign(zi ) ∗ ReLU(|zi | + bi ) (20)
where |zi | is the absolute value of the real number zi .
We empirically find that this nonlinearity function performs
the best. We believe that this function possibly also serves
as a forgetting filter that removes the noise using the bias
threshold.
5. Experimental tests of our method
In this section, we compare the performance of our Efficient Unitary Recurrent Neural Network (EURNN) with
1. an LSTM RNN (Hochreiter & Schmidhuber, 1997),
2. a Partial Space URNN (Arjovsky et al., 2015), and
3. a Projective full-space URNN (Wisdom et al., 2016).
All models are implemented in both Tensorflow and
Theano, available from https://github.com/
jingli9111/EUNN-tensorflow and https:
//github.com/iguanaus/EUNN-theano.
5.1. Copying Memory Task
We compare these networks by applying them all to the
well defined Copying Memory Task (Hochreiter & Schmidhuber, 1997; Arjovsky et al., 2015; Henaff et al., 2016).
The copying task is a synthetic task that is commonly used
to test the network’s ability to remember information seen
T time steps earlier.
Specifically, the task is defined as follows (Hochreiter &
Schmidhuber, 1997; Arjovsky et al., 2015; Henaff et al.,
2016). An alphabet consists of symbols {ai }, the first n of
which represent data, and the remaining two representing
“blank” and “start recall”, respectively; as illustrated by the
following example where T = 20 and M = 5:
Input: BACCA--------------------:---Output: -------------------------BACCA
In the above example, n = 3 and {ai } = {A, B, C, −, :}.
The input consists of M random data symbols (M = 5
above) followed by T − 1 blanks, the “start recall” symbol
and M more blanks. The desired output consists of M + T
blanks followed by the data sequence. The cost function
C is defined as the cross entropy of the input and output
sequences, which vanishes for perfect performance.
We use n = 8 and input length M = 10. The symbol
for each input is represented by an n-dimensional one-hot
vector. We trained all five RNNs for T = 1000 with the
same batch size 128 using RMSProp optimization with a
learning rate of 0.001. The decay rate is set to 0.5 for EURNN, and 0.9 for all other models respectively. (Fig. 2).
This results show that the EURNN architectures introduced
in both Sec.4.2 (EURNN with N=512, selecting L=2) and
Sec.4.3 (FFT-style EURNN with N=512) outperform the
LSTM model (which suffers from long term memory problems and only performs well on the copy task for small time
delays T ) and all other unitary RNN models, both in-terms
of learnability and in-terms of convergence rate. Note that
the only other unitary RNN model that is able to beat the
baseline for T = 1000 (Wisdom et al., 2016) is significantly slower than our method.
Moreover, we find that by either choosing smaller L or by
using the FFT-style method (so that W spans a smaller unitary subspace), the EURNN converges toward optimal performance significantly more efficiently (and also faster in
wall clock time) than the partial (Arjovsky et al., 2015) and
projective (Wisdom et al., 2016) unitary methods. The EURNN also performed more robustly. This means that a fullcapacity unitary matrix is not necessary for this particular
task.
5.2. Pixel-Permuted MNIST Task
The MNIST handwriting recognition problem is one of the
classic benchmarks for quantifying the learning ability of
neural networks. MNIST images are formed by a 28×28
grayscale image with a target label between 0 and 9.
To test different RNN models, we feed all pixels of the
MNIST images into the RNN models in 28×28 time steps,
where one pixel at a time is fed in as a floating-point number. A fixed random permutation is applied to the order
of input pixels. The output is the probability distribution
quantifying the digit prediction. We used RMSProp with a
learning rate of 0.0001 and a decay rate of 0.9, and set the
batch size to 128.
As shown in Fig. 3, EURNN significantly outperforms
LSTM with the same number of parameters. It learns faster,
in fewer iteration steps, and converges to a higher classifi-
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
Table 1. Performance comparison of four Recurrent Neural Network algorithms: URNN (Arjovsky et al., 2015), PURNN (Wisdom et al.,
2016), and EURNN (our algorithm). T denotes the RNN length and N denotes the hidden state size. For the tunable-style EURNN, L
is an integer between 1 and N parametrizing the unitary matrix capacity.
Model
URNN
PURNN
EURNN (tunable style)
EURNN (FFT style)
Time complexity of one
online gradient step
O(T N log N )
O(T N 2 + N 3 )
O(T N L)
O(T N log N )
number of parameters
in the hidden matrix
O(N )
O(N 2 )
O(N L)
O(N log N )
Transition matrix
search space
subspace of U(N )
full space of U(N )
tunable space of U(N )
subspace of U(N )
Table 2. MNIST Task result. EURNN corresponds to our algorithm, PURNN corresponds to algorithm presented in (Wisdom et al.,
2016), URNN corresponds to the algorithm presented in (Arjovsky et al., 2015).
Model
LSTM
URNN
PURNN
EURNN (tunable style)
EURNN (FFT style)
number of
parameters
16k
16k
16k
13.3k
9.0k
Copying Memory Task, delay time T=1000
EURNN with N=512 L=2
EURNN with N=512 FFT
PURNN with N=128
URNN with N=512
LSTM with N=80
baseline
0.08
0.07
0.06
0.05
0.04
0.03
0.02
validation
accuracy
0.908
0.942
0.922
0.940
0.928
0.9
0.8
0.7
0.6
EURNN with N=1024 L=2
EURNN with N=512 FFT
LSTM with N=80
0.01
0.00
0
500
1000
1500 2000 2500
Training iterations
3000
3500
4000
Figure 2. Copying Task for T = 1000. EURNN corresponds to
our algorithm, projective URNN corresponds to algorithm presented in (Wisdom et al., 2016), URNN corresponds to the algorithm presented in (Arjovsky et al., 2015). A useful baseline performance is that of the memoryless strategy, which outputs M +T
blanks followed by M random data symbols and produces a cross
entropy C = (M log n)/(T + 2 ∗ M ). [Note that each iteration
for PURNN takes about 32 times longer than for EURNN models, for this particular simulation, so the speed advantage is much
greater than apparent in this plot.]
cation accuracy. In addition, the EURNN reaches a similar
accuracy with fewer parameters. In Table. 2, we compare
the performance of different RNN models on this task.
test
accuracy
0.902
0.933
0.921
0.937
0.925
Permuted-Pixel MNIST Task
1.0
Accuracy
Cross Entropy
hidden size
(capacity)
80
512
116
1024 (2)
512 (FFT)
0.5
0
5000
10000
15000 20000
Training iterations
25000
30000
Figure 3. Pixel-permuted MNIST performance on the validation
dataset.
5.3. Speech Prediction on TIMIT dataset
We also apply our EURNN to real-world speech prediction task and compare its performance to LSTM. The main
task we consider is predicting the log-magnitude of future
frames of a short-time Fourier transform (STFT) (Wisdom
et al., 2016; Sejdi et al., 2009). We use the TIMIT dataset
(Garofolo et al., 1993) sampled at 8 kHz. The audio .wav
file is initially diced into different time frames (all frames
have the same duration referring to the Hann analysis window below). The audio amplitude in each frame is then
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
Table 3. Speech Prediction Task result. EURNN corresponds to our algorithm, projective URNN corresponds to algorithm presented in
(Wisdom et al., 2016), URNN corresponds to the algorithm presented in (Arjovsky et al., 2015).
Model
LSTM
LSTM
EURNN (tunable style)
EURNN (tunable style)
EURNN (tunable style)
EURNN (FFT style)
hidden size
(capacity)
64
128
128 (2)
128 (32)
128 (128)
128 (FFT)
number of
parameters
33k
98k
33k
35k
41k
34k
MSE
(validation)
71.4
55.3
63.3
52.3
51.8
52.3
MSE
(test)
66.0
54.5
63.3
52.7
51.9
52.4
Figure 4. Example spectrograms of ground truth and RNN prediction results from evaluation sets.
Fourier transformed into the frequency domain. The logmagnitude of the Fourier amplitude is normalized and used
as the data for training/testing each model. In our STFT
operation we uses a Hann analysis window of 256 samples (32 milliseconds) and a window hop of 128 samples
(16 milliseconds). The frame prediction task is as follows:
given all the log-magnitudes of STFT frames up to time t,
predict the log-magnitude of the STFT frame at time t + 1
that has the minimum mean square error (MSE). We use
a training set with 2400 utterances, a validation set of 600
utterances and an evaluation set of 1000 utterances. The
training, validation, and evaluation sets have distinct speakers. We trained all RNNs for with the same batch size 32
using RMSProp optimization with a learning rate of 0.001,
a momentum of 0.9 and a decay rate of 0.1.
The results are given in Table. 3, in terms of the meansquared error (MSE) loss function. Figure. 4 shows prediction examples from the three types of networks, illustrat-
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
ing how EURNNs generally perform better than LSTMs.
Furthermore, in this particular task, full-capacity EURNNs
outperform small capacity EURNNs and FFT-style EURNNs.
References
6. Conclusion
Bahdanau, Dzmitry, Cho, Kyunghyun, and Bengio,
Yoshua. Neural machine translation by jointly learning
to align and translate. arXiv preprint arXiv:1409.0473,
2014.
We have presented a method for implementing an Efficient
Unitary Neural Network (EUNN) whose computational
cost is merely O(1) per parameter, which is O(log N )
more efficient than the other methods discussed above. It
significantly outperforms existing RNN architectures on
the standard Copying Task, and the pixel-permuted MNIST
Task using a comparable parameter count, hence demonstrating the highest recorded ability to memorize sequential
information over long time periods.
It also performs well on real tasks such as speech prediction, outperforming an LSTM on TIMIT data speech prediction.
We want to emphasize the generality and tunability of our
method. The ordering of the rotation matrices we presented
in Fig. 1 are merely two of many possibilities; we used it
simply as a concrete example. Other ordering options that
can result in spanning the full unitary matrix space can be
used for our algorithm as well, with identical speed and
memory performance. This tunability of the span of the
unitary space and, correspondingly, the total number of parameters makes it possible to use different capacities for
different tasks, thus opening the way to an optimal performance of the EUNN. For example, as we have shown, a
small subspace of the full unitary space is preferable for the
copying task, whereas the MNIST task and TIMIT task are
better performed by EUNN covering a considerably larger
unitary space. Finally, we note that our method remains
applicable even if the unitary matrix is decomposed into a
different product of matrices (Eq. 12).
This powerful and robust unitary RNN architecture also
might be promising for natural language processing because of its ability to efficiently handle tasks with long-term
correlation and very high dimensionality.
Acknowledgment
We thank Hugo Larochelle and Yoshua Bengio for helpful
discussions and comments.
This work was partially supported by the Army Research
Office through the Institute for Soldier Nanotechnologies
under contract W911NF-13-D0001, the National Science
Foundation under Grant No. CCF-1640012 and the Rothberg Family Fund for Cognitive Science.
Arjovsky, Martin, Shah, Amar, and Bengio, Yoshua. Unitary evolution recurrent neural networks. arXiv preprint
arXiv:1511.06464, 2015.
Bengio, Yoshua, Simard, Patrice, and Frasconi, Paolo.
Learning long-term dependencies with gradient descent
is difficult. IEEE transactions on neural networks, 5(2):
157–166, 1994.
Berglund, Mathias, Raiko, Tapani, Honkala, Mikko,
Kärkkäinen, Leo, Vetek, Akos, and Karhunen, Juha T.
Bidirectional recurrent neural networks as generative
models. In Advances in Neural Information Processing
Systems, pp. 856–864, 2015.
Cho, Kyunghyun, Van Merriënboer, Bart, Bahdanau,
Dzmitry, and Bengio, Yoshua. On the properties of neural machine translation: Encoder-decoder approaches.
arXiv preprint arXiv:1409.1259, 2014.
Clements, William R., Humphreys, Peter C., Metcalf, Benjamin J., Kolthammer, W. Steven, and Walmsley, Ian A.
An optimal design for universal multiport interferometers, 2016. arXiv:1603.08788.
Collobert, Ronan, Weston, Jason, Bottou, Léon, Karlen,
Michael, Kavukcuoglu, Koray, and Kuksa, Pavel. Natural language processing (almost) from scratch. Journal
of Machine Learning Research, 12(Aug):2493–2537,
2011.
Donahue, Jeffrey, Anne Hendricks, Lisa, Guadarrama,
Sergio, Rohrbach, Marcus, Venugopalan, Subhashini,
Saenko, Kate, and Darrell, Trevor. Long-term recurrent convolutional networks for visual recognition and
description. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 2625–
2634, 2015.
Garofolo, John S, Lamel, Lori F, Fisher, William M, Fiscus,
Jonathon G, and Pallett, David S. Darpa timit acousticphonetic continous speech corpus cd-rom. nist speech
disc 1-1.1. NASA STI/Recon technical report n, 93, 1993.
Henaff, Mikael, Szlam, Arthur, and LeCun, Yann. Orthogonal rnns and long-memory tasks. arXiv preprint
arXiv:1602.06662, 2016.
Hinton, Geoffrey, Deng, Li, Yu, Dong, Dahl, George E,
Mohamed, Abdel-rahman, Jaitly, Navdeep, Senior, Andrew, Vanhoucke, Vincent, Nguyen, Patrick, Sainath,
Tunable Efficient Unitary Neural Networks (EUNN) and their application to RNNs
Tara N, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four
research groups. IEEE Signal Processing Magazine, 29
(6):82–97, 2012.
Sutskever, Ilya, Vinyals, Oriol, and Le, Quoc V. Sequence to sequence learning with neural networks. In
Advances in neural information processing systems, pp.
3104–3112, 2014.
Hochreiter, Sepp.
Untersuchungen zu dynamischen
neuronalen netzen. Diploma, Technische Universität
München, pp. 91, 1991.
Wisdom, Scott, Powers, Thomas, Hershey, John, Le Roux,
Jonathan, and Atlas, Les. Full-capacity unitary recurrent neural networks. In Advances In Neural Information
Processing Systems, pp. 4880–4888, 2016.
Hochreiter, Sepp and Schmidhuber, Jürgen. Long shortterm memory. Neural computation, 9(8):1735–1780,
1997.
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E.
Imagenet classification with deep convolutional neural
networks. In Advances in neural information processing
systems, pp. 1097–1105, 2012.
Le, Quoc V, Jaitly, Navdeep, and Hinton, Geoffrey E. A
simple way to initialize recurrent networks of rectified
linear units. arXiv preprint arXiv:1504.00941, 2015.
LeCun, Yann, Bengio, Yoshua, and Hinton, Geoffrey. Deep
learning. Nature, 521(7553):436–444, 2015.
Mathieu, Michael and LeCun, Yann. Fast approximation of rotations and hessians matrices. arXiv preprint
arXiv:1404.7195, 2014a.
Mathieu, Michal and LeCun, Yann. Fast approximation of
rotations and hessians matrices. CoRR, abs/1404.7195,
2014b.
URL http://arxiv.org/abs/1404.
7195.
Reck, Michael, Zeilinger, Anton, Bernstein, Herbert J.,
and Bertani, Philip.
Experimental realization of
any discrete unitary operator.
Phys. Rev. Lett.,
73:58–61, Jul 1994.
doi: 10.1103/PhysRevLett.
73.58. URL http://link.aps.org/doi/10.
1103/PhysRevLett.73.58.
Saxe, Andrew M, McClelland, James L, and Ganguli,
Surya. Exact solutions to the nonlinear dynamics of
learning in deep linear neural networks. arXiv preprint
arXiv:1312.6120, 2013.
Sejdi, Ervin, Djurovi, Igor, and Jiang, Jin. Timefrequency feature representation using energy concentration: An overview of recent advances. Digital Signal Processing, 19(1):153 – 183, 2009. ISSN 10512004.
doi: http://dx.doi.org/10.1016/j.dsp.2007.12.
004. URL http://www.sciencedirect.com/
science/article/pii/S105120040800002X.
Shen, Yichen, Harris, Nicholas C, Skirlo, Scott, Prabhu,
Mihika, Baehr-Jones, Tom, Hochberg, Michael, Sun,
Xin, Zhao, Shijie, Larochelle, Hugo, Englund, Dirk,
et al. Deep learning with coherent nanophotonic circuits.
arXiv preprint arXiv:1610.02365, 2016.
| 9cs.NE
|
Under review as a conference paper at ICLR 2017
N EURO -S YMBOLIC P ROGRAM S YNTHESIS
Emilio Parisotto1,2 , Abdel-rahman Mohamed1 , Rishabh Singh1 ,
Lihong Li1 , Dengyong Zhou1 , Pushmeet Kohli1
1
Microsoft Research, USA 2 Carnegie Mellon University, USA
[email protected] , {asamir,risin,lihongli,denzho,pkohli}@microsoft.com
arXiv:1611.01855v1 [cs.AI] 6 Nov 2016
A BSTRACT
Recent years have seen the proposal of a number of neural architectures for the
problem of Program Induction. Given a set of input-output examples, these architectures are able to learn mappings that generalize to new test inputs. While
achieving impressive results, these approaches have a number of important limitations: (a) they are computationally expensive and hard to train, (b) a model has
to be trained for each task (program) separately, and (c) it is hard to interpret or
verify the correctness of the learnt mapping (as it is defined by a neural network).
In this paper, we propose a novel technique, Neuro-Symbolic Program Synthesis,
to overcome the above-mentioned problems. Once trained, our approach can automatically construct computer programs in a domain-specific language that are
consistent with a set of input-output examples provided at test time. Our method
is based on two novel neural modules. The first module, called the cross correlation I/O network, given a set of input-output examples, produces a continuous
representation of the set of I/O examples. The second module, the RecursiveReverse-Recursive Neural Network (R3NN), given the continuous representation
of the examples, synthesizes a program by incrementally expanding partial programs. We demonstrate the effectiveness of our approach by applying it to the
rich and complex domain of regular expression based string transformations. Experiments show that the R3NN model is not only able to construct programs from
new input-output examples, but it is also able to construct new programs for tasks
that it had never observed before during training.
1
I NTRODUCTION
The act of programming, i.e., developing a procedure to accomplish a task, is a remarkable demonstration of the reasoning abilities of the human mind. Expectedly, Program Induction is considered
as one of the fundamental problems in Machine Learning and Artificial Intelligence. Recent progress
on deep learning has led to the proposal of a number of promising neural architectures for this problem. Many of these models are inspired from computation modules (CPU, RAM, GPU) (Graves
et al., 2014; Kurach et al., 2015; Reed & de Freitas, 2015; Neelakantan et al., 2015) or common
data structures used in many algorithms (stack) (Joulin & Mikolov, 2015). A common thread in this
line of work is to specify the atomic operations of the network in some differentiable form, allowing
efficient end-to-end training of a neural controller, or to use reinforcement learning to make hard
choices about which operation to perform. While these results are impressive, these approaches
have a number of important limitations: (a) they are computationally expensive and hard to train, (b)
a model has to be trained for each task (program) separately, and (c) it is hard to interpret or verify
the correctness of the learnt mapping (as it is defined by a neural network). While some recently
proposed methods (Kurach et al., 2015; Gaunt et al., 2016; Riedel et al., 2016; Bunel et al., 2016)
do learn interpretable programs, they still need to learn a separate neural network model for each
individual task.
Motivated by the need for model interpretability and scalability to multiple tasks, we address the
problem of Program Synthesis. Program Synthesis, the problem of automatically constructing programs that are consistent with a given specification, has long been a subject of research in Computer
Science (Biermann, 1978; Summers, 1977). This interest has been reinvigorated in recent years on
1
Under review as a conference paper at ICLR 2017
the back of the development of methods for learning programs in various domains, ranging from
low-level bit manipulation code (Solar-Lezama et al., 2005) to data structure manipulations (Singh
& Solar-Lezama, 2011) and regular expression based string transformations (Gulwani, 2011).
Most of the recently proposed methods for program synthesis operate by searching the space of
programs in a Domain-Specific Language (DSL) instead of arbitrary Turing-complete languages.
This hypothesis space of possible programs is huge (potentially infinite) and searching over it is a
challenging problem. Several search techniques including enumerative (Udupa et al., 2013), stochastic (Schkufza et al., 2013), constraint-based (Solar-Lezama, 2008), and version-space algebra based
algorithms (Gulwani et al., 2012) have been developed to search over the space of programs in the
DSL, which support different kinds of specifications (examples, partial programs, natural language
etc.) and domains. These techniques not only require significant engineering and research effort to
develop carefully-designed heuristics for efficient search, but also have limited applicability and can
only synthesize programs of limited sizes and types.
In this paper, we present a novel technique called Neuro-Symbolic Program Synthesis (NSPS) that
learns to generate a program incrementally without the need for an explicit search. Once trained,
NSPS can automatically construct computer programs that are consistent with any set of input-output
examples provided at test time. Our method is based on two novel module neural architectures. The
first module, called the cross correlation I/O network, produces a continuous representation of any
given set of input-output examples. The second module, the Recursive-Reverse-Recursive Neural
Network (R3NN), given the continuous representation of the input-output examples, synthesizes a
program by incrementally expanding partial programs. R3NN employs a tree-based neural architecture that sequentially constructs a parse tree by selecting which non-terminal symbol to expand
using rules from a context-free grammar (i.e., the DSL).
We demonstrate the efficacy of our method by applying it to the rich and complex domain of regularexpression-based syntactic string transformations, using a DSL based on the one used by FlashFill (Gulwani, 2011; Gulwani et al., 2012), a Programming-By-Example (PBE) system in Microsoft
Excel 2013. Given a few input-output examples of strings, the task is to synthesize a program built
on regular expressions to perform the desired string transformation. An example task that can be
expressed in this DSL is shown in Figure 1, which also shows the DSL.
Our evaluation shows that NSPS is not only able to construct programs for known tasks from new
input-output examples, but it is also able to construct completely new programs that it had not observed during training. Specifically, the proposed system is able to synthesize string transformation
programs for 63% of tasks that it had not observed at training time, and for 94% of tasks when
100 program samples are taken from the model. Moreover, our system is able to learn 38% of 238
real-world FlashFill benchmarks.
To summarize, the key contributions of our work are:
• A novel Neuro-Symbolic program synthesis technique to encode neural search over the
space of programs defined using a Domain-Specific Language (DSL).
• The R3NN model that encodes and expands partial programs in the DSL, where each node
has a global representation of the program tree.
• A novel cross-correlation based neural architecture for learning continuous representation
of sets of input-output examples.
• Evaluation of the NSPS approach on the complex domain of regular expression based string
transformations.
2
P ROBLEM D EFINITION
In this section, we formally define the DSL-based program synthesis problem that we consider in this
paper. Given a DSL L, we want to automatically construct a synthesis algorithm A such that given
a set of input-output example, {(i1 , o1 ), · · · , (in , on )}, A returns a program P ∈ L that conforms
to the input-output examples, i.e.,
∀j : 1 ≤ j ≤ n P (ij ) = oj .
2
(1)
Under review as a conference paper at ICLR 2017
1
2
3
4
5
Input v
William Henry Charles
Michael Johnson
Barack Rogers
Martha D. Saunders
Peter T Gates
String e
Substring f
Output
Charles, W.
Johnson, M.
Rogers, B.
Saunders, M.
Gates, P.
Position p
Direction Dir
Regex r
(a)
:=
:=
|
:=
|
:=
:=
Concat(f1 , · · · , fn )
ConstStr(s)
SubStr(v, pl , pr )
(r, k, Dir)
ConstPos(k)
Start | End
s | T1 · · · | Tn
(b)
Figure 1: An example FlashFill task for transforming names to lastname with initials of first name,
and (b) The DSL for regular expression based string transformations.
The syntax and semantics of the DSL for string transformations is shown in Figure 1(b) and Figure 7
respectively. The DSL corresponds to a large subset of FlashFill DSL (except conditionals), and
allows for a richer class of substring operations than FlashFill. A DSL program takes as input a
string v and returns an output string o. The top-level string expression e is a concatenation of a
finite list of substring expressions f1 , · · · , fn . A substring expression f can either be a constant
string s or a substring expression, which is defined using two position logics pl (left) and pr (right).
A position logic corresponds to a symbolic expression that evaluates to an index in the string. A
position logic p can either be a constant position k or a token match expression (r, k, Dir), which
denotes the Start or End of the k th match of token r in input string v. A regex token can either be a
constant string s or one of 8 regular expression tokens: p (ProperCase), C (CAPS), l (lowercase), d
(Digits), α (Alphabets), αn (Alphanumeric), ∧ (StartOfString), and $ (EndOfString). The semantics
of the DSL programs is described in the appendix.
A DSL program for the name transformation task shown in Figure 1(a) that is consistent with the examples is: Concat(f1 , ConstStr(“, ”), f2 , ConstStr(“.”)), where f1 ≡
SubStr(v, (“ ”, −1, End), ConstPos(−1)) and f2 ≡ SubStr(v, ConstPos(0), ConstPos(1)). The
program concatenates the following 4 strings: i) substring between the end of last whitespace and
end of string, ii) constant string “, ”, iii) first character of input string, and iv) constant string “.”.
3
OVERVIEW OF OUR A PPROACH
We now present an overview of our approach. Given a DSL L, we learn a generative model of programs in the DSL L that is conditioned on input-output examples to efficiently search for consistent
programs. The workflow of our system is shown in Figure 2, which is trained end-to-end using a
large training set of programs in the DSL together with their corresponding input-output examples.
To generate a large training set, we uniformly sample programs from the DSL and then use a rulebased strategy to compute well-formed input strings that satisfy the pre-conditions of the programs.
The corresponding output strings are obtained by running the programs on the input strings.
A DSL can be considered a context-free grammar with a start symbol S and a set of non-terminals
with corresponding expansion rules. The (partial) grammar derivations or trees correspond to (partial) programs. A naı̈ve way to perform a search over the programs in a DSL is to start from the start
symbol S and then randomly choose non-terminals to expand with randomly chosen expansion rules
until reaching a derivation with only terminals. We, instead, learn a generative model over partial
derivations in the DSL that assigns probabilities to different non-terminals in a partial derivation and
corresponding expansions to guide the search for complete derivations.
Our generative model uses a Recursive-Reverse-Recursive Neural Network (R3NN) to encode partial trees (derivations) in L, where each node in the partial tree encodes global information about
every other node in the tree. The model assigns a vector representation for every symbol and every
expansion rule in the grammar. Given a partial tree, the model first assigns a vector representation
to each leaf node, and then performs a recursive pass going up in the tree to assign a global tree
representation to the root. It then performs a reverse-recursive pass starting from the root to assign
a global tree representation to each node in the tree.
3
Under review as a conference paper at ICLR 2017
{
{
{
p1
DSL
Program
Sampler
Input Gen Rules
DSL
i1 – o 1
i2 – o 2
…
ik – o k
…
pj
i1 – o 1
i2 – o 2
…
ik – o k
…
pn
i1 – o 1
i2 – o 2
…
ik – o k
DSL
pj,0
i1 – o1
i2 – o2
…
ik – ok
R3NN
DSL
I/O Encoder
R3NN
..
DSL
pj,1
R3NN
DSL
I/O Encoder
pj,2
R3NN
R3NN
..
DSL
R3NN
pj
Learnt program
(a) Training Phase
(b) Test Phase
Figure 2: An overview of the training and test workflow of our synthesis appraoch.
The generative process is conditioned on a set of input-output examples to learn a program that is
consistent with this set of examples. We experiment with multiple input-output encoders including
an LSTM encoder that concatenates the hidden vectors of two deep bidirectional LSTM networks
for input and output strings in the examples, and a Cross Correlation encoder that computes the cross
correlation between the LSTM tensor representations of input and output strings in the examples.
This vector is then used as an additional input in the R3NN model to condition the generative model.
4
T REE -S TRUCTURED G ENERATION M ODEL
We define a program t-steps into construction as a partial program tree (PPT) (see Figure 3 for a
visual depiction). A PPT has two types of nodes: leaf (symbol) nodes and inner non-leaf (rule)
nodes. A leaf node represents a symbol, whether non-terminal or terminal. An inner non-leaf node
represents a particular production rule of the DSL, where the number of children of the non-leaf
node is equivalent to the arity of the RHS of the rule it represents. A PPT is called a program tree
(PT) whenever all the leaves of the tree are terminal symbols. Such a tree represents a completed
program under the DSL and can be executed. We define an expansion as the valid application of
a specific production rule (e → e op2 e) to a specific non-terminal leaf node within a PPT (leaf
with symbol e). We refer to the specific production rule that an expansion is derived from as the
expansion type. It can be seen that if there exist two leaf nodes (l1 and l2 ) with the same symbol
then for every expansion specific to l1 there exists an expansion specific to l2 with the same type.
4.1
R ECURSIVE -R EVERSE -R ECURSIVE N EURAL N ETWORK
In order to define a generation model over PPTs, we need an efficient way of assigning probabilities
to every valid expansion in the current PPT. A valid expansion has two components: first the production rule used, and second the position of the expanded leaf node relative to every other node in
the tree. To account for the first component, a separate distributed representation for each production rule is maintained. The second component is handled using an architecture where the forward
propagation resembles belief propagation on trees, allowing a notion of global tree state at every
node within the tree. A given expansion probability is then calculated as being proportional to the
inner product between the production rule representation and the global-tree representation of the
leaf-level non-terminal node. We now describe the design of this architecture in more detail.
The R3NN has the following parameters for the grammar described by a DSL (see Figure 3):
1. For every symbol s ∈ S, an M -dimensional representation φ(s) ∈ RM .
2. For every production rule r ∈ R, an M −dimensional representation ω(r) ∈ RM .
3. For every production rule r ∈ R, a deep neural network fr which takes as input a vector
x ∈ RQ·M , with Q being the number of symbols on the RHS of the production rule r,
and outputs a vector y ∈ RM . Therefore, the production-rule network fr takes as input a
concatenation of the distributed representations of each of its RHS symbols and produces
a distributed representation for the LHS symbol.
4. For every production rule r ∈ R, an additional deep neural network gr which takes as
input a vector x0 ∈ RM and outputs a vector y 0 ∈ RQ·M . We can think of gr as a reverse
4
Under review as a conference paper at ICLR 2017
(a) Recursive pass
(b) Reverse-Recursive pass
Figure 3: (a) The initial recursive pass of the R3NN. (b) The reverse-recursive pass of the R3NN
where the input is the output of the previous recursive pass.
production-rule network that takes as input a vector representation of the LHS and produces
a concatenation of the distributed representations of each of the rule’s RHS symbols.
Let E be the set of all valid expansions in a PPT T , let L be the current leaf nodes of T and N be
the current non-leaf (rule) nodes of T . Let S(l) be the symbol of leaf l ∈ L and R(n) represent the
production rule of non-leaf node n ∈ N .
4.1.1
G LOBAL T REE I NFORMATION AT THE L EAVES
To compute the probability distribution over the set E, the R3NN first computes a distributed representation for each leaf node that contains global tree information. To accomplish this, for every
leaf node l ∈ L in the tree we retrieve its distributed representation φ(S(l)) . We now do a standard
recursive bottom-to-top, RHS→LHS pass on the network, by going up the tree and applying fR(n)
for every non-leaf node n ∈ N on its RHS node representations (see Figure 3(a)). These networks
fR(n) produce a node representation which is input into the parent’s rule network and so on until we
reach the root node.
Once at the root node, we effectively have a fixed-dimensionality global tree representation φ(root)
for the start symbol. The problem is that this representation has lost any notion of tree position. To
solve this problem R3NN now does what is effectively a reverse-recursive pass which starts at the
root node with φ(root) as input and moves towards the leaf nodes (see Figure 3(b)).
More concretely, we start with the root node representation φ(root) and use that as input into the
rule network gR(root) where R(root) is the production rule that is applied to the start symbol in
T . This produces a representation φ0 (c) for each RHS node c of R(root). If c is a non-leaf node,
we iteratively apply this procedure to c, i.e., process φ0 (c) using gR(c) to get representations φ0 (cc)
for every RHS node cc of R(c), etc. If c is a leaf node, we now have a leaf representation φ0 (c)
which has an information path to φ(root) and thus to every other leaf node in the tree. Once the
reverse-recursive process is complete, we now have a distributed representation φ0 (l) for every leaf
node l which contains global tree information. While φ(l1 ) and φ(l2 ) could be equal for leaf nodes
which have the same symbol type, φ0 (l1 ) and φ0 (l2 ) will not be equal even if they have the same
symbol type because they are at different positions in the tree.
4.1.2
E XPANSION P ROBABILITIES
Given the global leaf representations φ0 (l), we can now straightforwardly acquire scores for each
e ∈ E. For expansion e, let e.r be the expansion type (production rule r ∈ R that e applies) and
let e.l be the leaf node l that e.r is applied to. ze = φ0 (e.l) · ω(e.r) The score of an expansion is
calculated using ze = φ0 (e.l) · ω(e.r). The probability of expansion e is simply the exponentiated
ze
normalized sum over all scores: π(e) = P 0 e eze0 .
e ∈E
An additional improvement that was found to help was to add a bidirectional LSTM to process the
global leaf representations right before calculating the scores. The LSTM hidden states are then
5
Under review as a conference paper at ICLR 2017
used in the score calculation rather than the leaves themselves. This serves primarily to reduce the
minimum length that information has to propagate between nodes in the tree. The R3NN can be
seen as an extension and combination of several previous tree-based models, which were mainly
developed in the context of natural language processing (Le & Zuidema, 2014; Paulus et al., 2014;
Irsoy & Cardie, 2013).
5
C ONDITIONING WITH I NPUT /O UTPUT E XAMPLES
Now that we have defined a generation process over tree-structured programs, we need a way of
conditioning this generation process on a set of input/output examples. The set of input/output
examples provide a nearly complete specification for the desired output program, and so a good
encoding of the examples is crucial to the success of our program generator. For the most part, this
example encoding needs to be domain-specific, since different DSLs have different inputs (some
may operate over integers, some over strings, etc.). Therefore, in our case, we use an encoding
adapted to the input-output strings that our DSL operates over. We also investigate different ways of
conditioning program search on the learnt example input-output encodings.
5.1
E NCODING INPUT / OUTPUT EXAMPLES
There are two types of information that string manipulation programs need to extract from inputoutput examples: 1) constant strings, such as “@domain.com” or “.”, which appear in all output
examples; 2) substring indices in input where the index might be further defined by a regular expression. These indices determine which parts of the input are also present in the output. To simplify the
DSL, we assume that there is a fixed finite universe of possible constant strings that could appear in
programs. Therefore we focus on extracting the second type of information, the substring indices.
In earlier hand-engineered systems such as FlashFill, this information was extracted from the inputoutput strings by running the Longest Common Substring algorithm, a dynamic programming algorithm that efficiently finds matching substrings in string pairs. To extract substrings, FlashFill runs
LCS on every input-output string pair in the I/O set to get a set of substring candidates. It then takes
the entire set of substring candidates and simply tries every possible regex and constant index that
can be used at substring boundaries, exhaustively searching for the one which is the most “general”,
where generality is specified by hand-engineered heuristics.
In contrast to these previous methods, instead of of hand-designing a complicated algorithm to
extract regex-based substrings, we develop neural network based architectures that are capable of
learning to extract and produce continuous representations of the likely regular expressions given
input/output strings.
5.1.1
BASELINE LSTM ENCODER
Our first I/O encoding network involves running two separate deep bidirectional LSTM networks for
processing the input and the output string in each example pair. For each pair, it then concatenates
the topmost hidden representation at every time step to produce a 4HT -dimensional feature vector
per I/O pair, where T is the maximum string length for any input or output string, and H is the
topmost LSTM hidden dimension.
We then concatenate the encoding vectors across all I/O pairs to get a vector representation of the entire I/O set. This encoding is conceptually straightforward and has very little prior knowledge about
what operations are being performed over the strings, i.e., substring, constant, etc., which might
make it difficult to discover substring indices, especially the ones based on regular expressions.
5.1.2
C ROSS C ORRELATION ENCODER
To help the model discover input substrings that are copied to the output, we designed an novel I/O
example encoder to compute the cross correlation between each input and output example representation. We used the two output tensors of the LSTM encoder (discussed above) as inputs to this
encoder. For each example pair, we first slide the output feature block over the input feature block
and compute the dot product between the respective position representation. Then, we sum over all
overlapping time steps. Features of all pairs are then concatenated to form a 2 ∗ (T − 1)-dimensional
6
Under review as a conference paper at ICLR 2017
vector encoding for all example pairs. There are 2 ∗ (T − 1) possible alignments in total between
input and output feature blocks. We also designed the following variants of this encoder.
Diffused Cross Correlation Encoder: This encoder is identical to the Cross Correlation encoder
except that instead of summing over overlapping time steps after the element-wise dot product, we
simply concatenate the vectors corresponding to all time steps, resulting in a final representation that
contains 2 ∗ (T − 1) ∗ T features for each example pair.
LSTM-Sum Cross Correlation Encoder: In this variant of the Cross Correlation encoder, instead
of doing an element-wise dot product, we run a bidirectional LSTM over the concatenated feature
blocks of each alignment. We represent each alignment by the LSTM hidden representation of the
final time step leading to a total of 2 ∗ H ∗ 2 ∗ (T − 1) features for each example pair.
Augmented Diffused Cross Correlation Encoder: For this encoder, the output of each character
position of the Diffused Cross Correlation encoder is combined with the character embedding at this
position, then a basic LSTM encoder is run over the combined features to extract a 4∗H-dimensional
vector for both the input and the output streams. The LSTM encoder output is then concatenated
with the output of the Diffused Cross Correlation encoder forming a (4∗H +T ∗(T −1))-dimensional
feature vector for each example pair.
5.2
C ONDITIONING PROGRAM SEARCH ON EXAMPLE ENCODINGS
Once the I/O example encodings have been computed, we can use them to perform conditional
generation of the program tree using the R3NN model. There are a number of ways in which the
PPT generation model can be conditioned using the I/O example encodings depending on where the
I/O example information is inserted in the R3NN model. We investigated three locations to inject
example encodings:
1) Pre-conditioning: where example encodings are concatenated to the encoding of each tree leaf,
and then passed to a conditioning network before the bottom-up recursive pass over the program
tree. The conditioning network can be either a multi-layer feedforward network, or a bidirectional
LSTM network running over tree leaves. Running an LSTM over tree leaves allows the model to
learn more about the relative position of each leaf node in the tree.
2) Post-conditioning: After the reverse-recursive pass, example encodings are concatenated to the
updated representation of each tree leaf and then fed to a conditioning network before computing
the expansion scores.
3) Root-conditioning: After the recursive pass over the tree, the root encoding is concatenated to
the example encodings and passed to a conditioning network. The updated root representation is
then used to drive the reverse-recursive pass.
Empirically, pre-conditioning worked better than either root- or post- conditioning. In addition,
conditioning at all 3 places simultaneously did not cause a significant improvement over just
pre-conditioning. Therefore, for the experimental section, we report models which only use preconditioning.
6
E XPERIMENTS
In order to evaluate and compare variants of the previously described models, we generate a dataset
randomly from the DSL. To do so, we first enumerate all possible programs under the DSL up to
a specific number of instructions, which are then partitioned into training, validation and test sets.
In order to have a tractable number of programs, we limited the maximum number of instructions
for programs to be 13. Length 13 programs are important for this specific DSL because all larger
programs can be written as compositions of sub-programs of length at most 13. The semantics of
length 13 programs therefore constitute the “atoms” of this particular DSL.
In testing our model, there are two different categories of generalization. The first is input/output
generalization, where we are given a new set of input/output examples as well as a program with a
specific tree that we have seen during training. This represents the model’s capacity to be applied
on new data. The second category is program generalization, where we are given both a previously
unseen program tree in addition to unseen input/output examples. Therefore the model needs to
7
Under review as a conference paper at ICLR 2017
I/O Encoding
LSTM
Cross Correlation (CC)
Diffused CC
LSTM-sum CC
Augmented diffused CC
Train
88%
67%
89%
90%
91%
Test
88%
65%
88%
91%
91%
Table 1: The effect of different input/output encoders on accuracy. Each result used 100 samples.
There is almost no generalization error in the results.
have a sufficient enough understanding of the semantics of the DSL that it can construct novel
combinations of operations. For all reported results, training sets correspond to the first type of
generalization since we have seen the program tree but not the input/output pairs. Test sets represent
the second type of generalization, as they are trees which have not been seen before on input/output
pairs that have also not been seen before.
In this section, we compare several different variants of our model. We first evaluate the effect of
each of the previously described input/output encoders. We then evaluate the R3NN model against a
simple recurrent model called io2seq, which is basically an LSTM that takes as input the input/output
conditioning vector and outputs a sequence of DSL symbols that represents a linearized program
tree. Finally, we report the results of the best model on the length 13 training and testing sets, as
well as on a set of 238 benchmark functions.
6.1
S ETUP AND HYPERPARAMETERS SETTINGS
For training the R3NN, two hyperparameters that were crucial for stabilizing training were the use
of hyperbolic tangent activation functions in both R3NN and cross-correlation I/O encoders and
the use of minibatches of length 8. Due to the difficulty of batching tree-based neural networks and
time-constraints, we were limited to 8 samples per batch but some preliminary experiments indicated
that increasing the batch size even further improved performance. Additionally, for all results, the
program tree generation is conditioned on a set of 10 input/output string pairs.
For each latent function and set of input/output examples that we test on, we report whether we had
a success after sampling 100 functions from the model and testing all 100 to see if one of these
functions is equivalent to the latent function. Here we consider two functions to be equivalent with
respect to a specific input/output example set if the functions output the same strings when run on
the inputs. Under this definition, two functions can have a different set of operations but still be
equivalent with respect to a specific input-output set.
6.2
E XAMPLE ENCODING
In this section, we evaluate the effect of several different input/output example encoders. To control
for the effect of the tree model, all results here used an R3NN with fixed hyperparameters to generate
the program tree. Table 1 shows the performance of several of these input/output example encoders.
We can see that the summed cross-correlation encoder did not perform well, which can be due to
the fact that the sum destroys positional information that might be useful for determining specific
substring indices. The LSTM-sum and the augmented diffused cross-correlation models did the
best. Surprisingly, the LSTM encoder was capable of finding nearly 88% of all programs without
having any prior knowledge explicitly built into the architecture. We use 100 samples for evaluating
the Train and Test sets. The training performance is sometimes slightly lower because there are
close to 5 million training programs but we only look at less than 2 million of these programs during
training. We sample a subset of only 1000 training programs from the 5 million program set to
report the training results in the tables. The test sets also consist of 1000 programs.
6.3
IO 2 SEQ
In this section, we motivate the use of the R3NN by testing whether a simpler model can also be
used to generate programs. The io2seq model is an LSTM whose initial hidden and cell states
8
Under review as a conference paper at ICLR 2017
Sampling
io2seq
Train
44%
Test
42%
Table 2: Testing the I/O-vector-to-sequence model. Each result used 100 samples.
Sampling
1-best
1-sample
10-sample
50-sample
100-sample
300-sample
Train
60%
56%
81%
91%
94%
97%
Test
63%
57%
79%
89%
94%
97%
Table 3: The effect of backtracking (sampling) multiple programs on accuracy. 1-best is deterministically choosing the expansion with highest probability at each step.
are a function of the input/output encoding vector. The io2seq model then generates a linearized
tree of a program symbol-by-symbol. An example of what a linearized program tree looks like is
(S (e (f (ConstStr “@”)ConstStr )f )e )S , which represents the program tree that returns the constant
string “@”. Predicting a linearized tree using an LSTM was also done in the context of parsing (Vinyals et al., 2015). For the io2seq model, we used the LSTM-sum cross-correlation I/O
conditioning model.
The results in Table 2 show that the performance of the io2seq model at 100 samples per latent test
function is far worse than the R3NN, at around 42% versus 91%, respectively. The reasons for that
could be that the io2seq model needs to perform far more decisions than the R3NN, since the io2seq
model has to predict the parentheses symbols that determine at which level of the tree a particular
symbol is at. For example, the io2seq model requires on the order of 100 samples for length 13
programs, while the R3NN requires no more than 13.
6.4
E FFECT OF BACKTRACKING SEARCH
For the best R3NN model that we trained, we also evaluated the effect that a different number of
samples per latent function had on performance. The results are shown in Table 3. The increase of
the model’s performance as the sample size increases hints that the model has a notion of what type
of program satisfies a given I/O pair, but it might not be that certain about the details such as which
regex to use, etc. By 300 samples, the model is nearing perfect accuracy on the test sets.
6.5
F LASH F ILL B ENCHMARKS
We also evaluate our learnt models on 238 real-world FlashFill benchmarks obtained from the Microsoft Excel team and online help-forums. These benchmarks involve string manipulation tasks
described using input-output examples. We evaluate two models – one with a cross correlation encoder trained on 5 input-output examples and another trained on 10 input-output examples. Both
the models were trained on randomly sampled programs from the DSL upto size 13 with randomly
generated input-output examples.
The distribution of the size of smallest DSL programs needed to solve the benchmark tasks is shown
in Figure 4(a), which varies from 4 to 63. The figure also shows the number of benchmarks for
which our model was able to learn the program using 5 input-output examples using samples of
top-2000 learnt programs. In total, the model is able to learn programs for 91 tasks (38.2%). Since
the model was trained for programs upto size 13, it is not surprising that it is not able to solve tasks
that need larger program size. There are 110 FlashFill benchmarks that require programs upto size
13, out of which the model is able to solve 82.7% of them.
The effect of sampling multiple learnt programs instead of only top program is shown in Figure 4(b).
With only 10 samples, the model can already learn about 13% of the benchmarks. We observe
a steady increase in performance upto about 2000 samples, after which we do not observe any
9
Under review as a conference paper at ICLR 2017
Number of Benchmarks
Number of FlashFill Benchmarks solved
50
45
40
35
30
25
20
15
10
5
0
Total
4
7
9
10
11
13
15
17
19
24
Solved
25
27
30
31
37
50
59
63
Size of smallest programs for FlashFill Benchmarks
Sampling
10
50
100
200
500
1000
2000
5000
(a)
Solved Benchmarks
13%
21%
23%
29%
33%
34%
38%
38%
(b)
Figure 4: (a) The distribution of size of programs needed to solve FlashFill tasks and the performance of our model, (b) The effect of sampling for trying top-k learnt programs.
Input v
[CPT-00350
[CPT-00340]
[CPT-114563]
[CPT-1AB02
[CPT-00360
Output
[CPT-00350]
[CPT-00340]
[CPT-114563]
[CPT-1AB02]
[CPT-00360]
(a)
Input v
Output
732606129
0x73
430257526
0x43
444004480
0x44
371255254
0x37
635272676
0x63
(b)
Input v
John Doyle
Matt Walters
Jody Foster
Angela Lindsay
Maria Schulte
(c)
Output
John D.
Matt W.
Jody F.
Angela L.
Maria S.
Figure 5: Some example solved benchmarks: (a) cleaning up medical codes with closing brackets,
(b) generating Hex numbers with first two digits, (c) transforming names to firstname and last initial.
significant improvement. Since there are more than 2 million programs in the DSL of length 11
itself, the enumerative techniques with uniform search do not scale well (Alur et al., 2015).
We also evaluate a model that is learnt with 10 input-output examples per benchmark. Surprisingly,
this model can only learn programs for about 29% of the FlashFill benchmarks. We hypothesize
that the space of consistent programs gets more constrained with additional input-output examples,
which makes it harder for R3NN to learn the desired program. Another possibility is that the inputoutput encoder gets more confused with the additional example pairs.
Our model is able to solve majority of FlashFill benchmarks that require learning programs with
upto 3 Concat operations. We now describe a few of these benchmarks, also shown in Figure 5. An Excel user wanted to clean a set of medical billing records by adding a missing “]”
to medical codes as shown in Figure 5(a). Our system learns the following program given these
5 input-output examples: Concat(SubStr(v,ConstPos(0),(d,-1,End)), ConstStr(“]”)). The program concatenates the substring between the start of the input string and the position of the last
digit regular expression with the constant string “]”. Another task that required user to transform some numbers into a hex format is shown in Figure 5(b). Our system learns the following
program: Concat(ConstStr(“0x”),SubStr(v,ConstPos(0),ConstPos(2))). For some benchmarks
with long input strings, it is still able to learn regular expressions to extract the desired substring, e.g. it learns a program to extract “NancyF” from the string “123456789,freehafer ,drew
,nancy,19700101,11/1/2007,[email protected],1230102,123 1st Avenue,Seattle,wa,09999”.
Our system is currently not able to learn programs for benchmarks that require 4 or more Concat operations. Two such benchmarks are shown in Figure 6. The task of combining names in
Figure 6(a) requires 6 Concat arguments, whereas the phone number transformation task in Figure 6(b) requires 5 Concat arguments. This is mainly because of the scalability issues in training
with programs of larger size. There are also a few interesting benchmarks where the R3NN models
gets very close to learning the desired program. For example, for the task “Bill Gates” → “Mr.
Bill Gates”, it learns a program that generates “Mr.Bill Gates” (without the whitespace), and for
the task “617-444-5454” → “(617) 444-5454”, it learns a program that generates the string “(617
444-5454”.
10
Under review as a conference paper at ICLR 2017
1
2
3
4
Input v
John James Paul
Tom Mike Bill
Marie Nina John
Reggie Anna Adam
Output
John, James, and Paul.
Tom, Mike, and Bill.
Marie, Nina, and John.
Reggie, Anna, and Adam.
(a)
1
2
3
4
Input v
(425) 221 6767
206.225.1298
617-224-9874
425.118.9281
Output
425-221-6767
206-225-1298
617-224-9874
425-118-9281
(b)
Figure 6: Some unsolved benchmarks: (a)Combining names by different delimiters. (b) Transforming phone numbers to consistent format.
7
R ELATED W ORK
We have seen a renewed interest in recent years in the area of Program Induction and Synthesis.
In the machine learning community, a number of promising neural architectures have been proposed to perform program induction. These methods have employed architectures inspired from
computation modules (Turing Machines, RAM) (Graves et al., 2014; Kurach et al., 2015; Reed &
de Freitas, 2015; Neelakantan et al., 2015) or common data structures such as stacks used in many
algorithms (Joulin & Mikolov, 2015). These approaches represent the atomic operations of the network in a differentiable form, which allows for efficient end-to-end training of a neural controller.
However, unlike our approach that learns comprehensible complete programs, many of these approaches learn only the program behavior (i.e., they produce desired outputs on new input data).
Some recently proposed methods (Kurach et al., 2015; Gaunt et al., 2016; Riedel et al., 2016; Bunel
et al., 2016) do learn interpretable programs but these techniques require learning a separate neural
network model for each individual task, which is undesirable in many synthesis settings where we
would like to learn programs in real-time for a large number of tasks. Liang et al. (2010) restrict
the problem space with a probabilistic context-free grammar and introduce a new representation
of programs based on combinatory logic, which allows for sharing sub-programs across multiple
tasks. They then take a hierarchical Bayesian approach to learn frequently occurring substructures
of programs. Our approach, instead, uses neural architectures to condition the search space of programs, and does not require additional step of representing program space using combinatory logic
for allowing sharing.
The DSL-based program synthesis approach has also seen a renewed interest recently (Alur et al.,
2015). It has been used for many applications including synthesizing low-level bitvector implementations (Solar-Lezama et al., 2005), Excel macros for data manipulation (Gulwani, 2011; Gulwani
et al., 2012), superoptimization by finding smaller equivalent loop bodies (Schkufza et al., 2013),
protocol synthesis from scenarios (Udupa et al., 2013), synthesis of loop-free programs (Gulwani
et al., 2011), and automated feedback generation for programming assignments (Singh et al., 2013).
The synthesis techniques proposed in the literature generally employ various search techniques including enumeration with pruning, symbolic constraint solving, and stochastic search, while supporting different forms of specifications including input-output examples, partial programs, program
invariants, and reference implementation.
In this paper, we consider input-output example based specification over the hypothesis space defined by a DSL of string transformations, similar to that of FlashFill (without conditionals) (Gulwani, 2011). The key difference between our approach over previous techniques is that our system
is trained completely in an end-to-end fashion, while previous techniques require significant manual
effort to design heuristics for efficient search. There is some work on guiding the program search using learnt clues that suggest likely DSL expansions, but the clues are learnt over hand-coded textual
features of examples (Menon et al., 2013). Moreover, their DSL consists of composition of about
100 high-level text transformation functions such as count and dedup, whereas our DSL consists of
tree structured programs over richer regular expression based substring constructs.
There is also a recent line of work on learning probabilistic models of code from a large number of
code repositories (big code) (Raychev et al., 2015; Bielik et al., 2016; Hindle et al., 2016), which
are then used for applications such as auto-completion of partial programs, inference of variable
and method names, program repair, etc. These language models typically capture only the syntactic
11
Under review as a conference paper at ICLR 2017
properties of code, unlike our approach that also tries to capture the semantics to learn the desired
program. The work by Maddison & Tarlow (2014) addresses the problem of learning structured
generative models of source code but both their model and application domain are different from
ours.
The R3NN model employed in our work is related to several tree and graph structured neural networks present in the NLP literature (Le & Zuidema, 2014; Paulus et al., 2014; Irsoy & Cardie, 2013).
The Inside-Outside Recursive Neural Network (Le & Zuidema, 2014) in particular is most similar to
the R3NN, where they generate a parse tree incrementally by using global leaf-level representations
to determine which expansions in the parse tree to take next.
8
C ONCLUSION
We have proposed a novel technique called Neuro-Symbolic Program Synthesis that is able to construct a program incrementally based on given input-output examples. To do so, a new neural
architecture called Recursive-Reverse-Recursive Neural Network is used to encode and expand a
partial program tree into a full program tree. Its effectiveness at example-based program synthesis
is demonstrated, even when the program has not been seen during training.
These promising results open up a number of interesting directions for future research. For example,
we took a supervised-learning approach here, assuming availability of target programs during training. In some scenarios, we may only have access to an oracle that returns the desired output given
an input. In this case, reinforcement learning is a promising framework for program synthesis.
R EFERENCES
Alur, Rajeev, Bodı́k, Rastislav, Dallal, Eric, Fisman, Dana, Garg, Pranav, Juniwal, Garvit, KressGazit, Hadas, Madhusudan, P., Martin, Milo M. K., Raghothaman, Mukund, Saha, Shamwaditya,
Seshia, Sanjit A., Singh, Rishabh, Solar-Lezama, Armando, Torlak, Emina, and Udupa, Abhishek. Syntax-guided synthesis. In Dependable Software Systems Engineering, pp. 1–25. 2015.
Bielik, Pavol, Raychev, Veselin, and Vechev, Martin T. PHOG: probabilistic model for code. In
ICML, pp. 2933–2942, 2016.
Biermann, Alan W. The inference of regular lisp programs from examples. IEEE transactions on
Systems, Man, and Cybernetics, 8(8):585–600, 1978.
Bunel, Rudy, Desmaison, Alban, Kohli, Pushmeet, Torr, Philip H. S., and Kumar, M. Pawan. Adaptive neural compilation. CoRR, abs/1605.07969, 2016. URL http://arxiv.org/abs/1605.07969.
Gaunt, Alexander L, Brockschmidt, Marc, Singh, Rishabh, Kushman, Nate, Kohli, Pushmeet, Taylor, Jonathan, and Tarlow, Daniel. Terpret: A probabilistic programming language for program
induction. arXiv preprint arXiv:1608.04428, 2016.
Graves, Alex, Wayne, Greg, and Danihelka, Ivo.
arXiv:1410.5401, 2014.
Neural turing machines.
arXiv preprint
Gulwani, Sumit. Automating string processing in spreadsheets using input-output examples. In
POPL, pp. 317–330, 2011.
Gulwani, Sumit, Jha, Susmit, Tiwari, Ashish, and Venkatesan, Ramarathnam. Synthesis of loop-free
programs. In PLDI, pp. 62–73, 2011.
Gulwani, Sumit, Harris, William, and Singh, Rishabh. Spreadsheet data manipulation using examples. Communications of the ACM, Aug 2012.
Hindle, Abram, Barr, Earl T., Gabel, Mark, Su, Zhendong, and Devanbu, Premkumar T. On the
naturalness of software. Commun. ACM, 59(5):122–131, 2016.
Irsoy, Orzan and Cardie, Claire. Bidirectional recursive neural networks for token-level labeling
with structure. In NIPS Deep Learning Workshop, 2013.
12
Under review as a conference paper at ICLR 2017
Joulin, Armand and Mikolov, Tomas. Inferring algorithmic patterns with stack-augmented recurrent
nets. In NIPS, pp. 190–198, 2015.
Kurach, Karol, Andrychowicz, Marcin, and Sutskever, Ilya. Neural random-access machines. arXiv
preprint arXiv:1511.06392, 2015.
Le, Phong and Zuidema, Willem. The inside-outside recursive neural network model for dependency
parsing. In EMNLP, pp. 729–739, 2014.
Liang, Percy, Jordan, Michael I., and Klein, Dan. Learning programs: A hierarchical Bayesian
approach. In ICML, pp. 639–646, 2010.
Maddison, Chris J and Tarlow, Daniel. Structured generative models of natural source code. In
ICML, pp. 649–657, 2014.
Menon, Aditya Krishna, Tamuz, Omer, Gulwani, Sumit, Lampson, Butler W., and Kalai, Adam. A
machine learning framework for programming by example. In ICML, pp. 187–195, 2013.
Neelakantan, Arvind, Le, Quoc V, and Sutskever, Ilya. Neural programmer: Inducing latent programs with gradient descent. arXiv preprint arXiv:1511.04834, 2015.
Paulus, Romain, Socher, Richard, and Manning, Christopher D. Global belief recursive neural
networks. pp. 2888–2896, 2014.
Raychev, Veselin, Vechev, Martin T., and Krause, Andreas. Predicting program properties from ”big
code”. In POPL, pp. 111–124, 2015.
Reed, Scott and de Freitas, Nando.
arXiv:1511.06279, 2015.
Neural programmer-interpreters.
arXiv preprint
Riedel, Sebastian, Bosnjak, Matko, and Rocktäschel, Tim. Programming with a differentiable forth
interpreter. CoRR, abs/1605.06640, 2016. URL http://arxiv.org/abs/1605.06640.
Schkufza, Eric, Sharma, Rahul, and Aiken, Alex. Stochastic superoptimization. In ASPLOS, pp.
305–316, 2013.
Singh, Rishabh and Solar-Lezama, Armando. Synthesizing data structure manipulations from storyboards. In SIGSOFT FSE, pp. 289–299, 2011.
Singh, Rishabh, Gulwani, Sumit, and Solar-Lezama, Armando. Automated feedback generation for
introductory programming assignments. In PLDI, pp. 15–26, 2013.
Solar-Lezama, Armando. Program Synthesis By Sketching. PhD thesis, EECS Dept., UC Berkeley,
2008.
Solar-Lezama, Armando, Rabbah, Rodric, Bodik, Rastislav, and Ebcioglu, Kemal. Programming by
sketching for bit-streaming programs. In PLDI, 2005.
Summers, Phillip D. A methodology for lisp program construction from examples. Journal of the
ACM (JACM), 24(1):161–175, 1977.
Udupa, Abhishek, Raghavan, Arun, Deshmukh, Jyotirmoy V., Mador-Haim, Sela, Martin, Milo
M. K., and Alur, Rajeev. TRANSIT: specifying protocols with concolic snippets. In PLDI, pp.
287–296, 2013.
Vinyals, Oriol, Kaiser, Lukasz, Koo, Terry, Petrov, Slav, Sutskever, Ilya, and Hinton, Geoffrey.
Grammar as a foreign language. In ICLR, 2015.
13
Under review as a conference paper at ICLR 2017
JConcat(f1 , · · · , fn )Kv
JConstStr(s)Kv
JSubStr(v, pl , pr )Kv
JConstPos(k)Kv
= Concat(Jf1 Kv , · · · , Jfn Kv )
= s
= v[Jpl Kv ..Jpr Kv ]
= k > 0? k : len(s) + k
J(r, k, Start)Kv
=
Start of k th match of r in v
from beginning (end if k < 0)
J(r, k, End)Kv
=
End of k th match of r in v
from beginning (end if k < 0)
Figure 7: The semantics of the DSL for string transformations.
A
D OMAIN - SPECIFIC L ANGUAGE FOR S TRING T RANSFORMATIONS
The semantics of the DSL programs is shown in Figure 7. The semantics of a Concat expression
is to concatenate the results of recursively evaluating the constituent substring expressions fi . The
semantics of ConstStr(s) is to simply return the constant string s. The semantics of a substring
expression is to first evaluate the two position logics pl and pr to p1 and p2 respectively, and then
return the substring corresponding to v[p1 ..p2 ]. We denote s[i..j] to denote the substring of string
s starting at index i (inclusive) and ending at index j (exclusive), and len(s) denotes its length.
The semantics of ConstPos(k) expression is to return k if k > 0 or return len + k (if k < 0).
The semantics of position logic (r, k, Start) is to return the Start of k th match of r in v from the
beginning (if k > 0) or from the end (if k < 0).
14
| 6cs.PL
|
ON STABLE COMMUTATOR LENGTHS OF DEHN TWISTS
ALONG SEPARATING CURVES
arXiv:1606.08643v1 [math.GT] 28 Jun 2016
NAOYUKI MONDEN AND KAZUYA YOSHIHARA
Abstract. We give new upper bounds on the stable commutator lengths of
Dehn twists along separating curves in the mapping class group of a closed
oriented surface. The estimates of these upper bounds are O(1/g), where g is
the genus of the surface.
1. Introduction
For x in the commutator subgroup [G, G] of a group G, we define the commutator
length clG (x) of x to be the smallest number of commutators in G whose product
is equal to x. The stable commutator length sclG (x) of x is the limit
clG (xn )
.
n→∞
n
sclG (x) = lim
Let Mg be the mapping class group of a closed oriented surface of genus g, and
let tc be the Dehn twist along a simple closed curve c on the surface. The natural
problem is to calculate sclMg (tc ). However, in general, computing scl is difficult.
Therefore, it makes sense to give upper and lower bounds on sclMg (tc ) (Note that
using Louwsma’s results [16], we find sclM1 (tc ) = 1/12).
The lower bounds on sclMg (tc ) were given by Endo-Kotschick [9] using Gauge
theory. For technical reasons, they showed that 1/(18g + 6) ≤ sclMg (tc ) only for
a separating curve c. This assumption is removed by Korkmaz [13]. He also gave
upper bounds on stable commutator lengths of Dehn twists in [13]. An upper bound
on sclMg (tc ) given in [13] (see also [17]) is constant for g. However, according to [15]
(see also [5]), there is an estimate sclMg (tc ) = O(1/g), that is, limg→∞ sclMg (tc ) =
0. Such an upper bound were given in [6] for a nonseparating curve c, and the
estimate is sclMg (tc ) ≤ 1/(4g + 6 + (2/g)) for g ≥ 1. On the other hand, to the best
of our knowledge, there is no such upper bound for a separating curve sh on Σg ,
where sh separates into two components with genera h and g − h For this reason,
we focus on separating curves. We prove the following.
Theorem 1. For
some integers k, h and r, we assume that genus g = kh + r
(h = 1, 2, . . . , g2 , r = 0, 1, . . . , h − 1). Then we have
1
h(2h + 1)(2g − 2h + 1)
sclMg (tsr ) + 1 .
sclMg (tsh ) ≤
(g + 1)(2g + 1) − (2g − 2h + 1)r 2r + 1
Note that ts0 = id in Mg since s0 bounds a disk, and therefore scl(ts0 ) = 0. As
a corollary, we obtain the following results.
Corollary 2. sclMg (ts1 ) ≤
3(2g − 1)
.
(g + 1)(2g + 1)
1
2
N. MONDEN AND K. YOSHIHARA
Corollary 3. If g = kh, then sclMg (tsh ) ≤
h(2h + 1)(2g − 2h + 1)
.
(g + 1)(2g + 1)
Using Theorem 1 and Corollary 2, we can inductively give an upper bound of
sclMg (tsh ) for any h.
Here is an outline of this paper. In Section 2.1, we review basic facts of stable
commutator lengths and quasi-morphisms. Especially, we present “Bavard’s duality
theorem” which gives a relationship between stable commutator lengths and quasimorphisms. Section 2.2 provides relations in mapping class groups. These relations
are need to prove Theorem 1. In the last section, we prove Theorem 1.
Remark 4. There are applications of giving lower and upper bounds on the stable
commutator length of a Dehn twist. Powell [20] showed that any element x in Mg
can be factorized as a product of commutators if g ≥ 3, that is, Mg is a perfect
group. The next problem is uniformly perfectness of Mg . Here, a group G is called
uniformly perfect if there is a natural number N such that clG (x) ≤ N for any
element x in G. It was conjectured by Morita [19] that Mg is not uniformly perfect,
more strongly, the map Hb2 (Mg ) → H 2 (Mg ; R) is not injective, where H 2 (Mg )
(resp. Hb2 (Mg )) is the second (resp. bounded) cohomology of Mg . Endo-Kotschick
[9] soloved them affirmatively by giving a lower bound of sclMg (tc ). Recently, the
first author [18] improved an upper bound on sclMg (tc ) for any simple closed curve
c and g ≥ 3 by giving an explicit factorization of tnc for any n. Using this result,
h (n)
he improved an upper bound on the limit of limn→∞ gn for odd g, where hg (n)
is the minimal h such that there exists a Σg -bundle over Σh with signature 4n.
Computing the limit is open problem in Problem 2.18 (B) of [8].
Remark 5. We introduce some background results on factorizations of some power
of a Dehn twist as a product of commutators. Korkmaz and Ozbagci [12] proved
that clMg (tc ) = 2 for any simple closed curve c if g ≥ 3. It was shown that
5
clMg (t10
d ) ≤ 2 for a nonseparating curve d if g ≥ 2 (see [12]) and that clM2 (ts1 ) ≤ 6
1
for a separating curve s1 (see [14]). For the mapping class group Mg of a compact
oriented surface of genus g ≥ 2 with a boundary curve δ, it was proved in [3] proved
that clM1g (tnδ ) = [n/2] for any n, so sclM1g (tδ ) = 1/2 (see also [2]). In the above
results, explicit factorizations of some power of a Dehn twist were given.
Remark 6. We explain a geometrical interpretation of (stable) commutator lengths
in Mg as Lefschetz fibrations, which play an important role in 4-dimensional topology. Note that by the works of [11] and [7], a 4-manifold is symplectic if and only if
it admits a Lefschetz fibration, up to blow-up. Let x be a product of right-handed
Qk
Dehn twists tv1 · · · tvn . If we give a relation x = i=1 [ai , bi ] in Mg , then we obtain
a Lefschetz fibration with fiber Σg over Σk with n singular fibers such that each
of them is obtained by collapsing simple closed curve vj on Σg to a point (The
Euler characteristic of the 4-manifold admitting this Lefschetz fibration is equal
to 4(g − 1)(k − 1) + n). Conversely, given a Lefschetz fibration, we get the above
relation. From this, computing clMg (x) means that calculating the minimal base
genus of such Lefschetz fibration and the minimal Euler characteristic of its total
space. Therefore, clMg (tnc ) gives the “smallest” Lefschetz fibraitons (in the sense
of the Euler characteristic) in the “simplest” ones (in the sense of singular fibers).
Thus, we can regard sclMg (tc ) as the growth rate of the genus of their bases.
ON STABLE COMMUTATOR LENGTHS OF DEHN TWISTS ALONG SEPARATING CURVES3
2. Preliminaries
2.1. Stable commutator lengths and quasi-morphisms. For a group G, let
[G, G] denote the commutator subgroup of G.
Definition 7. For x ∈ [G, G], we define the commutator length clG (x) of x to be
the smallest number of commutators in G whose product is equal to x. The stable
commutator length sclG (x) of x is the limit
clG (xn )
.
n→∞
n
By convention we define clG (x) = ∞ if x ∈ G is not in [G, G], and sclG (x) < ∞ if
and only if xm ∈ [G, G] for some integer m.
sclG (x) = lim
The limit exists since the non-negative function n → clG (xn ) is subadditive, and
clG and sclG are class functions. We need the theory of quasi-morphisms in order
to prove Theorem 1. In this section, we recall the definition and basic properties
of quasi-morphisms. We refer the reader to [5] for the details.
Definition 8. Let G be a group. A quasi-morphism on G is a function φ : G → R
for which there is a constant D(φ) ≥ 0 such that
|φ(xy) − φ(x) − φ(y)| ≤ D(φ)
for all x, y ∈ G. We call D(φ) a defect of φ. A quasi-morphism φ is called homogeneous if
φ(xn ) = nφ(x)
for all x ∈ G and all n ∈ Z.
We recall the following basic properties of homogeneous quasi-morphisms (for
example, see Section 5.5.2 of [5] and Lemma 2.1 (1) of [15] for proofs).
Lemma 9. Let φ : G → R be a homogeneous quasi-morphism on a group G. For
all x, y ∈ G, the following holds.
(a) φ(x) = φ(yxy −1 ),
(b) If xy = yx, then φ(xy) = φ(x) + φ(y).
Theorem 10 (Bavard’s Duality Theorem [1] ). Let Q be the set of all homogeneous
quasi-morphisms on a group G with positive defects. For any x ∈ [G, G], we have
sclG (x) =
sup
φ∈Q,D(φ)
|φ(x)|
.
2D(φ)
2.2. Relations in mapping class groups. Let Σg be closed connected oriented
surface of genus g ≥ 2, and let Mg be the mapping class group of Σg , that is the
group of isotopy classes of orientation preserving self-diffeomorphisms of Σg . Since
Mg /[Mg , Mg ] is isomorphic to Z12 if g = 1, Z10 if g = 2 (see [4]), and is trivial if
g ≥ 3 (see [20]), we can define sclMg (x) for any x ∈ Mg .
Let tc be the right-handed Dehn twist along a simple closed curve c on Σg . We
review some relations among Dehn twists. More details can be found in [10].
Lemma 11. For any two separating curves sh and s′h on Σg which separate into
two components with genera h and g − h, tsh is conjugate to ts′h and tsg−h .
Lemma 12. If c and d are two disjoint simple closed curves on Σg , then tc td = td tc .
4
N. MONDEN AND K. YOSHIHARA
Lemma 13 (The hyperelliptic involution). Let c1 , c2 , . . . , c2g+1 be nonseparating
curves in Σg such that ci and cj are disjoint if |i − j| ≥ 2, and that ci and ci+1
intersect at one point. Then, the product
ι := tc1 tc2 · · · tc2g t2c2g+1 tc2g · · · tc2 tc1 .
is the hyperelliptic involution. In particular, we have the relation ιtci = tci ι for
i = 1, 2, . . . , 2g + 1.
Lemma 14 (The even chain relation). For a positive integer h, let us consider
a sequence of simple closed curves c1 , c2 , . . . , c2h in Σg such that ci and cj are
disjoint if |i − j| ≥ 2, and that ci and ci+1 intersect at one point. Then, a regular
neighborhood of c1 ∪c2 ∪· · ·∪c2h is a subsurface of genus h with connected boundary,
denoted by dh . We then have
(tc1 tc2 · · · tc2h )4h+2 = tdh .
3. The proof of Theorem1
In this section, we prove Theorem 1. For this, we need the following Lemma.
Lemma 15. Let G be a group, and let x1 , x2 , . . . , xn be elements in G such that
xi xj = xj xi for |i−j| ≥ 2. Then, x1 x3 · · · x2m−1 ·x2 x4 · · · x2m (resp. x1 x3 · · · x2m+1 ·
x2 x4 · · · x2m ) is conjugate to x1 x2 · · · x2m if n = 2m (resp. n = 2m + 1).
Proof. We consider the case where n = 2m + 1. Set
X2k+1 = x1 x2 · · · x2k+1 · x2k+3 x2k+5 · · · x2m+1 · x2k+2 x2k+4 · · · x2m ,
X2k = x1 x2 · · · x2k · x2k+2 x2k+4 · · · x2m · x2k+1 x2k+3 · · · x2m+1 .
Note that X1 = x1 x3 · · · x2m+1 · x2 x4 · · · x2m and X2m+1 = x1 x2 · · · x2m+1 . Therefore, in order to prove Lemma 15, it is sufficient to show that Xi+1 is conjugate to
Xi . From the assumption, we have
(x2k+2 x2k+4 · · · x2m ) · X2k+1 · (x2k+2 x2k+4 · · · x2m )−1
= x2k+2 x2k+4 · · · x2m · x1 x2 · · · x2k−1 x2k x2k+1 · x2k+3 x2k+5 · · · x2m+1
= x1 x2 · · · x2k−1 · x2k x2k+2 · · · x2m · x2k+1 x2k+3 x2k+5 · · · x2m+1 = X2k .
Therefore, X2k+1 is conjugate to X2k . By a similar argument, we can show that
X2k is conjugate to X2k−1 . The proof for n = 2m is similar. This finishes the proof
of Lemma 15.
We now give the proof of Theorem 1.
The proof of Theorem 1. Let φ be a homogeneous quasi-morphism on Mg , and let
c1 , c2 , . . . , c2g+1 the simple closed curves in Lemma 13. For simplicity of notation,
we write ti instead of tci . For i = 2, 3, . . . , k − 1, we write
T1 := t21 t2 t3 · · · t2h ,
Tk := t2(k−1)h+1 t2(k−1)h+2 · · · t2(g−h) ,
Ti := t2(i−1)h+1 t2(i−1)h+2 · · · t2ih ,
Tk+1 := t2(g−h)+1 t2(g−h)+2 · · · t2g ,
Tk+2 := t22g+1 t2g t2g−1 · · · t2(g−h)+2 ,
S := t2(g−h)+1 t2(g−h) · · · t2 .
ON STABLE COMMUTATOR LENGTHS OF DEHN TWISTS ALONG SEPARATING CURVES5
Then, by Lemma 9 (a), 11 and 14,
1
φ(tsh ),
4h
1
φ(Ti ) = φ(Tk+1 ) =
φ(tsh ),
4h + 2
1
φ(Tk ) =
φ(tsr ),
4r + 2
1
φ(S) =
φ(tsh )
4(g − h) + 2
(1)
φ(T1 ) = φ(Tk+2 ) =
(2)
(3)
(4)
for i = 2, 3, . . . , k − 1. From the conjugate t1 ιt−1
1 and Lemma 13, we have
t21 t2 t3 · · · t2g t22g+1 t2g · · · t3 t2 = ι−1 = ι,
that is, T1 T2 · · · Tk+2 S = ι. In particular, we see that the relation
T1 T2 · · · Tk+2 = ι · S −1
(5)
holds in Mg . Since φ(ι) = 0 by the definition of homogeneous quasimorphisms,
from the equations (4), (5), Lemma 9 (b) and 13, we obtain
φ(T1 T2 · · · Tk+2 ) = −φ(S) = −
(6)
1
φ(tsh ).
4(g − h) + 2
Note that Ti Tj = Tj Ti if |i − j| ≥ 2 from Lemma 12. From this and Lemma 15, we
have
φ(T1 T3 · · · Tk+1 · T2 T4 · · · Tk+2 ) if k + 2 is even
φ(T1 T2 · · · Tk+2 ) =
φ(T1 T3 · · · Tk+2 · T2 T4 · · · Tk+1 ) if k + 2 is odd.
By this equation, the definition of quasimorphisms, Lemma 9 (b) and the equations
(1)-(3) and (6),
D(φ) ≥ |φ(T1 T2 · · · Tk+2 ) −
k+2
X
φ(Ti )|
i=1
= −
2
k−1
1
1
φ(tsh ) −
φ(tsh ) −
φ(tsh ) −
φ(tsr )
4(g − h) + 2
4h
4h + 2
4r + 2
This gives
(g + 1)(2g + 1) − (2g − 2h + 1)r
1
|φ(tsh )| ≤
|φ(tsr )| + D(φ)
2h(2h + 1)(2g − 2h + 1)
4r + 2
from r = g − kh. Hence, we obtain
h(2h + 1)(2g − 2h + 1)
|φ(tsh )|
≤
2D(φ)
(g + 1)(2g + 1) − (2g − 2h + 1)r
1 |φ(tsr )|
+1 .
2r + 1 2D(φ)
1
sclMg (tsr ) + 1 .
2r + 1
By Bavard’s Duality Theorem,
sclMg (tsh ) ≤
h(2h + 1)(2g − 2h + 1)
(g + 1)(2g + 1) − (2g − 2h + 1)r
This completes the proof of Theorem 1.
Acknowledgements. The first author was supported by Grant-in-Aid for Young
Scientists (B) (No. 13276356), Japan Society for the Promotion of Science. The
second author would like to thank Susumu Hirose, Noriyuki Hamada, and Takayuki
Okuda for helpful comments and invaluable advice on the mapping class groups of
6
N. MONDEN AND K. YOSHIHARA
the surfaces. Finally, he also would like to thank Professor Osamu Saeki for many
helpful suggestions and comments.
References
[1] C. Bavard, Longueur stable des commutateurs, Enseign. Math. (2) 37 (1991), no. 1-2,
109–150.
[2] R.I. Baykur, Flat bundles and commutator lengths, Michigan Math. J. 63 (2014), no. 2,
333–344.
[3] R.I. Baykur, M. Korkmaz and N. Monden, Sections of surface bundles and Lefschetz
fibrations, Transactions of the American Mathematical Society, 365, (2013), no. 11, 5999–
6016.
[4] J. Birman, H. Hilden, On mapping class groups of closed surfaces as covering spaces, In
Advances in the theory of Riemann surfaces, Ann. of Math. Studies 66 (1971), 81–115.
[5] D. Calegari, scl, MSJ Memoirs, 20. Mathematical Society of Japan, Tokyo, 2009.
[6] D. Calegari, N. Monden and M. Sato, On stable commutator length in hyperelliptic mapping class groups, Pacific Journal of Mathematics, 272 (2014), no. 2, 323–351.
[7] S. K. Donaldson; Lefschetz pencils on symplectic manifolds, J. Diff. Geom. 53 (1999),
205–236.
[8] R. Kirby, editor, Problems in low-dimensional topology, from: “Geometric topology
(Athens, GA, 1993)”, AMS/IP Stud. Adv. Math. 2, Amer. Math. Soc. (1997) 35–473.
[9] H. Endo and D. Kotschick, bounded cohomology and non-uniform perfection of mapping
class groups, Invent. Math. 144 (2001), no. 1, 169–175.
[10] B. Farb and D. Margalit, A primer on mapping class groups, Princeton Mathematical
Series, 49. Princeton University Press, Princeton,
[11] R. Gompf and A. Stipsicz; 4-manifolds and Kirby calculus, Graduate Studies in Mathematics, vol. 20, American Math. Society, Providence 1999.
[12] M. Korkmaz and B. Ozbagci, Minimal number of singular fibers in a Lefschetz fibration,
Proc. Amer. Math. Soc. 129 (2001), no. 5, 1545–1549.
[13] M. Korkmaz, Stable commutator length of a Dehn twist, Michigan Math. J. 52 (2004),
no. 1, 23–31.
[14] M. Korkmaz and A. Stipsicz, Lefschetz fibrations on 4-manifolds, Handbook of
Teichmüller theory. Vol. II, 271-296, IRMA. Lect. Math. Theor. Phys., 13, Eur. Math.
Soc., Zurich, 2009.
[15] D. Kotschick, Quasi-homomorphisms and stable lengths in mapping class groups, Proc.
Amer. Math. Soc. 132 (2004), no. 11, 3167–3175.
[16] J. Louwsma, Extremality of the Rotation Quasimorphism on the Modular Group, Dissertation (Ph.D.), California Institute of Technology.
[17] N. Monden, On upper bounds on stable commutator lengths in mapping class groups,
Topology Appl. 159 (2012), no. 4, 1085–1091.
[18] N. Monden, Signatures of surface bundles and stable commutator lengths of Dehn twists,
in preparation.
[19] S. Morita, Structure of the mapping class groups of surfaces: a survey and a prospect,
(English summary) Proceedings of the Kirbyfest (Berkeley, CA, 1998), 349–406 (electronic), Geom. Topol. Monogr., 2, Geom. Topol. Publ., Coventry, 1999.
[20] J. Powell, Two theorems on the mapping class group of a surface, Proc. Amer. Math. Soc.
68 (1978), no. 3, 347–350.
Department of Engineering Science, Osaka Electro-Communication University, Hatsucho 18-8, Neyagawa, 572-8530, Japan
E-mail address: [email protected]
| 4math.GR
|
MOEA/D with Angle-based Constrained Dominance
Principle for Constrained Multi-objective Optimization
Problems
arXiv:1802.03608v1 [cs.NE] 10 Feb 2018
Zhun Fana , Yi Fanga , Wenji Lia , Xinye Caib,∗, Caimin Weic , Erik Goodmand
a Department
of Electronic Engineering, Shantou University, Guangdong, 515063, China
of Computer Science and Technology, Nanjing University of Aeronautics and
Astronautics, Jiangsu, 210016, China
c Department of Mathematics, Shantou University, Guangdong, 515063, China
d BEACON Center for the Study of Evolution in Action, Michigan State University. East
Lansing, Michigan, USA
b College
Abstract
This paper proposes a novel constraint-handling mechanism named angle-based
constrained dominance principle (ACDP) embedded in a decomposition-based
multi-objective evolutionary algorithm (MOEA/D) to solve constrained multiobjective optimization problems (CMOPs). To maintain the diversity of the
working population, ACDP utilizes the information of the angle of solutions
to adjust the dominance relation of solutions during the evolutionary process.
This paper uses 14 benchmark instances to evaluate the performance of the
MOEA/D with ACDP (MOEA/D-ACDP). Additionally, an engineering optimization problem (which is I-beam optimization problem) is optimized. The
proposed MOEA/D-ACDP, and four other decomposition-based CMOEAs, including C-MOEA/D, MOEA/D-CDP, MOEA/D-Epsilon and MOEA/D-SR are
tested by the above benchmarks and the engineering application. The experimental results manifest that MOEA/D-ACDP is significantly better than the
other four CMOEAs on these test instances and the real-world case, which indicates that ACDP is more effective for solving CMOPs.
Keywords: Constraint-handling Mechanism, Angle-based Constrained
Dominance Principle (ACDP), Decomposition based Multi-objective
Algorithm (MOEA/D)
1. Introduction
Multi-objective optimization problems (MOPs) involve the optimization of
more than one objective function. In the real world, many optimization prob∗ Corresponding
author
Email address: [email protected] (Xinye Cai)
Preprint submitted to Elsevier
February 13, 2018
lems invariably involve a number of constraints and a multiple conflicting objectives. In general, a CMOP can mathematically be described as follows:
T
minimize F(x) = (f1 (x), . . . , fm (x))
subject to g (x) ≥ 0, i = 1, . . . , q
i
(1)
h
j (x) = 0, j = 1, . . . , p
x ∈ Rn
where F (x) = (f1 (x), f2 (x), . . . , fm (x))T ∈ Rm is an m-dimensional objective
vector, gi (x) ≥ 0 is the ith inequality constraint, and hj (x) = 0 is the j th
equality constraint. x ∈ Rn is an n-dimensional decision vector. The feasible
region S is defined as the set {x|gi (x) ≥ 0, i = 1, . . . , q and hj (x) = 0, j =
1, . . . , p}.
In CMOPs, there are usually more than one constraint. To demonstrate the
degree of constraint violation, these constraints are commonly summarized into
a scalar value as follows:
φ(x) =
q
X
| min(gi (x), 0)| +
i=1
p
X
|hj (x)|
(2)
j=1
when φ(x) = 0, the solution x is feasible, otherwise it is infeasible.
For any two feasible solutions xa ∈ S and xb ∈ S of a CMOP, it can be said
that xa dominates xb if the following condition is met:
∀i fi (xa ) ≤ fi (xb ) and ∃j fj (xa ) < fj (xb )
(3)
where i, j ∈ {1, 2, ..., m}. If there exists a solution x∗ ∈ S, which dominates
any other solutions in S, x∗ can be said as a Pareto optimal solution. The set
of all Pareto optimal solution belonging to S is called as Pareto set (PS). The
set of the mapping vectors of PS in the objective space is named as Pareto front
(PF), which can be defined in the form of P F = {F (x)| x ∈ P S}.
CMOPs are consist of a few objectives and constraints. To solve CMOPs,
the constraint-handling technique should be applied to the framework of multiobjective evolutionary algorithm (MOEA).
According to different selection strategies, MOEAs mainly can be classified
into three types: (1) Pareto-domination-based; (2) decomposition-based; (3)
indicator-based. The typical examples of domination-based MOEAs include
NSGA-II [1], MOGA [2], PAES-II [3], SPEA-II [4] and NPGA [5]. In recent
years, decomposition-based MOEAs attract a lot of attention. The most representative algorithm of this type is MOEA/D [6]. Some variants of MOEA/D include MOEA/D-DE [7], MOEA/D-M2M [8], EAG-MOEA/D [9] and MOEA/DSAS [10]. For indicator-based MOEAs, they use a scalar metric to assist the
selection, typical examples of this type are IBEA [11], SMS-EMOA [12], HypE
[13] and FV-MOEA [14].
To solve CMOPs, constraint-handling mechanisms are important. In recent
years, many different constraint-handling mechanisms have been proposed [15,
2
16]. According to [17], constraint-handling techniques can be generally classified
into (1) penalty functions; (2) special representations and operators; (3) repair
algorithms; (4) separate objectives and constraints; and (5) hybrid methods.
As a representative objective-constraint separating method, constraineddomination principle (CDP) proposed in [1] is widely used. It solves CMOPs by
treating constraints as the top priority. Stochastic ranking [18], ε-constrained
method [19] and non-greedy constraint-handling technique [20] are also in this
category.
Currently, CTPs [21] and CFs [22] are the most widely used CMOP test
instances. The common characteristic of these benchmarks is that they all
have large feasible regions in the objective space. However, constraint-handling
mechanisms do not work when the working population falls in feasible regions, so
these two series of instances are not actually suitable for testing the effectiveness
of constraint-handling mechanisms.
The rest of this paper is organized as follows: Section 2 briefly introduces
MOEA/D and four other decomposition-based CMOEAs. Section 3 introduces
the details of the angle-based constrained dominance principle embedded in
MOEA/D. Section 4 gives comprehensive experimental results of the proposed
algorithm MOEA/D-ACDP and four other CMOEAs on LIR-CMOPs and the
I-beam optimization problem. Finally, conclusions are made in section 5.
2. Relative Work
2.1. MOEA/D
In the original framework of MOEA/D [6], given a series of uniform distributed weight vectors, a MOP is decomposed into N scalar subproblems
(SOPs), and each SOP relates to one solution. In MOEA/D, a set of N uniformly spread weight vectors λ1 , . . . , λN is initially generated for N subproblems.
A weight vector λi satisfies the following conditions:
m
X
λik = 1
and λik ≥ 0
for each k ∈ {1, . . . , m}.
(4)
k=1
There are several approaches to decompose a MOP into a number of scalar
optimization subproblems [6, 23]. Three decomposition approaches, including
weighted sum [23], Tchebycheff [23] and boundary intersection approaches [6]
are commonly used. In this paper, Tchebycheff decomposition method is used
in the MOEA/D framework. The j-th subproblem is defined as follows:
)
(
1
∗
te
∗
|fi (x) − zi |
minimize g (x|λ, z ) = max
1≤i≤m
λji
subject to x ∈ S
(5)
∗
) is the ideal point, and zi∗ = min{fi (x)|x ∈ S}.
where z ∗ = (z1∗ , . . . , zm
3
2.2. Decomposition-based CMOEAs
Decomposition-based CMOEAs combine the MOEA/D with different constrainthandling mechanisms. In this paper, we introduce four representative decompositionbased CMOEAs including C-MOEA/D [24], MOEA/D-CDP [25], MOEA/DEpsilon [26], and MOEA/D-SR [25].
• C-MOEA/D [24] uses a variant of the epsilon constraint-handling technique. In this technique, the epsilon level is set to handle constraints
according to the constraint violation and the proportion of feasible solutions in the current population. When comparing any two solutions, if
overall constraint violations of the solutions are both less than the epsilon
level, the one with a better aggregation value dominates the other. Otherwise, the one with a smaller overall constraint violation dominates the
other.
• MOEA/D-CDP [25] uses CDP [1] to judge the domination relationship
between two arbitrary solutions. CDP can be simply summarized as following three rules:
1) When two feasible solutions are compared, the one with a better aggregation value dominates the other.
2) When a feasible solution is compared with an infeasible solution, the
feasible solution dominates the infeasible solution.
3) When two infeasible solutions are compared, the one with a smaller
degree of constraint violation dominates the other.
The second and the third rules can be combined as a single rule: When at
least one of two compared solutions is infeasible, the one with a smaller
degree of constraint violation dominates the other.
• MOEA/D-Epsilon [26] uses the original epsilon constraint-handling technique. The epsilon level setting can be referred to [19]. With the generation counter K increasing, the epsilon level will dynamically decrease.
• MOEA/D-SR [25] embeds the stochastic ranking method (SR) [18] in
MOEA/D to deal with constraints. A threshold parameter rf ∈ [0, 1] is
set to balance the selection between the objectives and the constraints
in MOEA/D-SR. When comparing two solutions, if a random number in
[0, 1] is less than rf , the one with a better aggregation value is retained
into the next generation. If the random number in [0, 1] is greater than rf ,
the whole framework of MOEA/D-SR is similar to that of MOEA/D-CDP.
In the case of rf = 0, MOEA/D-SR is equivalent to MOEA/D-CDP.
3. MOEA/D with Angle-based Constrained Dominance Principle
In this section, the definition of the proposed ACDP and the effectiveness of
this mechanism in MOEA/D are detailed.
4
3.1. Angle-based Constrained Dominance Principle
In the CDP approach [1], with its three basic rules, the overall constraint
violation is the most important factor during the evolutionary process, and some
useful information in the infeasible regions tends to be ignored.
The angle between any two solutions in the objective space is useful information, which is related to the similarity between these two solutions. The
definition of angle between any two solutions x1 and x2 is given as follows:
(F(x1 ) − z ∗ )T · (F(x2 ) − z ∗ )
1
2 ∗
(6)
angle(x , x , z ) = arccos
||F(x1 ) − z ∗ || · ||F(x2 ) − z ∗ ||
∗
where z ∗ = (z1∗ , . . . , zm
) is the ideal point, and zi∗ = min{fi (x|x ∈ S}. || · || is
the two-norm of a vector.
As shown in Fig. 1, given any two solutions x1 and x2 , the angle between
them in the objective space is θ12 . Obviously, the angle between any two solutions
is less than or equal to π/2, which means that the range of angle between any
two solutions belongs in [0, π/2].
Figure 1: Illustration of the angle between x1 and x2
Given any two solutions x1 and x2 , an threshold of angle θ, a random number
of Feasible Solutions
r and a parameter pf ( NumberPopulation
) which denotes the proportion
Size
of feasible solutions in the current population, the ACDP is defined as follows:
1. If x1 and x2 are both feasible, the one dominating the other is better.
2. If there is at lease one infeasible solution and angle(x1 , x2 , z ∗ ) ≤ θ, the
one with a smaller constraint violation dominates the other.
3. When there is at least one infeasible solution and angle(x1 , x2 , z ∗ ) > θ,
if r < pf , the one dominating the other is better, otherwise, they are
incomparable.
5
3.2. ACDP in the framework of MOEA/D
As we know, MOEA/D uses the value of decomposition function of a solution
to update its neighbors. In order to use ACDP to handle constraints in the
framework of MOEA/D, here we provide a version of ACDP which is suitable
to the algorithm.
Given a subproblem sp with the weight vector λ, for two solutions x1 and
2
x , their overall constraint violations are φ1 and φ2 , and their decomposition
values on the subproblem sp are g te (x1 |λ, z ∗ ) and g te (x2 |λ, z ∗ ). The ACDP
dominance θ in the framework of MOEA/D is defined as follows:
x1 θ x2 ⇔
Rule 1 if φ1 = 0, φ2 = 0 :
g te (x1 |λ, z ∗ ) < g te (x2 |λ, z ∗ );
1
2
Rule 2 if φ < φ :
angle(x1 , x2 , z ∗ ) ≤ θ;
Rule 3 otherwise :
angle(x1 , x2 , z ∗ ) > θ, r < pf ,
g te (x1 |λ, z ∗ ) < g te (x2 |λ, z ∗ ).
(7)
where θ is a threshold parameter, which is defined by users. In Eq. (7), the
constraint-handling method ACDP is equivalent to CDP [1] when θ ≥ π2 .
In Rule 1 of ACDP, when these two solutions are both feasible, the solution
with a lower aggregation value dominates the other, which is similar to the first
rule of CDP.
When at least one of x1 and x2 is infeasible, CDP only utilizes the constraint violations of these two solutions to compare, which is difficult to keep
the diversity of the working population when most of solutions in the population
are infeasible. Nevertheless, ACDP utilizes additional information to compare
the two solutions, which includes the angle between the two compared solutions
in the objective space and the proportion of feasible solutions in the current
population (pf ). More details of ACDP in this situation are listed as follows:
• In Rule 2 of ACDP, if the angle between x1 and x2 in the objective space
is smaller than the parameter θ, ACDP considers that these two solutions are similar and compares them according to their constraint violations. Because these two solutions are similar, based on the framework of
MOEA/D, they will be considered to relate to the same subproblem. In
this situation, using the constraint violations to compare the two solutions
will not cause the loss of the diversity.
• In Rule 3 of ACDP, if the angle between x1 and x2 in the objective space
is larger than the parameter θ, ACDP considers that these two solutions
are dissimilar, and the solution with a lower decomposition value will
dominate the other with a probability of pf . Some solutions with low
aggregation values but large constraint violations will have a chance to be
6
selected in the next generation, which can enhance the convergence of the
working population effectively.
• The probability in Rule 3 of ACDP is set to be the proportion of feasible
solutions in the current population. It keeps the balance of the exploration
of the working population between infeasible regions and feasible regions.
When pf is large, ACDP tends to explore infeasible regions. When pf is
small, ACDP tends to explore feasible regions.
3.3. Effectiveness of ACDP in MOEA/D
The evolutionary process of a CMOEA can be generally divided in three
stages according to the status of the working population.
In the first stage, a population is generated randomly, and most of the individuals are far away from the real PF as shown in Fig. 2 (a) and Fig. 2
(b).
In the second stage, the working population begins to explore the search
space. As shown in Fig. 2 (c), when using CDP in MOEA/D, the working population will be attracted to feasible regions and actually difficult to get across
infeasible regions. As shown in Fig. 2 (d), when ACDP is applied to MOEA/D,
the working population can maintain the diversity by using angle information.
Some individuals can enter infeasible regions, which can help the working population to go across infeasible regions effectively. Additionally, ACDP uses the
information of the proportion of feasible solutions in the current population as
the probability to select solutions, which can help to balance the search between
feasible and infeasible regions.
In the third stage, the working population will converge to its near feasible
regions. when using CDP, the population is trapped into local optimum, because
of the difficulty to get across infeasible regions in the second stage, as shown in
Fig. 2 (e). Conversely, when using ACDP, the working population can converge
to the real PF more completely as shown in Fig. 2 (f), because the population
can keep the diversity and explore infeasible regions in the second stage.
3.4. The Parameter Setting of Theta
In the early stage of evolutionary process, the population is commonly far
away from real PF. To prevent the population from being trapped into local optimum, the value of θ should be set small to maintain the diversity. In the later
stage of evolutionary process, the convergence to the feasible regions should be
emphasized, then the value of θ should become larger. Based on the above discussions, the threshold θ(k) should be dynamically increased with the generation
counter k increasing. A method of setting θ(k) is proposed as follows:
θ(k) =
(
θ0 1 +
k
Tmax
π
2
cp
, 1 ≤ k ≤ Tc
, Tc < k ≤ Tmax
7
(8)
(a) Stage 1 of CDP
(b) Stage 1 of ACDP
(c) Stage 2 of CDP
(d) Stage 2 of ACDP
(e) Stage 3 of CDP
(f) Stage 3 of ACDP
Figure 2: Illustrations of the evolutionary process of MOEA/D with CDP and ACDP.
8
where θ0 is an initial threshold value which is set as π/2N , N is the size of
population and Tmax is the maximum generation. α is a parameter, which is
set as 0.8. Tc = αTmax is the termination generation to control θ. Parameter
cp is initialized to log(N )/log(1 + α).
As we know, the uniform weight vectors defined in Eq. (4) decide that the
maximum angle between two vectors is π/2, and the average angle between
two adjacent vectors is π/2N , where N is the size of population. Then, θ0 is
initially set as π/2N . According to Eq. (8), when the generation counter k
reaches Tc , the value of θ(k) is π/2, and keeps constant afterwards. As shown
in Fig. 3, we assume that the population size N and the maximum generation
Tmax are set as 300 and 500, respectively. Meanwhile α is set as 0.8. We can
find that θ0 = π/2N . In the early stage of evolutionary process, θ(k) increases
continuously but slowly. It benefits the population to maintain diversity. When
k gets more and more closed to Tc , θ(k) rises faster, which helps to accelerate
the convergence speed to the feasible regions. When k reaches Tc , θ(k) is equal
to π/2, ACDP is transformed into CDP.
Figure 3: The changing trend of θ.
3.5. ACDP embedded in MOEA/D
The proposed MOEA/D-ACDP integrates the general framework of MOEA/D
and the angle-based constrained dominance principle.
The psuecode of MOEA/D-ACDP is listed in Algorithm 1. Lines 1-5 initialize some parameters in MOEA/D-ACDP. First, a CMOP is decomposed into
N subproblems which are associated with weight vectors λ1 , . . . , λN . Then the
population P , the initial increasing factor cp, the ideal point z ∗ and the neighbor
indexes B(i) are initialized. Lines 7-11 update the angle threshold value θ(k).
Line 12 updates the proportion of feasible solutions in the current population pf .
Lines 13-23 generate a set of new solutions and update the ideal point z ∗ . To be
more specific, lines 14-21 determine the set of neighboring solutions that may be
9
Algorithm 1: MOEA/D-ACDP
Input:
N : the number of subproblems.
Tmax : the maximal generation.
N weight vectors: λ1 , . . . , λN .
T : the size of the neighborhood.
δ: the selecting probability from neighbors.
nr : the maximal number of solutions replaced by a child.
θ0 , α: the parameters of ACDP method.
Output: N S : a set of feasible non-dominated solutions
1
N
1 Decompose a CMOP into N subproblems associated with λ , . . . , λ .
1
N
2 Generate an initial population P = {x , . . . , x }.
log(N )
3 Initialize cp to be log(1+α) .
∗
4 Initialize the ideal point z = (z1 , . . . , zm ).
i
i
5 For each i = 1, . . . , N , set B(i) = {i1 , . . . , iT }, where λ 1 , . . . , λ T are the
i
T closest weight vectors to λ .
6 for k ← 1 to Tmax do
7
if k ≤ αTmax then
k
8
Set θ(k) according to θ(k) = θ0 (1 + Tmax
)cp .
9
else
10
Set θ(k) to be equal to π2
11
end
12
Update pf in the current generation.
13
Generate a random permutation rp from {1, . . . , N }.
14
for i ← 1 to N do
15
Generate a random number r ∈ [0, 1].
16
j = rp(i).
17
if r < δ then
18
S = B(j)
19
else
20
S = {1, . . . , N }
21
end
22
Generate yj through DE and polynomial mutation operators.
23
Update the current ideal point.
24
Set c = 0.
25
while c 6= nr and S 6= ∅ do
26
select an index j from S randomly, S = S\{j}.
27
result = U pdateSubproblems(xj , yj , θ(k), pf )
28
if result == true then c = c + 1;
29
end
30
end
S
31
N S = NondominatedSelect(N S P )
32 end
10
updated by a newly generated solution yj . In line 22, the differential evolution
(DE) crossover operator is adopted to generate a new solution yj . Meanwhile,
yj is further mutated by the polynomial mutation operator. The ideal point
z ∗ is updated in line 23. Lines 24-39 update subproblems. In line 27, the subproblems are updated based on the ACDP approach whose detailed psuecode is
listed in Algorithm 2. At the end of each generation, non-dominated solutions
(N S) in the population are selected to update the external archive based on
non-dominated sorting in line 31.
Algorithm 2: Subproblem Update
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Function result = UpdateSubproblems(xj , yj , θ(k), pf )
result = f alse
if φ(yj ) == 0 and φ(xj ) == 0 then
if g te (yi |λj , z ∗ ) ≤ g te (xj |λj , z ∗ ) then
xj = yj
result = ture
end
else
if angle(F(yj ), F(xj ), z ∗ ) < θ(k) then
if φ(yj ) < φ(xj ) then
xj = y j
result = ture
end
else
if rand() < pf then
if g te (yi |λj , z ∗ ) ≤ g te (xj |λj , z ∗ ) then
xj = yj
result = ture
end
end
end
end
return result
end
In Algorithm 2, the algorithm updates a subproblem in terms of Eq. (7).
Lines 3-7 denote that when two feasible solutions xj and yj are compared, the
one with a better aggregation value is selected. Lines 9-12 denote that when at
least one of two solutions xj and yj is infeasible, if the angle between them in
the objective space is lower than θ, the solution with a lower constraint violation
is selected. Lines 13-17 denote that when at least one of two solutions xj and
yj is infeasible, if the angle between them in the objective space is larger than
θ, the solution with a lower aggregation value will be selected with a probability
of pf .
11
4. Experimental Study
4.1. Test Instances LIR-CMOPs
To evaluate the performance of the proposed MOEA/D-ACDP, 14 constrained multi-objective test problems with large infeasible regions in the objective space are used [27, 28].
The general characteristic of LIR-CMOPs is that their real PFs are blocked
by a number of large infeasible regions, and thus hard to be found during an
evolutionary process. Their constraint functions are comprised of controllable
shape functions and distance functions [29]. More specifically, the shape functions are used to turn the PF shapes as convex or concave, while the distance
functions are adopted to adjust the convergence difficulty for CMOEAs. Fig. 4
and Fig. 5 plot the feasible regions of LIR-CMOPs with two or three objectives,
respectively.
4.2. Real-world Engineering Optimization: I-beam
To evaluate the performance of MOEA/D-ACDP for solving real world optimization problems, an engineering optimization problem with two conflicting
objectives is studied.
As defined in [30], an optimization problem of I-beam is a bi-objective constrained optimization problem which needs to minimize the following objectives
simultaneously:
1. Cross section area of the beam;
2. Static deflection of the beam for the displacement under the force P .
To study the landscape in the objective space of the I-beam optimization problem, 1,000,000 sampling solutions are generated, where 850,000 solutions are generated randomly, and the other 150,000 solutions are generated
by MOEA/D-ACDP. In Fig. 6, it is observed that there exist a few infeasible
regions (the proportion of feasible solutions in all sampling solutions p = 0.5339,
which means that nearly a half of selected points are infeasible.) in the objective
space for the I-beam optimization problem.
4.3. Experimental Settings
To evaluate the performance of the proposed MOEA/D-ACDP, four popular CMOEAs (C-MOEA/D, MOEA/D-CDP, MOEA/D-Epsilon and MOEA/DSR), with differential evolution (DE) crossover operator, are adopted and tested
on LIR-CMOP1-14 and I-beam optimization problem. The detailed parameters
are listed as follows:
1. Mutation probability P m = 1/n (n is the number of decision variables)
and its distribution index is set to 20. For DE operator, CR = 1.0, f = 0.5.
2. Population size: N = 300. Neighborhood size: T = 30.
3. Stopping condition: each algorithm runs for 30 times independently, and
stops when 150,000 function evaluations are reached.
12
(a) LIR-CMOP1
(b) LIR-CMOP2
(c) LIR-CMOP3
(d) LIR-CMOP4
(e) LIR-CMOP5
(f) LIR-CMOP6
(g) LIR-CMOP7
(h) LIR-CMOP8
(i) LIR-CMOP9
(j) LIR-CMOP10
(k) LIR-CMOP11
(l) LIR-CMOP12
Figure 4: Illustrations of the feasible and infeasible regions of LIR-CMOP1-12.
(a) LIR-CMOP13
(b) LIR-CMOP14
Figure 5: Illustrations of the infeasible regions of LIR-CMOP13-14.
13
4. Probability of selecting individuals in the neighborhood: δ = 0.9.
5. The maximal number of solutions replaced by a child: nr = 2.
6. Parameter setting in MOEA/D-ACDP: α = 0.8 and θ0 = π/2N .
7. Parameter setting in MOEA/D-Epsilon: Tc = 400, cp = 2 and θ = 0.05N .
8. Parameter setting in MOEA/D-SR: Sr = 0.01.
4.4. Performance Metric
To measure the performance of MOEA/D-ACDP, C-MOEA/D, MOEA/DCDP, MOEA/D-Epsilon and MOEA/D-SR, two widely used metrics inverted
generation distance (IGD) [31] and hypervolume (HV ) [32] are adopted as
evaluation metrics. Their definitions are listed as follows.
• Inverted Generational Distance (IGD):
IGD is a metric which evaluates the performance related to convergence and
diversity simultaneously. Let P ∗ be a set of uniformly distributed points in the
ideal PF. Let A denote an approximate PF achieved by a certain CMOEA. The
metric IGD that represents average distance from P ∗ to A is defined as:
P
d(y ∗ , A)
∗
∗
y ∈P
IGD(P ∗ , A) =
|P ∗ |
(9)
p
Pm
∗
2
d(y ∗ , A) = min{
i=1 (yi − yi ) }
y∈A
In our experiment, for CMOPs with two objectives, 1000 points are sampled uniformly from the PF to constitute P ∗ . For CMOPs with three objectives, 10000
points are sampled uniformly from the PF to constitute P ∗ . A smaller IGD
represents a better performance regarding to both diversity and convergence.
• Hypervolume (HV ):
HV reflects the closeness between the non-dominated set achieved by a CMOEA
and the representative PF. The larger HV means that the corresponding nondominated set is closer to the true PF.
!
[
r
HV (S) = V OL
[f1 (x), z1r ] × ...[fm (x), zm
]
(10)
x∈S
r T
where V OL(·) is the Lebesgue measure, zr = (z1r , ..., zm
) is a reference point
in the objective space. For the test instances LIR-CMOPs, the reference point
is set as 1.4 times the nadir point of the real PF. The HV with a larger value
represents the better performance regarding to both diversity and convergence.
14
Figure 6: The distribution of the I-Beam problem.
As the real PF of the I-beam optimization problem is not known, IGD metric
can not be calculated. The experiment uses the HV metric [32] to measure the
performance of these CMOEAs. In the I-beam optimization case, the reference
point is set as z r = [850, 0.0615]T .
4.5. Discussion of Experimental Results
4.5.1. Performance Evaluation on LIR-CMOP Test Instances
The results of the IGD values on LIR-CMOP1-14 achieved by five CMOEAs
in 30 independent runs are shown in Table 1.
As discussed in Section 4, LIR-CMOP1-14 all have large infeasible regions
in their objective space. For LIR-CMOP3-14, MOEA/D-ACDP significantly
outperforms the other four compared algorithms in terms of the IGD metric.
For LIR-CMOP1-2, MOEA/D-ACDP significantly outperforms C-MOEA/D,
MOEA/D-CDP and MOEA/D-Epsilon.
The results of the HV values on LIR-CMOP1-14 achieved by five CMOEAs
in 30 independent runs are shown in Table 2. For LIR-CMOP2-14, MOEA/DACDP significantly outperforms the compared algorithms in terms of the HV
metric. For the LIR-CMOP1, MOEA/D-ACDP significantly outperforms CMOEA/D, MOEA/D-CDP and MOEA/D-Epsilon.
Fig. 7 (a) shows the final external archives achieved by MOEA/D-ACDP
and other four CMOEAs with the median IGD values on LIR-CMOP3 during
30 independent runs. It is obvious that MOEA/D-ACDP can almost converge
to the whole real PF and has the best diversity among the five CMOEAs.
In Fig. 7 (b), the results of each CMOEA with the median IGD values
on LIR-CMOP5 during 30 independent runs are shown. The external archive
achieved by MOEA/D-ACDP covers the real PF. However, the other four
CMOEAs are trapped into local optimum. As shown in Fig. 7 (c), for LIRCMOP10, MOEA/D-ACDP performs the best in terms of convergency. In Fig.
15
7 (d), for LIR-CMOP11, it shows that MOEA/D-ACDP can get the most of
the PF, but the other four algorithms can only achieve a few parts of the PF.
It is worthwhile to point out that for three-objective test instances (LIRCMOP13 and LIR-CMOP14), MOEA/D-ACDP also performs significantly better than the other four CMOEAs.
Based on the above performance comparison on the fourteen test instances
LIR-CMOP1-14, it is clear that MOEA/D-ACDP outperforms the other four
decomposition-based CMOEAs on most of cases.
A common feature of the above test instances LIR-CMOPs is that they all
have large infeasible regions in their objective space. The experimental results
demonstrate that the proposed ACDP method can deal with CMOPs well by
taking advantage of angle information of the working population.
Table 1: IGD results of MOEA/D-ACDP and the other four CMOEAs on LIR-CMOP1-14
test instances
Test Instances
mean
std
mean
LIR-CMOP2
std
mean
LIR-CMOP3
std
mean
LIR-CMOP4
std
mean
LIR-CMOP5
std
mean
LIR-CMOP6
std
mean
LIR-CMOP7
std
mean
LIR-CMOP8
std
mean
LIR-CMOP9
std
mean
LIR-CMOP10
std
mean
LIR-CMOP11
std
mean
LIR-CMOP12
std
mean
LIR-CMOP13
std
mean
LIR-CMOP14
std
LIR-CMOP1
MOEA/D-ACDP
5.159E-02
1.815E-02
2.269E-02
9.418E-03
4.659E-02
1.850E-02
2.784E-02
1.477E-02
1.771E-02
2.965E-02
1.757E-01
4.129E-02
1.408E-01
4.385E-02
1.812E-01
4.854E-02
3.595E-01
5.345E-02
1.388E-01
1.148E-01
1.318E-01
4.487E-02
1.497E-01
9.985E-03
7.414E-02
2.727E-03
6.732E-02
1.918E-03
C-MOEA/D
1.591E-01†
3.534E-02
1.462E-01†
4.141E-02
2.309E-01†
4.135E-02
2.080E-01†
4.197E-02
1.162E+00†
2.180E-01
1.265E+00†
3.067E-01
1.620E+00†
3.036E-01
1.607E+00†
2.680E-01
4.981E-01†
6.991E-02
3.775E-01†
7.446E-02
4.422E-01†
1.759E-01
3.597E-01†
1.074E-01
1.266E+00†
2.173E-01
1.235E+00†
1.209E-01
MOEA/D-CDP
1.348E-01†
5.996E-02
1.549E-01†
2.966E-02
2.268E-01†
4.403E-02
2.188E-01†
3.766E-02
1.207E+00†
1.660E-02
1.303E+00†
2.319E-01
1.623E+00†
2.905E-01
1.631E+00†
2.464E-01
4.868E-01†
5.372E-02
3.774E-01†
6.858E-02
4.662E-01†
1.439E-01
3.236E-01†
1.023E-01
1.289E+00†
6.321E-02
1.103E+00†
3.857E-01
MOEA/D-Epsilon
8.234E-02†
5.321E-02
4.708E-02†
1.339E-02
7.858E-02†
2.978E-02
5.662E-02†
3.366E-02
1.201E+00†
1.963E-02
1.231E+00†
3.602E-01
1.568E+00†
4.101E-01
1.577E+00†
3.767E-01
4.962E-01†
6.987E-02
3.257E-01†
9.833E-02
4.154E-01†
1.508E-01
3.680E-01†
8.664E-02
1.183E+00†
3.456E-01
1.127E+00†
3.329E-01
MOEA/D-SR
4.406E-02
3.360E-02
2.057E-02
1.072E-02
1.529E-01†
7.688E-02
2.038E-01†
7.907E-02
1.123E+00†
2.842E-01
1.175E+00†
3.967E-01
1.136E+00†
7.315E-01
1.369E+00†
5.735E-01
4.813E-01†
4.571E-02
2.821E-01†
1.135E-01
3.489E-01†
1.129E-01
3.012E-01†
8.989E-02
1.093E+00†
4.269E-01
1.143E+00†
3.002E-01
Wilcoxons rank sum test at a 0.05 significance level is performed between MOEA/D-ACDP
and each of the other four CMOEAs. † and ‡ denote that the performance of the
corresponding algorithm is significantly worse than or better than that of MOEA/D-ACDP,
respectively. The best mean is highlighted in boldface.
4.5.2. Performance Evaluation on I-beam Optimization Problem
The experimental results of HV values of MOEA/D-ACDP and the four
other CMOEAs on the I-beam optimization problem are shown in Table 3. It
can be observed that MOEA/D-ACDP significantly outperforms the compared
CMOEAs on this engineering problem.
16
Table 2: HV results of MOEA/D-ACDP and the other four CMOEAs on LIR-CMOP1-14
test instances
Test Instances
mean
std
mean
LIR-CMOP2
std
mean
LIR-CMOP3
std
mean
LIR-CMOP4
std
mean
LIR-CMOP5
std
mean
LIR-CMOP6
std
mean
LIR-CMOP7
std
mean
LIR-CMOP8
std
mean
LIR-CMOP9
std
mean
LIR-CMOP10
std
mean
LIR-CMOP11
std
mean
LIR-CMOP12
std
mean
LIR-CMOP13
std
mean
LIR-CMOP14
std
LIR-CMOP1
MOEA/D-ACDP
1.365E+00
2.493E-02
1.737E+01
1.306E-02
1.188E+00
4.929E-02
1.421E+00
1.946E-02
1.903E+00
5.658E-02
1.280E+00
4.613E-02
3.408E+00
1.409E-01
3.330E+00
1.461E-01
4.080E+00
9.501E-02
3.755E+00
2.208E-01
5.004E+00
1.564E-01
6.713E+00
05.874E-02
7.897E+00
2.943E-02
8.641E+00
1.546E-02
C-MOEA/D
9.499E-01†
7.038E-02
1.395E+01†
8.154E-02
7.558E-01†
5.730E-02
1.069E+00†
6.952E-02
1.192E-01†
3.352E-01
7.863E-02†
3.011E-01
2.990E-01†
6.927E-01
3.246E-01†
5.878E-01
3.715E+00†
2.079E-01
3.274E+00†
1.623E-01
3.937E+00†
6.479E-01
5.977E+00†
3.855E-01
6.444E-01†
1.317E+00
7.766E-01†
6.140E-01
MOEA/D-CDP
1.009E+00†
1.298E-01
1.374E+01†
6.160E-02
7.600E-01†
5.809E-02
1.051E+00†
5.462E-02
5.805E-02†
4.042E-04
4.312E-02†
2.362E-01
2.886E-01†
6.348E-01
2.695E-01†
5.297E-01
3.755E+00†
1.600E-01
3.268E+00†
1.416E-01
3.842E+00†
5.507E-01
6.134E+00†
3.617E-01
4.728E-01†
2.689E-01
1.627E+00†
2.473E+00
MOEA/D-Epsilon
1.353E+00†
4.417E-02
1.705E+01†
1.693E-02
1.184E+00†
2.898E-02
1.390E+00†
4.405E-02
5.829E-02†
2.022E-04
1.325E-01†
4.251E-01
4.055E-01†
8.879E-01
3.859E-01†
8.166E-01
3.724E+00†
2.033E-01
3.385E+00†
2.122E-01
4.038E+00†
5.727E-01
6.010E+00†
3.074E-01
1.092E+00†
2.0522E+00
1.430E+00†
2.095E+00
MOEA/D-SR
1.376E+00
3.974E-02
1.736E+01†
1.890E-02
9.313E-01†
1.620E-01
1.089E+00†
1.360E-01
1.707E-01†
4.442E-01
1.682E-01†
4.061E-01
1.313E+00†
1.567E+00
8.287E-01†
1.244E+00
3.752E+00†
1.142E-01
3.477E+00†
2.383E-01
4.274E+00†
4.463E-01
6.240E+00†
2.950E-01
1.513E+00†
2.422E+00
1.269E+00†
1.919E+00
Wilcoxons rank sum test at a 0.05 significance level is performed between MOEA/D-ACDP
and each of the other four CMOEAs. † and ‡ denotes that the performance of the
corresponding algorithm is significantly worse than or better than that of MOEA/D-ACDP,
respectively. The best mean is highlighted in boldface.
17
(a) LIR-CMOP3
(b) LIR-CMOP5
(c) LIR-CMOP10
(d) LIR-CMOP11
Figure 7: The non-dominated solutions achieved by each algorithm with the median IGD in
the 30 independent runs for LIR-CMOP3, LIR-CMOP5, LIR-CMOP10 and LIR-CMOP11.
18
To further study the superiority of the proposed method MOEA/D-ACDP,
the non-dominated solutions achieved by each CMOEA during the 30 independent runs are plotted in Fig. 8 (a)-(e).The non-dominated set of all the above
solutions generates a set of ideal reference points. It is clear that the external
archive obtained by MOEA/D-ACDP has a better performance of convergence.
The box plot of HV values of the five CMOEAs is shown in Fig. 8 (f), which
further verifies that MOEA/D-ACDP outperforms the other four CMOEAs on
the I-beam optimization problem.
(a) MOEA/D-ACDP
(b) C-MOEA/D
(c) MOEA/D-CDP
(d) MOEA/D-Epsilon
(e) MOEA/D-SR
(f) The box plots of each
CMOEA
Figure 8: The non-dominated solutions achieved by each algorithm during the 30 independent
runs are plotted in (a)-(e). In (f), the box plots of each CMOEA are plotted.
Table 3: HV results of MOEA/D-ACDP and the other four CMOEAs on the I-Beam optimization problem
Test Instances
mean
std
MOEA/D-ACDP
3.583E+01
1.950E-01
C-MOEA/D
3.481E+01†
2.968E-01
MOEA/D-CDP
3.518E+01†
2.078E-01
MOEA/D-Epsilon
3.514E+03†
2.034E-01
MOEA/D-SR
3.477E+03†
1.248E+00
Wilcoxons rank sum test at a 0.05 significance level is performed between MOEA/D-ACDP
and each of the other four CMOEAs. † and ‡ denote that the performance of the
corresponding algorithm is significantly worse than or better than that of MOEA/D-ACDP,
respectively. The best mean is highlighted in boldface.
5. Conclusions
This paper proposes a new constraint-handling mechanism named ACDP.
It utilizes the angle information of any two solutions to dynamically maintain
the diversity of the population during the evolutionary process. A set of CMOP
19
instances named LIR-CMOP1-14 are tested. All the test instances have large
infeasible regions in their objective space, which make general CMOEAs difficult
to achieve the real PFs. Compared with the other four popular CMOEAs, the
proposed algorithm can help the population to go across large infeasible regions
more effectively. Additionally, the experimental results demonstrate that the
proposed algorithm can work well in the real-world engineering problem. Thus,
we can conclude that MOEA/D-ACDP outperforms the other four CMOEAs
when CMOPs. In summary, MOEA/D-ACDP has following advantages:
• The proposed MOEA/D-ACDP utilizes the angle information of solutions
to maintain the diversity of the population for CMOPs.
• MOEA/D-ACDP enhances the convergence to PF by exploring feasible
and infeasible regions simultaneously during the evolutionary process, instead of wasting the useful information of the infeasible solutions.
Future work will focus on novel constraint-handling mechanisms to solve
CMOPs. A study on developing new mechanisms of mining more useful information during the evolutionary process to further improve the performance of
the proposed algorithm will be conducted.
Acknowledgment
This work was supported in part by the National Natural Science Foundation of China under Grant (61175073, 61300159, 61332002, 51375287) , the
Guangdong Key Laboratory of Digital signal and Image Processing, the Science
and Technology Planning Project of Guangdong Province (2013B011304002)
and the Project of Educational Commission of Guangdong Province, China
2015KGJHZ014).
References
References
[1] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic algorithm: Nsga-ii, IEEE Transactions on Evolutionary
Computation 6 (2) (2002) 182–197.
[2] T. Murata, H. Ishibuchi, Moga: Multi-objective genetic algorithms, in:
Evolutionary Computation, 1995., IEEE International Conference on,
Vol. 1, IEEE, 1995, p. 289.
[3] D. W. Corne, N. R. Jerram, J. D. Knowles, M. J. Oates, Pesa-ii: regionbased selection in evolutionary multiobjective optimization, in: Conference
on Genetic and Evolutionary Computation, 2001, pp. 283–290.
[4] E. Zitzler, M. Laumanns, L. Thiele, et al., Spea2: Improving the strength
pareto evolutionary algorithm (2001).
20
[5] J. Horn, N. Nafpliotis, D. E. Goldberg, A niched pareto genetic algorithm for multiobjective optimization, in: Evolutionary Computation,
1994. IEEE World Congress on Computational Intelligence., Proceedings
of the First IEEE Conference on, 2002, pp. 82–87 vol.1.
[6] Q. Zhang, H. Li, Moea/d: A multiobjective evolutionary algorithm based
on decomposition, IEEE Transactions on Evolutionary Computation 11 (6)
(2007) 712–731.
[7] H. Li, Q. Zhang, Multiobjective optimization problems with complicated
pareto sets, moea/d and nsga-ii, IEEE Transactions on Evolutionary Computation 13 (2) (2009) 284–302.
[8] H.-L. Liu, F. Gu, Q. Zhang, Decomposition of a Multiobjective Optimization Problem Into a Number of Simple Multiobjective Subproblems, IEEE
Transactions on Evolutionary Computation 18 (3) (2014) 450–455.
[9] X. Cai, Y. Li, Z. Fan, Q. Zhang, An External Archive Guided Multiobjective Evolutionary Algorithm Based on Decomposition for Combinatorial Optimization, IEEE Transactions on Evolutionary Computation 19 (4)
(2015) 508–523.
[10] X. Cai, Z. Yang, Z. Fan, Q. Zhang, Decomposition-Based-Sorting
and Angle-Based-Selection for Evolutionary Multiobjective and ManyObjective Optimization, IEEE Transactions on Cybernetics PP (99) (2016)
1–14.
[11] E. Zitzler, S. Künzli, Indicator-based selection in multiobjective search,
Lecture Notes in Computer Science 3242 (2004) 832–842.
[12] N. Beume, B. Naujoks, M. Emmerich, Sms-emoa: Multiobjective selection based on dominated hypervolume, European Journal of Operational
Research 181 (3) (2007) 1653–1669.
[13] J. Bader, E. Zitzler, Hype: an algorithm for fast hypervolume-based manyobjective optimization., Evolutionary Computation 19 (1) (2011) 45.
[14] S. Jiang, J. Zhang, Y.-S. Ong, A. N. Zhang, P. S. Tan, A Simple and
Fast Hypervolume Indicator-Based Multiobjective Evolutionary Algorithm,
IEEE Transactions on Cybernetics 45 (10) (2015) 2202–2213.
[15] X. Cai, Z. Hu, Z. Fan, A novel memetic algorithm based on invasive weed
optimization and differential evolution for constrained optimization, Soft
Computing 17 (10) (2013) 1893–1910.
[16] Z. Hu, X. Cai, Z. Fan, An improved memetic algorithm using ring neighborhood topology for constrained optimization, Soft Computing 18 (10)
(2013) 2023–2041.
21
[17] C. A. C. Coello, Theoretical and numerical constraint-handling techniques
used with evolutionary algorithms: a survey of the state of the art, Computer Methods in Applied Mechanics and Engineering 191 (11–12) (2002)
1245–1287.
[18] T. P. Runarsson, X. Yao, Stochastic ranking for constrained evolutionary
optimization, IEEE Transactions on evolutionary computation 4 (3) (2000)
284–294.
[19] T. Takahama, S. Sakai, Constrained optimization by the ε constrained
differential evolution with gradient-based mutation and feasible elites, in:
Evolutionary Computation, 2006. CEC 2006. IEEE Congress on, 2006, pp.
1–8.
[20] H. K. Singh, T. Ray, W. Smith, C-psa: Constrained pareto simulated annealing for constrained multi-objective optimization, Information Sciences
180 (13) (2010) 2499–2513.
[21] K. Deb, D. Kalyanmoy, Multi-Objective Optimization Using Evolutionary
Algorithms, John Wiley & Sons, Inc, 2001.
[22] Q. Zhang, A. Zhou, S. Zhao, P. N. Suganthan, W. Liu, S. Tiwari, Multiobjective optimization test instances for the cec 2009 special session and
competition, University of Essex, Colchester, UK and Nanyang technological University, Singapore, special session on performance assessment of
multi-objective optimization algorithms, technical report 264.
[23] K. Miettinen, Nonlinear multiobjective optimization, volume 12 of international series in operations research and management science (1999).
[24] M. Asafuddoula, T. Ray, R. Sarker, K. Alam, An adaptive constraint handling approach embedded moea/d, in: Evolutionary Computation, 2012,
pp. 1–8.
[25] M. A. Jan, R. A. Khanum, A study of two penalty-parameterless constraint
handling techniques in the framework of MOEA/D, Applied Soft Computing 13 (1) (2013) 128–148.
[26] Z. Yang, X. Cai, Z. Fan, Epsilon constrained method for constrained multiobjective optimization problems: some preliminary results, in: Companion
Publication of the 2014 Conference on Genetic and Evolutionary Computation, 2014, pp. 1181–1186.
[27] Z. Fan, W. Li, X. Cai, H. Li, K. Hu, Q. Zhang, K. Deb, E. D. Goodman,
Difficulty adjustable and scalable constrained multi-objective test problem
toolkit, arXiv preprint arXiv:1612.07603.
[28] Z. Fan, W. Li, X. Cai, H. Huang, Y. Fang, Y. You, J. Mo, C. Wei, E. Goodman, An improved epsilon constraint-handling method in moea/d for cmops
with large infeasible regions, arXiv preprint arXiv:1707.08767.
22
[29] S. Huband, P. Hingston, L. Barone, L. While, A review of multiobjective
test problems and a scalable test problem toolkit, IEEE Transactions on
Evolutionary Computation 10 (5) (2006) 477–506.
[30] A. Osyczka, Multicriteria optimization for engineering design, in: Design
Optimization, 1985, pp. 193–227.
[31] P. A. N. Bosman, D. Thierens, The balance between proximity and diversity in multiobjective evolutionary algorithms, IEEE Transactions on
Evolutionary Computation 7 (2) (2003) 174–188.
[32] E. Zitzler, L. Thiele, Multiobjective evolutionary algorithms: a comparative case study and the strength pareto approach, IEEE Transactions on
Evolutionary Computation 3 (4) (1999) 257–271.
23
| 9cs.NE
|
Working Draft: This is work in progress!
Datasheets for Datasets∗
Timnit Gebru1 , Jamie Morgenstern2 , Briana Vecchione3 ,
Jennifer Wortman Vaughan1 , Hanna Wallach1 ,
Hal Daumé III1,4 , and Kate Crawford1,5
arXiv:1803.09010v1 [cs.DB] 23 Mar 2018
1
Microsoft Research: {timnit.gebru, hal3, jenn, wallach, kate}@microsoft.com
2
Georgia Tech: [email protected]
3
[email protected]
4
University of Maryland
5
AI Now Institute
March 28, 2018
Abstract
Currently there is no standard way to identify how a dataset was created, and what
characteristics, motivations, and potential skews it represents. To begin to address
this issue, we propose the concept of a datasheet for datasets, a short document to
accompany public datasets, commercial APIs, and pretrained models. The goal of
this proposal is to enable better communication between dataset creators and users,
and help the AI community move toward greater transparency and accountability.
By analogy, in computer hardware, it has become industry standard to accompany
everything from the simplest components (e.g., resistors), to the most complex microprocessor chips, with datasheets detailing standard operating characteristics, test
results, recommended usage, and other information. We outline some of the questions
a datasheet for datasets should answer. These questions focus on when, where, and
how the training data was gathered, its recommended use cases, and, in the case of
human-centric datasets, information regarding the subjects’ demographics and consent as applicable. We develop prototypes of datasheets for two well-known datasets:
Labeled Faces in The Wild [33] and the Pang & Lee Polarity Dataset [45].
1
Introduction
Artificial Intelligence is rapidly moving from a purely academic discipline to a technology
that is embedded in everyday products. Our cars, homes, and computers are regularly
controlled by machine learning algorithms. These algorithms are also used by state and
federal agencies and corporations to predict our behavior: law enforcement uses facial
recognition software to catch suspects [30, 52], the US criminal justice system uses risk
assessment scores to estimate a person’s likelihood of committing a crime [5], and companies use models to filter job applications before requesting interviews [37, 32]. Critical
components of our world’s infrastructure rely on machine learning models, for example to
∗
Working Draft: This is work in progress, so writing may be less polished than usual and
is subject to revision. New versions forthcoming. We’d love to incorporate your feedback,
so if you have comments for us, please let us know!
1
Working Draft: This is work in progress!
monitor and manage our water systems [41] and power grids [11]. Since 2011, a majority
of financial trades made in US on our stock exchanges are made automatically [35].
By definition, all such machine learning models are trained from some data: indeed,
the datasets on which they are trained are an integral in determining their functionality.
The ubiquity with which data is gathered and models are trained and tested on that data
allows for incredible innovation, but also poses a number of risks and questions: Under
what circumstances do we expect such models to perform well? In different circumstances,
how will the model’s performance degrade? Is this dataset fixed, or is more data gathered
over time? Is it possible to observe and ameliorate any bias perpetuated and exacerbated
by these models when making choices affecting people? All of these questions depend
upon how the data was gathered, cleaned and processed, and how a model incorporates
that data. Without this information, even experts trained in machine learning cannot hope
to say with any certainty when a model might fail badly. Moreover, there is currently no
standard practice for carefully testing or describing these datasets, APIs, or freely available
models. This is of particular concern when we consider high-stakes uses of these models
or datasets.
In this paper, we argue that the first step towards transparency must involve attaching
significant, standardized information about any dataset, API, or pretrained model. We
concentrate specifically on datasets, and recommend a standardized description format
with standard questions. Even if the eventual goal is to understand pretrained models or
APIs, understanding datasets is an important first step: regardless of how they are trained,
models and APIs are typically evaluated against fixed datasets, and understanding the
characteristics of that evaluation is of paramount importance. We draw inspiration from
standardized forms of sharing information and rigorous testing conducted in the much
more mature field of hardware. This document, called henceforth a datasheet, would be
a short (typically 3-5 page) document detailing any tests that have been conducted on a
dataset, its recommended usage, its collection procedure, any regulations governing its use,
and so on. A datasheet for a pretrained model would contain information pertaining to
the model’s training data and tested behavior.
We structure this paper as follows. Section 2 discusses the motivation and context for
our work, Section 3 briefly discusses the evolution of safety standards in other industries
to draw a comparison with AI. Section 4 discusses the concept of datasheets in hardware, and Section 5 outlines the most important questions that should be answered by a
datasheet. Two prototypes of datasheets for datasets can be found in Appendix B. The
paper ends with a discussion of challenges and future work in Section 6.
2
Context
Many datasets have little or no documentation describing the settings under which they
were gathered, the characteristics of the training sets, how representative the dataset is of
a particular population, or recommended usage. This lack of information makes it difficult
to assess what contexts models and APIs trained on these datasets can be used for and
what scenarios should be avoided. Most of these APIs are not accompanied by detailed
information describing recommended usage, standard operating characteristics, and tests
performed to verify these conditions.
Of particular concern is the recently discovered extent to which AI systems exhibit and
amplify biases. Buolamwini and Gebru [8] showed that commercial gender classification
APIs have near perfect performance for light skinned males while error rates for darker
2
Working Draft: This is work in progress!
skinned females can be as high as 33%.1 In a different setting, Bolukbasi et al. [6] showed
that word embeddings trained on news articles exhibit gender bias, finishing the analogy
“Man is to computer programmer as woman is to X” with “home maker,” a stereotypical
role for women. Caliskan et al. [9] showed that racial biases also exist in word embeddings:
traditional European-American names, for example, are more closely related to words that
are considered pleasant (e.g., joy); those associated with African-American names however,
are closer to words such as agony. These biases can have dire consequences that might not
be easily discovered. For example, Bolukbasi et al. [6] argue that bias in word embeddings
can result in hiring discrimination. Much like a faulty resistor or a capacitor in a circuit,
the effects of a biased AI-based component can propagate throughout a system making
them difficult to track down.
One of the biggest challenges in the deployment of AI technology is deploying systems,
built using datasets or pretrained models, in unsuitable environments. Such systems can
exhibit poor or unpredictable performance when deployed: the models’ behavior on some
benchmark may say very little about their behavior in a different setting. Such unpredictability is of concern when these systems make high-stake real-time decisions, such as
controlling a large amount of trading volume on the stock exchanges or allocating valuable
resources to different markets. Unpredictability is of even greater concern when the system
is used to interact with or make decisions directly influencing humans.
This problem of unintentional misuse of datasets is often exacerbated when the users
are not domain experts. We believe that this problem can be mitigated, at least partially,
by accompanying datasets with datasheets that describe their creation, strengths and limitations. While this is not the same as making everyone a domain expert, it gives an
opportunity for domain experts to easily communicate what they know about the dataset
and its limitations to whomever may be using it. This is particularly important today,
when companies such as Google, Amazon, Microsoft, and Facebook are moving toward
“democratizing AI” by creating toolboxes such as Cloud AutoML that can be trained by
those with little-to-no machine learning expertise and domain knowledge [12]. Pre-trained
models and standard datasets are freely available for use with open source tools such as
Caffe [22], Tensorflow [2], and PyTorch [48]. As powerful machine learning tools become
available to a much broader set of users, it becomes even more important to enable those
users to understand all the implications of their work.
Part of the educational gap, outside traditional “ivory tower” education, is being covered
by organizations like Coursera [20], Udacity [53], Fast.ai [26], and others, which offer online
AI courses to those with no prior experience. A goal of these educational platforms is to
enable people around the world to use AI to solve problems in their communities. For
instance, one of Fast.ai’s missions is “to get deep learning into the hands of as many people
as possible, from as many diverse backgrounds as possible” [26]. These readily available
AI courses and toolkits belong to a movement whose laudable goal is to enable many
people to integrate AI into their systems. Coupling this educational strategy, which often
(in particular in the case of Fast.ai) includes explicit training in dataset bias and ethics
(topics sometimes lacking even in the “ivory tower”), with datasheets that can explain the
context and biases in existing datasets, can much more quickly enable progress by both
domain experts and AI experts.
1
The evaluated APIs provided the labels of female and male, failing to address the complexities of
gender beyond binary.
3
Working Draft: This is work in progress!
3
The Evolution of Safety Standards in Other Industries
We briefly discuss the evolution of safety standards in 3 different industries: automobiles,
health care and electronics. Understanding the dangers that were posed by the proliferation
of new technology in these industries, and the safety measures that were put in place to
combat them, can help us carve out a path forward for AI.
3.1
The Automobile Industry
Similar to our current hopes for AI to positively transform society, the introduction of
automobiles promised to expand mobility and provide additional recreational, social, and
economic opportunities. However, much like current AI technology, the automobile was
introduced with few safety checks or regulations in place. When cars first became available
in the United States, there were no speed limits, stop signs, traffic lights, driver education,
or regulations pertaining to seat belts or drunk driving [10]. Thus, the early 1900s saw
many deaths and injuries due to collisions, speeding, and reckless driving [31]. Much like
current debates regarding the future of AI, there were courtroom discussions and news
paper editorials outlining the possibility that the automobile was inherently evil [1].
It took many years in different parts of the world for laws, traffic signals, driver education and fines to be enacted. In the United States, for instance, driver’s licenses were
not fully implemented across all states until 1954 [43]. By the 1930s, a variety of technical safety responses were introduced such as four-wheel hydraulic brakes, shatter-resistant
windshields, and all-steel bodies [38]. Despite these safety standards, the automobile industry fell victim to similar “bad dataset” problems faced by AI technology, in particular
in the construction of the crash-test dummies used to evaluate vehicle safety. Although
crash-test dummies became a mandated part of U.S. safety standards in 1973, almost all the
crash-test dummies in use were modeled after prototypical male physiology [4]. It wasn’t
until almost four decades later that, in 2011, using “female” crash-test dummies became
mandatory for frontal crash tests [3]. Subsequent studies suggest that male-centric engineering design is responsible for disparate rates of vehicle injuries by sex: A safety study
of automobiles manufactured between 1998 and 2008 concluded that women wearing seat
belts were 47% more likely to be seriously injured than males in similar accidents [7].
Automobile safety standards in the US have continued to evolve since their introduction
in the early 1920s, and there are many countries without adequate road safety measures.
Road accidents are still the biggest global killer of teenagers [44]. In the case of seat belts
in particular, it was only in 1968 that features like padded dashboards and seat belts were
made mandatory in the United States [49]. Still, most motorists were reluctant to use
seat belts, and safety campaigns had to be sponsored to promote adoption. By analogy,
an “AI solution” is likely to require both laws and regulations (in particular in high-stakes
environments) and also social campaigns to promote best practices. This underscores the
need to aggressively work towards standardization of best practices for the creation and
proliferation of datasets and APIs in AI, and have a mechanism by which these practices
can continue to evolve.
3.2
Clinical Trials in Medicine
We can draw lessons from studying the evolution of safety measures for clinical trials, and
some of the harms that were caused by inadequate standards and unethical procedures.
Like the need for large scale data collection and experimentation before the deployment
of an AI system, clinical trials are an important step in any drug development. However,
4
Working Draft: This is work in progress!
the US legal system viewed clinical trials as a form of medical malpractice until well
into the 20th century, making standardized large-scale evaluations difficult [23]. After the
acceptance of certain clinical trials came various standards which were to be followed, most
of which were spurred by some atrocity or another committed in the name of science. For
example, the US government ran a number of experiments on its citizens without their
consent, from a public health study on syphilis where participants were not informed of
their disease [21], to radiation experiments [25, 39]. The poor, the imprisoned, minority
groups, pregnant women, and children comprised a majority of these study groups.
Currently, in the US, participants in a drug trial must be informed that the drug is
experimental and not proven to be effective, and subjects must be willing participants.
Prior to a drug being tested in a clinical trial, an Institutional Review Board and the Food
and Drug Administration must approve an application which shows evidence of the drug’s
relative safety (in terms of basic chemistry and animal testing results) and lays out the
design of the trial (including who will participate in the trial) [29].
The closest legal analog in AI are laws like the European Union’s General Data Protection Regulation (GDPR), which are starting to be deployed to ensure that people consent
to having their data used in the training of AI based models [47]. Data collection standards
are now a central topic of concern for scientific research broadly, and clinical trials are no
exception. The US National Institute of Health has a body of guidelines for how their
funded projects are to gather, store, and share human subject data [42].
Finally, the lack of diversity in clinical trial participants has led to the development
of drugs that do not work well for many groups of people. For example, eight out of ten
drugs pulled from circulation between 1997 and 2001 had more adverse effects for women,
suggesting clinical trials without representative samples did not accurately display risks
for those drugs for women [36]. As late as 2013, a majority of federally-funded clinical
trials still did not break down their results by sex [40]. In 2014, the FDA promoted an
action plan to make results of clinical trials broken down by subpopulation more easily
available [27]. In the late 1980s, the FDA moved to require different age groups participate
in clinical trials [28]. Not until 1998 was a regulation stating that safety and efficacy
data be provided broken down by sex, race, and age. These progressions parallel some
recent results showing disparities in accuracy of various AI based models by subpopulation
(e.g., [8]), and calls for more diverse datasets, inclusive testing, and standards in place to
reduce these disparities.
3.3
Electrical and Electronic Technologies
Similar to the current proliferation of AI within everyday products, electrical and electronic components are used in devices that are incredibly widespread. They are designed
into devices ranging from those used in communication (TV, radio, phones), transportation (automobiles, trains, planes), healthcare, military equipment, energy production, and
transmission. With the move towards smart homes and the internet of things [54], soon
one may be hard-pressed to find a synthetic object without electronic components.
Like datasets and the models trained on them, electronic components, such as resistors
or capacitors, are designed into a system whose larger goal may be far removed from the
task of that component. Thus, small deviations that may seem insignificant while studying
a component in isolation can result in dire consequences due to its interactions with the
rest of the system.
For example, while all types of resistors can be abstracted into an idealized mathematical model, different non-idealities are important depending on the context. The operating
temperature range of a power resistor meant for operation under high voltage conditions
5
Working Draft: This is work in progress!
is much more crucial than that for low power thin film resistors in a motherboard [55].
Even still, within the power resistor family, the safety considerations for those used in
power plants for instance are different from those in consumer electronics [55]. Thus,
the electronic component community has developed standards that specify ideal operation
characteristics, tests and manufacturing conditions for components manufactured with different tasks in mind.
Many of these standards are specified by the International Electrotechnical Comission
(IEC). According to the IEC, “Close to 20,000 experts from industry, commerce, government, test and research labs, academia and consumer groups participate in IEC Standardization work” [13]. After the International System of Electrical and Magnetic Units was
agreed to at the first International Electrical Congress in 1881, it became clear that many
other questions of standardization would arise [34]. The IEC was founded in 1906 and in
1938, it published an international vocabulary to unify terminology relating to electrical,
electronic and related technologies [13, 17]. There are currently 9, 000 IEC standards in
use today, with over 10 standards pertaining to different types of resistors alone [18]. For
example IEC 60195:2016 describes the recommended procedures to assess the magnitude
of current noise in fixed resistors of any type, IEC 60115-2:2014 provides standards for
leaded fixed low-power film resistors for use in electronic equipment, and IEC 60322 states
operating conditions and testing methodology for power resistors used in railway applications [14, 15, 16]. We argue that standards with similar detail and scope can be set in
place for datasets and pre-trained AI models used in different scenarios.
4
Datasheets for Electronic Components
We take our inspiration from the standardization of datasheets for electronic components.
All electronic components, ranging from the cheapest and most ubiquitous resistors, to
highly complex integrated circuits (like CPUs), are accompanied by datasheets characterizing their recommended operating conditions and other detailed technical characteristics.
While the information contained in any given datasheet depends on the specific product,
there are several aspects that are common to all datasheets (such as the manufacturer’s
name, and the product name and number).
Most datasheets start with a description of the component’s function and features, and
contain other specifications like the absolute maximum and minimum operating voltages.
In addition to the technical characteristics, datasheets contain physical details of the component (such as size and pin connections), and list of available packages. Datasheets can
also contain liability disclaimers to protect the manufacturer (e.g., in case the component
is used in high stakes environments like nuclear power plants or life support systems). If
the component design, manufacturing and testing adheres to some standard (e.g., an IEC
standard), this is typically also stated in the datasheet.
When one navigates to a product webpage, there is a datasheet associated with each
product. Figure 4 shows a screenshot of the the product page for a KEMET Corporation
tantalum capacitor [24]. The datasheet for this product (indicated by the red arrow), is
prominently featured along with other documentation. An example of a datasheet for a
miniature aluminum electrolytic capacitor is shown in Appendix A. Like many datasheets
for similar components, this one contains:
• A short description of the component’s function and notable features (including the
component’s compliance with the RoHS directive adopted by the European Union
restricting the use of certain hazardous materials [46])
6
Working Draft: This is work in progress!
Figure 1: A screenshot of the product webpage for a KEMET Corporation tantalum capacitor with the datasheet for the component highlighted in red.
• Various standard operating characteristics, such as its operating temperature range
and capacitance tolerance
• A diagram of the component showing its dimensions and pin connections
• Plots showing the change in various characteristics vs. time, temperature and frequency; for example, it is well known that capacitance decreases over time, and the
first graph measures this change across 1000 hours
Some examples of other datasheets are those for semiconductors [51] (56 pages), resistors [50] (2 pages), and other components [19] (18 pages).
4.1
What has driven the use of datasheets in hardware?
To better understand how we might hope for proliferation of datasheets for datasets, it is
useful to note some of the potential reasons for their standardization in hardware.
De facto industry standard. The “highest order bits” (one to two phrase summary describing a component, such as conductance or resistance) are nowhere near enough to
understand in what settings a component should/could be used, how it will behave
in a variety of settings, how robust it is, what size it is, and so forth. In order to
make informed purchasing decisions, one needs to know additional information as
provided in a datasheet. In the AI setting, many dataset characteristics need to be
outlined to understand their use cases, potential biases, and limitations.
Product Liability. The seller of a hardware component wants to clearly outline the appropriate settings for using their component. If the component fails during some use
and causes damage or injury, they could be liable if they have not clearly specified
that type of use as outside the component’s operating characteristics. In the AI
setting, many failure modes are not as clearly visible to all, and the liability may not
be easily traced back to datasets. This could make the adoption of semi-standardized
datasheets more difficult.
7
Working Draft: This is work in progress!
Market forces. A hardware builder would never buy a component that does not have a
datasheet since they would have no way of knowing how it would operate. In the
AI setting, practitioners are training custom models with specific datasets without
understanding the limitations of these datasets. The introduction of datasheets can
encourage the AI community to perform a thorough analysis of dataset characteristics
before applying them to specific settings.
Understanding out-of-spec behavior. Any of the testing done on components can be
described in the datasheet, and different tests are performed on different types of
components. However, there are a set of minimal tests that are usually included
in all datasheets. In the AI setting, datasets containing the same types of instances
(examples) meant for use in different settings will need different specifications. For
example, a dataset of faces used to train a model recognizing people from around the
world should contain a representative sample of people around the world.
5
Datasheets for Datasets
In the context of artificial intelligence and data science, datasets play a central role in both
training and evaluation. This is the case regardless of whether the dataset is used to build
a predictor that will be deployed as part of a system, or used to ask scientific questions
and reach scientific conclusions. In both cases, the specific properties of a dataset can
have profound impact on the quality of a learned predictor, or the quality of scientific
conclusions. Akin to how it is important to understand the operating characteristics of a
resistor when “debugging” a microcontroller, it is also important to understand the specific
properties of a dataset to understand how it fits into the larger data ecosystem.
Below we have proposed sample questions that a datasheet should arguably contain.
The prototypes in the appendix of this paper are provided as examples of how these might
be answered in practice. Several fundamental objectives drove our formation of these
questions. First, a practitioner should be able to decide, from reading this datasheet, how
appropriate this dataset is for their task, what its strengths and limitations are, and how
it fits into the broader dataset ecosystem. Second, the creators of a dataset should be able
to use the questions on a datasheet to help them think about aspects of data creation that
may not have otherwise occurred to them. Third, users should be able to understand—
based on the performance of a model or API on a dataset—what that performance measure
actually means, and when to be comfortable using such models.
The set of questions we provide here is not intended to be definitive. Instead, we
hope it will initiate a larger conversation about how data provenance, ethics, privacy,
and documentation might be handled by the community of data curators. Below are
our proposed questions, which include details about the gathering, cleaning, testing, and
releasing of a dataset. Not all questions will be applicable to all datasets, in which case
they can simply be left out.
Appendix B includes prototypes of datasheets for two datasets: Labeled Faces in The
Wild [33] and the Pang & Lee Polarity Dataset [45]. (In the creation of these datasheets,
sometimes information was unknown; this is marked in red text.)
8
Working Draft: This is work in progress!
Motivation for Dataset Creation
Who was involved in the data collection process?
(e.g., students, crowdworkers) and how were they comWhy was the dataset created? (e.g., was there a spe- pensated (e.g., how much were crowdworkers paid)?
cific task in mind? was there a specific gap that needed
to be filled?)
Over what time-frame was the data collected? Does
the collection time-frame match the creation timeframe of the instances?
What (other) tasks could the dataset be used for?
Has the dataset been used for any tasks already? If How was the data associated with each instance acso, where are the results so others can compare (e.g., quired? Was the data directly observable (e.g., raw
text, movie ratings), reported by subjects (e.g., survey
links to published papers)?
responses), or indirectly inferred/derived from other
data (e.g., part of speech tags; model-based guesses
Who funded the creation of the dataset?
for age or language)? If the latter two, were they validated/verified and if so how?
Any other comments?
Does the dataset contain all possible instances? Or is
it a sample (not necessarily random) of instances from
a larger set?
Dataset Composition
What are the instances? (that is, examples; e.g., documents, images, people, countries) Are there multiple If the dataset is a sample, then what is the populatypes of instances? (e.g., movies, users, ratings; peo- tion? What was the sampling strategy (e.g., deterministic, probabilistic with specific sampling probabilities)?
ple, interactions between them; nodes, edges)
Is the sample representative of the larger set (e.g., geAre relationships between instances made explicit in ographic coverage)? If not, why not (e.g., to cover a
the data (e.g., social network links, user/movie ratings, more diverse range of instances)? How does this affect
possible uses?
etc.)?
Is there information missing from the dataset and
How many instances are there? (of each type, if apwhy? (this does not include intentionally dropped inpropriate)?
stances; it might include, e.g., redacted text, withheld
documents) Is this data missing because it was unavailWhat data does each instance consist of? “Raw” able?
data (e.g., unprocessed text or images)?
Features/attributes? Is there a label/target associated Any other comments?
with instances? If the instances related to people, are
subpopulations identified (e.g., by age, gender, etc.)
and what is their distribution?
Data Preprocessing
What preprocessing/cleaning was done? (e.g., discretization or bucketing, tokenization, part-of-speech
tagging, SIFT feature extraction, removal of instances)
Is everything included or does the data rely on external resources? (e.g., websites, tweets, datasets) If
external resources, a) are there guarantees that they
will exist, and remain constant, over time; b) is there
an official archival version; c) are there access restrictions or fees?
Was the “raw” data saved in addition to the preprocessed/cleaned data? (e.g., to support unanticipated
future uses)
Are there recommended data splits and evaluation
Is the preprocessing software available?
measures? (e.g., training, development, testing; accuracy or AUC)
Does this dataset collection/processing procedure
achieve the motivation for creating the dataset stated
What experiments were initially run on this dataset? in the first section of this datasheet? If not, what are
Have a summary of those results.
the limitations?
Any other comments?
Any other comments?
Data Collection Process
Dataset Distribution
How was the data collected? (e.g., hardware appara- How will the dataset be distributed? (e.g., tarball on
tus/sensor, manual human curation, software program, website, API, GitHub; does the data have a DOI and
software interface/API)
is it archived redundantly?)
9
Working Draft: This is work in progress!
Legal & Ethical Considerations
When will the dataset be released/first distributed?
What license (if any) is it distributed under? Are there If the dataset relates to people (e.g., their attributes)
or was generated by people, were they informed about
any copyrights on the data?
the data collection? (e.g., datasets that collect writing, photos, interactions, transactions, etc.)
Are there any fees or access/export restrictions?
If it relates to people, were they told what the dataset
would be used for and did they consent? If so, how?
Were they provided with any mechanism to revoke their
consent in the future or for certain uses?
Any other comments?
Dataset Maintenance
Who is supporting/hosting/maintaining the dataset?
If it relates to people, could this dataset expose people to harm or legal action? (e.g., financial social or
otherwise) What was done to mitigate or reduce the
potential for harm?
Will the dataset be updated? If so, how often and by
whom?
If it relates to people, does it unfairly advantage or
disadvantage a particular social group? In what ways?
How will updates be communicated? (e.g., mailing How was this mitigated?
list, GitHub)
If it relates to people, were they provided with privacy
guarantees? If so, what guarantees and how are these
Is there an erratum?
ensured?
If the dataset becomes obsolete how will this be comDoes the dataset comply with the EU General Data
municated?
Protection Regulation (GDPR)? Does it comply with
any other standards, such as the US Equal Employment
Is there a repository to link to any/all papers/systems Opportunity Act?
that use this dataset?
If others want to extend/augment/build on this
dataset, is there a mechanism for them to do so? If
so, is there a process for tracking/assessing the quality
of those contributions. What is the process for communicating/distributing these contributions to users?
Any other comments?
6
Does the dataset contain information that might be
considered sensitive or confidential? (e.g., personally
identifying information)
Does the dataset contain information that might be
considered inappropriate or offensive?
Any other comments?
Challenges and Future Work
Our proposal for a set of standardized datasheets faces several challenges in implementation; we outline the most pressing of these below and urge the machine learning community
to make progress on these in future work. These challenges fall into several categories: how
to converge on the format and content of the datasheet, the incentives required to encourage
datasheet production and the need to overcome inertia, and the communication with outside experts that is necessary to properly address ethical and demographic considerations
of datasets containing data about people.
As a community, we will need to come to some consensus about what should be included
in a datasheet, and how that data can be most effectively solicited and communicated. Just
as in the context of hardware datasheets (Section 4), the most relevant information regarding each dataset will likely be context-specific; just as hardware has different categories
of components with differing relevant characteristics, datasets comprised of photographs
of human faces will have different relevant documentation needs than datasets of health
or weather records. We should not expect this consensus to come easily; researchers and
10
Working Draft: This is work in progress!
practitioners who work in an individual domain might first want to agree upon on a small
number of critical domain-specific attributes for their own datasets. However, there are
also universal questions relevant to all datasets (e.g., who paid for the collection of the data,
or whether there were human subjects generating the dataset), upon which the broader
community should perhaps come to an agreement. This will likely be part of a larger
conversation, both at a high level (datasheets for all types of datasets), as well as in a
domain-specific manner. Along the lines of the former, a group at the MIT Media Lab recently publicized a “Data Nutrition Label” idea2 , which, at the time of writing, has similar
goals to our proposal, though the details are not yet available.
It is also unclear to what extent a datasheet should delve into ethical questions such as
bias or privacy. Questions regarding ethical considerations should be framed in a manner
that encourages practitioners to use ethical procedures to gather data, without discouraging
them from providing as much information as possible about the process.
While this paper outlines questions that a datasheet for datasets would ideally answer,
a similar endeavor needs to be undertaken for datasheets pertaining to models that have
been pre-trained and their APIs. In particular, what are the important questions to ask
about the behavior of these models, and how should these be measured and communicated,
especially when the models are built based on multiple datasets, together with expert
knowledge and other sources of input? Institutions that produce such models should
iterate with customers and developers to arrive at the right set of questions and guidelines
in a “datasheet for models” that would parallel our proposal for datasets.
There will be overhead in creating datasheets, some of which we can mitigate by carefully designing an interactive survey that would automatically produce a datasheet based
on answers to questions. Moreover, hopefully a carefully crafted datasheet will, in the
long run, reduce the amount of time the dataset creators will need to spend answering
one-off questions about their data. Both large and small organizations will face hurdles
in producing these datasheets. For instance, extra details in a datasheet may result in an
organization being exposed to legal or PR risks, or such details might contain proprietary
information which give the organization a competitive edge. Organizations may also delay
releasing any datasheet—even an imperfect one—in order to “complete” it. Small organizations might consider the overhead in preparing a datasheet more onerous than large
organizations. On the other hand, datasheets also provide an opportunity for smaller organizations to differentiate themselves as more transparent than larger, more established
players. Ultimately, we believe the work involved in collecting and preparing a model or
dataset for public use far exceeds the cost of creating a datasheet, which can improve the
usefulness of this work for other users.
Finally, a large chunk of the work which remains in implementing datasheets for machine learning systems will be to communicate with experts in other areas. One example of
this need comes from considering demographic information for datasets related to people
(how the data is collected, collated, and analyzed). Other fields (such as anthropology)
are well-versed in the difficulties arising from demography, and we should avail ourselves of
that resource. It will also be important to consider that datasets are rarely gathered “from
the ground up” in such a way that it would be theoretically possible to gather all sorts
of additional information about the elements of the dataset. Instead, in many settings
datasets are scraped from some source without the ability to gather additional features,
demographic information, or consent. Some contextual information might still be available
(e.g., for the Enron email dataset, we might not have demographic information on a peremployee basis, but some demographic information about the employees of the company
2
See http://datanutrition.media.mit.edu/.
11
Working Draft: This is work in progress!
as a whole may be available). Again, other industries have considered these difficulties (as
discussed in Section 3) and we should learn from their best practices.
7
Acknowledgements
We would like to thank Meredith Whittaker, Erik Learned-Miller, Yoshua Bengio, Joy Buolamwini, Margaret Mitchell, Miroslav Dudík, and Rachel Thomas for valuable discussions
and feedback.
References
[1] Lewis v. amorous.
https://groups.google.com/forum/#!topic/alt.lawyers/
I-sU32WypvQ, 1907. [3 Ga.App. 50, 59 S.E. 338 (1907). Online accessed 18-March-2018].
[2] Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean,
Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. Tensorflow: A
system for large-scale machine learning. In OSDI, volume 16, pages 265–283, 2016.
[3] National Highway Traffic Safety Administration. Final regulatory evaluation, amendment
to federal motor vehicle safety standards 208. https://www.nhtsa.gov/sites/nhtsa.dot.
gov/files/dummy_milestones_812189.pdf, 2006. [Online; accessed 18-March-2018].
[4] National Highway Traffic Safety Administration. Milestones for nhtsa’s crash test dummies. https://www.gpo.gov/fdsys/pkg/FR-2006-08-31/pdf/06-7225.pdf, 2015. [Online;
accessed 18-March-2018].
[5] Don A Andrews, James Bonta, and J Stephen Wormith. The recent past and near future of
risk and/or need assessment. Crime & Delinquency, 52(1):7–27, 2006.
[6] Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai.
Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In
Advances in Neural Information Processing Systems, pages 4349–4357, 2016.
[7] Dipan Bose, Maria Segui-Gomez, and Jeff R. Crandall. Vulnerability of female drivers involved
in motor vehicle crashes: an analysis of us population at risk. American Journal of Public
Health, 2011.
[8] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in
commercial gender classification. In Conference on Fairness, Accountability and Transparency,
pages 77–91, 2018.
[9] Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. Semantics derived automatically
from language corpora contain human-like biases. Science, 356(6334):183–186, 2017.
[10] Bill Canis. Issues with federal motor vehicle safety standards. https://fas.org/sgp/crs/
misc/R44800.pdf, 2017. [Online; accessed 18-March-2018].
[11] Glennda Chui. Project will use ai to prevent or minimize electric grid failures. https:
//phys.org/news/2017-09-ai-minimize-electric-grid-failures.html, 2017. [Online;
accessed 14-March-2018].
[12] Google Cloud. Cloud automl. https://cloud.google.com/automl/, 2018. [Online; accessed
14-March-2018].
[13] International Electrotechnical Commission. About the iec: Overview. https://basecamp.
iec.ch/download/welcome-to-the-iec/, 2017.
12
Working Draft: This is work in progress!
[14] International Electrotechnical Commission. Iec 60195:2016. https://webstore.iec.ch/
publication/24478, 2018.
[15] International Electrotechnical Commission. Iec 60322:2001. https://webstore.iec.ch/
publication/768, 2018.
[16] International Electrotechnical Commission. Iec 60322:2001. https://webstore.iec.ch/
publication/1462, 2018.
[17] International Electrotechnical Commission. Welcome to the iec international electrotechnical
commission. http://www.iec.ch/about/history/overview/, 2018.
[18] International Electrotechnical Commission.
searchform&q=resistor, 2018.
Iec webstore.
https://webstore.iec.ch/
[19] KEMET Electronic Components. Surface mount multilayer ceramic chip capacitors (smd
mlccs. http://www.mouser.com/ds/2/212/KEM_C1035_C0G_PULSE_SMD-1103961.pdf, 2018.
[20] Coursera. Coursera. http://www.coursera.org/, 2017.
[21] William J Curran. The tuskegee syphilis study, 1973.
[22] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and
Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition.
In International conference on machine learning, pages 647–655, 2014.
[23] Harry F Dowling. The emergence of the cooperative clinical trial. Transactions & studies of
the College of Physicians of Philadelphia, 43(1):20–29, 1975.
[24] Mouser Electronics.
Kemet t520b107m006ate040.
https://www.mouser.com/
ProductDetail/KEMET/T520B107M006ATE040/?qs=A%2FUypIgi4aDLVIJ%2FABrGAw==, 2018.
[25] Ruth R Faden, Susan E Lederer, and Jonathan D Moreno. Us medical researchers, the nuremberg doctors trial, and the nuremberg code: A review of findings of the advisory committee
on human radiation experiments. JAMA, 276(20):1667–1671, 1996.
[26] Fast.ai. Fast.ai. http://www.fast.ai/, 2017.
[27] Food and Drug Administration. Content and format of a new drug application (21 cfr
314.50 (d)(5)(v)). https://www.fda.gov/downloads/ScienceResearch/SpecialTopics/
WomensHealthResearch/UCM557761.pdf, 1985.
[28] Food and Drug Administration. Guidance for the study of drugs likely to be used in the
elderly, 1989.
[29] Food and Drug Administration. Fda clinical trials guidance documents. https://www.fda.
gov/RegulatoryInformation/Guidances/ucm122046.htm, 2018.
[30] Clare Garvie, Alvaro Bedoya, and Jonathan Frankle. The Perpetual Line-Up: Unregulated
Police Face Recognition in America. Georgetown Law, Center on Privacy & Technology, 2016.
[31] Ralph Hingson, Jonathan Howland, and Suzette Levenson. Effects of legislative reform to
reduce drunken driving and alcohol-related traffic fatalities. Public Health Reports, 1988.
[32] HireVue. Hirevue. https://www.hirevue.com/, 2018.
[33] Gary B Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the
wild: A database for studying face recognition in unconstrained environments. Technical
report, Technical Report 07-49, University of Massachusetts, Amherst, 2007.
[34] Randall K Kirschman. High temperature electronics. IEEE Press, 1999.
13
Working Draft: This is work in progress!
[35] Tom CW Lin. The new investor. UCLA L. Rev., 60:678, 2012.
[36] Katherine A Liu and Natalie A Dipietro Mager. Women’s involvement in clinical trials:
historical perspective and future implications. Pharmacy Practice (Granada), 14(1):0–0, 2016.
[37] G Mann and C O’Neil. Hiring algorithms are not neutral. Harvard Business Review, 2016.
[38] Clay McShane. The Automobile. Routledge, 2018.
[39] Jonathan D Moreno. Undue risk: secret state experiments on humans. Routledge, 2013.
[40] Martha R Nolan and Thuy-Linh Nguyen. Analysis and reporting of sex differences in phase iii
medical device clinical trials—how are we doing? Journal of Women’s Health, 22(5):399–401,
2013.
[41] Mary Catherine O’Connor. How ai could smarten up our water system. https://medium.
com/s/ai-for-good/how-ai-could-smarten-up-our-water-system-f965b87f355a, 2017.
[Online; accessed 14-March-2018].
[42] National Institute of Health. Nih sharing policies and related guidance on nih-funded research
resources. https://grants.nih.gov/policy/sharing.htm, 2018.
[43] U.S. Department of Transportation Federal Highway Administration. Year of first state driver
license law and first driver examination. https://www.fhwa.dot.gov/ohim/summary95/
dl230.pdf, 1997. [Online; accessed 18-March-2018].
[44] World Health Organization. Global status report on road safety 2015. Fillin, 2015.
[45] Bo Pang and Lillian Lee. A sentimental education: Sentiment analysis using subjectivity
summarization based on minimum cuts. In Proceedings of the 42nd annual meeting on Association for Computational Linguistics, page 271. Association for Computational Linguistics,
2004.
[46] The European Parliament and the Council of the European Union. Directive 2002/95/ec of
the european parliament and of the council of 27 january 2003 on the restriction of the use
of certain hazardous substances in electrical and electronic equipment. http://eur-lex.
europa.eu/LexUriServ/LexUriServ.do?uri=OJ:L:2003:037:0019:0023:EN:PDF, 2003.
[47] THE EUROPEAN PARLIAMENT and THE COUNCIL OF THE EUROPEAN UNION.
General data protection regulation - european council. http://data.consilium.europa.
eu/doc/document/ST-5419-2016-INIT/en/pdf, 2016.
[48] Adam Paszke, Sam Gross, Soumith Chintala, and Gregory Chanan. Pytorch, 2017.
[49] Sam Peltzman. The effects of automobile safety regulation. Journal of Political Economy,
1975.
[50] Arcol Resistors. Hs aluminium housed resistors. http://www.arcolresistors.com/
wp-content/uploads/2014/03/HS-Datasheet.pdf, 2008.
[51] Freescale Semiconductor. Mac7100 microcontroller family hardware specifications. https:
//www.nxp.com/docs/en/data-sheet/MAC7100EC.pdf, 2006.
[52] Doha Suppy Systems. Facial recognition. http://doha.co.za/facialrecognition.html,
2017. [Online; accessed 14-March-2018].
[53] Udacity. Udacity. http://www.udacity.org/, 2017.
[54] Rolf H Weber and Romana Weber. Internet of things, volume 12. Springer, 2010.
[55] Frank A Wolff. The so-called international electrical units. Journal of the Institution of
Electrical Engineers, 34(170):190–207, 1905.
14
Working Draft: This is work in progress!
A
An Example of a Datasheet for a Hardware Component
‘
Miniature Aluminum Electrolytic Capacitors
XRL Series
FEATURES
• Low impedance characteristics
• Case sizes are smaller than conventional general-purpose capacitors,
with very high performance
• Can size larger than 9mm diameter has safety vents on rubber end seal
• RoHS Compliant
CHARACTERISTICS
Item
Operating Temperature Range
Capacitance Tolerance
Leakage Current
Dissipation Factor
(Tan δ, at 20°C 120Hz)
Characteristics
-40°C ~ +85°C
±20% at 120Hz, 20°C
≤100V
I = 0.01CWV or 3µA whichever is greater after 2 minutes of applied rated DC working voltage at 20°C
Where: C = rated capacitance in µF; WV = rated DC working voltage
>100V
CWV ≤ 1000 µF: I= 0.03 CWV + 15uA;
CWV ≥ 1000 µF: I= 0.02 CWV + 25uA;
Working voltage (WV)
Tan δ
C= rated capacitance in uF
WV= rated DC working voltage in V
6.3
0.23
10
0.20
16
0.16
25
0.14
35
0.12
50
0.10
63
0.09
100
0.08
160
0.12
250
0.17
350
0.20
450
0.25
6.3
8
6.3
6
8
10
18
10
13
10
4
6
8
16
16
20
16
3
4
6
12
25
32
25
3
4
6
10
35
44
35
2
3
4
8
50
63
50
2
3
3
8
63
79
63
2
3
3
6
100
125
100
2
3
3
6
160
200
160
3
3
4
4
250
300
250
8
8
10
10
350
400
350
12
12
16
16
450
500
450
16
16
20
20
For capacitors whose capacitance exceeds 1,000µF, the specification of tan δ is increased by 0.02 for every addition of 1,000µF
Working voltage (WV)
Surge voltage (SV)
Working voltage (WV)
Z(-25°C)/Z(+20°C) øD<16
øD≥16
Z(-40°C)/Z(+20°C) øD<16
øD≥16
Surge Voltage
Low Temperature
Characteristics
(Imp. ratio @ 120Hz)
When returned to +20°C after 2,000 hours application of working voltage at +85°C, the capacitor will meet the following limits: Capacitance change is ≤ ±20% of initial value; tan δ is < 200% of specified value; leakage current is
within specified value
Load Test
When returned to +20°C after 1,000 hours at +85°C with no voltage applied, the capacitor will meet the following limits: Capacitance change is ≤ ±20% of initial value; tan δ is < 200% of specified value; leakage current is within
specified value
Shelf Life Test
PART NUMBERING SYSTEM
1
4
0
Prefix
X
R
L
1
Series
6
V
1
Voltage
Actual Value
0
0
Capacitance (µF)
Actual Value
R
C
Suffix
RoHS Compliant
RIPPLE CURRENT AND FREQUENCY MULTIPLIERS
Capacitance (µF)
<100
100 ~ 1000
>1000
60 (50)
0.70
120
1.0
0.75
0.80
Frequency (Hz)
500
1.30
1.0
1.0
1.20
1.10
1K
1.40
≥10K
1.50
1.30
1.12
1.35
1.15
RIPPLE CURRENT AND TEMPERATURE MULTIPLIERS
Temperature (°C)
Multiplier
<50
1.78
70
1.4
85
1.0
XICON PASSIVE COMPONENTS • (800) 628-0544
XC-600178
Date Revised: 1/8/07
Specifications are subject to change without notice. No liability or warranty implied by this information. Environmental compliance based on producer documentation.
15
Working Draft: This is work in progress!
Miniature Aluminum Electrolytic Capacitors
XRL Series
DIMENSIONS AND PERMISSIBLE RIPPLE CURRENT
Vinyl sleeve
ød
Vent ≥ 6.3ø
P ±0.5
L ±1.5 max.
Value
(µF)
.10
.22
.33
.47
1.0
2.2
3.3
4.7
10
22
33
47
100
220
330
470
1000
2200
3300
4700
6800
10000
Value
(µF)
10
øD x L
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
6.3 x 11
8 x 11.5
8 x 11.5
10 x 12.5
13 x 20
13 x 20
16 x 25
16 x 25
18 x 36
.47
1.0
2.2
3.3
4.7
10
øD x L
5 x 11
5 x 11
6.3 x 11
8 x 11.5
8 x 11.5
10 x 16
47
100
220
13 x 25
16 x 25
18 x 36
22
33
330
13 x 20
13 x 20
18 x 40
mA
54
70
84
100
145
250
350
415
650
1240
1420
1980
2220
2880
160
20 min.
5
min.
Lead Spacing and Diameter (mm)
øD
5
6.3
8
10
P
2.0
ød
øD ±1
max.
0.5
2.5
0.5
3.5
0.6
5.0
0.6
13
5.0
0.6
16
7.5
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
6.3 x 11
8 x 11.5
8 x 11.5
10 x 12.5
10 x 16
13 x 20
16 x 25
16 x 32
16 x 32
18 x 36
mA
øD x L
mA
20
30
41
49
75
90
110
180
300
370
520
785
1295
1840
2260
2520
3080
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
6.3 x 11
8 x 11.5
10 x 12.5
10 x 16
13 x 20
16 x 25
16 x 32
18 x 36
18 x 36
22 x 40
6.0
10
16
25
31
54
80
97
115
190
320
470
620
1090
1660
2070
2520
2880
3440
0.8
øD x L
mA
5 x 11
5 x 11
5 x 11
6.3 x 11
6.3 x 11
8 x 11.5
10 x 12
10 x 16
10 x 16
13 x 20
16 x 32
18 x 36
18 x 36
22 x 41
40
58
87
115
145
240
420
570
740
1145
1890
2430
2700
2900
øD x L
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
6.3 x 11
6.3 x 11
8 x 11.5
10 x 16
13 x 20
13 x 20
16 x 25
18 x 40
22 x 40
25 x 40
mA
1.5
3.5
5.0
7.0
15
29
35
42
65
95
136
165
260
490
635
860
1530
2231
2785
3300
øD x L
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
6.3 x 11
8 x 11.5
8 x 11.5
10 x 12
10 x 20
13 x 20
16 x 25
16 x 25
18 x 40
22x 40
25 x 40
mA
3.0
4.5
7.5
9.5
17
28
34
45
70
115
150
190
320
565
765
1050
1700
2385
3000
3560
Working Voltage (WV); Dimensions: øD x L (mm); Ripple Current: mA/RMS @ 120Hz, 85°C
250
350
mA
13
20
34
50
60
115
216
270
354
582
900
1010
øD x L
8 x 11.5
8 x 11.5
8 x 11.5
10 x 12.5
10 x 16
10 x 20
13 x 20
13 x 25
16 x 25
18 x 40
22 x 40
22
0.8
1.0
25
10
12.5
1.0
Tape and box is 5.0mm lead space.
Working Voltage (WV); Dimensions: øD x L (mm); Ripple Current: mA/RMS @ 120Hz, 85°C
16
25
35
50
63
øD x L
18
7.5
mA
21
32
49
70
93
150
255
348
468
øD x L
8 x 11.5
8 x 11.5
10 x 16
10 x 16
10 x 20
13 x 20
13 x 25
16 x 32
16 x 36
822
1134
18 x 40
mA
21
32
63
78
103
174
282
438
500
685
øD x L
10 x 12.5
10 x 12.5
10 x 16
10 x 20
13 x 20
13 x 25
16 x 25
18 x 36
18 x 40
22 x 45
100
øD x L
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
5 x 11
6.3 x 11
8 x 11.5
10 x 16
10 x 16
10 x 20
13 x 25
16 x 32
18 x 36
22 x 40
450
mA
3.0
5.8
8.8
12
22
33
40
48
80
135
195
255
370
675
972
1135
2600
mA
26
38
63
86
120
192
354
426
555
750
XICON PASSIVE COMPONENTS • (800) 628-0544
XC-600178
Date Revised: 1/8/07
Specifications are subject to change without notice. No liability or warranty implied by this information. Environmental compliance based on producer documentation.
16
Working Draft: This is work in progress!
Miniature Aluminum Electrolytic Capacitors
XRL Series
1000µF 16V
1µF 50V
TYPICAL PERFORMANCE CHARACTERISTICS
Life Test
Temperature Characteristics
Capacitance Change vs. Temperature
Capacitance Change (%)
Capacitance Change (%)
Capacitance Change vs. Time (at +85°C)
+20
+10
0
-10
+10
10
-10
-20
-20
-40
500
750
Time (Hours)
1000
Dissipation Factor (tan δ) 120 Hz
Dissipation Factor vs. Time (at +85°C)
0.4
0.3
0.2
-20
0
+20 +40 +60
Temperature (°C)
+80 +100
Dissipation Factor vs. Temperature
Dissipation Factor (tan δ) 120 Hz
250
0.1
1.0
1.0
0.01
250
500
750
Time (Hours)
1000
-40
-20
0
+20 +40 +60
Temperature (°C)
+80 +100
Impedance vs. Frequency
10
Leakage Current vs. Time (at +85°C)
2
20
Impedance (Ω)
Leakage Current
25
15
10
10
-40°C
1
+20°C
-40°C
5
250
500
750
Time (Hours)
1000
0.1
+20°C
102
103
104
Frequency (Hz)
105
106
XICON PASSIVE COMPONENTS • (800) 628-0544
XC-600178
Date Revised: 1/8/07
Specifications are subject to change without notice. No liability or warranty implied by this information. Environmental compliance based on producer documentation.
17
Working Draft: This is work in progress!
18
Working Draft: This is work in progress!
B
Prototypes of Datasheets for Datasets
A Database for Studying Face Recognition in Unconstrained Environments
Motivation for Dataset Creation
Why was the dataset created? (e.g., was there a specific task in mind?
was there a specific gap that needed to be filled?)
Labeled Faces in the Wild was created to provide images that
can be used to study face recognition in the unconstrained setting
where image characteristics (such as pose, illumination, resolution, focus), subject demographic makeup (such as age, gender,
race) or appearance (such as hairstyle, makeup, clothing) cannot
be controlled. The dataset was created for the specific task of pair
matching: given a pair of images each containing a face, determine whether or not the images are of the same person.1
What (other) tasks could the dataset be used for?
The LFW dataset can be used for the face identification problem.
Some researchers have developed protocols to use the images in
the LFW dataset for face identification.2
Has the dataset been used for any tasks already? If so, where are the
results so others can compare (e.g., links to published papers)?
Papers using this dataset and the specified evaluation protocol are
listed in http://vis-www.cs.umass.edu/lfw/results.html
Who funded the creation of the dataset?
The building of the LFW database was supported by a United
States National Science Foundation CAREER Award.
Dataset Composition
What are the instances? (that is, examples; e.g., documents, images,
people, countries) Are there multiple types of instances? (e.g., movies,
users, ratings; people, interactions between them; nodes, edges)
Each instance is a pair of images labeled with the name of the
person in the image. Some images contain more than one face.
The labeled face is the one containing the central pixel of the
image—other faces should be ignored as “background”.
Are relationships between instances made explicit in the data (e.g.,
social network links, user/movie ratings, etc.)?
There are no known relationships between instances except for
the fact that they are all individuals who appeared in news sources
on line, and some individuals appear in multiple pairs.
What data does each instance consist of? “Raw” data (e.g., unprocessed text or images)? Features/attributes? Is there a label/target associated with instances? If the instances related to people, are subpopulations identified (e.g., by age, gender, etc.) and what is their distribution?
Each instance contains a pair of images that are 250 by 250 pixels in JPEG 2.0 format. Each image is accompanied by a label
indicating the name of the person in the image. While subpopulation data was not available at the initial release of the dataset, a
subsequent paper3 reports the distribution of images by age, race
and gender. Table 2 lists these results.
Is everything included or does the data rely on external resources?
(e.g., websites, tweets, datasets) If external resources, a) are there guarantees that they will exist, and remain constant, over time; b) is there an
official archival version; c) are there access restrictions or fees?
Everything is included in the dataset.
Are there recommended data splits and evaluation measures? (e.g.,
training, development, testing; accuracy or AUC)
The dataset comes with specified train/test splits such that none
of the people in the training split are in the test split and vice
versa. The data is split into two views, View 1 and View 2. View
1 consists of a training subset (pairsDevTrain.txt) with 1100 pairs
of matched and 1100 pairs of mismatched images, and a test subset (pairsDevTest.txt) with 500 pairs of matched and mismatched
images. Practitioners can train an algorithm on the training set
and test on the test set, repeating as often as necessary. Final
performance results should be reported on View 2 which consists
of 10 subsets of the dataset. View 2 should only be used to test
the performance of the final model. We recommend reporting
performance on View 2 by using leave-one-out cross validation,
performing 10 experiments. That is, in each experiment, 9 subsets should be used as a training set and the 10th subset should be
used for testing. At a minimum, we recommend reporting the estimated mean accuracy, µ̂ and the standard error of the mean:
SE for View 2.
µ̂ is given by:
P10
pi
(1)
µ̂ = i=1
10
where pi is the percentage of correct classifications on View 2
using subset i for testing. SE is given as:
How many instances are there? (of each type, if appropriate)?
The dataset consists of 13,233 face images in total of 5749 unique
individuals. 1680 of these subjects have two or more images and
4069 have single ones.
1 All information in this datasheet is taken from one of five sources.
Any errors
that were introduced from these sources are our fault.
Original
paper:
http://www.cs.cornell.edu/people/pabo/
movie-review-data/;
LFW
survey:
http://vis-www.cs.umass.
edu/lfw/lfw.pdf;
Paper measuring LFW demographic characteristics
:
http://biometrics.cse.msu.edu/Publications/Face/HanJain
UnconstrainedAgeGenderRaceEstimation MSUTechReport2014.pdf;
LFW website: http://vis-www.cs.umass.edu/lfw/.
2 Unconstrained face recognition:
Identifying a person of interest
from a media collection:
http://biometrics.cse.msu.edu/Publications/
Face/BestRowdenetal UnconstrainedFaceRecognition TechReport
MSU-CSE-14-1.pdf
Labeled Faces in the Wild
σ̂
SE = √
10
(2)
Where σ̂ is the estimate of the standard deviation, given by:
s
P10
2
i=1 (pi − µ̂)
σ̂ =
(3)
9
The multiple-view approach is used instead of a traditional
train/validation/test split in order to maximize the amount of data
available for training and testing.
3 http://biometrics.cse.msu.edu/Publications/Face/HanJain
UnconstrainedAgeGenderRaceEstimation MSUTechReport2014.pdf
19
Working Draft: This is work in progress!
A Database for Studying Face Recognition in Unconstrained Environments
Training Paradigms: There are two training paradigms that
can be used with our dataset. Practitioners should specify the
training paradigm they used while reporting results.
•
Image-Restricted Training This setting prevents the experimenter from using the name associated with each image
during training and testing. That is, the only available information is whether or not a pair of images consist of the same
person, not who that person is. This means that there would
be no simple way of knowing if there are multiple pairs of
images in the train/test set that belong to the same person.
Such inferences, however, might be made by comparing image similarity/equivalence (rather than comparing names).
Thus, to form training pairs of matched and mismatched images for the same person, one can use image equivalence to
add images that consist of the same person.
Property
Unrestricted Training In this setting, one can use the names
associated with images to form pairs of matched and mismatched images for the same person. The file people.txt in
View 2 of the dataset contains subsets of of people along
with images for each subset. To use this paradigm, matched
and mismatched pairs of images should be formed from images in the same subset. In View 1, the files peopleDevTrain.txt and peopleDevTest.txt can be used to create arbitrary pairs of matched/mismatched images for each person. The unrestricted paradigm should only be used to create training data and not for performance reporting. The test
data, which is detailed in the file pairs.txt, should be used
to report performance. We recommend that experimenters
first use the image-restricted paradigm and move to the unrestricted paradigm if they believe that their algorithm’s performance would significantly improve with more training
data. While reporting performance, it should be made clear
which of these two training paradigms were used for particular test result.
What experiments were initially run on this dataset? Have a summary
of those results.
The dataset was originally released without reported experimental
results but many experiments have been run on it since then.
Value
Database Release Year
Number of Unique Subjects
Number of total images
Number of individuals with 2 or more images
Number of individuals with single images
Image Size
Image format
Average number of images per person
2007
5649
13,233
1680
4069
250 by 250 pixels
JPEG
2.30
Table 1. A summary of dataset statistics extracted from the original paper: Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik LearnedMiller. Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments. University of Massachusetts,
Amherst, Technical Report 07-49, October, 2007.
Demographic Characteristic
Percentage of female subjects
Percentage of male subjects
Percentage of White subjects
Percentage of Black subjects
Percentage of Asian subjects
Percentage of people between 0-20 years old
Percentage of people between 21-40 years old
Percentage of people between 41-60 years old
Percentage of people over 61 years old
The files pairsDevTrain.txt and pairsDevTest.txt support
image-restricted uses of train/test data. The file pairs.txt in
View 2 supports the image-restricted use of training data.
•
Labeled Faces in the Wild
Table 2. Demographic characteristics of the LFW dataset as measured by
Han, Hu, and Anil K. Jain. Age, gender and race estimation from unconstrained face images. Dept. Comput. Sci. Eng., Michigan State Univ.,
East Lansing, MI, USA, MSU Tech. Rep.(MSU-CSE-14-5) (2014).
Data Collection Process
How was the data collected? (e.g., hardware apparatus/sensor, manual
human curation, software program, software interface/API)
The raw images for this dataset were obtained from the Faces in
the Wild database collected by Tamara Berg at Berkeley4 . The
images in this database were gathered from news articles on the
web using software to crawl news articles.
Who was involved in the data collection process? (e.g., students,
crowdworkers) and how were they compensated (e.g., how much were
crowdworkers paid)?
Unknown
Over what time-frame was the data collected? Does the collection timeframe match the creation time-frame of the instances?
Unknown
Any other comments?
Table 1 summarizes some dataset statistics and Figure 1 shows
examples of images. Most images in the dataset are color, a few
are black and white.
4 Faces
20
Value
22.5%
77.5%
83.5%
8.47%
8.03%
1.57%
31.63%
45.58%
21.2%
in the Wild: http://tamaraberg.com/faceDataset/
Working Draft: This is work in progress!
A Database for Studying Face Recognition in Unconstrained Environments
How was the data associated with each instance acquired? Was the
data directly observable (e.g., raw text, movie ratings), reported by subjects (e.g., survey responses), or indirectly inferred/derived from other data
(e.g., part of speech tags; model-based guesses for age or language)? If
the latter two, were they validated/verified and if so how?
The names for each person in the dataset were determined by an
operator by looking at the caption associated with the person’s
photograph. Some people could have given incorrect names particularly if the original caption was incorrect.
Does the dataset contain all possible instances? Or is it a sample (not
necessarily random) of instances from a larger set?
The dataset does not contain all possible instances.
If the dataset is a sample, then what is the population? What was the
sampling strategy (e.g., deterministic, probabilistic with specific sampling
probabilities)? Is the sample representative of the larger set (e.g., geographic coverage)? If not, why not (e.g., to cover a more diverse range of
instances)? How does this affect possible uses?
The original Faces in the Wild dataset is a sample of pictures of
people appearing in the news on the web. Labeled Faces in the
Wild is thus also a sample of images of people found on the news
on line. While the intention of the dataset is to have a wide range
of demographic (e.g. age, race, ethnicity) and image (e.g. pose,
illumination, lighting) characteristics, there are many groups that
have few instances (e.g. only 1.57% of the dataset consists of
individuals under 20 years old).
Is there information missing from the dataset and why? (this does
not include intentionally dropped instances; it might include, e.g., redacted
text, withheld documents) Is this data missing because it was unavailable?
Unknown
Data Preprocessing
What preprocessing/cleaning was done? (e.g., discretization or bucketing, tokenization, part-of-speech tagging, SIFT feature extraction, removal
of instances)
The following steps were taken to process the data:
1. Gathering raw images: First the raw images for this
dataset were obtained from the Faces in the Wild dataset
consisting of images and associated captions gathered from
news articles found on the web.
2. Running the Viola-Jones face detector5 The OpenCV version 1.0.0 release 1 implementation of Viola-Jones face detector was used to detect faces in each of these images, using
the function cvHaarDetectObjects, with the provided Haar
classifier—cascadehaarcascadefrontalfacedefault.xml. The
scale factor was set to 1.2, min neighbors was set to 2, and
the flag was set to CV HAAR DO CANNY PRUNING.
3. Manually eliminating false positives: If a face was detected and the specified region was determined not to be a
face (by the operator), or the name of the person with the
5 Paul
Viola and Michael Jones. Robust real-time face detection. IJCV, 2004
21
Labeled Faces in the Wild
detected face could not be identified (using step 5 below),
the face was omitted from the dataset.
4. Eliminating duplicate images: If images were determined
to have a common original source photograph, they are defined to be duplicates of each other. An attempt was made to
remove all duplicates but a very small number (that were not
initially found) might still exist in the dataset. The number
of remaining duplicates should be small enough so as not
to significantly impact training/testing. The dataset contains
distinct images that are not defined to be duplicates but are
extremely similar. For example, there are pictures of celebrities that appear to be taken almost at the same time by different photographers from slightly different angles. These
images were not removed.
5. Labeling (naming) the detected people: The name associated with each person was extracted from the associated
news caption. This can be a source of error if the original news caption was incorrect. Photos of the same person were combined into a single group associated with one
name. This was a challenging process as photos of some
people were associated with multiple names in the news captions (e.g.“Bob McNamara” and “Robert McNamara”). In
this scenario, an attempt was made to use the most common
name. Some people have a single name (e.g. “Madonna” or
“Abdullah”). For Chinese and some other Asian names, the
common Chinese ordering (family name followed by given
name) was used (e.g. “Hu Jintao”).
6. Cropping and rescaling the detected faces: Each detected
region denoting a face was first expanded by 2.2 in each dimension. If the expanded region falls outside of the image,
a new image was created by padding the original pixels with
black pixels to fill the area outside of the original image.
This expanded region was then resized to 250 pixels by 250
pixels using the function cvResize, and cvSetImageROI as
necessary. Images were saved in JPEG 2.0 format.
7. Forming pairs of training and testing pairs for View 1
and View 2 of the dataset: Each person in the dataset was
randomly assigned to a set (with 0.7 probability of being in
a training set in View 1 and uniform probability of being in
any set in View 2). Matched pairs were formed by picking
a person uniformly at random from the set of people who
had two or more images in the dataset. Then, two images
were drawn uniformly at random from the set of images of
each chosen person, repeating the process if the images are
identical or if they were already chosen as a matched pair).
Mismatched pairs were formed by first choosing two people uniformly at random, repeating the sampling process if
the same person was chosen twice. For each chosen person,
one image was picked uniformly at random from their set of
images. The process is repeated if both images are already
contained in a mismatched pair.
Working Draft: This is work in progress!
A Database for Studying Face Recognition in Unconstrained Environments
Labeled Faces in the Wild
Was the “raw” data saved in addition to the preprocessed/cleaned
data? (e.g., to support unanticipated future uses)
What license (if any) is it distributed under? Are there any copyrights
on the data?
The raw unprocessed data (consisting of images of faces and
names of the corresponding people in the images) is saved.
The crawled data copyright belongs to the news papers that the
data originally appeared in. There is no license, but there is
a request to cite the corresponding paper if the dataset is used:
Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik LearnedMiller. Labeled Faces in the Wild: A Database for Studying Face
Recognition in Unconstrained Environments. University of Massachusetts, Amherst, Technical Report 07-49, October, 2007.
Is the preprocessing software available?
While a script running a sequence of commands is not available,
all software used to process the data is open source and has been
specified above.
Does this dataset collection/processing procedure achieve the motivation for creating the dataset stated in the first section of this
datasheet? If not, what are the limitations?
There some potential limitations in the dataset which might bias
the data towards a particular demographic, pose, image characteristics etc.
•
•
•
•
The Viola-Jones detector can have systematic errors by race,
gender, age or other categories
Due to the Viola-Jones detector, there are only a small number of side views of faces, and only a few views from either
above or below
The dataset does not contain many images that occur under
extreme (or very low) lighting conditions
The original images were collected from news paper articles.
These articles could cover subjects in limited geographical
locations, specific genders, age, race, etc. The dataset does
not provide information on the types of garments worn by
the individuals, whether they have glasses on, etc.
There are no fees or restrictions.
Dataset Maintenance
Who is supporting/hosting/maintaining the dataset?
The dataset is hosted at the University of Massachusetts
and all and comments can be sent to: Gary Huang - [email protected].
Will the dataset be updated? If so, how often and by whom?
Unknown
How will updates be communicated? (e.g., mailing list, GitHub)
All changes to the dataset will be announced through the LFW
mailing list. Those who would like to sign up should send an
email to [email protected].
Is there an erratum?
Errata are listed under the “Errata” section of http://vis-www.cs.
umass.edu/lfw/index.html
If the dataset becomes obsolete how will this be communicated?
•
The majority of the dataset consists of White males
•
There are very few images of people who under 20 years old
•
Are there any fees or access/export restrictions?
The proposed train/test protocol allows reuse of data between View 1 and View 2 in the dataset. This could potentially introduce very small biases into the results
All changes to the dataset will be announced through the LFW
mailing list.
Is there a repository to link to any/all papers/systems that use this
dataset?
Papers using this dataset and the specified training/evaluation
protocols are listed under “Methods” section of http://vis-www.cs.
umass.edu/lfw/results.html
Dataset Distribution
How will the dataset be distributed? (e.g., tarball on website, API,
GitHub; does the data have a DOI and is it archived redundantly?)
The dataset can be downloaded from http://vis-www.cs.umass.edu/
lfw/index.html#download. The images can be downloaded as a
gzipped tar file.
When will the dataset be released/first distributed?
The dataset was released in October, 2007.
22
If others want to extend/augment/build on this dataset, is there a
mechanism for them to do so? If so, is there a process for tracking/assessing the quality of those contributions. What is the process for
communicating/distributing these contributions to users?
Unknown
Working Draft: This is work in progress!
A Database for Studying Face Recognition in Unconstrained Environments
Labeled Faces in the Wild
Figure 1. Examples of images from our dataset (matched pairs)
Legal & Ethical Considerations
If the dataset relates to people (e.g., their attributes) or was generated by people, were they informed about the data collection? (e.g.,
datasets that collect writing, photos, interactions, transactions, etc.)
No. The data was crawled from public web sources, and the individuals appeared in news stories. But there was no explicit informing of these individuals that their images were being assembled into a dataset.
If it relates to people, were they told what the dataset would be used
for and did they consent? If so, how? Were they provided with any
mechanism to revoke their consent in the future or for certain uses?
No (see first question).
If it relates to people, could this dataset expose people to harm or legal action? (e.g., financial social or otherwise) What was done to mitigate
or reduce the potential for harm?
There is minimal risk for harm: the data was already public.
If it relates to people, does it unfairly advantage or disadvantage a
particular social group? In what ways? How was this mitigated?
Unknown
If it relates to people, were they provided with privacy guarantees? If
so, what guarantees and how are these ensured?
No. All subjects in the dataset appeared in news sources so the
images that we used along with the captions are already public.
Does the dataset comply with the EU General Data Protection Regulation (GDPR)? Does it comply with any other standards, such as the US
Equal Employment Opportunity Act?
The dataset does not comply with GDPR because subjects were
not asked for their consent.
Does the dataset contain information that might be considered sensitive or confidential? (e.g., personally identifying information)
The dataset does not contain confidential information since all
information was scraped from news stories.
Does the dataset contain information that might be considered inappropriate or offensive?
No. The dataset only consists of faces and associated names.
23
Working Draft: This is work in progress!
Movie Review Polarity
Thumbs Up? Sentiment Classification using Machine Learning Techniques
Motivation for Dataset Creation
Why was the dataset created? (e.g., was there a specific task in mind?
was there a specific gap that needed to be filled?)
The dataset was created to enable research on predicting sentiment polarity: given a piece of (English) text, predict whether it
has a positive or negative affect or stance toward its topic. It was
created intentionally with that task in mind, focusing on movie reviews as a place where affect/sentiment is frequently expressed.1
these are words that could be used to describe the emotions of john sayles’
characters in his latest , limbo . but no , i use them to describe myself after
sitting through his latest little exercise in indie egomania . i can forgive many
things . but using some hackneyed , whacked-out , screwed-up * non * ending on a movie is unforgivable . i walked a half-mile in the rain and sat
through two hours of typical , plodding sayles melodrama to get cheated by a
complete and total copout finale . does sayles think he’s roger corman ?
Figure 1. An example “negative polarity” instance, taken from the file
neg/cv452 tok-18656.txt.
What (other) tasks could the dataset be used for?
The dataset could be used for anything related to modeling or
understanding movie reviews. For instance, one may induce a
lexicon of words/phrases that are highly indicative of sentiment
polarity, or learn to automatically generate movie reviews.
Has the dataset been used for any tasks already? If so, where are the
results so others can compare (e.g., links to published papers)?
At the time of publication, only the original paper http://xxx.lanl.
gov/pdf/cs/0409058v1. Between then and 2012, a collection of papers that used this dataset was maintained at http://www.cs.cornell.
edu/people/pabo/movie%2Dreview%2Ddata/otherexperiments.html.
Who funded the creation of the dataset?
Funding was provided though five distinct sources: the National
Science Foundation, the Department of the Interior, the National
Business Center, Cornell University, and the Sloan Foundation.
Dataset Composition
What are the instances? (that is, examples; e.g., documents, images,
people, countries) Are there multiple types of instances? (e.g., movies,
users, ratings; people, interactions between them; nodes, edges)
The instances are movie reviews extracted from newsgroup postings, together with a sentiment rating for whether the text corresponds to a review with a rating that is either strongly positive
(high number of stars) or strongly negative (low number of stars).
The polarity rating is binary {positive,negative}. An example instance is shown in Figure 1.
Are relationships between instances made explicit in the data (e.g.,
social network links, user/movie ratings, etc.)?
None explicitly, though the original newsgroup postings include
poster name and email address, so some information could be
extracted if needed.
How many instances are there? (of each type, if appropriate)?
There are 1400 instances in total in the original (v1.x versions)
and 2000 instances in total in v2.0 (from 2014).
What data does each instance consist of? “Raw” data (e.g., unprocessed text or images)? Features/attributes? Is there a label/target asso1 Information in this datasheet is taken from one of five sources; any errors
that were introduced are our fault. http://www.cs.cornell.edu/people/pabo/
movie-review-data/; http://xxx.lanl.gov/pdf/cs/0409058v1; http://www.cs.
cornell.edu/people/pabo/movie-review-data/rt-polaritydata.README.1.
0.txt; http://www.cs.cornell.edu/people/pabo/movie-review-data/poldata.
README.2.0.txt.
24
Features
unigrams (freq.)
unigrams
unigrams+bigrams
bigrams
unigrams+POS
adjectives
top 2633 unigrams
unigrams+position
corrected
NB
79.0
81.5
80.5
77.3
81.5
76.8
80.2
80.8
——–in paper——–
NB
ME
SVM
78.7
81.0
80.6
77.3
81.5
77.0
80.3
81.0
n/a
80.4
80.8
77.4
80.4
77.7
81.0
80.1
72.8
82.9
82.7
77.1
81.9
75.1
81.4
81.6
Table 1. Results on the original dataset (first column is after data repair
specified in the erratum, later).
ciated with instances? If the instances related to people, are subpopulations identified (e.g., by age, gender, etc.) and what is their distribution?
Each instance consists of the text associated with the review, with
obvious ratings information removed from that text (some errors
were found and alter fixed). The text was down-cased and HTML
tags were removed. Boilerplate newsgroup header/footer text was
removed. Some additional unspecified automatic filtering was
done. Each instance also has an associated target value: a positive (+1) or negative (-1) rating based on the number of stars that
that review gave (details on the mapping from number of stars to
polarity is given below in “Data Preprocessing”).
Is everything included or does the data rely on external resources?
(e.g., websites, tweets, datasets) If external resources, a) are there guarantees that they will exist, and remain constant, over time; b) is there an
official archival version; c) are there access restrictions or fees?
Everything is included.
Are there recommended data splits and evaluation measures? (e.g.,
training, development, testing; accuracy or AUC)
The instances come with a “cross-validation tag” to enable replication of cross-validation experiments; results are measured in
classification accuracy.
What experiments were initially run on this dataset? Have a summary
of those results.
Several experiments are reported in the README for baselines
on this data, both on the original dataset (Table 1) and the
cleaned version (Table 2). In these results, NB=Naive Bayes,
ME=Maximum Entropy and SVM=Support Vector Machine. The
feature sets include unigrams (with and without counts), bigrams,
part of speech features, and adjectives-only.
Working Draft: This is work in progress!
Movie Review Polarity
Features
unigrams (freq.)
unigrams
unigrams+bigrams
bigrams
unigrams+POS
adjectives
top 2631 unigrams
unigrams+position
Thumbs Up? Sentiment Classification using Machine Learning Techniques
# features
NB
ME
SVM
16162
16162
32324
16162
16688
2631
2631
22407
79.0
81.0
80.7
77.3
81.3
76.6
80.9
80.8
n/a
80.2
80.7
77.5
80.3
77.6
81.3
79.8
73.0
82.9
82.8
76.5
82.0
75.3
81.2
81.8
Data Preprocessing
What preprocessing/cleaning was done? (e.g., discretization or bucketing, tokenization, part-of-speech tagging, SIFT feature extraction, removal
of instances)
Table 2. Results on the cleaned dataset (first column is the number of
unique features).
Data Collection Process
How was the data collected? (e.g., hardware apparatus/sensor, manual
human curation, software program, software interface/API)
The data was collected by downloading reviews from the IMDb
archive of the rec.arts.movies.reviews newsgroup, at
http://reviews.imdb.com/Reviews.
Who was involved in the data collection process? (e.g., students,
crowdworkers) and how were they compensated (e.g., how much were
crowdworkers paid)?
Unknown
Over what time-frame was the data collected? Does the collection timeframe match the creation time-frame of the instances?
Unknown
How was the data associated with each instance acquired? Was the
data directly observable (e.g., raw text, movie ratings), reported by subjects (e.g., survey responses), or indirectly inferred/derived from other data
(e.g., part of speech tags; model-based guesses for age or language)? If
the latter two, were they validated/verified and if so how?
The data was mostly observable as raw text, except the labels
were extracted by the process described below.
Does the dataset contain all possible instances? Or is it a sample (not
necessarily random) of instances from a larger set?
The dataset is a sample of instances.
If the dataset is a sample, then what is the population? What was the
sampling strategy (e.g., deterministic, probabilistic with specific sampling
probabilities)? Is the sample representative of the larger set (e.g., geographic coverage)? If not, why not (e.g., to cover a more diverse range of
instances)? How does this affect possible uses?
The sample of instances collected is English movie reviews from
the rec.arts.movies.reviews newsgroup, from which a
“number of stars” rating could be extracted. The sample is limited
to forty reviews per unique author in order to achieve broader
coverage by authorship. Beyond that, the sample is arbitrary.
Is there information missing from the dataset and why? (this does
not include intentionally dropped instances; it might include, e.g., redacted
text, withheld documents) Is this data missing because it was unavailable?
No data is missing.
25
Instances for which an explicit rating could not be found
were discarded. Also only instances with strongly-positive
or strongly-negative ratings were retained. Star ratings were
extracted by automatically looking for text like “**** out of
*****” in the review, using that as a label, and then removing
the corresponding text. When the star rating was out of five stars,
anything at least four was considered positive and anything at
most two negative; when out of four, three and up is considered
positive, and one or less is considered negative. Occasionally half
stars are missed which affects the labeling of negative examples.
Everything in the middle was discarded. In order to ensure that
sufficiently many authors are represented, at most 20 reviews
(per positive/negative label) per author are included.
In a later version of the dataset (v1.1), non-English reviews were
also removed.
Some preprocessing errors were caught in later versions. The following fixes were made: (1) Some reviews had rating information
in several places that was missed by the initial filters; these are
removed. (2) Some reviews had unexpected/unparsed ranges and
these were fixed. (3) Sometimes the boilerplate removal removed
too much of the text.
Was the “raw” data saved in addition to the preprocessed/cleaned
data? (e.g., to support unanticipated future uses)
Yes.
Is the preprocessing software available?
No.
Does this dataset collection/processing procedure achieve the motivation for creating the dataset stated in the first section of this
datasheet? If not, what are the limitations?
The overarching goal of this dataset is to study the task of sentiment analysis. From this perspective, the current dataset represents a highly biased sample of all texts that express affect. In
particular: the genre is movie reviews (as opposed to other affective texts), the reviews are all in English, they are all from
the IMDb archive of the rec.arts.movies.reviews newsgroup, and all from a limited time frame. As mentioned above,
at most forty reviews were retained per author to ensure better
coverage of authors. Due to all these sampling biases, it is unclear whether models trained on this dataset should be expected
to generalize to other review domains (e.g., books, hotels, etc.)
or to domains where affect may be present but where affect is not
the main point of the text (e.g., personal emails).
Working Draft: This is work in progress!
Movie Review Polarity
Thumbs Up? Sentiment Classification using Machine Learning Techniques
Dataset Distribution
How will the dataset be distributed? (e.g., tarball on website, API,
GitHub; does the data have a DOI and is it archived redundantly?)
The dataset is distributed on Bo Pang’s webpage at Cornell: http:
If others want to extend/augment/build on this dataset, is there a
mechanism for them to do so? If so, is there a process for tracking/assessing the quality of those contributions. What is the process for
communicating/distributing these contributions to users?
//www.cs.cornell.edu/people/pabo/movie-review-data.
Others may do so and should contact the original authors about
incorporating fixes/extensions.
When will the dataset be released/first distributed?
Legal & Ethical Considerations
The dataset does
not have a DOI and there is no redundant archive.
The dataset was first released in 2002.
What license (if any) is it distributed under? Are there any copyrights
on the data?
If the dataset relates to people (e.g., their attributes) or was generated by people, were they informed about the data collection? (e.g.,
datasets that collect writing, photos, interactions, transactions, etc.)
The crawled data copyright belongs to the authors of the reviews
unless otherwise stated. There is no license, but there is a request
to cite the corresponding paper if the dataset is used: Thumbs up?
Sentiment classification using machine learning techniques. Bo
Pang, Lillian Lee, and Shivakumar Vaithyanathan. Proceedings
of EMNLP, 2002.
No. The data was crawled from public web sources, and the authors of the posts presumably knew that their posts would be public, but there was no explicit informing of these authors that their
posts were to be used in this way.
Are there any fees or access/export restrictions?
No (see first question).
No.
If it relates to people, were they told what the dataset would be used
for and did they consent? If so, how? Were they provided with any
mechanism to revoke their consent in the future or for certain uses?
If it relates to people, could this dataset expose people to harm or legal action? (e.g., financial social or otherwise) What was done to mitigate
or reduce the potential for harm?
Dataset Maintenance
There is minimal risk for harm: the data was already public, and
in the preprocessed version, names and email addresses were removed.
Who is supporting/hosting/maintaining the dataset?
Bo Pang is supporting/maintaining the dataset.
Will the dataset be updated? If so, how often and by whom?
Since its initial release (v0.9) there have been three later releases
(v1.0, v1.1 and v2.0).
How will updates be communicated? (e.g., mailing list, GitHub)
If it relates to people, does it unfairly advantage or disadvantage a
particular social group? In what ways? How was this mitigated?
Unknown
If it relates to people, were they provided with privacy guarantees? If
so, what guarantees and how are these ensured?
Updates are listed on the dataset web page.
Is there an erratum?
There is not an explicit erratum, but updates and known errors are specified in higher version README and diff files.
There are several versions of these: v1.0: http://www.cs.cornell.
edu/people/pabo/movie-review-data/README;
v1.1:
http://www.
cs.cornell.edu/people/pabo/movie%2Dreview%2Ddata/README.1.1
and
http://www.cs.cornell.edu/people/pabo/movie-review-data/diff.txt;
v2.0: http://www.cs.cornell.edu/people/pabo/movie%2Dreview%2Ddata/
poldata.README.2.0.txt. (This datasheet largely summarizes these
sources.)
No; however, while most names have been removed from the
preprocessed/tokenized versions of the data, the original data includes names and email addresses, which were also present on
the IMDb archive.
Does the dataset comply with the EU General Data Protection Regulation (GDPR)? Does it comply with any other standards, such as the US
Equal Employment Opportunity Act?
The preprocessed dataset may comply with GDPR; the raw data
does not because it contains personally identifying information.
Does the dataset contain information that might be considered sensitive or confidential? (e.g., personally identifying information)
If the dataset becomes obsolete how will this be communicated?
The raw form of the dataset contains names and email addresses,
but these are already public on the internet newsgroup.
This will be posted on the dataset webpage.
Is there a repository to link to any/all papers/systems that use this
dataset?
There is a repository, maintained by Pang/Lee through April
2012, at http://www.cs.cornell.edu/people/pabo/movie%2Dreview%
2Ddata/otherexperiments.html.
26
Does the dataset contain information that might be considered inappropriate or offensive?
Some movie reviews might contain moderately inappropriate or
offensive language, but we do not expect this to be the norm.
| 2cs.AI
|
arXiv:1803.04095v2 [math.GT] 4 Apr 2018
Action dimensions of some simple complexes
of groups
Michael W. Davis
Giang Le
Kevin Schreve
April 5, 2018
Abstract
The action dimension of a discrete group G is the minimum dimension
of contractible manifold that admits a proper G-action. We compute
the action dimension of the direct limit of a simple complex of groups
for several classes of examples including: 1) Artin groups, 2) graph
products of groups, and 3) fundamental groups of aspherical complements of arrangements of affine hyperplanes.
AMS classification numbers. Primary: 20F36, 20F55, 20F65,
57S30, 57Q35, Secondary: 20J06, 32S22
Keywords: action dimension, Artin group, Coxeter group, complex
of groups, graph product, hyperplane arrangement
Introduction
Suppose G is a discrete, torsion-free group with classifying space BG. Its
geometric dimension, gdim(G), is the smallest dimension of a model for its
classifying space BG by a CW complex. This number is equal to cd G, the
cohomological dimension of G (provided cd G 6= 2). Its action dimension,
actdim G, is the smallest dimension of a model for BG by a manifold. In other
words, actdim G is the minimum dimension of a thickening of a CW model
for BG to a manifold, possibly with boundary. It follows that gdim G ≤
actdim G with equality if and only if BG is homotopy equivalent to a closed
manifold. From general principles, actdim G ≤ 2 gdim G.
A common method of constructing groups and their classifying spaces
is to use the notion of a “complex of groups” (cf. [6]). Here we will only
1
use the easier notion of a simple complex of groups over a poset Q. By
definition, this means a functor, GQ, from Q to the category of groups
and monomorphisms. So, GQ is the following data: a collection of groups
{Gσ }σ∈Q and monomorphisms φστ : Gτ → Gσ , defined when τ < σ and
satisfying φστ φτ µ = φσµ when µ < τ < σ. Suppose we have models for the
BGσ and realizations for the homomorphisms φστ by maps φστ : BGτ →
BGσ . One then can glue together the {BGσ }σ∈Q (or more precisely iterated
mapping cylinders of the φστ ) to form a space BGQ called the aspherical
realization of GQ. If the geometric realization of Q is simply connected, then
it follows from van Kampen’s Theorem that π1 (BGQ) is the direct limit, G,
of the system of groups {Gσ , φστ }. The space BGQ may or may not be
aspherical. This is the “K(π, 1)-Question” for GQ. When the answer is
affirmative, BGQ is a model for BG. In this case, we get an upper bound for
gdim G in terms of the geometric dimensions of the Gσ . Similarly, if each BGσ
is modeled by a manifold with boundary, Mσ , and each φστ is homotopic to
an embedding fστ : Mτ → ∂Mσ , then we can glue together suitably thickened
versions of the Mσ to get a thickening of BGQ to a manifold with boundary
M. So, provided the K(π, 1)-Question for GQ has a positive answer, one
gets an upper bound for the action dimension of G in terms of the action
dimensions of the {Gσ }σ∈Q
In [4], Bestvina, Kapovich, and Kleiner define a number, obdim G, called
the “obstructor dimension” of G. It is a lower bound for the action dimension
of G. It is based on the classical van Kampen obstruction, vkn (K), for
embedding a finite simplicial complex K into Rn . This obstruction is a
cohomology class with Z2 coefficients in the configuration space of unordered
pairs of distinct points in K. Suppose EG denotes the universal cover of
BG. The idea of [4] is to find a complex K and a coarse embedding of
Cone∞ K into EG, where Cone∞ K means the cone of infinite radius on K.
It is proved in [4] that vkn (K) is also an obstruction to a coarse embedding
of Cone∞ K into any contractible (n + 1)-manifold; hence, when vkn (K) 6=
0, actdim G ≥ n + 2. In [4] the obstructor dimension of G is defined to
be n + 2, where n is the largest integer so that there exists a complex K
with vkn (K) 6= 0, together with a coarse embedding of Cone∞ K into EG.
We also shall have occasion to use a variation of this notion due to Yoon
[36], called the “proper obstructor dimension of G.” His idea is to consider
coarse embeddings T → EG where T is a contractible simplicial complex,
not necessarily of the form T = Cone∞ K.
In this paper we use the following two techniques to compute the action
2
dimension for certain groups which are direct limits of simple complexes of
groups.
(I) (Gluing). Construct a thickening of BGQ by gluing together manifolds
with boundary that are models for the BGσ , hence, establishing an
upper bound for actdim G. The pieces that are to be glued together
will have the form Mσ ×Dσ where Mσ ∼ BGσ and where Dσ is a “dual
disk.” Two such pieces will be glued together along a piece that is a
common codimension-0 submanifold of both boundaries. (The details
of this method are described in Section 2.)
(II) (Obstructors). Lower bounds for obdim G (and hence, for actdim G)
are established by finding obstructors for G. In most of our examples
the coarse obstructor will be a finite union of contractible manifolds,
containing a common basepoint, each of which is the universal cover
of some closed aspherical manifold. These contractible manifolds are
called sheets. When a sheet can be compactified to a disk, it is homeomorphic to the cone on a sphere and the coarse obstructor has the
form Cone∞ K where K is some configuration of spheres (K will often
be a “polyhedral join” of spheres, cf. Definition 4.6). So, such cases
reduce to calculating van Kampen obstructions vkn (K).
A necessary condition for BGQ to be aspherical is that the geometric
realization |Q| of the poset Q is contractible (see Remark 1.6). Often this will
be automatic, since if Q has a minimum element for which the corresponding
local group is the trivial group, then |Q| is a cone. If Q has such a minimum
element, then at the final stage of (I) we will need to glue a disk onto the
result of previous gluings. This will entail that each of previous manifolds
Mσ has nonempty boundary. However, if Q has no such minimum, then for
any minimal element σ of Q one can allow Mσ to be a closed manifold. We
shall return to this point later in the Introduction.
In most applications Q will be the poset S(L) of simplices in some simplicial complex L, including the empty simplex (so |S(L)| will be a cone).
A prototypical example is the case where G = AL , the right-angled Artin
group (or “RAAG”) associated to a d-dimensional flag complex L. Then
the simple complex of groups is the Artin complex, AS(L) = {Aσ }σ∈S(L) of
spherical Artin subgroups, BAL is the standard model for its classifying space
as a union of tori, and gdim AL = d + 1. It turns out that the appropriate
3
obstructor is the polyhedral join of 0-spheres O1 L, called the octahedralization
of L. The following two results are proved in [1, Theorems 5.1 and 5.2].
(i) If Hd (L, Z2 ) 6= 0, then vk2d (O1 L) 6= 0. Hence, actdim AL = obdim AL =
2d + 2.
(ii) If Hd (L, Z2 ) = 0 and d 6= 2, then actdim AL ≤ 2d + 1.
We discuss below three generalizations of RAAGs: 1) general Artin groups,
2) graph products of fundamental groups of closed aspherical manifolds, and
3) fundamental groups of aspherical complements of affine hyperplane arrangements. In each case we prove results similar to (i) and (ii) above. In all
three cases the relevant obstructor will be a polyhedral join Om L of (m − 1)spheres. Although we conjecture that a result similar to (ii) holds in both
cases 1) and 2), the proof of (ii) for RAAGs uses a special argument. (AL
is a subgroup of the right-angled Coxeter group corresponding to O1 L; if
vk2d (O1L) = 0, then O1 L embeds in a flag triangulation of S 2d and hence,
the Coxeter group is a subgroup of a Coxeter group that acts cocompactly
on a contractible (2d + 1)-manifold.) When the group is not a RAAG, we
instead use gluing methods to prove the analog of (ii) in the case where L
embeds in a contractible simplicial complex of the same dimension d. We
abbreviate this condition by saying that L is EDCE. (Note: if L is EDCE,
then Hd (L; Z2 ) = 0.)
1) (General Artin groups). Suppose AL is the Artin group, where the simplicial complex L is the nerve of the associated Coxeter system. The Artin
complex AS(L) is the simple complex of groups {Aσ }σ∈S(L) of spherical Artin
subgroups. The group AL is the direct limit of the system of groups defined
by AS(L). Conjecturally, the K(π, 1)-Question has an affirmative answer
for all Artin complexes. This is known in many cases (see [7]). When the
answer is affirmative, BAS(L) is covered by the “Salvetti complex” of AL
and gdim AL = dim BAS(L) = d + 1. The next result is proved in [15] and
[30]. We shall give the details of the arguments in Sections 2 and 4.
Theorem A (cf. [15] and Proposition 3.13). Suppose the K(π, 1)-Question
has a positive answer for AS(L).
(i) If Hd (L; Z2 ) 6= 0, then actdim AL = obdim AL = 2d + 2 = 2 gdim AL .
(ii) If L is EDCE, then actdim AL ≤ 2d + 1.
4
In [15] it is proved that the relevant obstructor for the Artin complex
is a polyhedral join of 0-spheres, O1 L⊘ , where L⊘ is a certain subdivision
of L whose simplices index the “standard free abelian subgroups” of AL .
As before, Cone(O1 L⊘ ) coarsely embeds in EAL . The calculation of the
van Kampen obstruction in part (i) of Theorem A is then the same as its
calculation for O1 L in the case of a RAAG.
2) (Graph products). Suppose {Gv }v∈V is a collection of groups indexed by
the vertex set V of a simplicial graph L1 . The graph product G is the quotient
of the free product of the Gv by the relations that Gv and Gw commute
whenever {v, w} is an edge of L1 . Let L be the flag complex associated to L1 .
For each simplex σ ∈ S(L), Gσ denotes the direct product of the Gv over the
vertex set of σ. The graph product complex GS(L) is the simple complex of
groups {Gσ }σ∈S(L) where the monomorphisms φστ are the natural inclusions.
Obviously, G is the direct limit of the Gσ . Moreover, the K(π, 1)-Question for
GS(L) always has a positive answer (see 1.3). One can essentially determine
actdim G in terms of the actdim Gv and we do so in Sections 2 and 4. The
most interesting case is when each Gv is the fundamental group of a closed
aspherical manifold Mv . A RAAG is the special case where each Mv = S 1 .
To simplify the statements of our results, assume each vertex manifold Mv
has the same dimension m. Then gdim G = m(d + 1). Using appropriate
thickenings of the Mσ we can use gluing technique (I) to construct a manifold
M of dimension (m + 1)(d + 1) which is a model for BG. By gluing together
standard lifts of the Mσ , we get a coarsely embedded Cone∞ (Om−1 L) in EG.
We then get the following analog of Theorem A.
Theorem B (cf. Corollaries 3.3, 4.25 and 3.15). Suppose, as above, that G
is the graph product of fundamental groups of closed, aspherical m-manifolds
Mv over a d-dimensional flag complex L.
(i) If Hd (L, Z2 ) 6= 0, then actdim G = obdim G = (m + 1)(d + 1).
(ii) If L is EDCE, then actdim G ≤ (m + 1)(d + 1) − 1.
A mild generalization of this is the case where each local group Gσ is
the fundamental group of a closed aspherical manifold Mσ of dimension
m(dim(σ)+1) and where the φστ are realized by embeddings fστ : Mτ ֒→ Mσ .
In Subsection 3.1, we call such a system {Mσ }σ∈S(L) a simple complex of
closed aspherical manifolds. When L is a flag complex, the K(π, 1)-Question
5
for GS(L) has a positive answer (cf. Theorem 1.19) and the conclusion of
Theorem B holds without change.
3) (Aspherical complements of hyperplane arrangements). Suppose A is an
arrangement
of affine hyperplanes in Cn . Let M(A) denote the complement
S
n
C − A. The relevant poset Q is the intersection poset of A. Its elements
are the proper subspaces σ of Cn which are intersections of hyperplanes in
A, ordered by reverse inclusion. The minimal elements of Q are a family of
parallel subspaces, and the arrangement is essential if these are zero dimensional. For each σ ∈ Q, there is a central arrangement Aσ in the subspace
normal to σ in Cn . Put G = π1 (M(A)) and Gσ = π1 (M(Aσ )). This is the
data for a simple complex of groups GQ. Of course, M(A) need not be aspherical; however, if it is, then so are the M(Aσ ). Let us assume that M(A)
is aspherical.
Since M(A) is a 2n-manifold, actdim G ≤ 2n. If A is central (meaning that the hyperplanes are linear), then M(A) deformation retracts onto
the complement of the hyperplane arrangement in the unit sphere S 2n−1 in
Cn and hence, actdim G ≤ 2n − 1. If A is not essential, then the parallel
subspaces can be deformation retracted, and again actdim G ≤ 2n − 1. On
the other hand, for essential aspherical arrangements we have the following
theorem.
Theorem C (cf. Theorem 5.11). Let A be an essential arrangement of affine
hyperplanes in Cn . Suppose M(A) is aspherical. Let G = π1 (M(A)). If A
does not decompose as a product with a factor equivalent to a nontrivial
central arrangement, then actdim G = obdim G = 2n.
This theorem implies that if A decomposes as a product of irreducibles
and k is the number of factors which are irreducible central arrangements,
then actdim G = 2n − k.
To understand why this is an analog of the previous theorems, two points
require explanation. First, |Q(A)| is homotopy equivalent to a wedge of
(n − 1)-spheres and, when A is not central, then there is at least one sphere
in the wedge. In particular, Hn−1 (|Q|, Z2 ) 6= 0. Secondly, there is a simplicial complex IQ(A), called the “irreducible complex,” such that IQ(A) is
homotopy equivalent to |Q(A)|, and so that the simplices of IQ(A) index
the standard free abelian subgroups in G. So, IQ(A) is the analog to L⊘ .
As before, the relevant obstructor is the polyhedral join O1 (IQ(A)).
The computations are evidence for the following conjecture connecting
action dimension to L2 -cohomology.
6
The Action Dimension Conjecture. (Davis-Okun [18]). If the ith L2 Betti number of G is nonzero, then actdim G ≥ 2i.
For example, if A is an irreducible, essential, affine hyperplane arrangement in Cn such that M(A) is aspherical, then the L2 -Betti number of
π1 (M(A)) are zero in degrees 6= n and if the arrangement is irreducible
and not central, then the nth L2 -Betti number of π1 (M(A)) is nonzero, cf.
[16].
In a forthcoming paper, we will determine the action dimensions for some
more examples of simple complexes of groups. These computations provide
further evidence for the Action Dimension Conjecture.
This paper is organized as follows. In Section 1, we review simple complexes of groups and explain our main examples. In Section 2, we discuss
thickenings of simplicial complexes and explain our method of gluing together
manifolds. In Section 3, we perform this operation to give upper bounds on
action dimension for Artin groups, graph products and hyperplane complements. In Section 4, we review the van Kampen obstruction, introduce our
main example of an obstructor complex, and use it to give lower bounds
for the action dimension of graph products and other simple complexes of
groups. In Section 5, we again use obstructor complexes to compute the
action dimension of fundamental groups of hyperplane complements.
Acknowledgements. This material is based upon work supported by the
National Science Foundation under Grant No. DMS-1440140 while the authors were in residence at the Mathematical Sciences Research Institute in
Berkeley, California, during the fall semester of 2016. The first author was
partially supported by an NSA grant and he received support in the spring
of 2017 as a Simons Fellow at the Isaac Newton Institute. The third author
is partially supported by NSF grant DMS-1045119. This material is based
upon work supported by the National Science Foundation under Award No.
1704364.
1
1.1
Simple complexes of groups
The basic construction
We begin by reviewing the theory of simple complexes of groups as developed
in Bridson-Haefliger [6, II.12].
7
Let Q be a poset. As in the Introduction, a simple complex of groups GQ
over Q is a collection of groups {Gσ }σ∈Q and monomorphisms φστ : Gτ → Gσ
defined whenever τ < σ. The Gσ are the local groups. Furthermore, GQ must
be a functor from Q to the category of groups and monomorphisms in the
sense that φστ φτ µ = φσµ whenever µ < τ < σ. (In [6, II.12.11, p. 375] the
order relation on Q is reversed.) A simple morphism ψ = (ψσ ) from GQ
to a group G is a function which assigns to each σ ∈ Q a homomorphism
ψσ : Gσ → G such that ψτ = ψσ φστ whenever τ < σ. The simple morphism
ψ is injective on local groups if each ψσ is injective. Such a simple complex
of groups GQ = {Gσ , φστ }; has a direct limit, denoted lim GQ.
For each σ ∈ Q, there is a canonical homomorphism ισ : Gσ → lim GQ,
hence, a canonical simple morphism ι : GQ → lim GQ. The simple complex
of groups GQ is developable if ι is injective on local groups. The direct
limit has the universal property that for any group H and simple morphism
ψ : GQ → H, there is a unique homomorphism ψ̂ : lim GQ → H such that
ψσ = ψ̂ισ (see [6, II.12.13, p. 376]).
The order complex of a poset P is the simplicial complex whose simplices
are the totally ordered finite subsets {τ0 , . . . , τk } of elements of P (where τ0 <
· · · < τk ). The underlying topological space of order complex of P is denoted
|P| and is called the geometric realization of P. If P opp denotes the opposite
poset where the order relations are reversed, then |P opp | is isomorphic to |P|.
There are two natural stratifications of |P| both indexed by P:
|P|σ := |P≤σ | and |P|σ := |P≥σ |.
(1.1)
In the first stratification, inclusion of one stratum into another corresponds
to the original order relation on P; in the second one, the order relation
is reversed. So, we should regard {|P|σ } as being indexed by P opp . Given
x ∈ |P|, let σ(x) be the index of the smallest stratum |P|σ containing x.
Given a simple complex of groups GQ and a simple morphism ψ : GQ →
G that is injective on local groups, one can define a poset D(Q, ψ) equipped
with a G-action, as well as, a space D(|Q|,
`ψ) also equipped with a G-action.
The poset D(Q, ψ) is the disjoint union Q G/Gσ , i.e., it is the set of pairs
(gGσ , σ), where σ ∈ Q and g ∈ G. (Here we are identifying ψσ (Gσ ) with
Gσ .) The order relation is defined by inclusion of cosets, i.e.,
(gGσ , σ) < (g ′Gσ′ , σ ′ ) ⇐⇒ σ < σ ′ and (g ′)−1 g ∈ Gσ′
The space D(|Q|, ψ) is called the basic construction or the development of
(|Q|, ψ). It is the quotient of G × |Q| by the equivalence relation ∼ defined
8
by
(g, x) ∼ (g ′, x′ ) ⇐⇒ x = x′ and gGσ(x) = g ′Gσ(x) .
(1.2)
One checks that D(|Q|, ψ) is the geometric realization of D(Q, ψ). Let [g, x]
denote the equivalence class of (g, x). The natural G-action on D(|Q|, ψ)
is induced from the action of G on itself by left translation; the isotropy
subgroup at [g, x] is gGσ(x) g −1 , and projection onto the second factor identifies the orbit space with |Q|. The orbit projection D(|Q|, ψ) → |Q| has
a a section i : |Q| → D(|Q|, ψ) defined by i : x 7→ [1, x]. Thus, i(|Q|) is
a strict fundamental domain for the G-action on D(|Q|, ψ). (To say that a
closed subspace of a G-space is a strict fundamental domain means that it
intersects each orbit in exactly one point.) Note that this implies that |Q| is
a retract of D(|Q|, ψ) (the orbit projection is the retraction).
Remark 1.1. Suppose G acts on a space D with a strict fundamental domain
Y so that the stratification of Y by closures of points of the same orbit type
{Y σ }σ∈Q is indexed by some poset Qopp and so that for any point y ∈ Yσ , the
isotropy subgroup Gy contains Gσ . This gives the data for a simple complex
of groups GQ over Q, where the local group Gσ is the isotropy subgroup at
a generic point of Y σ . There is a canonical simple morphism ψ = (ψσ ) from
GQ to G corresponding to the inclusions Gσ ֒→ G. As before, one defines
the development of Y with respect to ψ by D(Y, ψ) = (G × Y )/ ∼, where
the equivalence relation ∼ is defined as in (1.2). Moreover, the inclusion
Y ֒→ D extends to a G-equivariant homeomorphism D(Y, ψ) → D (see
[6, Prop. 12.20 (1), II.12]). In other words, D is determined by the strict
fundamental domain Y , the group G and the simple morphism from GQ to
G. (In previous work of the first author, e.g., [13, §5.1], the basic construction
is denoted by U(G, Y ) rather than by D(Y, ψ).)
Next, we recall a basic lemma, which can be found as [6, Prop. 12.20 (4),
II.12] or [35, Thms. 6, 10, pp. 32, 39].
Lemma 1.2. Let GQ be a developable simple complex of groups over Q. Let
ψ be a simple morphism from GQ to a group G and let ψ̂ : lim GQ → G be the
induced homomorphism. Suppose |Q| is simply connected. Then D(|Q|, ψ)
is simply connected if and only if ψ̂ is an isomorphism. (In particular, if
G = lim GQ, then D(|Q|, ι) is simply connected.)
Remark 1.3. Although we will not define the concepts of the “universal
cover” or the “fundamental group” for a general complex of groups, we will
9
give definitions for simple complexes of groups which agree with the more
general definitions in [6]. Put G = lim GQ and let ι : GQ → G be the
canonical simple morphism. First suppose |Q| is simply connected. Then,
since D(|Q|, ι) is simply connected (by Lemma 1.2), D(|Q|, ι) is the universal
cover of GQ and π1 (GQ) = G (cf. bh [6, III.C.3.11(1), p. 551]). If |Q| is
not connected, then each component of |Q| gives its own simple complex of
groups and can be treated separately. So, suppose |Q| is connected, but not
necessarily simply connected. Put π = π1 (|Q|). The poset structure on |Q|
lifts to a poset structure P on the universal cover of |Q| (so that the universal
cover of |Q| is |P|). The group of deck transformations π acts on |P| and on
P. Let p : P → Q be the projection. We have a simple complex of groups
GP defined by Gσ̃ = Gσ where σ̃ ∈ P lies above σ. Put G̃ = lim GP and let
ι̃ : GP → G̃ be the canonical simple morphism. By Lemma 1.2, the basic
construction D(|P|, ι̃) is simply connected. The group π acts on GP and
hence, on G̃. So, the semidirect product G̃ ⋊ π acts on D(|P|, ι̃) with orbit
space |P|/π = |Q|. Therefore, D(|P|, ι̃) is the universal cover of GQ and
π1 (GQ) = G̃ ⋊ π.
Aspherical realizations. The classifying space of a discrete group H is
denoted BH. The universal cover of BH is denoted EH. A simple complex of
groups GQ gives the data for a poset of spaces {BGσ , φστ }, where τ < σ ∈ Q
and where φστ : BGτ → BGσ is the map induced by the monomorphism
φστ : G`
τ → Gσ . Using this data we can glue together the disjoint union of
spaces |Q|σ × BGσ by using iterated mapping cylinders. For each τ < σ,
|Q|σ is a subcomplex of |Q|τ and one glues the subspace |Q|σ × BGτ of
|Q|τ × BGτ to |Q|σ × BGσ via the map:
I × φστ : |Q|σ × BGτ → |Q|σ × BGσ ,
where I denotes the identity map on |Q|σ . The resulting space BGQ is the
aspherical realization of GQ. It is well-defined up to homotopy equivalence.
If |Q| is connected and simply connected, then it follows from van Kampen’s
Theorem that π1 (BGQ) = lim GQ. We note that
dim BGQ = sup{(dim BGσ + dim |Q|σ ) | σ ∈ Q}
(1.3)
Proposition 1.4. Suppose GQ is a simple complex of groups over Q with
|Q| simply connected. Let G = lim GQ and let D = D(|Q|, ι) be the basic
construction. When GQ is developable, BGQ is homotopy equivalent to the
Borel construction EG ×G D.
10
Proof. Projection on the second factor induces a projection p : EG ×G D →
D/G = |Q| so that the inverse image of the vertex σ is homotopy equivalent
to BGσ . This uses the fact that GQ → G is injective on local groups,
otherwise, p−1 (σ) is homotopy equivalent to BGσ , where Gσ means the image
of Gσ in G. It follows that EG ×G D is an aspherical realization of GQ.
Recall that the K(π, 1)-Question for GQ is the following.
The K(π, 1)-Question. Is BGQ aspherical?
Corollary 1.5. Suppose GQ is a developable simple complex of groups with
|Q| simply connected. Then the K(π, 1)-Question for GQ has a positive
answer if and only if D is contractible.
Proof. By Proposition 1.4, BGQ ∼ EG ×G D. Hence, BGQ is aspherical if
and only if the universal cover of EG ×G D is contractible. This universal
cover is EG × D, which yields the corollary.
Remark 1.6. Since |Q| is a retract of D, a necessary condition for D to be
contractible is that |Q| is contractible.
In the construction of BGQ we can assume that for each σ, dim BGσ =
gdim Gσ . There is a simple condition which implies that the maximum value
of the quantity dim BGσ + dim |Q|σ in (1.3) occurs when σ is a maximal
element of Q, i.e., when |Q|σ is a point. It is:
gdim Gσ > gdim Gτ ,
whenever σ > τ .
(1.4)
Since a k-simplex in |Q|τ corresponds to a chain τ = τ0 < · · · < τκ , condition
(1.4) implies gdim Gτk ≥ gdim Gτ +k ≥ gdim Gτ +dim |Q|τ ; so the maximum
value occurs when σ is maximal.
Proposition 1.7. Suppose GQ is a developable simple complex of groups
and that the K(π, 1)-Question for GQ has a positive answer. Then
gdim G ≤ dim BGQ = sup{gdim Gσ + dim |Q|σ }.
σ∈Q
If (1.4) holds, then gdim G = supσ∈Q {gdim Gσ }.
Proof. Since gdim G ≤ dim BGQ, the first formula follows from (1.3). So, if
condition (1.4) holds, gdim G ≤ sup{gdim Gσ | σ ∈ Q}. Since EG/Gσ is a
model for BGσ , gdim G ≥ gdim Gσ , so the previous inequality must be an
equality.
11
1.2
Examples where Q is a poset of simplices
Given a simplicial complex L, its poset of simplices (including the empty
simplex) is denoted by S(L). Its geometric realization |S(L)| is the cone on
the barycentric subdivision of L. (The vertex corresponding to the empty
simplex is the cone point.) Most of this paper concerns simple complexes
of groups over posets Q of the form S(L). Given such a simple complex
of groups GS(L), we often will write G for lim GS(L) and D for the basic
construction D(|S(L)|, ι). In this subsection we introduce examples of main
interest of such complexes of groups coming from Coxeter groups, Artin
groups, and graph products.
Example 1.8. (Coxeter groups) Suppose (W, S) is a Coxeter system (cf. [5]
or [13]). This means that W is a group, that S is a distinguished set of
generators and that W has a presentation of the form
W := hsi ∈ S|s2i = (si sj )mij = 1i
For any subset T of S, the subgroup generated by T is denoted WT and called
the special subgroup corresponding to T . It is a standard fact that (WT , T )
also is a Coxeter system (cf. [5, pp.12-13]). The subset T is spherical if WT is
finite, in this case WT is a spherical special subgroup. Let S(W, S) denote the
poset of spherical subsets of S. There is a simplicial complex L (= L(W, S)),
called the nerve of (W, S). Its vertex set is S and a subset T ≤ S spans a
simplex of L if and only if T is spherical. Thus, S(L) = S(W, S), the poset of
spherical subsets. This gives a simple complex of groups over S(L), denoted
W S(L), and called the complex of spherical Coxeter groups (or simply the
Coxeter complex ). The local group Wσ at a simplex σ ∈ S(L) is the spherical
subgroup generated by the vertices of σ. The direct limit of W S(L) is the
Coxeter group W .
Example 1.9. (Artin groups). Given a Coxeter system (W, S) there is an
associated Artin group A. This group has one generator xs for each s ∈ S;
its relations are the braid relations:
xs xt · · · = xt xs · · ·,
| {z } | {z }
mst terms
mst terms
where both sides of the equation are alternating words in xs and xt , where
{s, t} ranges over the edges of the nerve L, and where mst denotes the order
12
of st in W . The matrix (mst ) is called the Coxeter matrix. For any T ≤ S,
let AT denote the Artin group corresponding to the Coxeter system (WT , T ).
As with Coxeter groups, AT can be identified with the special subgroup of
A generated by {xt }t∈T . (When T is a spherical subset of S, this is proved
in [20, Thèoréme 4.13 (iii)] and in general in [29, Theorem 4.14].) The
Artin group AT is spherical if T is a spherical subset of S. As with Coxeter
groups, let S(L) be the poset of spherical subsets of S. The spherical special
subgroups of A give a simple complex of groups AS(L) called the Artin
complex. When σ is a simplex of L and T = Vert σ, we shall often write
Aσ instead of AT . It is clear that the direct limit lim AS(L) is A. Since Aσ
is isomorphic to a subgroup of A, AS(L) is developable. Note that if L′ is
any subcomplex of the full simplex on S with the same 1-skeleton as L, then
lim AS(L′ ) = A.
By Deligne’s Theorem in [20], any spherical Artin group Aσ has a classifying space BAσ which is a finite CW complex of dimension one greater
than dim σ. (There is a specific model for BAσ called the Salvetti complex,
see [8] or Subsection 3.2). The “K(π, 1)-Conjecture” for Artin groups is the
conjecture that the K(π, 1)-Question has a positive answer for any Artin
poset AS(L). By Proposition 1.5 this is equivalent to the conjecture that
D(|AS(L)|, A) is contractible. It is the most important unsolved problem
concerning general Artin groups. A detailed discussion can be found in [7],
where the conjecture is proved whenever L is a flag complex (we generalize
this in Theorem 1.19 below).
The discussion in Example 1.9 yields the following proposition.
Proposition 1.10. Let A be an Artin group such that the nerve L of its
associated Coxeter system is d-dimensional. If the K(π, 1)-Question for the
associated Artin poset has a positive answer, then gdim A = d + 1.
Proof. Since the aspherical realization BAS(L) is formed by gluing together
the BAσ × |S(L)|σ with σ ∈ S(L) and since dim BAσ = dim σ + 1, we have
dim BAS(L) = d + 1. Since BAS(L) is a model for BA, gdim A ≤ d + 1. On
the other hand, for any d-simplex σ ∈ S(L), the spherical Artin group Aσ
contains a free abelian subgroup of rank d+1 (e.g., see [15]). So, cd AL ≥ d+1.
Hence, d+1 ≥ gdim AL ≥ cd AL ≥ d+1; so all inequalities are equalities.
Definition 1.11. A simplicial complex L is a flag complex if it satisfies the
following: if T is any finite set of vertices of L which are pairwise connected
13
by edges, then T spans a simplex of L. A simplicial graph L1 determines a
flag complex L: the simplices of L are the cliques in L1 . (This is also called
the “clique complex” of L1 .)
Definition 1.12. Suppose L1 is a simplicial graph with vertex set V and
edge set E. Let {Gv }v∈V be aQcollection of groups indexed by V . The graph
product of the Gv , denoted L1 Gv , is the quotient of the free product of
the Gv , v ∈ V , by the normal subgroup generated by all commutators of the
form, [gv , gw ], where {v, w} ∈ E, gv ∈ Gv and gw ∈ Gw .
Q
Example 1.13. (The graph product complex ). Suppose L1 Gv is a graph
product and that L is the flag complex determined by L1 . There is a simple
complex of groups GS(L) over S(L) called the graph product complex. It is
defined by putting Gσ equal to the direct product,
Y
Gv ,
v∈Vert σ
for each simplex σ ∈ L and letting φστQ: Gτ → Gσ be the natural inclusion
whenever τ < σ. It is immediate that L1 Gv = lim GS(L).
Remark 1.14. Another approach to the graph product complex is given
in [12] or [14] using the notion of a “polyhedral product”. Suppose L is a
simplicial complex with vertex set V and that we are given a collection of
pairs of spacesQ(X, Y ) = {(Xv , Yv )}v∈V together with a choice of basepoint
∗v ∈ Yv . Let v∈V Xv denote the subspace of the Cartesian product consisting of all V -tuples (xv )v∈V such that xv = ∗v for all but finitely many v.
Given a V -tuple, x = (xv ), put σ(x) = {v Q
∈ V | xv ∈ Xv − Yv }. The polyL
hedral product (X, Y ) is the subspace of v∈V Xv consisting of all x such
that σ(x) is a simplex in L. (Usually, we only shall be concerned with the
case where each Yv equals the basepoint ∗v , in which case the notation will
be simplified to X L .) For any τ ∈ S(L),
Q the set of all x with σ(x) ≥ τ , can
be identified with the product, X τ := v∈Vert τ Xv . Thus, X L is the union of
the X τ , with τ ∈ S(L).
′
Next let L1 be any simplicial graph and LQ
any simplicial complex
with
Q
1
1-skeleton = L . Let G be the graph product L1 Gv and H = v∈V Gv be
the direct product. Let ι : GS(L1 ) → G and ψ : GS(L1 ) → G be the natural
simple morphisms. Consider the polyhedral product:
′
Z(L′ ) = (Cone Gv , Gv )L .
14
Then Z(L′ ) is locally isomorphic to a product of cones on discrete sets.
Moreover, Z(L′ ) can be identified with the development D(|S(L′ )|, ψ). The
fundamental group of Z(L′ ) can be identified with the kernel of the natural
epimorphism G → H. So, the H-action on Z(L′ ) lifts to a G-action on the
universal cover Z̃(L′ ) with the same strict fundamental domain. It follows
from Remark 1.1 that Z̃(L′ ) = D(|S(L′ )|, ι). When L′ = L, Z̃(L) can be
identified with the standard realization of a right-angled building (cf. [14,
§2.2]). Therefore, Z̃(L) has a CAT(0) structure and hence, is contractible.
A corollary to these observations is the following result.
Q
Proposition 1.15. (cf. [14, Theorem 2.22] and [17]). Suppose G = L1 Gv
is the graph product of nontrivial groups over a simplicial graph L1 , and let
L be the flag complex determined by L1 .. Then the K(π, 1)-Question for the
graph product complex GS(L) has a positive answer.
Since BG1 × BG2 is a model for B(G1 × G2 ), it is obvious that
gdim(G1 × G2 ) ≤ gdim G1 + gdim G2 .
On the other hand, by using certain torsion-free subgroups of RACGs, Dranishnikov [24] showed that there are groups G1 and G2 for
Q which the inequality
is strict (cf.P[13, Example 8.5.9]). Hence, for Gσ = v∈σ Gv , we have that
gdim Gσ ≤ v∈σ gdim Gv and the inequality can be strict.
A corollary to Proposition 1.15 is the following calculation of the geometric dimension of any graph product of groups.
Q
Corollary 1.16. Suppose G = L1 Gv is the graph product of nontrivial
groups over a simplicial graph L1 . Let L be the flag complex determined by
L1 . Then gdim G = sup{gdim Gσ | σ ∈ S(L)}.
1.3
Flag complexes and the K(π, 1)-Question
We will use the following result from [35].
Proposition 1.17. ([35, Prop. 3, p. 6]). Let Gi be a collection of groups
with common subgroup A and let A Gi denote the amalgamated product. Let
Hi ⊂ Gi be subgroups and suppose the intersection B = Hi ∩A is independent
of i. Then the natural homomorphism B Hi → A Gi is injective.
∗
∗
15
∗
As usual, GS(L) is a simple complex of groups. For any full subcomplex
L of L, put GL′ = lim GS(L′ ) and let GL′ denote the image of GL′ in GL .
For any σ ∈ S(L), let Gσ denote the image of Gσ in GL . (If GS(L) is
developable, then Gσ → Gσ is an isomorphism.) The intersection of local
groups condition for simplices σ and τ of L is the following:
′
Gσ ∩ Gτ = Gσ∩τ .
(1.5)
Lemma 1.18. Suppose GS(L) is developable and that (1.5) holds for all
σ, τ ∈ S(L). If L is a flag complex, then for any full subcomplex L′ of L, the
natural map GL′ → GL is an injection.
We first prove this in a special case where L is the cone on another flag
complex. We will use the notation St(u) for the cone and Lk(u) for the base
of the cone; regarding u as the cone vertex.
Special Case: We will prove the following:
(i) The natural map GLk(u) → GSt(u) is injective.
(ii) Suppose L0 , L1 are full subcomplexes of Lk(u) with L0 ⊂ L1 . Then
GL1 ∩ GL0 ∗u = GL0 .
Proof of the Special Case. The proof is by induction on the number of vertices of St(u). When St(u) is a simplex, (i) and (Ii) follow from (1.5). If
St(u) is not a simplex, then, since Lk(u) is a flag complex, there is a vertex
v in Lk(u) which is not connected by an edge to some other vertex of Lk(u).
Denote the star of v in St(u) simply by St(v). So, St(v) is a proper subcomplex of St(u). By induction on the number of vertices in the cone, (i) holds
for St(v), i.e., GLk(v) → GSt(v) is injective, where Lk(v) := Lk(v, St(v)). By
induction on the number of vertices, the natural maps,
GLk(u)−v → GSt(u)−v , GSt(v,Lk(u)) → GSt(v) , and GLk(v,Lk(u)) → GLk(v) ,
are injective. We have decompositions as amalgamated products:
GSt(u) = GSt(u)−v ∗GLk(v) GSt(v)
GLk(u) = GLk(u,St(u)−v) ∗GLk(v,Lk(u)) GLk(u,St(v))
By induction on the number of vertices, we have that (ii) holds in two cases:
first with L1 = Lk(u, St(u) − v) = Lk(u) − v and L0 = Lk(v, Lk(u)) and
16
Lk(u, Lk(v))
Lk(u)
Lk(u, St(v)) = St(v, Lk(u))
St(u)
u
v
St(v)
Figure 1: Setup for Lemma 1.19
second with L1 = St(v) and L0 = Lk(v, Lk(u)). This yields
GLk(u,St(u)−v) ∩ GLk(v,Lk(u))∗u = GLk(v,Lk(u)) ,
GSt(v) ∩ GLk(v,Lk(u))∗u = GLk(v,Lk(u)) .
Applying Proposition 1.17, we see that GLk(u) → GSt(u) is injective.
Completion of Proof. We can now complete the proof of Lemma 1.18. The
argument is again by induction on the number of vertices. Let L′ be a full
subcomplex of L. If L = L′ ∗ v, we are done by the special case. Otherwise,
there is a vertex v such that L 6= St(v) and v ∈
/ L′ . There is an amalgamated
product decomposition:
GL = GL−v ∗GLk(v) GSt(v) .
By inductive hypothesis for the proper subcomplexes L − v and St(v), we
have inclusions GLk(v) → GL−v and GLk(v) → GSt(v) . Since both factors inject
into the amalgamated product, GL−v → GL is an injection. Continuing by
deleting one vertex at a time, we see the same holds for GL′ .
Theorem 1.19. (cf. [7, Remark on p. 619]). Suppose L is a flag complex,
that GS(L) is developable, and that condition (1.5) holds for all τ , σ in S(L).
Then the K(π, 1)-Question for GS(L) has a positive answer.
17
Proof. The proof is by induction on the number of maximal simplices in
L. The base case is when L is a single simplex σ. Then Gσ = lim GS(σ)
and BGS(σ) ∼ BGσ . So, the answer to the K(π, 1)-Question for GS(σ) is
positive.
Suppose L is not a simplex. Since L is a flag complex, there are distinct
vertices v1 , v2 in V = Vert L so that v1 and v2 are not joined by an edge of L.
Let L1 , L2 and L0 be the full subcomplexes of L spanned by V −{v1 }, V −{v2 }
and V − {v1 , v2 }, respectively, and let G1 , G2 and G0 be the respective direct
limits. Since GL is a direct limit, it is the amalgamated product GL =
G1 ∗G0 G2 . By Lemma 1.18, the natural maps from G0 → Gi and Gi → GL
are injections for i = 1, 2. A lemma of Whitehead states that if two aspherical
complexes are glued together along an aspherical subcomplex so that the
fundamental group injects into either side, then the result of the gluing is
also aspherical (for example, see [13, Thm. E.1.15]). (This is a special case
of theorem that we are proving.) By inductive hypothesis, BGS(Li ) ∼ BGi
and, by its definition, BGS(L) is the union of BGS(L1 ) and BGS(L2 ) along
BGS(L0 ). Hence, Whitehead’s Lemma shows that BGS(L) ∼ BGL .
Remark 1.20. Theorem 1.19 shows that if GS(L) is developable, (1.5)
holds and L is a flag complex, then D(|S(L)|, GL ) is contractible. Moreover, D(|S(L′ )|, GL′ ) is contractible for any full subcomplex L′ ≤ L. For
graph product complexes, these developments are right-angled buildings and
hence, are CAT(0). This raises the question of nonpositive curvature in the
context of Theorem 1.19. This is equivalent to the question of whether the
link of each simplex in D(|S(L)|, GL ) is CAT(1), (see [6]). For bounded curvature to make sense, we should first put a piecewise spherical structure on L.
Since L is already assumed to be a flag complex we might as well assume that
each simplex in L is an all right spherical simplex so that D(|S(L)|, GL ) becomes a cube complex, see [13]. The link, Lkσ , corresponding to a nonempty
simplex σ in L is the development of the simplex of groups GS(∂σ) with
respect to the natural simple morphism ψ : GS(L) → Gσ , so the question
becomes whether Lkσ = D(|S(∂σ)|, ψ) is CAT(1). By Gromov’s Lemma (cf.
[13, Appendix I.6, pp.516-517]), this is equivalent to the question of whether
it is a flag complex. Note that this is implied by the condition that the development of the simplex of groups, D(|S(∂σ)|, G∂σ ), is a flag complex. (Here
G∂σ = lim GS(∂σ)).
Similar considerations led R. Charney and the first author to conjecture
in [7] that the development of any Artin complex in Example 1.9 can be
18
given CAT(0) structure. This would imply the K(π, 1)-Conjecture for all
Artin groups. The piecewise spherical structure on L should be the natural
one in which σ is isometric to a fundamental spherical simplex in the round
sphere on which Wσ acts as reflection group; in other words, the spherical
simplex which makes the spherical realization of the Coxeter complex for Wσ
into a round sphere. In the Artin complex, Lkσ is the Deligne complex for
Aσ . So, the conjecture of [7] is that the natural piecewise spherical metric
on the Deligne complex for Aσ is CAT(1).
1.4
Q is the poset of nonempty simplices in L
Suppose L is a full subcomplex of another simplicial complex L̄. Then every
simplex σ̄ ∈ S(L̄) can be decomposed as a join σ̄ = σ ∗ τ , where σ = σ̄ ∩ L
and τ is a simplex whose vertices lie in L̄ − L. Suppose GS(L) and GS(L̄)
are simple complexes of groups such that GS(L) is the restriction of GS(L̄)
to S(L). Then GS(L̄) is a trivial extension of GS(L) if Gσ̄ = Gσ , whenever
σ̄ = σ ∗ τ decomposes as a join as above.
When L is a subcomplex of L̄, we can replace L̄ by a subdivision relative
to L so that L becomes a full subcomplex of L̄. If L is a flag complex,
then the subdivision L̄ can be assumed to be flag. So, any simple complex
of groups GS(L) over S(L) has a trivial extension to a simple complex of
groups over S(L̄) as above. The following lemma is clear.
Lemma 1.21. Suppose GS(L̄) is a trivial extension of GS(L). Then
(i) lim GS(L̄) = lim GS(L).
(ii) GS(L̄) is developable if and only if GS(L) is developable.
(iii) BGS(L̄) is homotopy equivalent to BGS(L).
For a given simplicial complex L, let S o (L) denote its poset of nonempty
simplices, i.e., S o (L) = S(L)>∅ . The geometric realization of S o (L) is the
barycentric subdivision of L. Let GS(L) be a simple complex of groups
over S(L) and let GS o (L) denote its restriction to S o (L). For example,
S o (Cone L) can be identified with S(L) and any simple complex of groups
GS(L) can be identified with a simple complex of groups over S o (Cone L).
Since we shall always assume that the local group G∅ attached to the empty
simplex is the trivial group, then simple complexes of groups GS o (L) and
19
GS(L) have the same direct limit, which we denote by G. The next lemma
is immediate.
Lemma 1.22. The complex of groups GS o (L) is developable if and only if
GS(L) is developable.
Lemma 1.23. Suppose L is connected. Then the following statements are
equivalent
(i) L is contractible.
(ii) The inclusion D(|S o (L)|, G) ֒→ D(|S(L)|, G) of developments is a Gequivariant homotopy equivalence.
(iii) The inclusion BGS o (L) ֒→ BGS(L) of aspherical realizations is a homotopy equivalence.
(iv) The K(π, 1)-Questions for GS o (L) and for GS(L) have the same answers.
Proof. By Remark 1.3, the following three conditions are equivalent:
• L is simply connected.
• D(|S o (L)|, G) is simply connected.
• π1 (BGS o (L)) = G.
To simplify notation, put D = D(|S(L)|, G) and D o = D(|S o (L)|, G). Let
C = |S(L)| − |S o (L)| be the open dual cone of the vertex in |S(L)| corresponding to ∅. The inverse image of C in D(|S(L)|, G) consists of G copies
of C. Hence,
M
H∗ (Cone L, L), and
H∗ (D, D o) ∼
=
G
H∗ (EGS(L), EGS o (L)) ∼
=
M
H∗ (Cone L, L).
G
The equivalence of conditions (i), (ii) and (iii) follows. The equivalence of
condition (iv) follows from Corollary 1.5.
20
We will use the above lemma in the following way. Given GS(L), first
embed L as a full subcomplex of a contractible simplicial complex Lc and
then take a trivial extension of GS(L) to GS(Lc ). By Lemmas 1.21 and
1.23, there are homotopy equivalences BGS(L) ∼ BGS(Lc ) ∼ BGS o (Lc ),
i.e., BGS o (Lc ) is another model for BGS(L). If dim Lc = dim L, then
dim(|S o (L)|) = dim(|S o (L)|) − 1 and we will sometimes be able to use this
to reduce upper bounds on the action dimension or the geometric dimension
by 1.
Definition 1.24. A d-dimensional simplicial complex L is equidimensionally, contractibly embeddable (abbreviated EDCE) if it can be embedded in
a contractible complex Lc of the same dimension d.
Remark 1.25. For d 6= 2 the condition that L be EDCE is equivalent to
the following two conditions:
Hd−1 (L, Z) is free abelian and Hd (L, Z) = 0.
By the Universal Coefficient Theorem, the above conditions are equivalent
to the condition that H d (L, Z) = 0. Indeed, when these conditions hold, one
can use standard methods to attach cells of dimension ≤ d to kill all the
homology of L. When d = 2, we will only end up with an Lc which is acyclic.
A conjecture of Kervaire asserts that one cannot kill a nontrivial group by
adding the same number of generators and relations. So, in the many situations where Kervaire’s Conjecture is known to hold, it is not possible to
obtain a contractible complex by adding 1 and 2-cells to an acyclic complex.
2
Gluing
Suppose we are given a collection of manifolds with boundary {Mτ }τ ∈Q indexed by a poset Q. Further suppose that whenever τ < σ, we have an
embedding iτ σ : Mτ ֒→ ∂Mσ with trivial normal bundle. This means, in particular, that whenever σ is not a minimal element of the poset, ∂Mσ must be
nonempty. The poset of manifolds {Mτ }τ ∈Q is n-dimensional if for each maximal σ, dim Mσ = n. Henceforth, we assume this. Let c(τ ) = n − dim Mτ
and let Dτ be a disk of dimension c(τ ). The basic idea in this section is
that we can glue together the Mτ × Dτ along codimension-zero submanifolds of their boundaries to obtain M, an n-manifold with boundary. Let
21
Gτ := π1 (Mτ ). We assume each inclusion Mτ ֒→ Mσ is π1 -injective, and that
GQ = {Gτ }τ ∈Q is a simple complex of groups. If |Q| is simply connected,
then π1 (M) = lim GQ is the direct limit of the π1 (Mτ ). So, if each Mτ is
aspherical, M will be a model for BGQ. In practice Q will be S(L) or S o (L)
for some L.
The main work in Subsection 2.1 is to describe the “dual disk” Dµ , for
each µ ∈ S(L) and a decomposition of its boundary sphere into pieces Tµ (τ ).
For τ > µ, Mµ × Dµ will be glued onto Mτ × Dτ along Mµ × Tµ (τ ). The dual
disk is a thickening of the dual cone, Cone(Lµ ), where Lµ means the normal
link of µ ∈ L (i.e., Lµ is the geometric realization of S(L)>µ ). If Dµ is a
k-disk, then it should be attached to other pieces along a submanifold Tµ of
codimension 0 in ∂Dµ (= S k−1). The manifold Tµ is a thickening of Lµ in
S k−1 (Tµ is the “thick link”). The thick link Tµ is further decomposed into
pieces Tµ (τ ) on which the piece corresponding to τ is to be attached.
Here is a picture to keep in mind. Suppose L is a simplicial graph, so
that L embeds in S 3 . The dual disk D∅ is D 4 and T∅ is a thickening of L
in S 3 to a 3-manifold with boundary. For each vertex µ, T∅ (µ) is a 3-ball
neighborhood of µ and for each edge τ , T∅ (τ ) is a tubular neighborhood of
the edge as a solid cylinder. (See Figure 2.1.) The dual disk to a vertex µ
is a 2-disk. If the degree of µ is p, then Lµ consists of p points and Tµ is a
thickening to p intervals in S 1 . The dual cell to an edge (i.e. to a maximal
simplex) is a 0-disk.
The simplicial complex |S(L)| is equal to Cone L′ , where L′ means the
barycentric subdivision of L. So, a k-simplex in Cone L′ is a chain {τ0 , . . . , τk }
of length k + 1 in S(L). We denote the geometric realization of this simplex
by [τ0 , . . . , τk ]. (We will always write such chains in increasing order, i.e.,
τ0 < · · · < τk .) A chain of length 1, {τ }, is either the cone point [∅] or it
corresponds to a vertex [τ ] of L′ (thought of as the barycenter of τ ). Thus,
[τ0 , . . . , τk ] is the k-simplex in Cone L′ spanned by the vertices [τ0 ], . . . , [τk ].
Given a simplex α = [τ0 , . . . τk ] in Cone L′ , its minimum vertex is defined by
min α = τ0 .
To understand the decompositions of dual disks, one first needs to understand the stratification of |S(L)| into “dual cones.” In this section we shall
often use the notation K (or K∅ ) instead of |S(L)|. Similarly, Kσ is |S(L)|≥σ
and called the dual cone of σ. (Here we are reversing the use of superscripts
and subscripts from notation in (1.1).) We will use ∂Kσ or L′σ for |S(L)|>σ ,
the barycentric subdivision of the link of σ. Thus, a simplex α = [τ0 , . . . τk ]
of K lies in Kσ (resp., ∂Kσ ) if and only if min α ≤ σ (resp., min α < σ).
22
2.1
Thick links
Suppose L is a finite simplicial complex of dimension d. We recall a method
for thickening L to a manifold with boundary T . First, piecewise linearly
embed L into a sphere S n−1 , where n ≥ 2d + 2. Thus, L is a full subcomplex
of some PL triangulation S of S n−1 . Denote the barycentric subdivisions of L
and S by L′ and S ′ , respectively. Let T denote the first derived neighborhood
of L in S. In other words, T is the union of simplices in S ′ which have
nonempty intersection with L′ .
If ρ0 < · · · < ρk is a chain of simplices in S(S), then [ρ0 , . . . , ρk ] denotes
the simplex in S ′ spanned by their barycenters.
For each vertex ν in L, let T (ν) denote the closed star of [ν] in S ′ , i.e.,
[
T (ν) = {γ ∈ S ′ | ν ≤ min γ}.
S
Then T (ν) is an (n − 1)-disk. Moreover, T = T (ν), where the union is over
all vertices ν ∈ L. If τ is a simplex of L, then let T (τ ) denote the normal
star of τ in S, i.e.,
[
T (τ ) = {γ ∈ S ′ | τ ≤ min γ} = [τ ] ∗ Lk(τ, S)′ .
So, T (τ ) is the cone on Lk(τ, S)′. Since Lk(τ, S) is a sphere of dimension
n−k−2, where k = dim τ , we see that T (τ ) is PL homeomorphic to a (n−k−
1)-disk. Moreover, if Vert τ = {ν0 , . . . , νk }, then T (τ ) = T (ν0 ) ∩ · · · ∩ T (νk ).
So, T is an (n − 1)-manifold with boundary embedded in S n−1 = ∂D n
Next we want to apply this construction to links in L. For each simplex
µ ∈ S(L), let Lµ = Lk(µ, L). We want to thicken Lµ in a sphere Sµ := S c(µ)−1
of an appropriate dimension c(µ) − 1 to obtain a manifold with boundary
Tµ called a thick link, embedded in a disk Dµ = D c(µ) , called the dual disk.
Similarly, whenever µ < τ , we will have Tµ (τ ), the normal star of τ in Sµ .
Thus, Tµ (τ ) is a thickening of ∂Kµ (τ ). In particular, when µ is the empty
simplex we will have T∅ = T , D∅ = D n and T∅ (τ ) = T (τ ).
To specify the dimensions of the dual disks suppose we are given a function
c : S(L) → N so that
• For any maximal simplex σ of L, c(σ) = 0.
(2.1)
• If τ < σ, then c(τ ) − c(σ) ≥ 2 codim(τ, σ)
(2.2).
23
µ
τ
σ
T (µ)
Tµ (σ)
T (τ )
[µ]
[τ ]
[σ]
Figure 2: A thickening of a graph in S 3 , and some of the associated dual
disks.
Put n = c(∅). (If we are using S o (L) instead of S(L), then we don’t
require the second condition for τ = ∅.) Thus, the dual disk Dµ will be a
thickening of the dual cone Kµ and the thick link Tµ will be a thickening of
∂Kµ (= L′µ ).
The link L′µ naturally is a subcomplex of L′∅ and, in fact, whenever µ < τ ,
L′τ is a subcomplex of L′µ . Similarly, if Sµ′ means the barycentric subdivision
of S c(µ)−1 , then we want to arrange that Sτ′ is a subcomplex of Sµ′ . This
amounts to requiring that Sτ′ is PL embedded in Lk(τ, S c(µ)−1 )′ ⊂ Sµ′ . For
example, if µ = ∅ and L∅ is a graph with a vertex τ , then dim L′∅ = 1 and
dim L′τ = 0. We can thicken L′∅ in S 3 = S∅ , while L′τ should be thickened in
Sτ′ = S 1 , see Figure 2.
Next we want to explain how dual disks can be regarded as manifolds
with corners. Recall that an n-manifold with boundary, P , is a smooth
manifold with corners if it is locally differentiably modeled on the simplicial
cone [0, ∞)n . This can be extended to a definition for topological manifolds
by requiring that the overlap maps preserve the stratification of [0, ∞)n by
intersections with coordinate subspaces. The stratification of [0, ∞)n by its
faces then induces a stratification of P . A codimension-one stratum of P is
24
called a facet.
Lemma 2.1. For each µ ∈ S(L), the dual disk Dµ is a c(µ)-manifold with
corners. The facets are {Tµ (τ )} where µ is a codimension one face of τ (i.e.,
where [τ ] is a vertex of L′µ ) together with ∂Dµ − Tµ .
The facet ∂Dµ − Tµ is called a boundary piece; the other facets are ordinary facets.
Proof. If µ is a codimension one face of τ , then Tµ (τ ) ⊂ ∂Dµ is a disk
of codimension one in Dµ . In general, if µ is a codimension-k face of a
simplex σ in L and {[τ0 ], . . . , [τk ]} are the vertices of σ in L′µ , then Tµ (σ) =
Tµ (τ0 ) ∩ · · · ∩ Tµ (τ0 ) is a disk of codimension k in Dµ . Hence, the Tµ (τ )
intersect in the same fashion as the facets of the simplicial cone [0, ∞)c(µ) .
2.2
Gluing complexes of manifolds with boundary
A complex of manifolds with boundary over S(L) is a collection of manifolds
with boundary, {Mτ }τ ∈S(L) , together with embeddings iστ : Mτ ֒→ ∂Mσ defined whenever τ < σ. In particular, we must have that ∂Mσ is nonempty
whenever σ > ∅; however, the minimum manifold M∅ can have empty boundary, and we usually assume this. In other words, the manifolds are indexed by
the vertices of Cone(L′ ) (the cone on the barycentric subdivision of L), while
the embeddings iστ are indexed by the edges of Cone(L′ ). In addition, we
require that there are certain (k − 1)-parameter families of isotopies between
the iστ which are indexed by the k-simplices of Cone(L′ ). We eventually
will want to require that the codimension of Mτ in Mσ is c(τ ) − c(σ) where
c : S(L) → N is a function as in Subsection 2.1, and that the image of Mτ
in ∂Mσ has trivial normal bundle. Before giving the precise requirements let
us mention that we also will be using the notion of a complex of manifolds
over S 0 (L), where the empty simplex is not needed and where Cone(L′ ) is
replaced by L′ .
We want to describe how to glue together the Mτ × Dτ , where Dτ is the
dual disk defined in Subsection 2.1. It is easier to first describe how to glue
together the Mτ × Kτ where Kτ is the dual cone of τ (i.e., Kτ = [τ ] ∗ L′τ
is a subcomplex of the |S(L)| = [∅] ∗ L′ ). The gluing is accomplished using
various embeddings:
hα∗µ : Mµ × α ֒→ ∂Mmin α × α,
25
[σ]
Kτ
[τ ]
...
...
Kµ
[µ]
Figure 3: Neighborhoods in the barycentric subdivision of L, the link of ∅ in
S(L).
where α is a simplex in L′ and µ ∈ S(L) is a proper face of min α. The
embedding hα∗µ will be used to glue Mµ × (α ∗ [µ]) onto Mmin α × Kmin α .
For the gluing to be well-defined the hα∗µ must satisfy certain compatiblity
relations which we will now describe.
First suppose α is a vertex of L′ , i.e., α = [τ ], where τ is a simplex of
L. Then h[τ ]∗[µ] := iτ µ × I[τ ] : Mµ × [τ ] → ∂Mτ × [τ ]. Essentially, h[τ ]∗[µ]
is iτ µ . Next, suppose α = [τ, σ] is an edge in L′ so that h[τ,σ]∗[µ] : Mµ ×
[τ, σ] → Mτ × [τ, σ] is an isotopy. The restriction of this isotopy to Mµ × [τ ]
is h[τ ]∗[µ] : Mµ × [τ ] → Mτ × [τ ]. Its restriction to the other end when
precomposed with h[σ]∗[τ ] should equal h[σ]∗[µ] , that is,
(h[σ]∗[τ ] )(h[τ,σ]∗[µ] |Mµ ×[σ] ) = h[σ]∗[µ] .
(2.3)
In other words, the composition of the two gluing maps defined on the left
hand side of (2.3) is equal to the gluing map on the right hand side.
Given a k-simplex α = [τ0 , . . . , τk ] in L′ , let αi = [τ0 , . . . , τˆi , . . . , τk ] denote
the face opposite [τi ]. Then hα∗[µ] : Mµ × α → ∂Mτ0 × α is a k-parameter
isotopy. For i 6= 0, we require:
hα∗[µ] |Mµ ×αi = hαi ∗[µ] ,
(2.4)
(hα0 ∗[τ0 ] )(hα∗[µ] |Mµ ×α0 ) = hα0 ∗[µ]
(2.5)
while for i = 0,
26
There is one further condition which our isotopies should satisfy. If τ =
min α and if µ and µ′ are two faces of τ and ρ = µ ∩ µ′ , then we require that
Im(hα∗[µ] ) ∩ Im(hα∗[µ′ ] ) = Im(hα∗[ρ] )
(2.6)
In other words, on the complement of Im(hα∗[ρ] ), the embeddings hα∗[µ] :
Mµ → Mτ and hα∗[µ′ ] : Mµ′ → Mτ must have disjoint images.
Next we describe how to glue together {Mτ × Kτ }τ ∈S(L) to obtain a space
X
` together with a projection map p : X → K. Start with the disjoint union
Mτ × Kτ . We will construct X(0) ⊂ · · · ⊂ X(k) ⊂ · · · ⊂ X(d + 1) = X so
that X(k) will be the inverse image of `
the k-skeleton of K in X. The space
X(0) is defined to be the disjoint union Mσ ×[σ]. Next, given an edge [τ, σ]
in [∅] ∗ L′ , we glue Mτ × [τ, σ] to Mσ × [σ] via h[σ]∗[τ ] : Mτ × [σ] → ∂Mσ × [σ].
(Recall that h[σ]∗[τ ] = iστ .) After doing this gluing for each edge [τ, σ], we
′
obtain X(1). Notice that if τ is a (d−1)-simplex of L, then the
τ , is the
`link, LS
disjoint union of the vertices [σ] where τ < σ and Kτ = [τ ] ∗ [σ] = [τ, σ];
hence, after building X(1) we will have glued Mτ × Kτ onto X(0). Next,
consider a 2-simplex [µ, τ, σ]. Glue Mµ × [µ, τ, σ] onto Mτ × [τ, σ] using
h[τ,σ]∗[µ] : Mµ × [τ, σ] → ∂Mτ × [τ, σ]. By (2.3) this is compatible with the
previously defined gluing map Mµ × [σ] → ∂Mσ × [σ]. Hence, the union of
the two gluing maps gives a well-defined map Mµ × [τ, σ] → (∂Mτ × [τ, σ]) ∪
(∂Mσ × [σ]) which we can use to glue Mµ × [µ, τ, σ] onto X(1).
Continue by induction. Suppose X(k) has been defined over the kskeleton of K. For each k-simplex α in L′ and µ ∈ S(L) with µ < min α, we
have hα∗µ : Mµ × α → ∂Mmin α × α. By (2.4) and (2.5), this is compatible
with previously defined gluing maps hβ∗[µ] , where β is a face of α. So, for
α = [τ0 , . . . , τk ], we get a well-defined map from Mµ × α onto the image of
(∂Mτ0 × [τ0 , . . . , τk ]) ∪ (∂Mτ1 × [τ1 , . . . , τk ]) ∪ · · · ∪ (∂Mτk × [τk ])
in X(k).
We can summarize the above as follows. The union of the hα∗[µ] , with
min α = τ , define an embedding:
Hτ µ : Mµ × ∂Kµ (τ ) → ∂Mτ × Kτ .
(2.7)
(Recall ∂Kµ (τ ) = Kτ .) Formulas (2.3), (2.4) and (2.5) imply that whenever
µ < τ < σ:
(Hστ |∂Mτ ×∂Kτ (σ) )(Hτ µ |∂Kµ (σ) ) = Hσµ .
(2.8)
27
h[σ]∗[τ ] (Mτ × [σ])
Mσ
∂Mσ
Mτ
Mσ
Mτ
h[τ ]∗[µ] (Mµ × [τ ])
∂Mτ
Mµ
Mµ
X(0)
h[σ]∗[µ] (Mµ × [σ])
X(1)
Mσ
Mτ
h[τ,σ]∗[µ] (Mµ × [τ, σ])
Mµ
X(2)
Figure 4: The first stages of our gluing procedure before thickening
Hence, if X>µ = p−1 (∂Kµ ), then the union of the Hτ µ fit together to give
a well-defined map Hµ : Mµ × ∂Kµ → X>µ , that specifies the gluing of
Mµ × ∂Kµ onto X>µ . By (2.6) Hµ is an embedding.
Remark 2.2. If each Mτ is aspherical, then X is a model for the aspherical
realization BGS(L).
Next we want to replace dual cones by dual disks. We suppose that
(a) M∅ is a point.
(b) For all maximal simplices σ in L, the manifolds Mσ all have the same
dimension n.
(c) For each τ < σ, dim Mσ − dim Mτ ≥ 2 codim(τ, σ).
The conditions in (c) are called the codimension ≥ 2 conditions. The
dimensions of the Mτ give us the data for a function c : S(L) → N as in
Subsection 2.1, defined by
c(τ ) = n − dim Mτ .
28
As in Subsection 2.1, we can use the function c to define the thick link Tµ
for each µ ∈ S(L).
Whenever µ is a codimension-one face of τ , Tµ (τ ) is a regular neighborhood of [τ ] in Sµ′ .
Moreover, Tµ (τ ) is homeomorphic to E × Dτ where E is a disk of dimension equal to the codimension of Mµ in ∂Mτ . Since the normal bundle
of Mµ in ∂Mτ is trivial, Mµ × E can be embedded as a codimension-zero
submanifold of ∂Mτ and hence, the embedding Hτ µ defined by (2.7) extends
to an embedding:
Jτ µ : Mµ × Tµ (τ ) ֒→ ∂Mτ × Dτ .
(2.9)
When codim(µ, τ ) > 1, Tµ (τ ) remains a thickening of ∂Kµ (τ ) (as well as a
thickening of Dτ ) and the embedding Hτ µ again extends to an embedding
Jτ µ : Mµ × Tµ (τ ) → ∂Mτ × Dτ . Moreover, the Jτ µ satisfy the analogous
formulas to (2.8).
Finally, we build an n-manifold with boundary M in exactly the same
fashion we constucted the space X above, namely,
a
M :=
Mτ × Dτ / ∼
(2.10)
τ ∈S(L)
where the equivalence relation ∼ is defined as before except that we use as
gluing maps the Jτ µ rather than the Hτ µ .
At this point we should explain why M is a n-manifold with boundary.
Since Dµ is a c(µ)-manifold with corners (cf. Lemma 2.1) and Mµ is a (n −
c(µ))-manifold with boundary, Mµ × Dµ is a n-manifold with corners. Each
facet is either the product of Mµ with a facet of Dµ or it has the form
∂Mµ × Dµ . If µ < τ is a codimension-one face, then Mµ × Dµ is glued onto
Mτ × Dτ by the embedding Jτ µ : Mµ × Tµ (τ ) → ∂Mτ × Dτ from a facet of
Mµ × Dµ to a facet of Mτ × Dτ . Similarly, if {[τ0 ], . . . , [τk ]} is the vertex set
of σ in L′µ , then Mµ × Dµ is glued onto Mσ × Dσ via a map from the stratum
Mµ × Tµ (σ) to an intersection of strata in ∂Mσ → Dσ .
Let {Mτ }τ ∈S(L) be a complex of manifolds with boundary over S(L). Let
Gτ = π1 (Mτ ). As usual suppose the induced homomorphisms φστ = (iστ )∗ :
Gτ → Gσ are injective and the system GS(L) = {Gτ }τ ∈S(L) is a simple
complex of groups. Let G = lim GS(L) be the direct limit. Also suppose
• GS(L) is developable, and
29
• each Mτ is homotopy equivalent to BGτ ,
Theorem 2.3. Suppose {Mτ }τ ∈S(L) is a complex of manifolds with boundary
over S(L) satisfying the above conditions and let M be the manifold with
boundary defined by (2.10). Then M is a thickening of BGS(L). If the
K(π, 1)-Question for GS(L) has a positive answer, then M ∼ BG, so
actdim G ≤ dim M.
2.3
Complexes of manifolds with boundary over S o (L)
We turn to the case where the manifolds Mv associated to the vertices v of L
are allowed to have empty boundary. When this happens our algorithm does
not tell us how to glue on the final disk M∅ × D∅ . (If ∂Mv = ∅, there is no
place to embed M∅ × T∅ .) One can still try to accomplish the gluings without
the final disk. For each higher-dimensional simplex τ , we still require ∂Mτ
to be nonempty. By excluding the empty simplex, we consider a complex of
manifolds with boundary over S o (L) where the embeddings iστ and gluing
maps Hτ µ , Jτ µ satisfy the same conditions as in the previous subsection.
We can glue together the pieces together as before; however, the resulting
manifold will usually not be a model for BG.
So, let us consider a complex of manifolds with boundary {Mτo }τ ∈S o (L) ,
where ∂Mvo is allowed to be empty for v ∈ Vert L, but ∂Mτo 6= ∅ if dim τ > 0.
The other conditions in the previous section are satisfied mutatis mutandis.
The complex is n-dimensional if dim Mσo = n for each maximal simplex σ.
As in (2.10), we can glue together the {Mτo × Dτ }τ ∈S o (L) to obtain an nmanifold with boundary M o . As before, M o will a model for BGS o (L);
however, as explained in Subsection 1.4, BGS o (L) will not be homotopy
equivalent to BG unless L is contractible (see Lemma 1.23). When L is
EDCE (see Definition 1.24), we can use the methods of Subsection 1.4 to
conclude that actdim G ≤ n.
Given the n-dimensional system {Mτo }τ ∈S o (L) over S o (L), there is an easy
way to extend it to an (n + 1)-dimensional system over S(L): simply take
the product of each manifold Mτo with the unit interval. In other words, put
Mτ = Mτo × [0, 1]. So, the vertex manifold Mvo is replaced by the manifold
with boundary Mv = Mvo × [0, 1]. Let M∅ be a point and iτ ∅ : M∅ → ∂Mτ an
appropriate embedding.
Since π1 (Mτo ) = π1 (Mτ ), the systems of fundamental groups {π1 (Mτ )}
and {π1 (Mτo )} define the same simple complex of groups GS(L) over S(L)
30
given by Gτ = π1 (Mτ ). Let G = lim GS(L). Let M o and M denote, respectively, the results of gluing together the systems {Mτo }τ ∈S o (L) and {Mτ }τ ∈S(L) .
We assume each Mτ is connected, so that π1 (M) = G. If L is simply connected, then π1 (M o ) is also equal to G. (In general, π1 (M o ) is the semidirect
product described in Remark 1.3.)
From Lemma 1.23 we get the following.
Lemma 2.4. The inclusion M o ֒→ M is a homotopy equivalence if and only
if L is contractible.
Suppose L is a full subcomplex of another simplicial complex L̄. In Subsection 1.4 we defined the notion of a trivial extension of a simple complex
of groups over S(L) to one over S(L̄). Similarly, we can define the notion of
a trivial extension of a complex of manifolds with boundary over S(L). Let
{Mτ }τ ∈S(L) be such a system. For simplicity, suppose that the dimension of
Mτ depends only on the dimension of the simplex τ , i.e., if dim Mτ = n(k)
for each k-simplex τ . Recall any simplex σ̄ can be decomposed as a join
σ̄ = σ ∗ τ , where σ ∈ S(L) and where vertices of τ lie in L̄ − L. A complex of manifolds with boundary {Nτ }τ ∈S(L̄) is called a trivial extension of
{Mτ }τ ∈S(L) if Nσ̄ = Mσ × D n(k) , whenever σ̄ = σ ∗ τ as above and D n(k) is
the disk Nτ . A trivial extension of a complex over S o (L) is defined similarly.
The next lemma follows immediately from Lemma 1.21.
Lemma 2.5. Suppose {Mσ }σ∈S(L) is a complex of manifolds and {Nσ̄ }σ̄∈S(L̄)
is a trivial extension of it over L̄. Let M and N be the result of gluing these
systems together. Then M is homotopy equivalent to N.
Proposition 2.6. Suppose L is a full subcomplex of a contractible complex Lc
of the same dimension as L. Let {Mτo }τ ∈S o (L) be an n-dimensional complex
of manifolds with boundary over S o (L) and {Nτo }τ ∈S o (Lc ) a trivial extension
of it to Lc . Let N o be the result of gluing together {Nτo } and let M and N
be the result of gluing the corresponding (n + 1)-dimensional complexes over
S(L) and S(Lc ), respectively. Then M is homotopy equivalent to N o
Proof. By Lemma 2.5, M is homotopy equivalent to N and by Lemma 2.4,
N is homotopy equivalent to N o .
For the next theorem we suppose that {Mσo } is a n-dimensional complex
of manifolds with boundary over S o (L) where the vertex manifolds Mvo are allowed to have empty boundary. Further suppose that the conditions for The31
orem 2.3 are satisfied, i.e., each Mσo is aspherical, the associated simple complex of groups GS(L) is developable and BGS(L) ∼ BG for G = lim GS(L)
(so the K(π, 1)-Question has a positive answer).
Theorem 2.7. Let {Mσo }σ∈S o (L) be an n-dimensional complex of aspherical
manifolds with boundary. Then, with hypotheses as above:
(i) actdim G ≤ n + 1.
(ii) If L is EDCE, then actdim G ≤ n.
Proof. Statement (i) follows from Theorem 2.3 applied to the (n+1)-dimensional
system {Mσ }, where Mσ = Mσo × [0, 1]. When L is EDCE, L embeds in a
contractible complex Lc of the same dimension; so, by Proposition 2.6, M is
homotopy equivalent to the n-manifold N o . Hence, actdim G ≤ n.
Remark 2.8. Most of the constructions in Subsections 2.2 and 2.3 work
if one replaces the poset of simplices S(L) by a more arbitrary poset Q.
Specifically, suppose Q is a poset with a minimum element m and that Qo =
Q − {m}. As before, one can define the notion of a posets of manifolds with
boundary over Q and Qo , respectively, and prove versions of Theorems 2.3
and 2.7.
3
3.1
Examples
Simple complexes of closed aspherical manifolds
We begin by discussing graph products. Notation is continued from Definition 1.12 and Example 1.13: L1 is a simplicial graph with vertex set
V , L is the associated
flag complex, {Gv }v∈V is a collection of groups of
Q
type FQ
, G = L1 Gv denotes their graph product and, for each σ ∈ S(L),
Gσ = v∈Vert σ Gv is the direct product. As in Example 1.13, this defines
a simple complex of groups, GS(L) = {Gσ }σ∈S(L) . By Corollary 1.16, the
K(π, 1)-Question has a positive answer for GS(L). For each v ∈ V , let Mv
be a modelQfor BGv by a manifold of minimum dimension mv = actdim Gv .
Let Mσ = v∈Vert σ Mv .
Suppose each Mv P
has nonempty boundary and has dimension at least 2.
Let mσ = dim Mσ = v∈Vert σ dim Mv and put
n = sup{mσ | σ ∈ S(L)}
32
(3.1)
By taking products with disks of suitable dimensions, we can assume that
for each maximal simplex σ of L, dim Mσ = n. It is straightforward to
give {Mσ }σ∈S(L) the structure of an n-dimensional complex of manifolds
with boundary satisfying the conditions in Subsection 2.2. Here are the
details. Choose basepoints xv ∈ ∂Mv . Whenever µ is a face of a simplex
τ ∈QS(L), put V (τ µ) = Vert τ − Vert µ and xτ µ = (xv )v∈V (τ µ) be a basepoint
in v∈V (τ µ) ∂Mv . This gives an inclusion
iτ µ : Mµ ֒→ Mµ × xτ µ ⊂ ∂Mτ
with trivial normal bundle. For a k-simplex α = [τ0 , τ1 , . . . , τk ] ∈ L′ , define
h[α∗τ0 ] : Mµ × α → ∂Mτ0 × α
by h[α∗τ0 ] (z, x) = (iτ µ (z), x). The h[α∗τ0 ] (z, x) obviously satisfy the conditions
in Subsection 2.2; so, {Mσ′ }σ∈S(L) is a complex of manifolds with boundary.
Using Theorem 2.3, we get an n-dimensional manifold M which is a model
for BG. This gives the following.
Proposition 3.1. (cf. Theorem 2.3 and Corollary 1.16). Suppose each Gv is
the fundamental
Q group of an aspherical manifold Mv with nonempty boundary. Let G = L1 Gv be the graph product. Then
actdim G ≤ sup{mσ | σ ∈ S(L)},
where mσ = dim Mσ .
We turn to the case where each Mv is a closed aspherical manifold not
equal to a point. (For example, when G is a RAAG each Mv is a circle.)
We can convert this case into the first case by the simple expedient of taking
the product of each Mv with [0, 1]. Then Mv′ = Mv × [0, 1] is a manifold
with boundary of dimension mv + 1. Proposition 3.1 then has the following
corollary.
Corollary 3.2. Suppose each G
Qv is the fundamental group of a closed aspherical manifold Mv . Let G = L1 Gv be the graph product. Then
actdim G ≤ sup{m′σ | σ ∈ S(L)},
where Mσ′ = Mσ × [0, 1]Vert σ and m′σ = dim Mσ′ = dim Mσ + dim σ + 1.
33
The argument in Proposition 3.1 and Corollary 3.2 is easily modified to
give a sharp upper bound for actdim G in the case when some Mv have empty
boundary and some do not.
When each Mv is closed and of the same dimension m, then Corollary 3.2
has the following corollary.
Corollary 3.3. Suppose each Gv is the fundamental group of a closed aspherical
Q m-manifold Mv , that the flag complex L has dimension d, and that
G = L1 Gv is the graph product.Then
actdim G ≤ (m + 1)(d + 1)
Proof. If σ is a d-simplex, then dim Mσ′ = (m + 1)(d + 1).
The arguments above for the graph product complex of fundamental
groups of closed aspherical m-manifolds can be generalized to a simple complex of fundamental groups of closed aspherical manifolds as defined below.
Definition 3.4. A simple complex of closed manifolds over S(L) is a collection of connected, closed manifolds {Mσ }σ∈S(L) and for each τ < σ a
π1 -injective embedding iστ : Mτ ֒→ Mσ as a submanifold. There are a few
more requirements:
• M∅ is a point.
• If Gσ = π1 (Mσ ) and φστ : Gτ → Gσ is the homomorphism defined by
iστ , then the system {Gσ , φστ } is a simple complex of groups GS(L).
• If τ1 , . . . , τk are faces of σ, then the Mτi intersect transversely in Mσ .
• If τ is a face of σ, then Mτ has trivial normal bundle in Mσ .
For Gσ = π1 (Mσ ), let GS(L) = {Gσ }σ∈S(L) be the associated simple complex
of groups. If each Mσ is a model for BGσ , then {Mσ }σ∈S(L) is a simple
complex of closed aspherical manifolds.
Definition 3.5. The system {Mσ , iστ } satisfies codimension-m conditions if
Mτ has codimension m in Mσ whenever τ is a codimension one face of σ.
(This implies that for any face τ of σ, codim(Mτ , Mσ ) = m codim(τ, σ).)
34
Note that since dim M∅ = 0, the codimension-m conditions imply that
dim Mσ = m(dim σ + 1).
For the graph product complex, if each vertex manifold Mv is a closed
aspherical m-manifold, then {Mσ }σ∈S(L) is a simple complex of closed aspherical manifolds satisfying the codimension-m conditions.
In the more general case of a complex of closed manifolds, follow the
same procedure as for graph products and replace each manifold Mσ by the
manifold with boundary Mσ′ = Mσ ×[0, 1]Vert σ . Then {Mσ′ }σ∈S(L) is a complex
of manifolds with boundary as in Subsection 2.2. Theorem 2.3 allows us glue
together the {Mσ′ } to get a manifold M ′ .
Suppose GS(L) is developable, that the intersection of local groups condition (1.5) holds and that L is a flag complex (so that the K(π, 1)-Question
for GS(L) has a positive answer). Then we get an upper bound for the action
dimension as follows.
Theorem 3.6. Suppose that {Mσ }σ∈S(L) is a simple complex of closed aspherical manifolds over S(L) satisfying the codimension-m conditions. Let
G = lim GS(L) be the direct limit of the π1 (Mσ ). Also suppose as above
that L is a d-dimensional flag complex and that the K(π, 1)-Question has a
positive answer for GS(L). Then
actdim(G) ≤ (m + 1)(d + 1).
In section 4, we will prove that if Hd (L, Z2 ) 6= 0, then actdim(G) =
(m + 1)(d + 1).
Intersecting submanifolds. One way to get examples of simple complexes of closed manifolds is to start with an ambient manifold M together
with a collection of connected, smooth submanifolds (e.g., hypersurfaces),
{M(i)}i∈I , so that the M(i) intersect
T transversely and so that for each subset J of I, the intersection M(J) = i∈J M(i) is nonempty. In order to make
the indexing compatible with previous notation we must replace subsets of I
by their complements. So, we will write v(i) for I − {i} and V for {v(i)}i∈I ,
the set of complements of singletons. A subset J ⊂ I corresponds to the
complementary subset σ(J) := {v(i)}i∈I−J of V . For example, in the case
of the Q
graph product complex, all
Q manifolds are subspaces of the manifold
M = v∈V Mv , while M(i) = j∈I−{i} Mv(j) . The case of a RAAG is a
further specialization: M is the T I , the torus on I, the T (i) are coordinate
35
subtori of codimension one, the Tv(i) are coordinate circles and the Tσ are
coordinate subtori.
To simplify the discussion, suppose that I = {1, . . . , p}, that dim M = mp
and that each M(i) has the same codimension m in M. Since the intersections
are transverse and nonempty, the intersection M(I) of all the M(i) is a
nonempty finite set of points. Choose one, x0 , as the basepoint. For each
σ = σ(J) ⊂ V , let Mσ be the component of M(J) containing x0 . Thus,
dim Mσ = m(dim σ + 1).
If M is locally CAT(0) and each M(i) is totally geodesic, then each
Mσ is aspherical with fundamental group Gσ := π1 (Mσ ). By uniqueness
of geodesics, for each τ < σ the inclusion Mτ → Mσ is π1 -injective. For the
same reason, the intersection of local subgroups condition (1.5) holds. Thus,
for any simplicial complex L, GS(L) is a simple complex of groups and when
L is a flag complex, the K(π, 1)-Question has a positive answer.
Example 3.7. (Complexes of hyperbolic manifolds). First we use the above
technique to construct a system√of hyperbolic manifolds satisfying the codimension1 conditions. Suppose K = Q( d) is a totally real quadratic extension of Q
and A is the ring of algebraic integers in K. Choose ε ∈ A so that ε > 0 and
ε < 0. Define a quadratic form ϕ : Ap+1 × Ap+1 → A by
(
δij , if (i, j) 6= (0, 0),
ϕ(ei , ej ) =
−ε, if (i, j) = (0, 0),
where {ei }0≤i≤p is the standard basis. Let O(ϕ) ⊂ GLp+1 (A) be the subgroup
which preserves the quadratic form ϕ. The signature of ϕ on Ap+1 ⊗A R ∼
=
p+1
R
is (p, 1); so, over R, the group of isometries of ϕ is identified with
O(p, 1) and O(ϕ) is a uniform lattice in O(p, 1). Let Γ be a normal torsionfree subgroup of O(ϕ) (for example, for almost any prime ideal, we could take
Γ to be the corresponding congruence subgroup of√O(ϕ)). Then M p = Hp /Γ
is a closed hyperbolic manifold. The image of ( ε, 0, . . . , 0) in M p is the
basepoint. For 1 ≤ i ≤ p, let ri ∈ O(ϕ) be the reflection which sends ei
to −ei . There is an induced involution of r i on M p . Its fixed point set
is a totally geodesic submanifold of codimension one and the component
containing the basepoint is the manifold M(i). If we require Γ to be a
subgroup of SO(ϕ), then M(i) will be orientable and have trivial normal
bundle in M. For any flag complex L with vertex set in V we then get
a simple complex of closed hyperbolic manifolds {Mσ }S(L) . This system
36
behaves similarly to a complex of coordinate tori in Tp . It is easy to modify
this construction to get a subsystem of {Mσ } satisfying the codimension-m
conditions. Let p = pm. Group the first m commuting involutions together
to get an involution s1 = r1 r2 . . . rm . Continue in this fashion, defining
sj = rm(j−1)+1 · · · rmj . The fixed set of sj on M is a hyperbolic submanifold
M(j) of codimension-m. If L is a flag complex with p vertices we can use the
intersections of the M(j) as above to get a system of hyperbolic manifolds
satisfying codimension-m
conditions.
S
We have that Mσ is a model for BGS(L). The union is a piecewise
hyperbolic space. Since L is a flag complex, its all right, piecewise spherical
metric is CAT(1). Since all intersections in M were orthogonal, the relevant
S
links of the Mσ in the union are full subcomplexes of L. It follows that Mσ
is a CAT(−1) space. Hence, G = lim GS(L) is word hyperbolic.
3.2
Models for spherical Artin groups
Suppose L is a nerve of a Coxeter system (W, S) and let AS(L) be the
associated Artin complex as in Example 1.9. For any simplex σ of L, there is
a corresponding spherical Coxeter group Wσ and a spherical Artin group Aσ .
In this subsection we construct a model for BAσ by a manifold with boundary.
The group Wσ is an orthogonal linear reflection group on Rd(σ)+1 , where
d(σ) = dim σ. By complexification, Wσ acts on Cd(σ)+1 . If Aσ denotes the
arrangement of reflecting hyperplanes in Cd(σ)+1 , then by Deligne’s Theorem
in [20], the arrangement complement M(Aσ ) is a manifold model for the
classifying space of the pure Artin group P Aσ (where P Aσ denotes the kernel
of Aσ → Wσ ). Hence, M(Aσ )/Wσ is a manifold model for BAσ . Similarly,
if S(Aσ ) := S 2d(σ)+1 ∩ M(Aσ ), where S 2d(σ)+1 denotes the unit sphere in
Cd(σ)+1 , then S(Aσ )/Wσ is a model for BAσ by a manifold of dimension
2d(σ) + 1. Our actual approach will be to define a certain Wσ -invariant
bordification of S(Aσ ) (which we will denote by the same symbol) and then
use Nσ := S(Aσ )/Wσ .
Hyperplane arrangements. A hyperplane arrangement A is a finite
T collection of affine hyperplanes in Cn . The arrangement is central if H∈A H
is nonempty. Its rank, rk(A) is the maximum codimension of any nonempty
intersection of hyperplanes in A. An arrangement is essential if its rank
is n. A subspace of A is either the ambient space Cn or a nonempty intersection of hyperplanes. The set of subspaces of A, partially ordered by
37
N?
reverse inclusion, is denoted Q(A) and is called the intersection poset. So, if
F, E ∈ Q(A), then F < E ⇐⇒ F ⊃ E.
′
′′
Suppose A′ and A′′ are arrangements in Cn and Cn , respectively. Define
′
′′
A′ × A′′ to be the arrangement in Cn × Cn consisting of all hyperplanes
′′
′
of the form H ′ × Cn or Cn × H ′′ for H ′ ∈ A, H ′′ ∈ A′′ . An arrangement
A is reducible if it is isomorphic to one which admits a nontrivial product
decomposition as above. Otherwise, it is irreducible. Note that the product
of two central arrangements is central.
The codimension c(E) of a subspace E in Q(A) is the complex dimension of a complementary subspace E ⊥ ⊂ Cn . Given E ∈ Q(A) its normal
arrangement AE is defined by
AE := {H | H ∈ A and H ≤ E}.
(Often we will identify AE with the essential, central arrangement in E ⊥
obtained by intersecting the hyperplanes with the orthogonal complement
E ⊥ of E in Cn .) There is also an arrangement AE in E, called the restriction
of A to E, defined by
AE := {H ∩ E | H ∩ E is a hyperplane in E}.
Arrangement complements and their bordifications. The arrangement complement is
[
M(A) := Cn −
A.
Suppose A is an essential, central arrangement in Cn with {0} the maximum
element of Q(A).
Put
S(A) := S 2n−1 ∩ M(A) and D(A) := D 2n ∩ M(A),
where S 2n−1 and D 2n denote the unit sphere and disk in Cn . Next we want
to attach boundaries to each of these manifolds to obtain a manifold with
corners. The idea is to remove a tubular neighborhood of each E ∈ Q(A),
starting with the E of smallest dimension. There is a canonical way to do
this which we describe below.
Suppose V → E is a vector bundle over a manifold E. If s : E → V
denotes the 0-section, define the associated sphere bundle S(V ) and cylinder
bundle C(V ) by
S(V ) := (V − s(E))/R+ ,
and C(V ) := (V − s(E)) ×R+ [0, ∞),
38
where R+ is the group of positive real numbers. So, if V → E has fiber Rm ,
then the fiber of C(V ) → E is the cylinder S m−1 × [0, ∞). Note that C(V ) is
a manifold with boundary with interior V − s(E). Next, suppose that X is
a manifold, that E ⊂ X is a submanifold and that VE is the normal bundle.
Let f : VE → X be a tubular neighborhood. Define X ⊙ E := C(V ) ∪ X − E
where C(V ) is glued onto X − E via the restriction of the tubular map to
the open subset VE − s(E). The manifold with boundary X ⊙ E is called the
blowup of X along E; it is formed from X − E by adding the sphere bundle
S(VE ) as boundary.
Next we define a bordification of M(A), called the blowup of Cn along
A. Start with Cn and then blowup the subspaces E of minimum dimension
to obtain a manifold with boundary. Each element of F ∈ Q(A)<E is also
blown up to a submanifold with boundary. We continue by blowing up
subspaces of increasing dimension to obtain a manifold with corners, which
we continue to denote by M(A). (A similar procedure is described in [11, Ch.
IV].) Bordifications of S(A) and D(A) are defined in the same fashion. The
boundary of M(A) is a union of manifolds with boundary, ∂E M(A), indexed
by the proper subspaces E ∈ Q(A), where ∂E M(A) means the part of the
boundary which results from blowing up E, i.e.,
∂E M(A) = S(AE ) × Ê,
(3.2)
where S(AE ) is the blowup of the normal arrangement in the unit sphere
of E ⊥ and Ê = M(AE ) is the blowup of E along AE . (Note that the right
hand side of (3.2) is a product since the normal bundle of E in Cn is trivial.)
Thus, S(AE ) is a submanifold of ∂M(A) and for each F < E, S(AF ) ∩ E is
also a submanifold of ∂F (Ê).
If the central arrangement A decomposes as A = A1 × · · · × Ak , then
there is an obvious homotopy equivalence S(A) ∼ S(A1) × · · · × S(Ak ).
Reflection arrangements. Now suppose (Wσ , Sσ ) is a spherical Coxeter
system, where Sσ = Vert(σ). Let ARσ be the associated real hyperplane
arrangement in Rd(σ)+1 and Aσ its complexification. The intersection posets
Q(ARσ ) and Q(A) are canonically identified. As usual, S(σ) denotes the
face poset of σ. Since σ is the dual of the fundamental simplex for Wσ on
S d(σ)+1 , each face τ ≤ σ is dual to a face which corresponds to a subspace
E R (τ ) ⊆ Rd(σ)+1 or equally well to E(τ ) ⊆ Cd(σ)+1 . This gives an orderpreserving injection τ 7→ E(τ ) from S(σ) to Q(Aσ ). The subspace E(τ ) is the
39
subspace of Cd(σ)+1 fixed by Wτ ; hence, Wτ acts as a reflection group on the
normal space E(τ )⊥ . By Deligne’s Theorem, S(Aσ )/Wσ is a model for BAσ
by a manifold with boundary of dimension 2d(σ) + 1. Hence, actdim Aσ ≤
2d(σ) + 1. When (Wσ , Sσ ) is reducible this estimate can be improved. So,
suppose its irreducible components are (Wi , Si ) = (Wσi , Sσi ), where 1 ≤ i ≤
k. Put Ai = Aσi and ni = dim σi + 1. Then W = W1 × · · · × Wk and
S(A1 )/W1 × · · · × S(Ak )/Wk is a manifold model for BAσ . This gives the
following.
Proposition 3.8. Suppose the decomposition of a spherical Coxeter group
Wσ into irreducibles is given by Wσ = W1 × · · · × P
Wk . Then the spherical
Artin group Aσ has action dimension ≤ 2n − k (= 2d(σi ) + 1).
The Artin complex. Consider an Artin complex AS(L) for which the
K(π, 1)-Question has a positive answer. Suppose the dimension of L is d. As
explained in Example 1.9, since the Salvetti complex for A is of dimension
d + 1, gdim A = d + 1. So, on general principles, actdim A ≤ 2d + 2. As we
explained below, one can use the gluing technique of Subsection 2.2 to obtain
the same estimate. To this end define a complex of aspherical manifolds with
boundary {Mσ }σ∈S(L) by putting
Mσ := M(Aσ )/Wσ .
(3.3)
It is a manifold with boundary of dimension 2d(σ) + 2 and a model for
BAσ . Next we need to define embeddings iστ : Mτ ֒→ ∂Mσ , whenever
τ < σ. For a fixed σ, Aσ is an arrangement in Cd(σ)+1 . If τ < σ, then
E(τ ) is a subspace of Aσ and its normal arrangement AE(τ ) (∼
= Aτ ) is an
⊥
arrangement in E(τ ) . Choose a basepoint x ∈ Ê(σ) and identify S(AE(τ ) )
with S(AE(τ ) ) × x ⊂ ∂E(τ ) S(Aσ ). (S(AE(τ ) ) is the fiber of the sphere bundle
of the normal bundle of S(Ê(τ )) in S(Aσ ).) Hence, we get an embedding
S(AE(τ ) ) ֒→ ∂S(Aσ ). Since M(Aτ ) = S(AE(τ ) ) × [0, ∞) and M(Aσ ) =
S(Aσ ) × [0, ∞) by taking the product with the identity map on [0, ∞), we
get the embedding M(Aτ ) ֒→ ∂M(Aσ ). Taking quotients by Wσ and Wτ (=
the stabilizer of E(τ ) in Wσ ), this induces an embedding iστ : Mτ ֒→ ∂Mσ .
Here a small remark is needed: if Wτ and Wτ ′ are conjugate subgroups of Wσ ,
then, as defined, Mτ and Mτ ′ are the same submanifold of ∂Mσ ; however,
if τ 6= τ ′ , we want their images to be disjoint. This is easily arranged by
picking a different point x′ ∈ Ê(τ ′ ) for τ ′ , so that Mτ and Mτ′ will be parallel
40
submanifolds in ∂Mσ . It is then straightforward to define the dual disk Dτ
to Mτ and isotopies as in (2.9) so that {Mτ }τ ∈S(L) becomes a complex of
manifolds with boundary. Applying Theorem 2.3 we get a manifold with
boundary M of dimension 2d + 2 which is a model for BA. This gives an
alternate proof for the following.
Proposition 3.9. Suppose the K(π, 1)-Question has a positive answer for
AS(L). Let A = lim AS(L) and d = dim L. Then actdim A ≤ 2d + 2.
Permutohedra. When L is EDCE and the manifolds over the vertices of
L are closed, it is necessary to use a trick with permutohedra in order to
apply the method of Subsection 2.3 to get sharp upper bounds for the action
dimensions. This trick is already needed in the case of Artin groups, indeed
for RAAGs. In the case of a RAAG, AL , our complex of groups is given by the
fundamental groups of a complex of tori {Mτ }τ ∈S(L) , where Mτ = (S 1 )Vert τ .
In Corollary 3.3 we produced a manifold model for AL of dimension 2d + 2
(where d = dim L) by using as a complex of manifolds with boundary Mτ′ :=
Mτ × I Vert τ . When L is EDCE the dimension can be decreased by one using
the complex of manifolds with boundary, Nτ = Mτ × P (τ ) where P (τ ) is a
certain d(τ )-dimensional polytope called a “permutohedron”. We shall see
in Subsection 3.3 below that the same trick works for any complex of closed
aspherical manifolds.
Definition 3.10. Suppose σ is a d-simplex. The permutohedron on σ, denoted by P (σ), is the d-dimensional convex simple polytope obtained by
truncating the proper nonempty faces of σ. The facets of P (σ) are indexed
by the elements of the interval (∅, σ) in S(σ). (A facet is a face of codimension
one). Denote the facet corresponding to τ by ∂τ P (σ). Alternatively, P (σ)
can be defined by blowing up the proper faces of σ by a procedure similar to
that described in the paragraph on bordifications in Subsection 3.2.
Whenever τ < σ, there is a natural inclusion iστ : P (τ ) → ∂τ P (σ).
Permutohedra arise naturally in blowups of hyperplane complements.
Example 3.11. Suppose A denotes the coordinate hyperplane arrangement
in Cd+1 , defined by zi = 0, 0 ≤ i ≤ d. Let ∆ be the spherical d-simplex
defined by ∆ = [0, ∞)d+1 ∩ S d and define p : S 2d+1 → ∆ by (z0 , . . . , zk ) 7→
(|z1 |2 , . . . , |zk |2 ). As in subsection 3.2, let S(A) denote the blowup of the
coordinate hyperplane arrangement in the unit sphere of Cd+1 . The map p
41
induces a map p̂ : S(A) → P (∆) which is the projection map of a trivial
bundle with fiber T d+1 . Hence, the coordinate hyperplane complement S(A)
is diffeomorphic as a manifold with corners to T d+1 × P (∆).
Example 3.12. More generally, suppose A is a central hyperplane arrangement in Cn and that A = A0 × · · · × Ak is its decomposition into irreducible
components. Since M(A) ∼
= M(A0 ) × · · · × M(Ak ), we see that S(A) is homotopy equivalent to S(A0 )×· · ·×S(Ak ). In fact, by using a permutahedron
we get a corresponding diffeomorphism of manifolds with corners,
∼
=
S(A) −→ [S(A0 ) × · · · × S(Ak )] × P (∆),
(3.4)
where P (∆) is a permutohedron of dimension k. If Ai is an arrangement in
Cn(i) , then Cn = Cn(0) ×· · ·×Cn(k) . A vector in Cn is given by (z(0), . . . , z(k)),
where z(i) ∈ Cn(i) . Define p : S 2n−1 → ∆ by (z(0), . . . , z(k)) 7→ (|z(0)|2 , . . . , |z(k)|2 ).
There is an induced map on blowups, p̂ : S(A) → P (∆). Using p̂ and the
various projections we get (3.4). For example, suppose Wσ is a spherical
Coxeter group and that Wσ = W0 × · · · × Wk is its decomposition into irreducibles. As in the paragraph above on reflection arrangements, let Ai be
the reflection arrangement corresponding to Wi and Aσ the arrangement for
W . As in (3.3), put Mσ = S(Aσ )/Wσ and Mi = S(Ai )/Wi . By (3.4) we have
a diffeomorphism of manifolds with corners:
∼
=
Mσ −→ [M0 × · · · × Mk ] × P (∆).
3.3
L is EDCE
When L is EDCE (cf. Definition 1.24) we can decrease by 1 our estimates
of upper bounds for the action dimension in Propositions 3.1 and 3.9, Theorem 3.6 and Corollaries 3.2 and 3.3. In each case we apply Theorem 2.7 of
Subsection 2.3. We assume throughout this subsection that dim L = d.
We first consider the Artin group case. When L is EDCE we can improve
the upper bound in Proposition 3.9 to get the following result, first proved
in [30].
Proposition 3.13. Suppose the K(π, 1)-Question has a positive answer for
AS(L). If L is EDCE, then actdim A ≤ 2d + 1 (= 2 gdim A − 1).
Sketch of Proof. To prove this we essentially use the same complex of manifolds with boundary as in the paragraph on the Artin complex in Subsection 3.2, except that M(Aσ )/Wσ is replaced by S(Aσ )/Wσ . So, we start with
42
the complex of aspherical manifolds {Mσo }σ∈S o (L) , where Mσo = S(Aσ )/Wσ .
By taking products with other disks, we can arrange that for all maximal
simplices σ, each Mσo has dimension 2d + 1. Next embed L in a contractible
simplicial complex Lc of the same dimension d. As in Proposition 2.6, there
is a trivial extension of this to a complex of manifolds with boundary over
S o (Lc ). Finally, apply Theorem 2.7 to get the result.
Essentially the same argument gives the following improvement of Theorem 3.6.
Theorem 3.14. Suppose that {Mσ }σ∈S(L) is a simple complex of closed aspherical manifolds over S(L) satisfying the codimension-m conditions. Let
G = lim GS(L) be the direct limit of the π1 (Mσ ). Also suppose as before that
L is a d-dimensional flag complex and that the K(π, 1)-Question for GS(L)
has a positive answer. If L is EDCE, then
actdim(G) ≤ (m + 1)(d + 1) − 1.
Sketch of Proof. For each σ ∈ S o (L), let Nσo = Mσ × P (σ), where P (σ) is
a permutohedron (cf. Definition 3.10). By taking products with other disks
we can arrange that for all maximal simplices σ the dimensions of the Nσo are
equal. Then the complex of manifolds with boundary {Nτo }τ ∈S o (L) is a complex of manifolds with boundary satisfying the conditions in Subsection 2.3.
The proof is finished exactly as in the proof of the previous proposition: embed L in a contractible simplicial complex Lc of the same dimension d; there
is a trivial extension of {Nσo } to a complex of manifolds with boundary over
S o (Lc ); then use Theorem 2.7 to complete the proof.
For the graph product complex of fundamental groups, Theorem 3.14 has
the following corollary (cf. Corollary 3.3), which gives part (ii) of Theorem
B in the Introduction.
Corollary 3.15. Suppose L is a flag complex, that for each v ∈ Vert L,
Gv isQthe fundamental group of a closed aspherical m-manifold Mv and that
G = L1 Gv is the graph product. If L is EDCE, then
actdim G ≤ (m + 1)(d + 1) − 1
The case of a RAAG is where m = 1. Then, it is proved in [1] that for
d 6= 2, the EDCE condition, can be replaced with the weaker condition that
Hd (L, Z2 ) = 0.
43
Remark 3.16. On the other hand, Corollaries 3.3 and 3.15 have advantages
over the results of [1]. For example, suppose dim L = d and that G is
the graph product over L1 of free abelian groups of rank m. Then gdim G =
m(d+1). The group G also is the RAAG associated to the simplicial complex
L̄, which is the polyhedral join over L of (m − 1)-simplices. (The notion of
a “polyhedral join” is defined in Definition 4.6 below.) So, dim L̄ = m(d +
1) − 1. Corollary 3.3 yields, actdim G ≤ (m + 1)(d + 1), while [1] only gives,
actdim G ≤ 2 dim L̄ + 2 = 2m(d + 1). (When L is EDCE both estimates can
be improved by 1: Corollary 3.15 gives actdim G ≤ (m + 1)(d + 1) − 1 while
[1] gives actdim G ≤ 2m(d + 1) − 1).
4
Obstructors
˜
The ordered 2-point configuration space C(X)
of a space X is the space of
ordered pairs of distinct points in X, i.e.,
˜
C(X)
= (X × X) − D,
where D denotes the diagonal (in many situations it will be better to remove
a neighborhood of the diagonal). The 2-point configuration space C(X) is
the space of unordered pairs of distinct points:
˜
C(X) := C(X)/Z
2,
˜
where Z2 acts by switching the factors. The double cover C(X)
→ C(X)
∞
is classified by a map c : C(X) → RP = BZ2 . Let w1 denote the nontrivial element of H 1 (RP ∞ , Z2 ), so that (w1 )n is the nontrivial element of
H n (RP ∞ , Z2 ).
Definition 4.1. The Z2 -valued van Kampen obstruction for X in dimension
n, vkn (X) ∈ H n (C(X); Z2 ), is defined by vkn (X) = c∗ (w1 )n .
If vkn (X) 6= 0, then we say X is an n-obstructor.
Remark. One can also define a Z-valued van Kampen obstruction by replacing w1 by the nontrivial element of H 1 (RP ∞ ; Z− ) where Z− means twisted
integer coefficients (cf. [1]). We will not use this refinement in this paper.
44
Definition 4.2. A map f : X → Y between metric spaces is a coarse
embedding if there exist two nondecreasing functions ρ+ , ρ− : R+ → R+ such
that limt→∞ ρ− (t) = ∞ and
ρ− (d(x, y)) ≤ d(f (x), f (y)) ≤ ρ+ (d(x, y)).
Given a finite simplicial complex K, let
Cone∞ K := K × [0, ∞)/K × {0}
denote the cone of infinite radius on K. Equip Cone∞ K with a proper metric
so that for each pair of disjoint simplices σ and τ , the distance between
σ × [t, ∞) and τ × [t, ∞) goes to infinity as t → ∞.
Suppose, for simplicity, that the group G is of type F . The method of [4]
consists of the following two steps.
(1) Find a coarse embedding Cone∞ K → EG for a suitable complex K.
(2) Compute the van Kampen obstruction of K in degree n. (By the Linking Lemma of [4, p. 223], this is an obstruction to coarsely embedding
Cone∞ K in a contractible (n + 1)-manifold).
Putting this together, we have the following definition from [4].
Definition 4.3. The obstructor dimension of G, denoted obdim G, is the
maximal n + 2 such that there exists a complex K with nonzero van Kampen
obstruction in degree n and a coarse embedding of Cone∞ (K) → EG.
Remarks 4.4. (i) Since Bestvina-Kapovich-Kleiner [4] want to define obstructors for a general group G, instead of using (1) they consider proper,
expanding, Lipschitz maps from the 0-skeleton of Cone∞ K to G. When
G acts cocompactly on EG this amounts to finding a coarse embedding
Cone∞ (K) → EG.
(ii) For a finite complex K, the cohomology class vkn (K) is the classical obstruction for finding a PL embedding of K into Rn . Indeed, if K
embeds in Rn , then C(K) embeds in C(Rn ) ∼ RP n−1, i.e., C(K) classifies
into RP n−1 and hence, vkn (K) = c∗ (w1 )n = 0. In the converse direction,
suppose F : K → Rn is a PL map in general position and that Σ is a mod
two n-cycle in C(K). Then the result of evaluating the cohomology class
on the cycle, hvkn (K), Σi, counts the self-intersections of F which lie in Σ.
45
(This is explained in [4, §2.1].) In Subsection 4.4 we will use this method
of calculating self-intersections of general position maps to prove certain van
Kampen obstructions are nonzero.
(iii) If G is a word-hyperbolic group or if EG is a CAT(0)-space, then EG
has a Z-set compactification EG. Put ∂G := EG − EG. If a finite simplicial
complex K is a subspace of ∂G, we get a coarse embedding Cone∞ K → EG
by choosing a basepoint in EG and coning off K.
The coarse van Kampen obstruction. A generalization of (1) is used by
Yoon in [36]. He considers instead coarse embeddings f : T → EG for some
contractible CW complex T with a proper metric into EG. We are assuming
BG is a finite complex and that EG has a length metric induced from a
length metric on BG. As in [4], Yoon considers the van Kampen obstruction
for embedding T in a contractible (n+1)-manifold. The 2-point configuration
space of a contractible (n + 1)-manifold W is homotopy equivalent to RP n .
˜ ) ⊂ C(W
˜ ) and hence, c : C(T ) → RP ∞ factors
If T embeds in W , then C(T
n+1
n
through RP . So, if vk (T ) 6= 0, then T does not embed in a contractible
(n + 1)-manifold. The following lemma of [4, Lemma 8] is important, at least
psychologically, to understanding the case T = Cone∞ K.
Lemma 4.5. (The Cone Lemma of [4]). If K is an n-obstructor, then
Cone K is a (n + 1)-obstructor, i.e., vkn+1 (Cone K) = vkn (K).
One actually needs to show the stronger result that T does not coarsely
embed in any contractible (n + 1)-manifold. For T = Cone∞ K this is proved
in [4] by using the Linking Lemma. In the case of a more general contractible
complex T , Yoon considers the “deleted configuration spaces” C˜r (T ) := [(T ×
T )−Nr (D)], where Nr (D) means a Z2 -stable r-neighborhood of the diagonal.
One then needs to show that the van Kampen obstruction remains nonzero
on the quotients, Cr (T ) := C˜r (T )/Z2 . So, we define the coarse van Kampen
obstruction to be the image of vkn+1 (T ) in limr→∞ H n+1 (Cr (T ); Z2 ). We
define the proper obstructor dimension of G, denoted pobdim(G), to be the
maximal n+1 such that there is a coarse embedding of a contractible complex
T into EG such that the coarse van Kampen obstruction of T in dimension
n is nonzero. Yoon shows that pobdim(G) ≤ actdim(G).
46
4.1
Configurations of subgroups and sheets
Definition 4.6. Suppose {Xv }v∈V is a collection of spaces indexed by the
vertex set V of a simplicial complex L. For σ ∈ S o (L), let X(σ) denote the
join v∈Vert σ Xv . Define the polyhedral join over L of the {Xv }v∈V by
[
X(σ).
(4.1)
L Xv :=
∗
∗
σ∈S o (L)
As in Subsection 1.1, suppose that GQ = {Gσ }σ∈Q is a developable complex of groups with |Q| simply connected, that BGQ is its aspherical realization and that G = π1 (BGQ) = lim GQ. Put
[
G :=
Gσ
(4.2)
σ∈Q
and call it a configuration of standard subgroups.
Suppose each finite subset
T
{σ1 , . . . σk } has a greatest lower bound, σi . With the word metrics, each
inclusion Gσ ֒→ G is a coarse embedding. Furthermore, given any two subgroups H1 and H2 of G, the coarse intersection of H1 and H2 in G is coarsely
equivalent to their actual intersection
T H1 ∩ H2 (see [32] for precise definitions
of the coarse intersection). Since Gσi = GT σi , this implies the inclusion of
the configuration G into G is also a coarse embedding.
Let EGQ denote the universal cover of BGQ. For each σ, choose a
basepoint b′σ ∈ BGσ and a path connecting it to the basepoint b′ ∈ BGQ
(BGσ is a subcomplex of BGQ). Choose a lift of b′ to a basepoint b ∈ EGQ.
The path from b′ to b′σ lifts to a path from b to a point bσ . Let EGσ denote
the component of the inverse image of BGσ in EGQ containing the basepoint
bσ (so, EGσ is a copy of the universal cover of BGσ ). Put
[
EG :=
EGσ
(4.3)
σ∈Q
and call it a standard configuration of sheets in EGQ. Identify the Gσ -orbit
of bσ with the group Gσ . If each Gσ is type F , then EG is quasi-isometric
to the union of orbits Gb. Hence, EG ֒→ EGQ also is a coarse embedding in
the type F case.
Example 4.7. (RAAGs). As in Example 1.9, suppose AL is the RAAG
associated to a flag complex L and that AS(L) is the Artin complex. For
47
each σ ∈ S(L), let Rσ denote the Euclidean space with basis Vert σ, let
Zσ ⊂ Rσ be the integer lattice and let T σ = Rσ /Zσ be the torus. The local
group Aσ is Zσ . The spaces BAσ and EAσ are identified with T σ and Rσ ,
respectively. Let d(σ) = dim σ. The octahedron on σ, denoted Oσ, is the
(d(σ) + 1)-fold join of S 0 ’s, where the copies of S 0 are indexed by Vert σ. So,
each standard sheet Rσ is identified with Cone∞ Oσ and the configuration of
standard sheets EG is identified with Cone∞ OL, where, as in Definition 4.6,
OL denotes the polyhedral join of 0-spheres, i.e.,
[
OL := L S 0 =
Oσ.
(4.4)
∗
σ∈S o (L)
Sometimes we shall consider finer configurations of subgroups. Suppose
each local group Gσ contains a simple complex of groups Q⊘ (σ), where the
poset Q⊘ (σ) has geometric realization homeomorphic to that of Q≤σ . Moreover, the union of these posets will be the poset of simplices in a simplicial
complex Q⊘ with the same geometric realization as Q. For each α ∈ QS
⊘ (σ),
the corresponding local group Hα is a subgroup of Gσ . S
Then H(σ) := Hα
is a configuration of subgroups in Gσ and EH(σ) := EHα is a configuration of sheets in EGσ . Taking
S the union over all σ ∈ Q, we get a configuration
S of subgroups H := H(σ) ⊂ G and a configuration of sheets
EH := EH(σ) ⊂ EG. The archetype is the case of a general Artin group
A considered in [15]. This is explained in the next example.
Example 4.8. (The configuration of standard free abelian subgroups in an
Artin group). As in Example 1.9, let AS(L) denote the Artin complex associated to a Coxeter system (W, S) with nerve L. Let AL be the associated
Artin group. For each σ ∈ S(L), Wσ is the corresponding spherical Coxeter
group and Dσ is its Coxeter diagram. As explained in [15], there is a subdivision σ⊘ of σ whose vertices correspond to the connected subdiagrams of
Dσ . In other words the vertices of σ⊘ are the irreducible special subgroups of
Wσ . Corresponding to each vertex we have the infinite cyclic group generated
by the element ∆2σ in the pure spherical Artin group P Aσ . So, the group
corresponding to a vertex s of L is the square of the corresponding Artin
generator xs . The simplices of σ⊘ index the standard free abelian subgroups
of Aσ . These subdivisions of simplices fit together to give a subdivision L⊘
of L. For example, the edge {s, t} of L is subdivided into two edges exactly
when 3 ≤ mst < ∞. This leads to the configuration of standard free abelian
48
subgroups in AL and a configuration of standard flats in AS(L):
[
[
Zα ⊂ AL and
Rα ⊂ EAS(L).
α∈S(L⊘ )
α∈S(L⊘ )
Moreover, the configuration of standard flats is isometric to Cone∞ OL⊘ ,
where, as in (4.4), OL⊘ , is a polyhedral join of 0-spheres. So, when the
K(π, 1)-Question for AS(L) has a positive answer, Cone∞ OL⊘ is coarsely
embedded in EAL .
It follows from [15] that the inclusion of standard infinite subgroups corresponding to the vertices of L⊘ defines a homomorphism Φ : AL⊘ → AL
from the RAAG AL⊘ to the Artin group AL . This leads us to the following.
Conjecture 4.9. The homomorphism Φ : AL⊘ → AL is injective.
Conjecture 4.9 is related to a conjecture of Tits which was proved by
proved by Paris and Crisp in [10]. Let Γ denote the subgraph of L1 consisting
of the edges labeled 2. Let AΓ be the RAAG defined by Γ. There is a
homomorphism AΓ → AL which sends each generator for AΓ to the square
of the corresponding generator for AL . Since the flag complex determined by
Γ is a full subcomplex of L⊘ , the Tits Conjecture amounts to the conjecture
that the homomorphism Φ in Conjecture 4.9 is injective. Conjecture 4.9 is
still open even for spherical Artin groups. For example, if A is the braid group
B4 , the Tits Conjecture states that there is an injective homomorphism from
Z2 ∗ Z into B4 , whereas our conjecture predicts an injective homomorphism
from the RAAG, ACone(C5 ) to B4 , where C5 is a five-cycle.
There are similar configurations of abelian subgroups for affine hyperplane
complements (cf. section 5). The configurations of free abelian and nilpotent
groups in [3] as well as the configurations of free abelian subgroups generated
by Dehn twists and“Mess subgroups” in the mapping class group in [23]
follow similar lines.
4.2
The complex Om L
Given a d-dimensional flag complex L, let Om L denote the polyhedral join
of (m − 1)-spheres, as in Definition 4.6, i.e.,
[
Om L := L S m−1 :=
Om σ,
∗
σ∈S o (L)
49
where Om σ denotes the (d(σ) + 1)-fold join of (m − 1)-spheres, Sv , indexed
by Vert σ (eventually each of these (m − 1)-spheres will be given a simplicial
structure). Thus, Om σ is a sphere of dimension m(d(σ) + 1) − 1. So, the
dimension of Om L is m(d + 1) − 1. We denote this dimension by δm (L) (or
simply by δ):
δ = δm (L) := dim Om L = m(d + 1) − 1
(4.5)
Let OL = O1 L. In [1] the van Kampen obstruction of OL was computed in
many cases.
Theorem 4.10. ([1]). Let L be any d-dimensional flag complex. If Hd (L; Z2 ) 6=
0, then vk2d (OL) 6= 0. Therefore, actdim AL = 2 gdim AL = 2d + 2.
The idea for the proof of this in [1] was to construct a specific 2d-cycle
Ω in the 2-point configuration space C(OL) so that vk2d (OL) evaluates nontrivially on Ω. In the following subsections, we will generalize this to Om L.
The main result in this section is the following.
Theorem 4.11. Suppose L is a d-dimensional flag complex and that δ =
dim Om L = m(d + 1) − 1. If Hd (L; Z2 ) 6= 0, then Om L is a (δ + d)-obstructor.
The proof of Theorem 4.11 will occupy Subsection 4.5. A similar result
holds for polyhedral joins of spheres over L when the spheres are allowed be
of different dimensions, and the proof is essentially the same as the proof of
Theorem 4.11. Note that δ + d = m(d + 1) − 1 + d = (m + 1)(d + 1) − 2.
Before proving Theorem 4.11, we note in the following proposition that
the vanishing of the van Kampen obstruction of Om L in degrees higher than
δ + d follows from the gluing constructions as in Corollaries 3.3 and 3.15.
Proposition 4.12. Suppose L is a d-dimensional flag complex and that δ =
dim Om L.
(i) vkδ+d+1 (Om L) = 0.
(ii) If L is EDCE, then vkδ+d (Om L) = 0.
Proof. For each v ∈ Vert L, choose
Q a closed, nonpositively curved m-manifold
Mv . Put Gv = π1 (Mv ) and let L Gv be the graph product of the {Gv }. As
before, BGS(L) is the polyhedral product of the Mv . Then Cone∞ Om L can
be identified with the configuration of standard sheets in EG. By Proposition 3.1, BG thickens to a manifold of dimension (m + 1)(d + 1) = δ + d + 2.
50
So, Cone∞ Om L coarsely embeds into a contractible (δ + d + 2)-manifold
and hence, vkn (Om L) = 0 for n ≥ δ + d + 1, giving statement (i). Similarly, if L is EDCE, then BG thickens to a (δ + d + 1)-manifold, so that
vkδ+d (Om L) = 0.
Remark 4.13. One can give a different argument for Proposition 4.12 (i)
by showing directly that Om L has an embedding of codimension (d + 1) in
Euclidean space and hence, that vkδ+d+1 (Om L) = 0.
4.3
Sheets of contractible manifolds
As in Definitions 3.4 and 3.5, suppose {Mσ }σ∈S(L) is a simple complex of
closed aspherical manifolds over S(L) satisfying the codimension-m conditions. Let GS(L) be the associated simple complex of groups. Put G =
lim GS(L). Assume the K(π, 1)-Question has a positive answer for GS(L).
fσ be the copy of
Each Mσ is a subspace of BG (= BGS(L)). Let M
S
fσ
its universal cover containing a given basepoint b ∈ EG. Let EG = M
fσ is CAT(0) and if its visual boundary,
be the union of sheets. If each M
fσ , is homeomorphic to the sphere Om σ, then M
fσ is homeomorphic to
∂∞ M
Cone∞ (Om σ) = Rm(d(σ)+1) . Hence, EG is homeomorphic to Cone∞ (Om L).
fτ intersect transversely; so that for
This uses the assumption that the M
fτ is identified with the standard subsphere
τ < σ, the visual sphere of M
Om τ ⊂ Om σ. By Theorem 4.11, if Hd (L; Z2 ) 6= 0, then vkδ+d (Om L) 6= 0 and
hence, obdim G ≥ δ + d + 2.
Our goal in this subsection is to show how to generalize this without the
fσ has a Z-set compactification with boundary a sphere. In
assumption that M
fσ is simply connected at infinity.
particular, we do not need to assume that M
So, assuming Theorem 4.11 (which will be proved in the next subsection),
we here prove the following.
Theorem 4.14. Suppose that {Mσ }σ∈S(L) is a simple complex of closed aspherical manifolds over S(L), where L is a d-dimensional flag complex. Take
hypotheses and notation as above. If Hd (L, Z2 ) 6= 0, then the coarse van
Kampen obstruction of EG in degree δ + d + 1 is nonzero. So, pobdim G ≥
δ + d + 2. Hence, actdim G = δ + d + 2.
This theorem applies, for example, when G is the graph product of fundamental groups of closed aspherical manifolds.
51
f
To simplify notation write
S Eσ instead ofSMσ and Cσ instead of Cone(Om σ).
Also, write E for EG = Eσ and C for Cσ . We can identify Cσ with an
open neighborhood of the basepoint in Eσ . Thus, C is an open neighborhood of the basepoint b in E and C × C is an open neighborhood of (b, b) in
E × E. The inclusion C × C ֒→ E × E takes the diagonal to the diagonal
˜
˜
so we have a Z2 -equivariant inclusion C(C)
֒→ C(E)
inducing an inclusion of
2-point configuration spaces, i : C(C) ֒→ C(E).
˜
˜
Lemma 4.15. The inclusions C(C)
֒→ C(E)
and i : C(C) ֒→ C(E) are both
homotopy equivalences.
We will assume that each Eσ comes with a proper Gσ -invariant metric
so that the inclusions Eτ ֒→ Eσ are isometries. Extend these metrics to a
metric on E by taking the induced path metric. This implies that if x and
y are points in E such that x ∈ Eσ , y ∈ Eτ , then there is a point z ∈ Eσ∩τ
with d(x, z) + d(y, z) = d(x, y). As in [36], let Nr be an r-neighborhood of
the diagonal D in E × E. Write C˜r (E) for (E × E) − Nr and Cr (E) for its
quotient by the free Z2 -action.
Lemma 4.16. The inclusions Cr (E) ֒→ C(E) induce an isomorphism on
cohomology:
∼
=
H ∗ (C(E))−→ lim H ∗ (Cr (E)),
r→∞
where lim means direct limit.
Both Lemmas 4.15
S and 4.16 have similar proofs.
Note that E = Eσ is a poset of contractible manifolds
over S(L). In
S
particular, that Eσ ∩ Eτ = Eσ∩τ . Similarly, E × E = (σ,τ ) Eσ × Eτ is a poset
of contractible manifolds over S(L)×S(L). The diagonal D(E) intersects the
terms in this decomposition as follows: (Eσ ×Eτ )∩D(E) = Dσ∩τ , where Dσ∩τ
denotes the diagonal in Eσ∩τ × Eσ∩τ . It is a properly embedded, contractible
submanifold with trivial normal bundle in the contractible manifold Eσ∩τ ×
Eσ∩τ . Thus,
[
˜
C(E)
:= (E × E) − D(E) =
(Eσ × Eτ ) − Dσ∩τ .
(σ,τ )
The manifold (Eσ × Eτ ) − Dσ∩τ is homotopy equivalent to a normal sphere
S c(σ,τ )−1 , where c(σ, τ ) is the codimension of Dσ∩τ in Eσ × Eτ . The normal
vector space of Dσ∩τ in Eσ ×Eτ decomposes as V a + V b + V d , where V a is the
52
normal space of Eσ∩τ in Eσ , V b is the normal space of Eσ∩τ in Eτ , and V d is
the normal space of Dσ∩τ in Eσ∩τ . Thus, S c(σ,τ )−1 has a join decomposition
as S a−1 ∗S b−1 ∗S d−1 . The involution (switching factors) maps V a and V b into
different factors and acts by the antipodal map on V d . It follows that the
image of the normal sphere to Dσ∩τ in (Eσ × Eτ ) in the 2-point configuration
space C(E) is homotopy equivalent to S a−1 ∗S b−1 ∗RP d−1 , i.e., to a suspension
of projective space.
The previous
S paragraph goes through, mutatis mutandis, for the union
of cones C = Cσ , as well as, for its ordered 2-point configuration space,
˜
C(C).
˜
˜
Proof of Lemma 4.15. Both C(C)
and C(E)
are posets of spaces over (S(L)×
S(L))>(∅,∅) . The geometric realization of this poset is the join, L ∗ L. The
˜
˜
relative homology groups of (C(E),
C(C))
can be computed from a spectral
sequence for the poset of spaces, e.g., see [19]. Its E 1 -page is
1
Epq
= Cp (L ∗ L; Hq ((Eσ × Eτ ) − D, (Cσ × Cτ ) − D)).
(The coefficients in this spectral sequence are not locally constant). Since
Cσ ×Cτ −D ֒→ Eσ ×Eτ −D is a homotopy equivalence, Hq (Eσ ×Eτ −D, Cσ ×
˜
˜
Cτ −D) = 0 for all q. Hence, H∗ (C(E),
C(C))
vanishes in all degrees. When L
˜
is connected, L ∗ L is simply connected; so, by van Kampen’s Theorem, C(C)
˜
˜
and C(E) are simply connected. Hence, by Whitehead’s Lemma, C(C) ֒→
˜
C(E)
is a homotopy equivalence. A more careful analysis yields the same
statement even when L is not connected. We shall not give the argument
since all we need is that the map induces an isomorphism on homology.
˜
˜
Since C(C)
֒→ C(E)
is Z2 -equivariant it induces a homotopy equivalence
C(C) → C(E).
To prove Lemma 4.16, first note that C˜r (E) also has a decomposition as
a poset of contractible manifolds:
[
C˜r (E) =
(Eσ × Eτ ) − Nr (Dσ∩τ ),
(σ,τ )
where Nr (Dσ∩τ ) means the r-neighborhood of Dσ∩τ in Eσ × Eτ . By coarse
Alexander duality (e.g., see [36]), (Eσ × Eτ ) − Nr (Dσ∩τ ) has the same prohomology type as (Eσ ×Eτ )−Dσ∩τ (which is homotopy equivalent to S c(σ,τ )−1 );
53
so, the inclusions induce an isomorphisms:
H ∗ (S c(σ,τ )−1 ) = H ∗ ((Eσ × Eτ ) − Dσ∩τ ) → lim H ∗ ((Eσ × Eτ ) − Nr (Dσ∩τ )).
r→∞
(4.6)
Proof of Lemma 4.16. The cohomology spectral sequences for the poset of
spaces give spectral sequences with E1 -pages:
E1pq = C p (L ∗ L; H q ((Eσ × Eτ ) − Dσ∩τ ))
E1pq (r) = C p (L ∗ L; H q ((Eσ × Eτ ) − Nr (Dσ∩τ )))
By (4.6), the inclusions induce an isomorphism, E1pq → lim E1pq (r) and hence,
by the comparison theorem for spectral sequences, an isomorphism,
∼
=
˜
−→ lim H ∗ (C˜r (E)).
H ∗ (C(E))
r→∞
There is a similar isomorphism for H ∗ (Cr (E)).
Proof of Theorem 4.14. By Theorem 4.11, vkδ+d (Om L) 6= 0. By Lemma 4.5,
this implies vkδ+d+1 (Cone(Om L)) 6= 0. By Lemma 4.15, its image vkδ+d+1 (E) ∈
H δ+d+1 (C(E)) also is not zero. Finally, by Lemma 4.16, the coarse van Kampen obstruction (i.e., the image of this class in lim H δ+d+1 (Cr (E))) is 6= 0.
Therefore, pobdim G ≥ δ + d + 2. Since, by Proposition 3.1, actdim G ≤
δ + d + 2, the last sentence of the theorem follows.
4.4
The van Kampen obstruction and general position
For a finite simplicial complex K there is an equivalent definition of the van
Kampen obstruction in terms of a general position map of K into Euclidean
space which we now describe. First, replace C(K) by the simplicial 2-point
configuration space of K:
C(K) = [(K × K) − D]/Z2 ,
(4.7)
where D = {(σ, τ ) ∈ K × K | σ ∩ τ 6= ∅} is a simplicial thickening of
the diagonal. Since the 2-point configuration space and the simplicial 2point configuration space are homotopy equivalent, we can denote both C(K)
without risking confusion.
54
Definition 4.17. Let K be a k-dimensional simplicial complex, and let f :
K → Rn be a general position map. This means, in particular, that if σ
and τ are two disjoint simplices of K with dim σ + dim τ = n, then the
images of σ and τ intersect in a finite number of points. The van Kampen
obstruction vkn (K) ∈ H n (C(K); Z2 ) is the cohomology class of the cocycle
νκ (= νκn (K)) defined by
hνκ, {σ, τ }i = |f (σ) ∩ f (τ )| mod 2,
where {σ, τ } means an unordered pair of disjoint simplices in K (we are using
|X| to denote the cardinality of a finite set X).
Generalizing [31, Appendix D], we give the following description of a
cocycle representing vkn (K). Given any total ordering of the vertices of K,
there is a general position map f from K to Rn defined by sending the ith
vertex in K to λ(i), where λ = (t, t2 , . . . , tn ) ∈ Rn is the moment curve,
and extending linearly. Suppose σ, τ ∈ K with dim σ + dim τ = n. The
convex hull of the union of vertices of f (σ) and f (τ ) is the cyclic polytope
C(n + 2, n). If σ and τ intersect, then neither of them can be contained
in faces of C(n + 2, n). The faces of C(n + 2, n) are completely determined
by Gale’s Evenness Condition, which in this case says that a set T of n
vertices of C(n + 2, 2) spans a face if and only if the two missing elements in
Vert(C(n + 2, n)) − T are separated by an even number of elements of T .
Two simplices σ and τ with dim(σ) + dim(τ ) = n are said to be meshed
if the order on their vertices is either
v0 < w0 < v1 < w1 < · · · < vn/2 < wn/2 , or
v0 < w0 < v1 < w1 < · · · < w(n−1)/2 < v(n+1)/2 .
Lemma 4.18. Two simplices σ and τ with dim(σ) + dim(τ ) = n intersect
under the map f if and only if they are meshed.
Proof. If σ and τ are not meshed, there are two vertices vi and vi+1 of σ with
no vertex of τ between them. In C(n + 2, n) the union of the vertices of f (τ )
and all other vertices of f (σ) except vi and vi+1 spans a face by the evenness
condition. Therefore, f (τ ) is contained in a face of the cyclic polytope and
so cannot intersect f (σ).
If σ and τ are meshed, then the evenness condition implies that f (σ) and
f (τ ) are not proper faces of C(n + 2, n).
55
Suppose that f (σ) and f (τ ) do not intersect. Let H be a hyperplane
separating f (σ) and f (τ ), so that H partitions the vertices of C(n + 2, n)
into Vert f (σ) and Vert f (τ ). If f (σ) (or f (τ )) is in the interior of C(n+2, n),
then another vertex of C(n + 2, n) is on the same side of H; hence, f (σ) ⊂
∂C(n + 2, n), a contradiction.
Note that if the difference between the dimensions of σ and τ is greater
than one, then f (σ) and f (τ ) are disjoint.
4.5
Proof of Theorem 4.11
We recall the construction in [1]. Suppose that L is a d-dimensional complex
with Hd (L; Z2 ) 6= 0, and suppose C is a d-cycle in L with coefficients in
Z2 . Identify C with its support. Choose a d-simplex ∆ ∈ C with vertices
v0 , . . . , vd . Let vi± denote the two vertices in OC lying above vi . Let D C (∆)
be the full subcomplex of OL containing C − and the vertices v0+ , . . . , vd+ of
∆+ . We say D C (∆) is C doubled over the simplex ∆. Suppose α, β are
disjoint d-simplices in D C (∆). Define a chain Ω ∈ C2d (C(D C (∆)); Z2 ) by
declaring the 2d-cell {α, β} of C(D C (∆)) to be in Ω if and only if
• α ∩ β = ∅, and
• Vert ∆ ⊂ p(Vert α) ∪ p(Vert β). (Here p : Vert OL → Vert L is the
natural projection.)
It is proved in [1] that Ω is a cycle and that νκ2d (D C (∆)) evaluates
nontrivially on Ω.
C
Next, we define a subcomplex Dm
(∆) of Om L. We assume that each
C
sphere in Om L is triangulated as the boundary of an m-simplex.
Let Dm
(∆)
S
be the full subcomplex of Om L containing Vert(C)×{1} Vert(∆)×{2, 3, . . . m}.
C
So, Dm
(∆) is constructed by replacing each vertex v ∈ ∆ with the boundary
of a m-simplex. As before, let δ = δm (L) denote the dimension of Om L.
Surprisingly, the above definition of Ω works in the case of Om L. Define
C
a chain Ωm in Cd+δ (C(Dm
(∆)); Z2 ) to be the union of all cells {σ, τ } such
that
• σ∩τ =∅
• Vert ∆ ⊂ p(Vert σ) ∪ p(Vert τ ).
56
It will be shown in Theorem 4.22 below that Ωm is a cycle. We first need
C
a few lemmas which restrict the possible (d + δ)-cells in C(Dm
(∆)).
Definition 4.19. For any w ∈ Vert ∆ and σ, τ ∈ Om L, let Mwστ be the
collection of missing vertices in p−1 (w), i.e., Mwστ is the set of vertices in
p−1 (w) that are not contained in σ ∪ τ .
Note that if p(σ) misses a vertex w in Vert ∆, then |Mwστ | ≥ 1 for any τ ,
since the preimage p−1 (w) does not span a simplex in Om (L).
Lemma 4.20. If {σ, τ } ∈ Ωm , then for any w ∈ Vert ∆, |Mwστ | ≤ 1.
Proof. The cardinality of Vert(σ ∪ τ ) is the same as that of Vert p−1 (∆).
Assume Vert(p(σ) ∪ p(τ )) includes l vertices not contained in ∆, so that
Vert(σ ∪ τ ) contains d + δ − l vertices in p−1 (∆). If Vert σ contains a vertex
outside∆, then p(σ) misses a vertex w of ∆, which implies that |Mwστ | ≥ 1.
Furthermore, if v and v ′ are distinct vertices in L − ∆ which are contained
in Vert σ ∪ τ , then there are distinct vertices w and w ′ so that |Mwστ | and
|Mwστ′ | are both ≥ 1. Otherwise, σ and τ would both miss a vertex w ∈ ∆
and {σ, τ } would not be contained in Ωm . Similarly, if there are l vertices
which are not contained in ∆, then there are l vertices w1 , . . . , wl of ∆ with
|Mwστi | ≥ 1. Therefore, if |Mwστ | > 1 for any w, the number of total missing
vertices in p−1 (∆) is greater than l, a contradiction.
Lemma 4.21. If {σ, τ } ∈ Ωm then p(σ) and p(τ ) are in L(d) .
Proof. Assume Vert(p(σ) ∪ p(τ )) includes l vertices not contained in ∆, so
that p(σ) ∪ p(τ ) contains d + l + 1 vertices. The proof of Lemma 4.20 implies
there are l vertices w1 , . . . , wl of ∆ such that |Mwστi | = 1. For the other
d + 1 − l vertices of Vert ∆, |Mwστ | = 0. Neither Vert σ nor Vert τ can contain
Vert p−1 (w), so each such w is contained in p(σ) and p(τ ). Thus, p(σ) ∩ p(τ )
contains at least d+1−l vertices. Since | Vert p(σ)| and | Vert p(τ )| is bounded
above by d + 1, the equality
| Vert p(σ) ∪ Vert p(τ )| = | Vert(p(σ) ∪ p(τ ))| + | Vert(p(σ) ∩ p(τ ))|
implies that | Vert p(σ)| and | Vert p(τ )| both equal d + 1.
The next two theorems are our computation of the van Kampen obstruction of Om L.
57
Theorem 4.22. Let L be a d-dimensional flag complex with Hd (L, Z2 ) 6= 0.
Let C be a d-dimensional cycle contained in L, and ∆ ⊂ C a d-simplex.
C
Then Ωm ∈ Cd+δ (C(Dm
(∆)); Z2 ) is a (d + δ)-cycle.
Proof. We assume that m > 1, since the m = 1 case was proved in [1] (this
C
slightly simplifies the argument). Let {σ, α} be a (d+δ−1)-cell in C(Dm
(∆)).
We claim the sum of the cardinality of the sets
C
V1 := {v ∈ Vert Dm
(∆)|{σ ∗ v, α} ∈ Ωm }
C
V2 := {v ∈ Vert Dm
(∆)|{σ, α ∗ v} ∈ Ωm }
C
is even. Note that some vertices of Dm
(∆) may be contained in both V1 and
V2 .
First, suppose p(σ) and p(α) are in C (d) . In this case, if v ∈ Vi , then
p(v) ∈ ∆. By Lemma 4.21, we can assume |Mwσα | = 0, 1 or 2 for all w ∈ ∆;
otherwise, V1 and V2 would be empty.
If Vert σ ∩Vert p−1 (w) 6= ∅ and Vert α ∩Vert p−1 (w) 6= ∅, then each vertex
of Mwσα is in V1 and V2 , hence w contributes an even number to the sum of
|V1 | and|V2 |. If Vert σ ∩ Vert p−1 (w) = ∅ and Vert α ∩ Vert p−1 (w) = ∅, then
V1 and V2 are empty.
Next, suppose that Vert σ ∩Vert p−1 (w) 6= ∅ and Vert α∩Vert p−1 (w) = ∅,
so that Mwσα makes no contribution to V2 . If |Mwσα | = 1, then the missing
vertex is not in V1 since p−1 (w) does not span a simplex in Om L. If |Mwσα | = 2,
then each vertex in Mwσα is contained in V1 , so again Mwσα contributes an even
number to |V1 |. The same argument works if Vert σ ∩ Vert p−1 (w) = ∅ and
Vert α ∩ Vert p−1 (w) 6= ∅.
Now, assume p(σ) is a d-simplex of L and p(α) is a (d − 1)-simplex of L.
In this case, V1 is empty by Lemma 4.21. Again, we consider the sets Mwσα
for w ∈ Vert ∆.
Note that |Mwσα | = 2 for at most one w ∈ Vert ∆ by Lemma 4.20, and
if |Mwσα | = 2 for some w ∈ ∆, then V2 is contained in p−1 (w). In this case,
there are 0 or 2 vertices in V2 , depending on whether or not w is in the link
6 2 for all w ∈ ∆. Since C is a cycle and p(α) is
of α. Suppose |Mwσα | =
(d − 1)-dimensional, the link LkC (p(α)) is an even number of vertices. For
each w ∈ LkC (p(α)) ∩ ∆, by assumption there is precisely one vertex in V2
which is in Mwασ (if there were zero vertices, then σ would contain all of
p−1 (w)).
Now, we claim that all the vertices in LkC (p(α)) − ∆ are in V2 . Such a
vertex v is not in V2 if and only if it is contained in σ. Since p(σ) ∪ p(α)
58
contains ∆, if v were in σ this would imply that ∆ ∪ v is a simplex in L,
which contradicts L being d-dimensional and flag. Therefore, each vertex in
LkC (p(α)) − ∆ is in V2 , and since |Mwασ | = 1 for all w ∈ LkC (p(α)) ∩ ∆,
the total cardinality of the set V2 is even (and equal to the cardinality of
LkC (p(α))).
C
Theorem 4.23. Let C be the support of a cycle in Hd (L; Z2 ), let Dm
(∆) be
C doubled over a d-simplex ∆ ⊂ C and let Ωm ∈ Zd+δ (C(Om L); Z2 ) be as
above. Then νκd+δ evaluates nontrivially on Ωm .
Proof. Order the vertices of ∆ so that v0 < · · · < vd+1 and then order the
other vertices of L so that each vertex of ∆ is < each vertex of L−∆. Extend
this to an ordering on the vertex set of Om L, by
1
m
v01 < v02 < · · · < v0m < v11 < v12 < · · · < v1m < · · · < vd+1
< · · · < vd+1
.
We have the following decomposition of the obstruction cocycle νκd+δ
evaluated on Ωm :
X
X
X
νκd+δ ({σ, τ }) =
νκd+δ ({σ, τ }).
{σ,τ }∈Ωm
{a,b}∈M
{σ,τ }∈Ωm
Vert ∆⊂a∪b p(σ)=a
p(τ )=b
If a = b = ∆, then there is exactly one meshed pair, because the union
of vertices of σ and τ is precisely the set of vertices of Om ∆, and for each
m-simplex in Dm (∆) there is a unique pair of meshed faces. Now, suppose
a 6= ∆. Let b such that Vert ∆ ⊂ a ∪ b, and let σ ∈ p−1 (a). If p(τ ) = b, then
{σ, τ } ∈ Ωm if and only if σ ∩ τ = ∅. For all w ∈ ∆ − p(σ), if τ contains
more than two vertices of p−1 (w), then σ and τ do not mesh by our choice of
ordering. On the other hand, the vertices of σ ∪ τ can omit at most 1 vertex
of p−1 (w) by Lemma 4.20; hence, no meshing can occur. Therefore, the only
contribution comes from the unique meshed pair with a = b = ∆, and hence,
νκd+δ evaluates nontrivially on Ωm .
Putting this all together, we get the following.
Theorem 4.24. Let L be a d-dimensional flag complex and let Om L be a
polyhedral join over L of (m − 1)-spheres. Let δ = dim Om L. If Hd (L; Z2 ) 6=
0, then vkd+δ (Om L) 6= 0.
Theorems 4.24 and 3.6 have the following corollary.
59
σ
Figure 5: D2 (∆) when L is a 1-cycle and each sphere is 1-dimensional.
Corollary 4.25. Suppose L is a d-dimensional flag complex, and G =
Q
L1 Gv is a graph product over L, where each Gv is the fundamental group
of a closed aspherical m-manifold. If Hd (L; Z2 ) 6= 0, then
obdim G = actdim G = (m + 1)(d + 1) = gdim G + (d + 1)
Combining this corollary with Theorem 3.14, gives Theorem B in the
Introduction.
Remark. It may be confusing why we chose to replace each vertex in Om L
with the boundary of an m-simplex. In fact, at first it seemed more natural
to us to replace each vertex with an (m − 1)-octahedron, as this would give
Om L a simple flag triangulation. However, we could not find a way to extend
the definition of [1] to this case. We’ll illustrate this with a simple example,
see Figure 5.
Suppose L is a 1-cycle, and replace two of the vertices with cellulated S 1 ’s.
We need to construct a 4-cycle β ∈ C4 (C(Dm (∆)); Z2 ). For this to be the case,
then for any 3-cell σ in C(D2 (∆)), the collection of 1-cells {τ |{σ, τ } ∈ β} must
form a 1-cycle. If we triangulate the S 1 ’s as the boundary of a 2-simplex,
then for each such σ there is a natural 1-cycle containing all the vertices
not contained in σ. On the other hand, if we replace each vertex with a
1-octahedron, then we have to make a choice of 1-cycle to pair with σ. We
could not find a way to do this consistently.
Remark 4.26. (Homology below the top dimension). In Section 5, we will
need a slight generalization of the previous arguments, where we consider
complexes with homology below the top dimension.
60
Suppose L is a d-dimensional flag complex, and let C be the support of a
cycle in Hk (L, Z2 ) for k < d. If C is a full subcomplex, then the arguments
in Theorem 4.22 and Theorem 4.23 generalize to show that vkk+δ (Om L) 6= 0.
If C is not full, then Ωm may not be a cycle if we choose ∆ incorrectly.
However, the argument generalizes if the following ∗-condition is satisfied,
see [1]:
For all σ, τ ∈ C with ∆0 ⊂ σ ∪ τ we have σ ∩ τ ⊂ ∆.
(∗)
We do not know an example of a d-dimensional complex L and a class
φ ∈ Hk (L; Z2 ) such that the ∗-condition fails for the support of every representative C for φ.
One instance where the ∗-condition is automatically satisfied is if the
cycle is in the top dimension.
Theorem 4.27. If L is a d-dimensional flag complex and Hd (L; Z2 ) 6= 0,
then vkd+δ (Om L) 6= 0.
5
Obstructors for hyperplane complements
In Subsection 4.1 we defined various configurations of standard subgroups
for simple complexes of groups. In this section, we will show that for any
finite arrangement A of affine hyperplanes in Cn there is a configuration of
abelian subgroups in the fundamental group of the complement π1 (M(A)),
indexed by the simplices in a certain simplicial complex, which is homeomorphic to the geometric realization |Q(A)| of the intersection poset. If A
satisfies certain conditions, this simplicial complex will satisfy the ∗-condition
in Remark 4.26 below. When these conditions hold, the obstructor dimension method will imply that if A is irreducible, essential, and not central,
then actdim(π1 (M(A))) ≥ 2n. In particular, when M(A) is aspherical,
actdim(π1 (M(A))) = 2n.
5.1
Free abelian subgroups
Many of the terms which we use in this subsection were defined earlier in
Subsection 3.2. Recall that the intersection poset Q(A) is ordered by reverse
inclusion. Given G ∈ Q(A), let AG := {H ∈ A | H ≤ G} be the induced
61
central subarrangement of hyperplanes containing G. We will use these central subarrangements to construct free abelian subgroups of π1 (M(A)). The
same construction of these free abelian groups is given in [22].
Lemma 5.1. For any H ≤ G ∈ Q(A), π1 (M(AH )) injects into π1 (M(AG )).
Proof. There is a natural inclusion j : M(AG ) → M(AH ). We define a
map f : M(AH ) → M(AG ) by first choosing a point x in H and a small
ball Bx that only intersects hyperplanes in M(AH ). We can deformation
retract M(AH ) to Bx and then compose with the inclusion Bx → M(AG ).
Clearly, j ◦ f is homotopic to the identity, and therefore f∗ : π1 (M(AH )) →
π1 (M(AG )) is injective.
Lemma 5.2. For any central arrangement A, π1 (M(A)) has an infinite
center.
Proof. There is a projectivization map p : Cn − {0} → CPn−1 with fiber C∗ .
The restriction to M(A) is a trivial bundle ([33, Proposition 5.1]). Then
γ = i(π1 (C∗ )) ⊂ M(A) is in the center of π1 (M(A)).
Of course, if the central arrangement A is reducible, then the center
of π1 (M(A)) has rank greater than one (take central elements from each
factor). If A is irreducible, central, and M(A) is aspherical, it turns out that
the center is infinite cyclic.
Next, suppose G ∈ Q(A) is such that AG is irreducible. By the previous
lemma, there is an element γG of infinite order in the center of π1 (M(AG )).
Furthermore, if G1 < G2 < · · · < Gn is a chain in Q(A) with each AGi
irreducible, then since
π1 (M(AG1 )) ⊂ π1 (M(AG2 )) ⊂ · · · ⊂ π1 (M(AGn )),
we obtain a free abelian group of rank n generated by γG1 , γG2 , . . . , γGn (its
rank is n by Theorem 5.6 below). On the other hand, suppose AG decomposes
as a product of irreducibles:
A(G) ∼
= A G1 × · · · × A Gk .
Then
M(AG ) ∼
= M(AG1 ) × M(AG2 ) × · · · × M(AGk )
and we obtain further free abelian groups as products of the free abelian
subgroups in the fundamental groups of the factors.
62
A
B
C
γD
D
γA
γB
γC
γA∩B∩C
Figure 6: A real line arrangement and an associated configuration of abelian
groups for its complexification.
We will produce a configuration of abelian groups based on a simplicial
complex with vertex set {G ∈ Q(A) | AG is irreducible}. Two elements G
and H of Q(A) are comparable if H < G or G < H. Distinct vertices vG and
vH are connected by an edge if and only if 1) G and H are comparable or 2)
AG ∩ H = AG × AH (see Figure 5.1).
To properly describe the simplicial complex and the corresponding configuration of abelian groups, we need the notion, introduced by De Concini
and Procesi [21], of a building set for Q(A).
5.2
Building sets
Given a collection of subspaces G in Q(A) and an element X ∈ Q(A), let
G≤X denote the set of elements in G that contain X. Let max G≤X be the set
of maximal elements of G≤X .
Definition 5.3. A collection G of subspaces in Q(A) is a building set if for
any X ∈ Q(A) such that max G≤X = {G1 , G2 , . . . Gk }, we have
AX ∼
= A G1 × A G2 × · · · × A Gk .
There are two canonical choices for a building set. Note that Q(A) is
itself a building set, since max Q(A)≤X = X. Also, note that the poset
V IQ(A) = {G ∈ Q(A) | AG is irreducible}
is a building set. The set V IQ(A) is called the set of irreducibles in Q(A);
eventually, V IQ(A) will be the vertex set of a simplicial complex IQ(A)
63
called the irreducible complex. In the case of V IQ(A), if we decompose AX
into irreducible components
AX ∼
= A G1 × A G2 × · · · × A Gk ,
then max V IQ(A)≤X = {G1 , G2 , . . . , Gk }. In fact, by considering X ∈ Q(A)
such that AX is irreducible, it is obvious that every building set must contain
V IQ(A).
Any building set determines a collection of “nested subsets”. These subsets will be the simplices of a simplicial complex on which our configuration
of abelian groups will be based.
Definition 5.4. Let G be a building set for an affine arrangement A. A
subset α ⊂ G is G-nested if for any subset {Xi } of α consisting of pairwise
incomparable elements Xi ∈ α, the intersection ∩Xi is nonempty and does
not belong to G.
Note that if X1 > X2 > · · · > Xn is any chain in G, then {X1 , . . . , Xn }
is a G-nested subset. It is also obvious from the definition that the nested
subsets form a simplicial complex, that is, if β ⊂ α and α is G-nested, then
β is G-nested. For example, if the building set is Q(A), then the nested
set complex is the barycentric subdivision |Q(A)|, as the nested subsets are
precisely chains in Q(A). Let IQ(A) denote the simplicial realization of the
V IQ(A)-nested subsets.
Lemma 5.5. Let α be a simplex in IQ(A). For each vertex of α corresponding to G ∈ V IQ(A), let γG denote the central element in π1 (M(AG ))
corresponding to the fiber of the Hopf fibration as defined in Lemma 5.1.
Then the subgroup generated by {γG }G∈Vert α is free abelian.
Proof. We claim if G and H are connected by an edge in V IQ(A), then
γG and γH commute. Two elements G and H of V IQ(A) are comparable if
G < H or H < G. A two element subset {G, H} with two elements is nested
if and only if H and G are comparable or if G ∩H 6= ∅ and G ∩H ∈
/ V IQ(A).
In the first case, γG and γH commute by Lemma 5.1. For the second case, we
have that AG∩H is reducible, hence AG∩H = AG1 ×AG2 ×· · ·×AGk . If AG and
AH are contained in a single AGi , this would imply that Gi ∈ G ∩ H, which
is a contradiction. This implies γG and γH commute. Since each γG ∼
= Z we
are done.
64
Theorem 5.6 ([22, Corollary 4.5]). For each G ∈ V IQ(A), the image γG of
γG in H1 (M(A), Z) satisfies the relation
X
γG =
γH .
H∈A
H≤G
Furthermore, for any simplex α of IQ(A), {γG | G ∈ Vert α} is linearly
independent.
Corollary 5.7. For any simplex α ∈ IQ(A), the free abelian subgroup constructed in Lemma 5.5 is free abelian of rank dim α + 1.
In general, the simplicial complex IQ(A) formed from the V IQ(A)-nested
subsets is not a flag complex. For example, if A is the complexification of
the real arrangement which consists of n general position lines in R2 , then
IQ(A) for the complexification is the one-skeleton of an n-simplex. Our
configuration of abelian groups will always be based on the flag completion
of IQ(A), where the flag completion is the unique flag complex with the
same one-skeleton as IQ(A).
Theorem 5.8. Let A be an affine arrangement, and let F IQ(A) be the flag
completion of IQ(A). Then π1 M(A) admits a configuration of free abelian
groups based on F IQ(A).
Proof. Assume that σ and τ are two simplices in F IQ(A), and Zσ and Zτ
the corresponding free abelian subgroups we have constructed. We have seen
above that the subgroups Zσ and Zτ have ranks equal to dim σ + 1 and
dim τ + 1, respectively. We must show that these subgroups intersect in the
subgroup corresponding to Zσ∩τ . Since each of these subgroups maps isomorphically onto its image in H1 (M(A), Z), it suffices to prove that the images
of the subgroups intersect correctly. However, this follows immediately from
Theorem 5.6: it is obvious that Zσ∩τ ⊂ Zσ ∩ Zτ and since each of the γG are
linearly independent we have Zσ ∩ Zτ ⊂ Zσ∩τ .
Remark 5.9. Note that the simplicial complex we construct does not capture all of the commuting relations between the standard abelian subgroups.
For example, for the arrangement depicted in Figure, 5.1, the elements γD
and γA∩B∩C commute, since γD commutes with γA , γB and γC . In this configuration of abelian groups it is only important that Zσ ∩ Zτ = Zσ∩τ .
65
5.3
The homotopy type of IQ(A)
Suppose A is a finite arrangement of affine hyperplanes in Cn of rank l. By a
theorem of Folkman [27], |Q(A)|, the geometric realization of the intersection
poset, is homotopy equivalent to a wedge of spheres, each of dimension (l−1).
The number of spheres in this wedge is an integer, βA . Alternatively, βA =
|χ(M(A))|, where χ(M(A)) is the Euler characterisitic of the complement
(see [33]). So, if
X
pA (t) =
bi (M(A))ti
i
is the Poincaré polynomial of M(A), then χ(M(A)) = pA (−1).
If A is central, then |Q(A)| is a cone and hence, βA = 0 and pA (t) has
(1 + t) as a factor. Conversely, by a theorem of Crapo [9], if βA = 0, then A
decomposes as A ∼
= A′ × A1 where A1 is a nontrivial central arrangement. It
follows that pA (t) has (1 + t)k as a factor if and only if A can be decomposed
as A′ × A1 × · · · × Ak , where Ai are nontrivial central arrangements.
Feichtner and Müller showed that if G and G ′ are two building sets
for a hyperplane arrangement, the simplicial complexes corresponding to
the nested subsets are homeomorphic via a series of stellar subdivisions
[26]. In particular, this implies that IQ(A) is a simplicial complex with
Hrk(A)−1 (IQ(A), Z2 ) 6= 0 if A is irreducible and has no central factor.
If A is a central arrangement, there is an affine arrangement d(A) ⊂ Cn−1 ,
called the “deconing” of A, which is obtained by projectivizing A and then
removing any projectivized hyperplane. The realization of the intersection
poset |Q(A)| is isomorphic to the cone on |Q(d(A))|, so therefore the Poincaré
polynomials satisfy pA (t) = (1 + t)pd(A) (t). If A is essential and irreducible,
this implies by Crapo’s theorem that d(A) is essential and has no central
factor.
5.4
Obstructor dimensions of arrangement complements
Suppose that A is an irreducible, essential, and noncentral arrangement in
Cn . We constructed a simplicial complex IQ(A) with Hn−1 (IQ(A), Z) 6= 0.
The configuration of abelian groups is obtained by taking the flag completion
F IQ(A). Since the dimension of |Q(A)| is always rk(A) − 1, it is a flag
simplicial complex with top dimensional homology if A is essential and has
no central factors.
66
Therefore, we can apply Theorem 4.11 if F IQ(A) is (n − 1)-dimensional.
We first show this holds for the case that M(A) is aspherical.
Lemma 5.10. Let A be a complex hyperplane arrangement such that M(A)
is aspherical. Then dim(F IQ(A)) ≤ n − 1.
Proof. It is a standard fact that M(A) is homotopy equivalent to an ndimensional complex. If M(A) is aspherical, then gdim(π1 (M(A))) ≤ n.
Therefore, the rank of any abelian subgroup of M is ≤ n, which implies
dim(F IQ(A)) ≤ n − 1.
This along with Theorem 4.11 and Theorem 5.8 immediately implies Theorem C from the introduction.
Theorem 5.11. If M(A) is an essential, aspherical arrangement with no
central factors, then actdim(π1 (M(A))) = 2n. In particular, M(A) is not
homotopy equivalent to a (2n − 1)-manifold.
A similar argument shows a general result for central arrangements.
Theorem 5.12. Let A be an aspherical, irreducible, essential, central arrangement in Cn . Then actdim(π1 (M(A))) = 2n − 1.
Proof. If A is essential and irreducible, then Theorem 5.16 implies that if
d(A) is the deconing of A, then obdim(π1 (M(d(A)))) = 2n − 2. Then since
M(A) ∼
= M(d(A)) × S 1 , we have π1 (M(A)) = Z × π1 (M(d(A))) and hence
obdim(π1 (M(d(A)))) = actdim(π1 (M(d(A)))) = 2n − 1.
The product formula for obstructor dimension gives the following corollary which is the general answer for the obstructor dimension of aspherical
hyperplane arrangements.
Corollary 5.13. Let A be an affine aspherical arrangement in Cn and suppose that
A∼
= A1 × A2 × · · · × Ak × A′
where each Ai is irreducible and central, and A′ has no central factor. Then
obdim(π1 (M(A))) = actdim(π1 (M(A))) = 2n − k.
Using Theorem 5.11, this gives the following computation of the action
dimension of any spherical Artin group (cf. Proposition 3.8).
67
Corollary 5.14 (Le [30]). Suppose Aσ is an irreducible spherical Artin group
of rank d + 1. Then obdim Aσ = actdim Aσ = 2d + 1. Therefore, the action
dimension of a spherical Artin group is the sum of the action dimensions of
its irreducible factors.
When M(A) is not aspherical, F IQ(A) could have much larger dimension than IQ(A). For example, we can take any arrangement and add new
hyperplanes in general position. Then IQ(A) for the new arrangement will
have arbitrarily high dimensional flag completion, as the new hyperplanes
themselves induce the 1-skeleton of a n-simplex in IQ(A). We now give
a condition on our arrangement that guarantees that F IQ(A) contains a
simplex which satisfies the ∗-condition from Subsection 4.26.
Definition 5.15. Let A be a complex hyperplane arrangement. A complete
chain of irreducibles is a chain of subspaces G0 > G1 > · · · > Gn such that
each AGi is irreducible.
We claim that the simplex σ := [G0 , G1 , . . . Gn ] in F IQ(A) satisfies the
∗-condition. Note that if a simplex does not satisfy the ∗-condition, then
there is a vertex in F IQ(A) that is connected to each of the vertices of that
simplex by an edge. So, suppose to the contrary that H is a subspace which
is connected to each Gi by an edge in F IQ(A). Now, since H is connected
to Gn in F IQ(A) and Gn is 0-dimensional, it must contain Gn . Let Gi be
the maximal subspace in the chain that is not contained in H. Since H and
Gi are connected by an edge in IQ(A), we must have AH∩Gi ∼
= A H × A Gi .
Since H contains Gi+1 we must have H ∩ Gi = Gi+1 . Therefore, the splitting
AH∩Gi ∼
= AH × AGi would contradict Gi ∈ V IQ(A).
Therefore, we have constructed a d-dimensional flag complex F IQ(A)
such that Hn−1 (F IQ(A), Z2 ) 6= 0. Since (F IQ(A), σ) satisfies the ∗-condition,
we can apply [1] to the complex O(F IQ(A)) to get the following:
Theorem 5.16. Let A be an arrangement in Cn that is essential and not
central. If A contains a complete chain of irreducibles, then
actdim(π1 (M(A))) ≥ obdim(π1 (M(A))) ≥ 2n.
If A is an inessential arrangement, then we can still compute lower bounds
for the action dimension. This is because M(A) splits as Ck × A′ , where
A′ is an essential arrangement in Cn−k , and if A is not central then A′
is not central. Conversely, our results say nothing about general position
68
hyperplane arrangements, though this is not very interesting in this context.
Hattori showed in [28] that the complement of a general position arrangement
has free abelian fundamental group and that the arrangement is homotopy
equivalent to a certain skeleton of a k-torus.
Example 5.17. We now describe another type of arrangement whose complement is always aspherical. First, an arrangement A is said to be strictly
linearly fibered over AG if G is a line and the restriction of the projection
πG : Cn → Cn /G to M(A) is a fiber bundle projection. A fiber-type arrangement is defined inductively: A ⊂ Cn is fiber-type if there is a line G
such that A is strictly linearly fibered over G and the induced arrangement
π(A) ⊂ Cn−1 is fiber-type. For example, the braid arrangement is fiber-type.
The action dimension of these examples was already known by work of Falk
and Randell in [25] and results in [4]. Indeed, Falk and Randell showed
that the fundamental group of the complement of a fiber-type arrangement
is an iterated semidirect product of free groups, so the computation of action
dimension followed from Corollary 27 in [4].
Example 5.18. A complex reflection is a periodic affine automorphism of
Cn whose fixed point set a complex hyperplane. A complex reflection group is
a finite group acting on Cn by complex reflections. For example, every finite
Coxeter group is a complex reflection group by complexification of the action
on Rn . There groups were completely classified by Shepard and Todd [34],
who showed that they fit into several infinite families depending on 3 parameters and 34 exceptional cases. The complement of the fixed hyperplanes is
a central arrangement, and the fundamental groups of such hyperplane complements can be thought of as generalizations of spherical Artin groups. It is
known that all such hyperplane complements are aspherical (the remaining
exceptional cases were resolved in [2]). Therefore, if the arrangement for a
finite reflection group is essential and irreducible, then the action dimension
of π1 (M(A)) is precisely 2n − 1.
6
Questions
Here are four questions that came up during our work. When the d-dimensional
flag complex L is EDCE and G is the fundamental group for graph product complex of closed aspherical m-manifolds (or more generally, the group
associated to a complex of closed aspherical manifolds), we showed that
69
actdim G ≤ (m + 1)(d + 1) − 1. On the other hand, in order to show that
the corresponding van Kampen obstruction is 0, we only need the weaker
assumption Hd (L, Z2 ) = 0.
Question 6.1. Is our upper bound for actdim G still valid when the hypothesis that L is EDCE is replaced by the hypothesis Hd (L, Z2 ) = 0?
Question 6.2. If OL piecewise linearly embeds in a sphere of codimension k,
does Om L piecewise linearly embed in a sphere of codimension k? Together
with the main theorem of [1] this would imply that if L is a flag complex with
Hn (L, Z2 ) = 0, then embdim(Om L) < d + δ. (Here embdim(Om L) means
the minimum dimension of a sphere into which there is a PL embedding of
Om L.)
Question 6.3. Let KL be a polyhedral join of simplicial complexes Ks over
L. Is there a formula for the van Kampen obstruction of KL in terms of the
van Kampen obstructions of the Ks and the homology of L?
Question 6.4. Suppose that A is an essential, noncentral arrangement which
admits a complete chain of irreducibles (see Definition 5.15). Is it possible
for M(A) to be homotopy equivalent to a (2n − 1)-manifold?
References
[1] G. Avramidi, M. W. Davis, B. Okun, and K. Schreve, Action dimension of right-angled
Artin groups, Bull. of the London Math. Society 48 (2015), no. 1, 115–126.
[2] D. Bessis, Finite complex reflection arrangements are K(π, 1)., Annals of Mathematics
181 (2015), no. 3, 809–904.
[3] M. Bestvina and M. Feighn, Proper actions of lattices on contractible manifolds, Invent. Math. 150 (2002), no. 2, 237–256.
[4] M. Bestvina, M. Kapovich, and B. Kleiner, Van Kampen’s embedding obstruction for
discrete groups, Invent. Math. 150 (2002), no. 2, 219–235.
[5] N. Bourbaki, Lie groups and Lie algebras. Chapters 4–6, Elements of Mathematics
(Berlin), Springer-Verlag, Berlin, 2002. Translated from the 1968 French original by
Andrew Pressley. MR1890629
[6] M. R. Bridson and A. Haefliger, Metric spaces of non-positive curvature, Grundlehren
der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 319, Springer-Verlag, Berlin, 1999. MR1744486
70
[7] R. Charney and M. W. Davis, The K(π, 1)-problem for hyperplane complements associated to infinite reflection groups, J. Amer. Math. Soc. 8 (1995), no. 3, 597–627,
DOI 10.2307/2152924. MR1303028
[8]
, Finite K(π, 1)s for Artin groups, Prospects in topology (Princeton, NJ, 1994),
Ann. of Math. Stud., vol. 138, Princeton Univ. Press, Princeton, NJ, 1995, pp. 110–
124. MR1368655
[9] H. Crapo, A Higher Invariant for Matroids, Journal of Combinatorial Theory 2
(1967), no. 2, 406-417.
[10] J. Crisp and L. Paris, The solution to a conjecture of Tits on the subgroup generated
by the squares of the generators of an Artin group, Inventiones 145 (2001), no. 1,
19-36.
[11] M. Davis, Smooth G-manifolds as collections of fiber bundles, Pacific J. Math. 77
(1978), no. 2, 315–363. MR510928
[12] M. W. Davis, Buildings are CAT(0), Geometry and cohomology in group theory (Durham, 1994), London Math. Soc. Lecture Note Ser., vol. 252, Cambridge
Univ. Press, Cambridge, 1998, pp. 108–123, DOI 10.1017/CBO9780511666131.009.
MR1709955
[13]
, The geometry and topology of Coxeter groups, London Mathematical Society Monographs Series, vol. 32, Princeton University Press, Princeton, NJ, 2008.
MR2360474
[14]
, Right-angularity, flag complexes, asphericity, Geom. Dedicata 159 (2012),
239–262, DOI 10.1007/s10711-011-9654-4. MR2944529
[15] M. W. Davis and J. Huang, Determining the action dimension of an Artin group by
using its complex of abelian subgroups, preprint, arXiv:1608.03572 (2016).
[16] M. W. Davis, T. Januszkiewicz, and I. J. Leary, The L2 -Cohomology of Hyperplane
Complements, Groups, Geometry and Dynamics 1 (2007), 301-309.
[17] M. W. Davis and P. H. Kropholler, Criteria for asphericity of polyhedral products:
corrigenda to “right-angularity, flag complexes, asphericity”, Geom. Dedicata 179
(2015), 39–44, DOI 10.1007/s10711-015-0066-8. MR3424656
[18] M. W. Davis and B. Okun, Vanishing theorems and conjectures for the ℓ2 -homology
of right-angled Coxeter groups, Geom. Topol. 5 (2001), 7–74.
[19]
, Cohomology computations for Artin groups, Bestvina-Brady groups, and
graph products, Groups Geom. Dyn. 6 (2012), no. 3, 485–531, DOI 10.4171/GGD/164.
MR2961283
[20] P. Deligne, Les immeubles des groupes de tresses généralisés, Invent. Math. 17 (1972),
273–302 (French). MR0422673
[21] C. De Concini and C. Procesi, Wonderful Models of Subspace Arrangements, Selecta
Mathematica 1 (1995), no. 3, 459–494.
71
[22] G. Denham, A. Suciu, and S. Yuzvinsky, Combinatorial Covers and Vanishing of
cohomology, Selecta Mathematica 22 (2016), no. 2, 561–594.
[23] Z. Despotovic, Action Dimension of Mapping Class Groups, Ph.D. Thesis, Department of Mathematics, University of Utah, 2006.
[24] A. N. Dranishnikov, On the virtual cohomological dimensions of Coxeter groups, Proc.
Amer. Math. Soc. 125 (1997), no. 7, 1885–1891, DOI 10.1090/S0002-9939-97-04106-3.
MR1422863
[25] M. Falk and R. Randall, The lower central series of a fiber-type arrangement, Invent.
Math. 82 (1985), 77–88.
[26] E. M. Feichtner and I. Müller, On the Topology of Nested Set Complexes, Proc. Amer.
Math. Soc. 131 (2003), 1695–1704.
[27] J. Folkman, The homology groups of a lattice, J. Math. Mech. 15 (1966), 631-636.
[28] A. Hattori, Topology of Cn minus a finite number of affine hyperplanes in general
position, J. Fac. Sci. Univ. Tokyo 22 (1975), 205-219.
[29] H. van der Lek, Extended Artin groups, Singularities, Part 2 (Arcata, Calif., 1981),
Proc. Sympos. Pure Math., vol. 40, Amer. Math. Soc., Providence, RI, 1983, pp. 117–
121. MR713240
[30] G. Le, The Action Dimension of Artin Groups, Ph.D. Thesis, Department of Mathematics, Ohio State University, 2016.
[31] J. Matoušek, M. Tancer, and U. Wagner, Hardness of embedding simplicial complexes
in Rd , J. Eur.Math.Soc.(JEMS) 13 (2011), no. 2, 259–295.
[32] L. Mosher, M. Sageev, and K. Whyte, Quasi-actions on trees I. Bounded valence,
Annals of Mathematics 158 (2003), 115–164.
[33] P. Orlik and H. Terao, Arrangements of hyperplanes, Springer-Verlag, Berlin Heidelberg GmbH, 1992.
[34] G. C. Shepard and J. A. Todd, Finite Unitary Reflection Groups, Canadian Journal
of Mathematics 6 (1954), 274-304.
[35] J.-P. Serre, Trees, Springer-Verlag, Berlin-New York, 1980. Translated from the French
by John Stillwell. MR607504
[36] S. Y. Yoon, A lower bound to the action dimension of a group, Algebr. Geom. Topol.
4 (2004), 273–296, DOI 10.2140/agt.2004.4.273. MR2059192
Michael W. Davis, Department of Mathematics, The Ohio State University, 231 W. 18th Ave., Columbus, Ohio 43210, [email protected]
Giang Le, Department of Mathematics, Oregon State University, 368 Kidder Hall, Corvallis, OR 97331, [email protected]
Kevin Schreve, University of Michigan-Ann Arbor, Department of Mathematical Sciences, PO Box 413, Ann Arbor, MI 48109, [email protected]
72
| 4math.GR
|
arXiv:1305.1422v4 [cs.DC] 9 Jun 2017
Somoclu: An Efficient Parallel Library for
Self-Organizing Maps
Peter Wittek1,2 , Shi Chao Gao3 , Ik Soo Lim4 , and Li Zhao3
1
2
University of Borås
ICFO-The Institute of Photonic Sciences
3
Tsinghua University
4
Bangor University
Abstract
Somoclu is a massively parallel tool for training self-organizing
maps on large data sets written in C++. It builds on OpenMP for
multicore execution, and on MPI for distributing the workload across
the nodes in a cluster. It is also able to boost training by using
CUDA if graphics processing units are available. A sparse kernel is
included, which is useful for high-dimensional but sparse data, such
as the vector spaces common in text mining workflows. Python, R
and MATLAB interfaces facilitate interactive use. Apart from fast
execution, memory use is highly optimized, enabling training large
emergent maps even on a single computer.
1
Introduction
Visual inspection of data is crucial to gain an intuition of the underlying
structures. As data often lies in a high-dimensional space, we use embedding
techniques to reduce the number of dimensions to just two or three.
Methods that rely on eigenvalue decomposition, such as multidimensional
scaling [4], achieve a global optimum for such an embedding: The global
topology of the space will be preserved.
1
Often the data points lie on a high-dimensional manifold that is curved
and nonlinear. These structures are difficult to find with eigenvalue decomposition. Manifold learning generalizes embedding further, assuming that
data in a high-dimensional space aligns to a manifold in a much lower dimensional space. For example, the algorithm Isomap finds a globally optimal
solution for an underlying nonlinear manifold and it builds on multidimensional scaling [21]. Isomap, however, fails to find nonconvex embeddings [28].
Nonconvex structures are one strong motivation to look at solutions that
are not globally optimal, but preserve the local topology of the space instead.
Self-organizing maps (SOMs) are a widespread visualization tool that embed
high-dimensional data on a two-dimensional surface – typically a section
of a plane or a torus – while preserving the local topological layout of the
original data [10]. These maps provide a visual representation of groups of
similar data instances, and they are also useful in analyzing the dynamics
of evolving data collections, as updates are easily made. Emergent selforganizing maps contain a much larger number of target nodes for embedding,
and thus capture the topology of the original space more accurately [24].
Training a map is computationally demanding, but a great advantage of
SOMs is that the computations are easy to parallelize. They have been accelerated on massively parallel graphics processing units (GPUs [14]). Tools
exist that scale to large data sets using cluster resources [20], and also combining GPU-accelerated nodes in clusters [29]. These tools focus on batch
processing. On the other end of the spectrum, interactive environments for
training SOMs are often single-core implementations [24], not making full
use of contemporary hardware.
We develop a common computational core that is highly efficient with the
available resources. This common core is used for a command-line interface
to enable batch processing on cluster resources, but the same core is used
as the computational back-end for environments popular in data analysis.
The tool, named Somoclu (originally standing for SOM on a cluster), has
the following improvements over other implementations:
• It is highly efficient in single-node multicore execution.
• Memory use is reduced by up to 50 per cent.
• Large emergent maps are feasible.
• A kernel for sparse data is introduced to facilitate text mining applications.
2
• Training time is reduced by graphics processing units when available.
• It improves the efficiency of distributing the workload across multiple
nodes when run on a cluster.
• An extensive command-line interface is available for batch processing.
• Python [25], R [17] and MATLAB [22] interfaces facilitate interactive
processing.
• Compatibility with Databionic ESOM Tools [24] ensures easy visualization.
The source code is available under GNU Public License Version 3 (http:
//peterwittek.github.io/Somoclu/). The Python version is also listed
in the Python package index (https://pypi.python.org/pypi/Somoclu/).
The R version is available from the Comprehensive R Archive Network (CRAN)
at https://CRAN.R-project.org/package=RSomoclu/.
2
Self-organizing maps in batch mode
The SOM training algorithm constructs a nonlinear topology preserving mapping of the input data set X = {x(t)|t ∈ {t0,...,tf }}, where t0 and tf are the
beginning and the end of the current training session, onto a set of neurons
M = {n1 , . . . , nk } of a neural network [10]. The neurons of the network are
arranged in a grid, with associated weight vectors
W = {w1 (t), . . . , wk (t)}
(1)
at a given time step t. W is known as the code book.
Each data point x(t) is mapped to its best matching unit
bm(x(t)) = nb ∈ M,
(2)
d(x(t), wb (t)) ≤ d(x(t), wj (t)) ∀wj (t) ∈ W,
(3)
such that
where d is the distance function on the data set in the feature space. The
neurons are arranged on a two dimensional map: Each neuron i has a two
coordinates in a grid. Next the weight vector of the best match neuron
3
and its neighbors are adjusted toward the input pattern using the following
equation:
wj (t + 1) = wj (t) + αhbj (t)(x(t) − wj (t)),
(4)
where 0 < α < 1 is the learning factor, and hbj (t) is the neighborhood
function that decreases for neurons further away from the best match neuron
in grid coordinates. A frequently used neighborhood function is the Gaussian:
−krb − rj k
,
(5)
hbj = exp
δ(t)
where rb and rj stand for the coordinates of the respective nodes. The width
δ(t) decreases from iteration to iteration to narrow the area of influence.
The training is repeated on the same data set to increase the fit, a training
cycle is called an epoch. Eventually, the neighborhood function decreases to
an extent that training might stop. The time needed to train an SOM grows
linearly with the data set size, and it grows linearly with the number of
neurons in the SOM. SOM has a batch formulation of updating the weights,
which is widely used in parallel implementations:
Ptf
0
0
t0 =t0 hbj (t )x(t )
.
(6)
wj (tf ) = P
tf
0)
h
(t
0
bj
t =t0
While not directly related to the training, it is worth mentioning the Umatrix associated with a SOM, which depicts the average Euclidean distance
between the code book vectors of neighboring neurons. Let N (j) denote the
immediate neighbors of a node j. Then the height value of the U-matrix for
a node j is calculated as
U (j) =
X
1
d(wi , wj ).
|N (j)|
(7)
i∈N (j)
The purpose of the U-matrix is to give a visual representation of the topology
of the network.
There is no shortage of implementations of SOM training algorithms. It
comes integrated in data mining suites such as RapidMiner [8]. Dedicated
visualization and training tools also exist, such as the Databionic ESOM
Tools [24]. Popular languages used in data analytics all have SOM modules,
including MATLAB [26], Python [7], and R [27]. Common to these tools is
4
that they seldom make use of parallel computing capabilities, although the
batch formulation of SOM training invites such implementations. Further
constraints include memory requirements, which increase fast with the size
of the neuron grid.
Lower level libraries address such scalability issues. Distributing the
workload across multiple nodes is feasible via standard methods such as MPI
and MapReduce [11, 20]. The inherent parallelism is easy to exploit with
graphics processing units (GPUs [14]). A method combining distributed
computing with parallelism on GPUs is also available [29].
With Somoclu, we bridge high-level languages and massively parallel architectures, while also simplifying the distribution of workload compared to
earlier implementations. We use a common core written in C++, supplemented by OpenMP [5], MPI [19], and CUDA [16]. This common core is
used by an extensive command-line interface, and wrappers in Python, R,
and MATLAB.
3
High-performance core for training
The starting point was a MapReduce-based distributed implementation of
SOM [20]. This implementation was later accelerated by GPUs [29]. We
found the MapReduce framework unnecessary, and derived a purely MPIbased implementation, that is more modular, faster, and includes a sparse
kernel. Design details are provided in Section 3.1. Distributing the workload
across multiple nodes is an extension of the parallel formulation (Section 3.2).
3.1
Parallelism
The dense CPU kernel is a straightforward implementation of the batch formulation in Equation 6, and it resembles the one implemented by Ref. [20].
For an overview of the parallel organization, refer to Figure 1. The MapReduce calls were replaced by MPI functions, leading to a more streamlined
code. Furthermore, single-node parallelism on multicore CPUs is redesigned
to use OpenMP instead of MPI. This strategy avoids duplicating the code
book: Each MPI process must have a copy of the code book, but OpenMP
threads can work on the same copy. The simplification leads to a minimum
fifty per cent reduction in memory even when only two threads are used.
5
A performance bottleneck in the original implementation was the accumulation of local weights into a new global code book by one single process
on the master node. This is parallelized by an OpenMP directive. Furthermore, the influence radius δ(t) of a best matching node is thresholded, which
translates to speed improvements without compromising the quality of the
trained map.
The GPU variant is more complex compared to the CPU kernel. The
complexity stems from the way the distance function is evaluated between the
nodes of the SOM and the training data. To maximize parallelism, the Gram
matrix is calculated, that is, a matrix of the distances between every data
instance and the nodes of the SOM. A naı̈ve approach would be to extend an
efficient matrix multiplication algorithm, replacing the dot product by the
distance function. Opting for a Euclidean distance, it is possible to derive an
alternative formulation of calculating the Gram matrix using linear algebra
operations [13]. Benchmarking the two approaches, we found that the latter
approach is a magnitude faster on the GPU, mainly due to a more favorable
memory access pattern.
We implemented the GPU kernel with Thrust [2], a C++ template library for CUDA, which has high-performance primitives, avoiding the need
to manually tune individual GPU calls. Compared to the implementation
by Ref. [29], the device memory use of the GPU code is reduced to approximately one-third, and the new implementation completely avoids costly
matrix transposing operations.
Further complexity arises from the disparity between the number of GPUs
and the number of cores in a computer. The CPU kernel achieves maximum
speed by running an MPI process on each available core, resulting in a far
lower number of data instances per core and a speedier local update of the
weights. For instance, if there are eight cores and two GPUs, than each
GPU has four times more data to process and its corresponding MPI thread
would have four times more data to update the local weights. While the
GPU handles the load efficiently, it would be highly inefficient to use a single
thread to update the local weights. We thus hybridized the kernel and rely
on OpenMP to parallelize the weight update. The GPU implementation runs
as many MPI processes on a node as there are GPUs, and uses all CPU cores
in the weight update.
The sparse kernel is a straightforward extension of the dense CPU kernel,
and its main virtue is the reduced memory use. A vector space coming from
a text processing pipeline typically contains 1–5% nonzero elements, leading
6
Code book
Data chunk
Best matching units
Memory structures
OpenMP/CUDA
threads
Process
Figure 1: Overview of the parallel organization of the batch training. The
global code book is replicated in each process, therefore it is more efficient to
use OpenMP-based parallelization than to rely on MPI on multicore CPUs.
If a GPU is available, it will overtake most data parallel operations from the
CPU, replacing OpenMP threads.
to a 20–100× reduction in memory use when using a sparse representation.
This kernel does not have a GPU implementation, as the irregular access
patterns that are inevitable with sparse data structures are not efficient on
streaming architectures.
3.2
Workload in distributed environment
The training procedure has a simple communication structure in a distributed
memory system. Finding the best matching unit in Equation 2 is independent
for every data instance, resembling the parallel version. This means that we
can distribute equally sized parts of the data to each node, without any
further communication of training data later on.
The update of the code book requires two way communication between a
master node and the slave nodes. Once all weight updates are calculated in a
slave node, the local updates are sent to the master node, which accumulates
the changes to the code book. The new code book is broadcast to all slave
7
nodes.
This communication structure ensures a near-linear scaling for smaller
code books. Emergent maps with high-dimensional data, however, scale
poorly. The code book is always a dense structure, even if the training data
is sparse, as there are hardly any zero entries. Storing the code book in
memory is the primary constraint for single node execution, and it is also
the key constraint in distributed workloads.
Compared to an earlier GPU-accelerated, distributed implementation of
SOM [29] that used MapReduce, we found that ordinary MPI calls are sufficient. Execution time did not improve significantly, but the overall code
simplified, and our implementation does not have to rely on a little-used,
MPI-based MapReduce library.
4
Interfaces
We regard the package libsvm for training support vector machines as a role
model: Using an efficient computational core, numerous wrappers were added
to the library to interface with other languages and environments [3]. The
package fastcluster is also similar in its approach [15].
We provide an extensive command-line interface, which is able to access all functionality provided by the computational core (Section 4.1). The
training functionality is exposed via an application programming interface,
a single function that performs one epoch of training (Section 4.2). This
function is wrapped to be called from Python, R, and MATLAB without
duplicating data structures (Section 4.3).
4.1
Command-line interface
One sparse and two dense data formats are supported. All of them are plain
text files. The entries can be separated by any white-space character. One
row represents one data instance across all formats. Comment lines starting
with a hash mark are ignored.
The basic dense format includes the coordinates of the data vectors, separated by a white-space. This file is parsed twice to get the basic dimensions
right. The second dense format is identical, but it includes a header that
contains information about the layout of the matrix. This second format is
compatible with Databionic ESOM Tools.
8
The sparse representation is similarly row-oriented, and it uses the same
sparse format as libsvm [3]. For instance, the vector [1.2 0 0 3.4] is represented
as the following line in the file: 0:1.2 3:3.4. The file is parsed twice: once to
get the number of instances and features, and the second time to read the
data in the individual threads.
Example files are available in the package. Extension to further matrix
types is simple to add.
The tool is used via the command line. Somoclu takes a plain text input
file – either dense or sparse data. The basic execution is as follows:
$ [mpirun -np NPROC] Somoclu [OPTIONs] INPUT_FILE OUTPUT_PREFIX
The mpirun is required only when multiple nodes are used, or when there
is more than one GPU in the system. The parameter INPUT_FILE is selfexplanatory. Instead of names of output files for the best matching units,
code books, and U-matrices, an output prefix is requested in OUTPUT_PREFIX.
The resulting files will be differentiated by the extension, and, if interim snapshots are requested, also by the indices of the epochs in which the snapshots
are taken.
The rest of the arguments are as follows.
-c FILENAME
The parameter specifies an initial code book for the map. The default is
random initialization.
-e NUMBER
The number of training epochs, that is, the number of times each data
instances will be presented to the learner.
-g TYPE
The type of the grid type; the default is square and the other option is
hexagonal.
-k NUMBER
This parameter defines the kernel type. The value 0 stands for the dense
CPU kernel, 1 is for the dense GPU kernel, and 2 is for the sparse CPU
kernel.
9
-m TYPE
The map type is either planar or toroid; the default is planar.
-n FUNCTION
This option chooses the neighborhood function, which can be Gaussian
or bubble; the default being the former.
-p NUMBER
If the neighborhood function is Gaussian, setting this parameter to 1 will
cut off the update of nodes beyond the current radius.
-t STRATEGY
Radius cooling strategy is either linear or exponential; the default is linear.
-r NUMBER
The parameter defines the start radius. The default value is half of the
map size in the smaller direction.
-R NUMBER
The final radius; it defaults to 1.
-T STRATEGY
Learning rate cooling strategy is either linear or exponential; the default
is linear.
-l NUMBER
The starting learning rate; with a default value of 1.0.
-L NUMBER
The final learning rate; the default is 0.01.
10
-s NUMBER
This parameter decides whether to save interim files. The default value is
0, which means that no interim files will be saved. Setting the parameter to
1 will calculate and save the U-matrix in each time step. If the value passed
is 2, then the code book and the best matching units are also saved in each
epoch. Since the code book is enormous for large emergent maps, passing a
value of 2 may significantly slow down each epoch.
-x, --columns NUMBER
This is the number of columns in map, that is, the size of the SOM in
direction x; the default value is 50.
-y, --rows
NUMBER
The number of rows in map is defined by this value, that is, the size of
the SOM in direction y; the default value is 50. Examples:
$ Somoclu data/rgbs.txt data/rgbs
$ mpirun -np 4 --hostfile hostfilename Somoclu -k 0 --rows 20 --columns 20 \
>
data/rgbs.txt data/rgbs
4.2
As an application programming interface
Designed to work with MPI, Somoclu was not conceived to be used as an
API. Yet, given sufficient preparation on the calling side, it is possible to
interface with Somoclu as an API. The primary entry point for training is
the following function, as specified in Somoclu.h:
void trainOneEpoch(int itask, float *data, svm_node **sparseData,
float *codebook, int *globalBmus, unsigned int nEpoch,
unsigned int currentEpoch, unsigned int nSomX, unsigned int nSomY,
unsigned int nDimensions, unsigned int nVectors,
unsigned int nVectorsPerRank, unsigned int radius0,
unsigned int radiusN, string radiusCooling, float scale0, float scaleN,
string scaleCooling, unsigned int kernelType, string mapType,
string gridType, bool compact_support, bool gaussian)
11
The parameters nSomX, nSomY, nEpoch, kernelType, and so on, are selfexplanatory, they are the same as in the command-line interface.
The parameter itask specifies the rank of the current MPI process. If
the calling environment does not use MPI, the value should be set as zero.
Data are either stored in data or sparseData, depending on whether the
data are dense or sparse. Only data belonging to the process is expected,
not the full matrix. The parameter nVectors stores the total number of
instances, whereas nVectorsPerRank the number of instances belonging to
one MPI thread.
4.3
Python, R and MATLAB interfaces
Solutions for scalable data processing in higher level languages tend to be ad
hoc. Focusing on R as an example, the R Installation and Administration
manual states that the language is ill-suited for working with data structures
larger than about 10–20% of the main memory [18]. The key problem is
with the overheads on the data structures. Parallel processing is scarcely
addressed. Some modules aim to overcome these barriers [9].
Distributed computing is also challenging for such languages. Most solutions work entirely on a case by case basis, such as the distributed text
mining module for R [23].
The aim of the interfaces is to address the problem of memory use and
parallel execution. The Python package easily installable from the Python
package index. The R package is available from CRAN. These packaged
variants rely on the OpenMP parallelism alone. To get the massive CUDA
parallelism through the interfaces, users with basic software building knowledge can easily follow the instructions on the project site. MATLAB users
need to follow the instructions on the project site to build the module as no
central package distribution site is available. Distributed workloads directed
from a high-level language remain for future development.
Starting with the Python interface, our primary concern is to avoid memory copies and duplicating large structures in memory. The native variable
system of Python is not strictly typed, hence we require the use of numpy.
Using numpy float32 data type for storing arrays, we map directly to the
single-precision floats used in the C++ code. We pass pointers between the
two languages, making the best use of memory and almost zero computational overhead. This interface encapsulates the training routines in a class
to simplify the workflow.
12
An example use of the interface is as follows:
>>>
>>>
>>>
>>>
>>>
>>>
import Somoclu
import numpy
data = numpy.loadtxt(’data/random.dat’)
n_rows, n_columns = 50, 50
som = Somoclu.Somoclu(n_columns, n_rows, data = data)
som.train()
The final code book is in the class variable Somoclu.codebook, the best
matching units are in Somoclu.bmus, and the U-matrix is also calculated in
Somoclu.umatrix.
The R interface uses the Rcpp package to cast pointers and provide type
conversions between R and C++ [6]. Since R uses double precision matrices
by default, and Somoclu uses single-precision floats internally, we must convert between double and float arrays of the input and output. This makes
the interface less efficient than the Python version. The interface integrates
with the popular R SOM package kohonen to facilitate use. The following is
an example on how to call the wrapper function from R:
R>
R>
R>
R>
R>
R>
R>
R>
R>
R>
R>
R>
+
library("RSomoclu")
data("rgbs", package = "RSomoclu")
input_data <- data.matrix(rgbs)
nSomX <- 50; nSomY <- 50
nEpoch <- 10
radius0 <- 0; radiusN <- 0
radiusCooling <- "linear"
scale0 <- 0; scaleN <- 0.01
scaleCooling <- "linear"
kernelType <- 0
mapType <- "planar"
res <- RSomoclu.train(input_data, nEpoch, nSomX, nSomY, radius0, radiusN,
radiusCooling, scale0, scaleN, scaleCooling, kernelType, mapType)
The code book is returned in res$codebook, the best matching units are
in res$globalBmus, and the calculated final U-matrix is in res$uMatrix.
Most R users will probably want to work with the map through the package
kohonen:
13
R> sommap <- RSomoclu.kohonen(input_data, res)
The MATLAB interface uses the official MEX-file mechanism to interface
with C++. As with R, MATLAB also defaults to double precision variables,
rendering some overhead of type conversion inevitable. We designed a similar training function call to that of the SOM Toolbox [26]. And provided
examples on visualization using both SOM Toolbox and Databionic ESOM
Tools. The following is an example call from the MATLAB interface.
>> data = importdata(’data/random.dat’);
>> msize = [50 50];
>> sMap = som_randinit(D, ’msize’, msize);
>> nEpoch = 10;
>> radius0 = 0;
>> radiusN = 0;
>> radiusCooling = ’linear’;
>> scale0 = 0;
>> scaleN = 0.01;
>> scaleCooling = ’linear’;
>> kernelType = 0;
>> mapType = ’planar’;
>> gridType = ’rectangular’;
>> compactSupport = false;
>> neighborhood = ’gaussian’;
>> [sMap, sTrain, globalBmus, uMatrix] = ...
Somoclu_train(sMap, data, ’msize’, msize, ’radius0’, radius0, ...
’radiusN’, radiusN, ’radiusCooling’, radiusCooling, ...
’scale0’, scale0, ’scaleN’, scaleN, ’scaleCooling’, scaleCooling, ...
’kernelType’, kernelType, ’mapType’, mapType, ...
’gridType’, gridType, ’compactSupport’, compactSupport, ...
’neighborhood’, neighborhood, ’nEpoch’, nEpoch);
The returned variables are identical to the Python version.
4.4
Visualization
The primary purpose of generating a map is visualization. Yet – apart from
the Python interface – Somoclu does not come with its own functions for visualization, since there are countless generic and SOM-specific tools that are
14
capable of plotting high-quality figures. The code book, the best matching
units, and the U-matrix are exported at the end of the training, or even after
each epoch.
The simplest procedure is to use a generic plotting library, such as gnuplot. For instance, the following simple gnuplot script plots the U-matrix in
PNG format:
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
gnuplot>
set datafile commentschars "#!%"
set autoscale
unset log
unset label
set lmargin at screen 0
set rmargin at screen 1
set bmargin at screen 0
set tmargin at screen 1
unset xtic
unset ytic
set notitle
set pm3d at b
set pm3d map
set nokey
set nocolorbox
unset surface
set term pngcairo transparent size 500,500
set output ’umatrix.png’
splot "filename" matrix
The dimensions of the picture should be adapted to the size of the network
to avoid distortion.
The Python interface is equipped with plotting routines. For instance, a
map can be plotted along with the labels or class colors:
>>> som.view_umatrix(bestmatches = True)
We may also zoom in to a specific part of the map:
>>> som.view_umatrix(bestmatches = True, zoom = ((n_rows//2, n_rows), (0, n_column
15
(a) The complete map.
(b) A close-up of region.
Figure 2: The U-matrix of a toroid map with labels trained on a toy example
plotted in Python.
(a) Codebook.
(b) U-matrix.
Figure 3: Visualizing a toy example trained by Somoclu and plotted by
kohonen in R.
Using a toy example included with the code, we obtained a visualization
shown in Figure 2.
The R interface allows visualization through the package kohonen. This
is entirely external to the library. A conversion to kohonen’s format is nec16
U−matrix
Variable1
1.13
1
0.564
0.5
6.46e−14
2.7e−10
d
Variable2
Variable3
1
1
0.5
0.5
1.01e−14
d
1.96e−05
d
SOM 07−Jan−2016
Figure 4: Visualizing a toy example trained by Somoclu and plotted by SOM
Toolbox in MATLAB.
essary:
R> sommap <- RSomoclu.kohonen(input_data, res)
R> plot(sommap, type = "codes", main = "Codes"))
R> plot(sommap, type = "dist.neighbours")
A toy example is shown in Figure 3.
The results of the MATLAB interface allows visualization through the
SOM Toolbox, which is also external to the library.
>> som_show(sMap);
The toy example is shown Figure 4.
More advanced, SOM-specific tools may also offer an interactive GUI to
select gradients and other parameters. One such tool is Databionic ESOM
Tools [24], with which the output formats of Somoclu are compatible. The
result of such visualization is shown in Section 5.3.
17
104
CPU
GPU
R
Running time (s)
Running time (s)
10
5
104
103
102
12500 25000
50000
Data instances
CPU
GPU
103
102
1250 2500
100000
(a) 50×50 map.
5000
Data instances
10000
(b) 200×200 emergent map.
Figure 5: Training time on a single node with CPU and GPU kernels and
the R package kohonen. The time axis is logarithmic. The data instances
had 1,000 dimensions.
5
Experimental results
To ensure replicability of the results, we benchmarked with publicly available
cluster GPU instances provided by Amazon Web Services. The instance type
was cg1.4xlarge,1 equipped with 22 GiB of memory, two Intel Xeon X5570
quad-core CPUs, and two NVIDIA Tesla M2050 GPUs, running Ubuntu
12.04.
5.1
Single-node performance
The most direct comparison should be with the implementations by Refs. [20,
29]. Unfortunately the former was not able to handle the matrix sizes we
were benchmarking with. The implementation by Ref. [29] is sensitive to
the size of map, and it did not scale to the map sizes benchmarked here,
and thus we left it out from the comparison. To compare with single-core
performance, we included the R package kohonen [27]. The number of data
instances ranged from 12,500 to 100,000, the number of dimensions was fixed
at 1,000 for a regular 50 × 50 self-organizing map. The data elements were
randomly generated, as we were interested in scalability alone. We also tested
an emergent map of 200 × 200 nodes, with the number of training instances
ranging from 1,250 to 10,000. This large map size with the largest data
1
https://aws.amazon.com/ec2/instance-types/
18
2500
Dense
Sparse
Maximum memory use
(MByte)
Running time (s)
3000
2000
1500
1000
500
0
12500 25000
50000
Data instances
100000
450
Dense
400
Sparse
350
300
250
200
150
100
50
0
12500 25000
50000
Data instances
(a) Running time.
100000
(b) Memory use.
Figure 6: Training time on a single node with dense and sparse kernels. The
data instances had 1,000 dimensions, with five per cent of the elements being
nonzero.
matrix filled the memory of a single GPU, hence giving an upper limit to
single-node experiments. Emergent maps in the package kohonen are not
possible, as the map is initialized with a sample from the data instances. If
the map has more nodes than data instances, kohonen exits with an error
message. The map size did not affect the relative speed of the different
kernels (Figure 5). The comparison was based on the command-line version
of Somoclu.
Compared to the R package, even the CPU version is at least ten times
faster. The difference increases with the data size, indicating serious overhead
problems in the R implementation.
The GPU results show at least a two-times speedup over the CPU version.
This is less than expected, but this result considers a single GPU in the
dual configuration, and the CPU is also a higher end model. These results,
nevertheless, show that the Thrust template library is not efficient on twodimensional data structures.
Comparing the sparse and dense kernels on a 50 × 50 map, we benchmarked with random data instances of 1,000 dimensions that contained five
per cent of nonzero elements (Figure 6). Execution time was about two
times faster with the sparse kernel. The reduction in memory use was far
more dramatic, the sparse kernel using only twenty per cent of the memory
of the dense one with 100,000 instances. Naturally, the difference with emergent maps would be less apparent, as the code book is always stored in a
dense format.
19
Maximum memory use
(MByte)
9000
C++
8000
MATLAB
7000
R
6000
Python
5000
4000
3000
2000
1000
0
12500 25000
50000
Data instances
100000
Figure 7: Memory overhead of the Python, R, and MATLAB interfaces
compared to the command-line version (indicated as C++).
Measuring the memory overhead of the interfaces compared to the native
version, we are not surprised to see that the Python variant is the closest
(Figure 7). As the data structures are not duplicated in this interface, it
is still counter-intuitive that the gap increases with larger data sets. The R
and MATLAB versions have predictably larger and growing gaps compared to
the command-line interface, as they both must duplicate all data structures.
Apart from the time spent on duplicating the data structures in the R and
MATLAB versions, the computational overhead is negligible in all interfaces.
5.2
Multi-node scaling
Using 100,000 instances and a map of 50 × 50 nodes, the calculations scale
in a linear fashion (Figure 8). This was expected, as there is little communication between nodes, apart from the weight updates. As calculations
cannot overlap with communication, we did not benchmark the GPU kernel
separately, as its scaling is identical to that of the CPU kernel.
5.3
Visualization on real data
We used the Reuters-21578 document collection for an example on text mining visualization [12]. We used Lucene 3.6.2 [1] to create an inverted index of
the document collection. Terms were stemmed and we discarded those that
occurred less than three times or were in the top ten per cent most frequent
ones. Thus we had 12,347 index terms, lying in an approximately twentythousand dimensional space. We trained a toroid emergent self-organizing
20
8
CPU
Speedup
7
6
5
4
3
2
1
1
2
4
Number of nodes
8
Figure 8: Speedup on multiple nodes with CPU kernel compared to a single
node. The data instances had 1,000 dimensions.
Figure 9: The U-matrix of a toroid emergent self-organizing map after ten
epochs of training on the feature space of sparse data. The individual dots
are neurons with a weight vector that match a data instance. The other
neurons reflect the distances in the original high-dimensional space.
map of 336 × 205 dimensions. The initial learning rate was 1.0, which decreased linearly over ten epochs to 0.1. The initial radius for the neighborhood was a hundred neurons, and it also decreased linearly to one. The
21
neighborhood function was a noncompact Gaussian.
We studied the U-matrix of the map. Visualizing this with the Databionic
ESOM Tools [24], we plotted the global structure of the map in Figure 9. The
map clearly shows dense areas where index terms are close and form tight
clusters. Other parts of the map are sparse, with large barriers separating
index terms into individual semantic regions.
6
Limitations
The most constraining limitation is the storing of the code book in the main
memory. While memory use has been optimized, and only the number of
computing nodes sets a limit to the amount of data to be processed, each node
keeps a full copy of the code book. This is not a problem for feature spaces of
a few thousand dimensions – even emergent maps are easy to compute. Yet,
if the feature space has over tens of thousands or more features, emergent
maps are no longer feasible.
Currently only the command-line interface is able to unlock all the capabilities of the computing core. The R, Python and MATLAB interfaces can
use capabilities of the OpenMP and CUDA core, but these wrappers cannot
use a distributed cluster for the calculations.
The R interface cannot support the GPU kernel easily on Windows due
to incompatibilities between the R supported compiler GCC and the CUDA
supported compiler Visual C++, while on Linux and OS X it can be built
easily.
7
Conclusions
Do we need another implementation of self-organizing maps? We believe the
answer is yes. Libraries in high-level languages do not have a scalable training
module, and even implementations on parallel and distributed architectures
could be improved on. Our solution scales from a single-thread execution
to massively parallel GPU-accelerated clusters using a common core. Batch
processing is encouraged by a command-line interface, whereas interactive
use is enabled by Python, R, and MATLAB interfaces. Even when called
interactively, execution is parallelized.
• Memory-efficient multicore execution.
22
• Purely MPI-based, highly efficient distributed implementation.
• Optimized hybrid CPU-GPU kernel for boosting training on dense
data.
• Sparse kernel.
• Common computational back-end for popular data-processing languages:
Python, R, and MATLAB.
• Windows, OS X, Linux support of the command-line interface, with
full parallel capability.
• Highly efficient on large emergent maps.
• Format compatibility with ESOM Tools for visualization.
As emergent maps are especially computationally intensive, we believe
that the primary usage scenario is a single-node, multi-GPU configuration for
data exploration, which was only feasible for small maps and small volumes
of data until now. Larger volumes of data are also easy to compute on a
cluster.
Acknowledgment
The first author was supported by the European Commission Seventh Framework Programme under Grant Agreement Number FP7-601138 PERICLES
and by the AWS in Education Machine Learning Grant award.
References
[1] Apache Software Foundation. Apache Lucene: Free and Open-Source
Information Retrieval Software Library, 2012. Version 3.6.2.
[2] Nathan Bell and Jared Hoberock. Thrust: A productivity-oriented library for CUDA. In GPU Computing Gems, volume 7. Morgan Kaufmann, 2011.
[3] Chih-Chung Chang and Chih-Jen Lin. libsvm: A Library for Support
Vector Machines, 2001.
23
[4] Trevor F. Cox and Michael A.A. Cox. Multidimensional Scaling. Chapman and Hall, 1994.
[5] Leonardo Dagum and Ramesh Menon. OpenMP: An industry standard
API for shared-memory programming. Computational Science & Engineering, 5(1):46–55, 1998.
[6] Dirk Eddelbuettel and Romain François. Rcpp: Seamless R and C++
integration. Journal of Statistical Software, 40(8):1–18, 2011.
[7] Michael Hanke, Yaroslav O Halchenko, Per B Sederberg, Stephen José
Hanson, James V Haxby, and Stefan Pollmann. PyMVPA: A Python
toolbox for multivariate pattern analysis of fMRI data. Neuroinformatics, 7(1):37–53, 2009.
[8] Markus Hofmann and Ralf Klinkenberg. RapidMiner: Data Mining Use
Cases and Business Analytics Applications. Chapman and Hall, 2013.
[9] Michael Kane, John W. Emerson, and Stephen Weston. Scalable strategies for computing with massive data. Journal of Statistical Software,
55(14):1–19, 2013.
[10] Teuvo Kohonen. Self-Organizing Maps. Springer-Verlag, 2001.
[11] Richard D. Lawrence, George S. Almasi, and Holly E. Rushmeier. A
scalable parallel algorithm for self-organizing maps with applications to
sparse data mining problems. Data Mining and Knowledge Discovery,
3(2):171–195, 1999.
[12] David D. Lewis. Reuters-21578 text categorization test collection distribution 1.0, 1999.
[13] Qi Li, Vojislav Kecman, and Raied Salman. A chunking method for
Euclidean distance matrix calculation on large dataset using multi-GPU.
In Proceedings of ICMLA-10, 9th International Conference on Machine
Learning and Applications, pages 208–213, Washington, DC, USA, 2010.
[14] Zhongwen Luo, Hongzhi Liu, Zhengping Yang, and Xincai Wu. Selforganizing maps computing on graphic process unit. In Proceedings of
ESANN-05, 13th European Symposium on Artificial Neural Networks,
2005.
24
[15] Daniel Müllner. fastcluster: Fast hierarchical, agglomerative clustering
routines for R and Python. Journal of Statistical Software, 53(9):1–18,
2013.
[16] NVidia Corporation. NVida Compute Unified Device Architecture Programming Guide 6.0, 2014.
[17] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2017.
[18] R Core Team. R Installation and Administration. R Foundation for
Statistical Computing, Vienna, Austria, 2017.
[19] Marc Snir, Steve W. Otto, Steven Huss-Lederman, David W. Walker,
and Jack Dongarra. MPI – The Complete Reference: The MPI Core,
volume 1. MIT Press, 1998.
[20] Seung-Jin Sul and Andrey Tovchigrechko. Parallelizing BLAST and
SOM algorithms with mapreduce-mpi library. In Proceedings of IPDPS11, 25th International Parallel and Distributed Computing Symposium,
pages 476–483, Anchorage, AK, USA, 2011.
[21] Joshua B. Tenenbaum, Vin de Silva, and John C. Langford. A global
geometric framework for nonlinear dimensionality reduction. Science,
290(5500):2319–2323, 2000.
[22] The MathWorks Inc. MATLAB: The language of technical computing,
version R2011b, 2011.
[23] Stefan Theußl, Ingo Feinerer, and Kurt Hornik. A tm plug-in for distributed text mining in R. Journal of Statistical Software, 51(5):1–31,
2012.
[24] Alfred Ultsch and Fabian Mörchen. ESOM-Maps: Tools for clustering,
visualization, and classification with emergent SOM. Technical report,
Data Bionics Research Group, University of Marburg, 2005.
[25] Guido van Rossum. Python Programming Language, 2014.
[26] Juha Vesanto, Johan Himberg, Esa Alhoniemi, and Juha Parhankangas.
Self-organizing map in MATLAB: The SOM Toolbox. Proceedings of the
MATLAB DSP Conference, pages 35–40, 1999.
25
[27] Ron Wehrens and Lutgarde M. C. Buydens. Self- and super-organizing
maps in R: The kohonen package. Journal of Statistical Software,
21(5):1–19, 2007.
[28] Kilian Q. Weinberger, Fei Sha, and Lawrence K. Saul. Learning a kernel
matrix for nonlinear dimensionality reduction. In Proceedings of ICML04, 21st International Conference on Machine Learning, pages 106–113,
Banff, Canada, 2004.
[29] Peter Wittek and Sándor Darányi. A GPU-accelerated algorithm for
self-organizing maps in a distributed environment. In Proceedings of
ESANN-12, 20th European Symposium on Artificial Neural Networks,
Computational Intelligence and Machine Learning, Bruges, Belgium,
2012.
26
| 9cs.NE
|
END-TO-END AUDIOVISUAL SPEECH RECOGNITION
Stavros Petridis1, Themos Stafylakis2, Pingchuan Ma1 , Feipeng Cai1
Georgios Tzimiropoulos2, Maja Pantic1
1
Dept. of Computing, Imperial College London, UK
Computer Vision Laboratory, University of Nottingham, UK
[email protected], [email protected]
arXiv:1802.06424v2 [cs.CV] 22 Feb 2018
2
ABSTRACT
Several end-to-end deep learning approaches have been recently presented which extract either audio or visual features
from the input images or audio signals and perform speech
recognition. However, research on end-to-end audiovisual
models is very limited. In this work, we present an end-toend audiovisual model based on residual networks and Bidirectional Gated Recurrent Units (BGRUs). To the best of our
knowledge, this is the first audiovisual fusion model which
simultaneously learns to extract features directly from the image pixels and audio waveforms and performs within-context
word recognition on a large publicly available dataset (LRW).
The model consists of two streams, one for each modality,
which extract features directly from mouth regions and raw
waveforms. The temporal dynamics in each stream/modality
are modeled by a 2-layer BGRU and the fusion of multiple
streams/modalities takes place via another 2-layer BGRU. A
slight improvement in the classification rate over an end-toend audio-only and MFCC-based model is reported in clean
audio conditions and low levels of noise. In presence of high
levels of noise, the end-to-end audiovisual model significantly
outperforms both audio-only models.
Index Terms— Audiovisual Speech Recognition, Residual Networks, End-to-End Training, BGRUs, Audiovisual
Fusion
1. INTRODUCTION
Traditional audiovisual fusion systems consist of two stages,
feature extraction from the image and audio signals and combination of the features for joint classification [1, 2, 3]. Recently, several deep learning approaches for audiovisual fusion have been presented which aim to replace the feature extraction stage with deep bottleneck architectures. Usually a
transform, like principal component analysis (PCA), is first
applied to the mouth region of interest (ROI) and spectrograms or concatenated Mel-Frequency Cepstral Coefficients
ACCEPTED TO ICASSP 2018
(MFCCs) and a deep autoencoder is trained to extract bottleneck features [4, 5, 6, 7, 8, 9]. Then these features are fed to
a classifier like a support vector machine or a Hidden Markov
Model.
Few works have been presented very recently which
follow an end-to-end approach for visual speech recognition. The main approaches followed can be divided into two
groups. In the first one, fully connected layers are used to extract features and LSTM layers model the temporal dynamics
of the sequence [10, 11]. In the second group, a 3D convolutional layer is used followed either by standard convolutional
layers [12] or residual networks (ResNet) [13] combined with
LSTMs or GRUs. End-to-end approaches have also been
successfully used for speech emotion recognition using 1D
CNNs and LSTMs [14].
However, work on end-to-end audiovisual speech recognition has been very limited. To the best of our knowledge,
there are only two works which perform end-to-end training
for audiovisual speech recognition [15, 16]. In the former, an
attention mechanism is applied to both the mouth ROIs and
MFCCs and the model is trained end-to-end. However, the
system does not use the raw audio signal or spectrogram but
relies on MFCC features. In the latter, fully connected layers together with LSTMs are used in order to extract features
directly from raw images and spectrograms and perform classification on the OuluVS database [17].
In this paper, we extend the work of [10], which mainly
works for small databases, using ResNets as proposed in
[13]. To the best of our knowledge, this is the first endto-end model which performs audiovisual word recognition
from raw mouth ROIs and waveforms on a large in-the-wild
database. The proposed model consists of two streams, one
per modality, which extract features directly from the raw
images and waveforms, respectively. Each stream consists of
a ResNet which extracts features from the raw inputs. This
is followed by a 2-layer BGRU network which models the
temporal dynamics in each stream. Finally, the information of
the different streams/modalities is fused via another 2-layer
BGRU which models the joint temporal dynamics. A similar
architecture has been proposed by [18] for audiovisual emo-
Target
Classes
Softmax (500)
BGRU (1024)
BGRU (1024)
(a)
(b)
Fig. 1. Example of mouth ROI extraction.
tion recognition. The main differences of our work are the
following: 1) we use a ResNet for the audio stream instead
of a rather shallow 2-layer CNN, 2) we do not use a pretrained ResNet for the visual stream but we train a ResNet
from scratch, 3) we use BGRUs in each stream which help
modeling the temporal dynamics of each modality instead of
using just one BLSM layer at the top and 4) we use a training
procedure which allows for efficient end-to-end training of
the entire network.
We perform classification of 500 words from the LRW
database achieving state-of-the-art performance for audiovisual fusion. The proposed system results in an absolute increase of 0.3% in classification accuracy over the end-to-end
audio-only model and an MFCC-based system. The end-toend audiovisual fusion model also significantly outperforms
(up to 14.1% absolute improvement) the audio-only models
under high levels of noise.
2. LRW DATABASE
For the purposes of this study we use the Lip Reading in the
Wild (LRW) database [19] which is the largest publicly available lipreading dataset in the wild. The database consists of
short segments (1.16 seconds) from BBC programs, mainly
news and talk shows. It is a very challenging set since it contains more than 1000 speakers and large variation in head pose
and illumination. The number of words, 500, is also much
higher than existing lipreading databases used for word recognition, which typically contain 10 to 50 words [20, 21, 17].
Another characteristic of the database is the presence of
several words which are visually similar. For example, there
are words which are present in their singular and plural forms
or simply different forms of the same word, e.g., America and
American. We should also emphasise that words appear in the
middle of an utterance and there may be co-articulation of the
lips from preceding and subsequent words.
3. END-TO-END AUDIOVISUAL SPEECH
RECOGNITION
The proposed deep learning system for audiovisual fusion is
shown in Fig. 2. It consists of two streams which extract
BGRU (1024)
BGRU (1024)
BGRU (1024)
BGRU (1024)
ResNet-34
ResNet-18
3D Conv.
Image Sequence
Audio Waveform
Fig. 2. Overview of the end-to-end audiovisual speech recognition system. Two streams are used for feature extraction
directly from the raw images and audio waveforms. The temporal dynamics are modelled by BGRUs in each stream. The
top two BGRUs fuse the information of the audio and visual
streams and jointly model their temporal dynamics.
features directly from the raw input images and the audio
waveforms, respectively. Each stream consists of two parts:
a residual network (ResNet) [22] which learns to automatically extract features from the raw image and waveform,
respectively and a 2-layer BGRU which models the temporal
dynamics of the features in each stream. Finally, 2 BGRU
layers on top of the two streams are used in order to fuse the
information of the audio and visual streams.
3.1. Visual Stream
The visual stream is similar to [13] and consists of a spatiotemporal convolution followed by a 34-layer ResNet and
a 2-layer BGRU. A spatiotemporal convolutional layer is capable of capturing the short-term dynamics of the mouth region and is proven to be advantageous, even when recurrent
networks are deployed for back-end [12]. It consists of a
convolutional layer with 64 3D kernels of 5 by 7 by 7 size
(time/width/height), followed by batch normalization and rectified linear units.
We use the 34-layer identity mapping version, which was
proposed for ImageNet [23]. The ResNet drops progressively
the spatial dimensionality until its output becomes a single dimensional tensor per time step. We should emphasize that we
did not make use of pretrained models, as they are optimized
for completely different tasks (e.g. static colored images from
ImageNet or CIFAR). Finally, the output of ResNet-34 is fed
to a 2-layer BGRU which consists of 1024 cells in each layer.
3.2. Audio Stream
The audio stream consists of an 18-layer ResNet followed by
two BGRU layers. There is no need to use a spatiotemporal
convolution front-end in this case as the audio waveform is an
1D signal. We use the standard architecture for the ResNet18 with the main difference being that we use 1D instead of
2D kernels which are used for image data. A temporal kernel
of 5ms with a stride of 0.25ms is used in the first convolutional layer in order to extract fine-scale spectral information.
The output of the ResNet is divided into 29 frames/windows
using average pooling in order to ensure the same frame rate
as the video is used. These audio frames are then fed to the
following ResNet layers which consist of the default kernels
of size 3 by 1 so deeper layers extract long-term speech characteristics. The output of the ResNet-18 is fed to a 2-layer
BGRU which consists of 1024 cells in each layer (using the
same architecture as in [13]).
3.3. Classification Layers
The BGRU outputs of each stream are concatenated and fed to
another 2-layer BGRU in order to fuse the information from
the audio and visual streams and jointly model their temporal
dynamics. The output layer is a softmax layer which provides
a label to each frame. The sequence is labeled based on the
highest average probability.
4. EXPERIMENTAL SETUP
4.1. Preprocessing
Video: The first step is the extraction of the mouth region of
interest (ROI). Since the mouth ROIs are already centered, a
fixed bounding box of 96 by 96 is used for all videos as shown
in Fig. 1. Finally, the frames are transformed to grayscale and
are normalized with respect to the overall mean and variance.
Audio: Each audio segment is z-normalised, i.e., has zero
mean and standard deviation one to account for variations in
different levels of loudness between the speakers.
4.2. Evaluation Protocol
The video segments are already partitioned into training, validation and test sets. There are between 800 and 1000 sequences for each word in the training set and 50 sequences
in the validation and test sets, respectively. In total there are
488766, 25000, and 25000 examples in the training, validation and test sets, respectively.
4.3. Training
Training is divided into 2 phases: first the audio/visual
streams are trained independently and then the audiovisual
network is trained end-to-end. During training data augmentation is performed on the video sequences of mouth ROIs.
This is done by applying random cropping and horizontal
flips with probability 50% to all frames of a given clip. Data
augmentation is also applied to the audio sequences. During
training babble noise at different levels (between -5 dB to 20
db) might be added to the original audio clip. The selection
of one of the noise levels or the use of the clean audio is done
using a uniform distribution.
4.3.1. Single Stream Training
Initialisation: First, each stream is trained independently.
Directly training end-to-end each stream leads to suboptimal
performance so we follow the same 3-step procedure as in
[13]. Initially, a temporal convolutional back-end is used instead of the 2-layer BGRU. The combination of ResNet and
temporal convolution (together with a softmax output layer)
is trained until there is no improvement in the classification
rate on the validation set for more than 5 epochs. Then the
temporal convolutional back-end is removed and the BGRU
back-end is attached. The 2-layer BGRU (again with a sotfmax output layer) is trained for 5 epochs, keeping the weights
of the 3D convolution front-end and the ResNet fixed.
End-to-End Training: Once the ResNet and the 2-layer
BGRU in each stream have been pretrained then they are put
together and the entire stream is trained end-to-end (using a
softmax output layer). The Adam training algorithm [24] is
used for end-to-end training with a mini-batch size of 36 sequences and an initial learning rate of 0.0003. Early stopping
with a delay of 5 epochs was also used.
4.3.2. Audiovisual Training
Initialisation: Once the single streams have been trained then
they are used for initialising the corresponding streams in
the multi-stream architecture. Then another 2-layer BGRU is
added on top of all streams in order to fuse the single stream
outputs. The top BGRU is first trained for 5 epochs (with a
softmax output layer), keeping the weights of the audio and
visual streams fixed.
End-to-End Training: Finally, the entire audiovisual network is trained jointly using Adam with a mini-batch size of
18 sequences and an initial learning rate of 0.0001. Early
stopping is also applied with a delay of 5 epochs.
5. RESULTS
Results are shown in Table 1. We report the performance of
the end-to-end audio-only, visual-only and audiovisual models. For comparison purposes, since there are no previous
100
A (End-to-End)
A (MFCC)
V (End-to-End)
V [13]*
V [15]
V [19]
A + V (End-to-End)
97.7
97.7
82.0
83.0
76.2
61.1
98.0
audio/audiovisual results on the LRW database we also compute the performance of a 2-layer BGRU network trained with
MFCC features which are the standard features for acoustic
speech recognition. We use 13 coefficients (and their deltas)
using a 40ms window and a 10ms step. The network is trained
in the same way as the BGRU networks in section 4.3 with
the only difference that it was trained for longer using early
stopping.
The end-to-end audio system results in a similar performance to MFCCs which is a significant result given that the
input to the system is just the raw waveform. However, we
should note that the effort required in order to train the endto-end system is significantly higher than the 2-layer BGRU
used with MFCCs. The end-to-end audiovisual system leads
to a small improvement over the audio-only models of 0.3%.
This is expected since the contribution of the visual modality
is usually marginal in clean audio conditions as reported in
previous works as well [1, 16].
In order to investigate the robustness to audio noise of the
audiovisual fusion approach we run experiments under varying noise levels. The audio signal for each sequence is corrupted by additive babble noise from the NOISEX database
[25] so as the SNR varies from -5 dB to 20 dB.
Results for the audio, visual and audiovisual models under noisy conditions are shown in Fig. 3. The video-only
classifier (blue solid line) is not affected by the addition of
the audio noise and therefore its performance remains constant over all noise levels. On the other hand, as expected, the
performance of the audio classifier (red dashed line) is significantly affected. Similarly, the performance of the MFCC
classifier (purple solid line) is also significantly affected by
noise. It is interesting to point out that although the MFCC
and end-to-end audio models result in the same performance
when audio is clean or under low levels of noise (10 to 20
dB), the end-to-end audio model results in much better performance under high levels of noise (-5 dB to 5 dB). It results
95
90
85
CR
Table 1. Classification Rate (CR) of the Audio-only (A),
Video-only (V) and audiovisual models (A + V) on the LRW
database. *This is a similar end-to-end model which uses a
different mouth ROI, computed based on tracked facial landmarks, in each video. In this work, we use a fixed mouth ROI
for all videos.
Stream
CR
80
75
70
V
A
AV
MFCC
65
60
-5
0
5
10
15
20
SNR (dB)
Fig. 3. Classification Rate (CR) as a function of the noise
level. A: End-to-End audio model. V: End-to-End visual
model, AV: End-to-End audiovisual model. MFCC: A 2-layer
BGRU trained with MFCCs.
in an absolute improvement of 0.9%, 3.5% and 7.5% over the
MFCC classifier, at 5 dB, 0 dB and -5 dB, respectively.
The audiovisual model (yellow dotted line) is more robust to audio noise than the audio-only models. It performs
slightly better under low noise levels (10 dB to 20 dB) but
it significantly outperforms both of them under high noise
levels (-5 dB to 5 dB). In particular, it leads to an absolute
improvement of 1.3%, 3.9% and 14.1% over the end-to-end
audio-only model at 5 dB, 0 dB and -5 dB, respectively.
6. CONCLUSION
In this work, we present an end-to-end visual audiovisual fusion system which jointly learns to extract features directly
from the pixels and audio waveforms and performs classification using BGRUs. Results on the largest publicly available database for within-context word recognition in the wild
show that the end-to-end audiovisual model slightly outperforms a standard MFCC-based system under clean conditions
and low levels of noise. It also significantly outperforms the
end-to-end and MFCC-based audio models in the presence of
high levels of noise. A natural next step would be to extend
the system in order to be able to recognise sentences instead
of isolated words. Finally, it would also be interesting to investigate in future work an adaptive fusion mechanism which
learns to weight each modality based on the noise levels.
7. ACKNOWLEDGEMENTS
This work has been funded by the European Community
Horizon 2020 under grant agreement no. 645094 (SEWA).
Themos Stafylakis has been partly funded by the European
Commission program Horizon 2020, under grant agreement
no. 706668 (Talking Heads).
8. REFERENCES
[1] G. Potamianos, C. Neti, G. Gravier, A. Garg, and A. W.
Senior, “Recent advances in the automatic recognition
of audiovisual speech,” Proceedings of the IEEE, vol.
91, no. 9, pp. 1306–1326, Sept 2003.
[2] S. Dupont and J. Luettin, “Audio-visual speech modeling for continuous speech recognition,” IEEE Trans. on
Multimedia, vol. 2, no. 3, pp. 141–151, 2000.
[3] S. Petridis and M. Pantic, “Prediction-based audiovisual
fusion for classification of non-linguistic vocalisations,”
IEEE Transactions on Affective Computing, vol. 7, no.
1, pp. 45–58, 2016.
[4] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y
Ng, “Multimodal deep learning,” in Proc. of ICML,
2011, pp. 689–696.
[5] D. Hu, X. Li, and X. Lu, “Temporal multimodal learning in audiovisual speech recognition,” in IEEE CVPR,
2016, pp. 3574–3582.
[6] H. Ninomiya, N. Kitaoka, S. Tamura, Y. Iribe, and
K. Takeda, “Integration of deep bottleneck features for
audio-visual speech recognition,” in Interspeech, 2015.
[7] Y. Mroueh, E. Marcheret, and V. Goel, “Deep multimodal learning for audio-visual speech recognition,” in
IEEE ICASSP, 2015, pp. 2130–2134.
[8] Y. Takashima, R. Aihara, T. Takiguchi, Y. Ariki, N. Mitani, K. Omori, and K. Nakazono, “Audio-visual speech
recognition using bimodal-trained bottleneck features
for a person with severe hearing loss,” Interspeech, pp.
277–281, 2016.
[9] S. Petridis and M. Pantic, “Deep complementary bottleneck features for visual speech recognition,” in ICASSP,
2016, pp. 2304–2308.
[10] S. Petridis, Z. Li, and M. Pantic, “End-to-end visual
speech recognition with LSTMs,” in IEEE ICASSP,
2017, pp. 2592–2596.
[11] M. Wand, J. Koutnik, and J. Schmidhuber, “Lipreading
with long short-term memory,” in IEEE ICASSP, 2016,
pp. 6115–6119.
[12] Y. M. Assael, B. Shillingford, S. Whiteson, and
N. de Freitas, “Lipnet: Sentence-level lipreading,” arXiv
preprint arXiv:1611.01599, 2016.
[13] T. Stafylakis and G. Tzimiropoulos, “Combining residual networks with LSTMs for lipreading,” in Interspeech, 2017, vol. 9, pp. 3652–3656.
[14] G. Trigeorgis, F. Ringeval, R. Brueckner, E. Marchi,
M. Nicolaou, B. Schuller, and S. Zafeiriou, “Adieu
features? end-to-end speech emotion recognition using a deep convolutional recurrent network,” in IEEE
ICASSP, 2016, pp. 5200–5204.
[15] J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman,
“Lip reading sentences in the wild,” IEEE CVPR, 2017.
[16] S. Petridis, Y. Wang, Z. Li, and M. Pantic, “End-toend audiovisual fusion with LSTMs,” in Auditory-Visual
Speech Processing Conference, 2017.
[17] I. Anina, Z. Zhou, G. Zhao, and M. Pietikäinen,
“Ouluvs2: A multi-view audiovisual database for nonrigid mouth motion analysis,” in IEEE FG, 2015, pp.
1–5.
[18] P. Tzirakis, G. Trigeorgis, M. A. Nicolaou, B. Schuller,
and S. Zafeiriou, “End-to-end multimodal emotion
recognition using deep neural networks,” IEEE Journal
of Selected Topics in Signal Processing 2017.
[19] J. S. Chung and A. Zisserman, “Lip reading in the wild,”
in ACCV. Springer, 2016, pp. 87–103.
[20] M. Cooke, J. Barker, S. Cunningham, and X. Shao, “An
audio-visual corpus for speech perception and automatic
speech recognition,” The Journal of the Acoustical Society of America, vol. 120, no. 5, pp. 2421–2424, 2006.
[21] E. Patterson, S. Gurbuz, Z. Tufekci, and J. Gowdy,
“Moving-talker, speaker-independent feature study, and
baseline results using the CUAVE multimodal speech
corpus,” EURASIP J. Appl. Signal Process., vol. 2002,
no. 1, pp. 1189–1201, Jan. 2002.
[22] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual
learning for image recognition,” in IEEE CVPR, 2016,
pp. 770–778.
[23] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings
in deep residual networks,” in ECCV. Springer, 2016,
pp. 630–645.
[24] D. Kingma and J. Ba, “Adam: A method for stochastic
optimization,” arXiv preprint arXiv:1412.6980, 2014.
[25] A. Varga and H. Steeneken, “Assessment for automatic
speech recognition: Ii. NOISEX-92: A database and
an experiment to study the effect of additive noise on
speech recognition systems,” Speech communication,
vol. 12, no. 3, pp. 247–251, 1993.
| 1cs.CV
|
New Algorithms for Heavy Hitters in Data Streams ∗
arXiv:1603.01733v1 [cs.DS] 5 Mar 2016
David P. Woodruff
IBM Research Almaden
[email protected]
Abstract
An old and fundamental problem in databases and data streams is that of finding the heavy
hitters, also known as the top-k, most popular items, frequent items, elephants, or iceberg
queries. There are several variants of this problem, which quantify what it means for an item
to be frequent, including what are known as the ℓ1 -heavy hitters and ℓ2 -heavy hitters. There
are a number of algorithmic solutions for these problems, starting with the work of Misra and
Gries, as well as the CountMin and CountSketch data structures, among others.
In this survey paper, accompanying an ICDT invited talk, we cover several recent results
developed in this area, which improve upon the classical solutions to these problems. In particular, with coauthors we develop new algorithms for finding ℓ1 -heavy hitters and ℓ2 -heavy
hitters, with significantly less memory required than what was known, and which are optimal
in a number of parameter regimes.
1
The Heavy Hitters Problem
A well-studied problem in databases and data streams is that of finding the heavy hitters, also
known as the top-k, most popular items, frequent items, elephants, or iceberg quries. These
can be used for flow identification at IP routers [21], in association rules and frequent itemsets
[1, 25, 26, 44, 47], and for iceberg queries and iceberg datacubes [7, 22, 24]. We refer the reader
to the survey [18], which presents an overview of known algorithms for this problem, from both
theoretical and practical standpoints.
There are various different flavors of guarantees for the heavy hitters problem. We start with
what is known as the ℓ1 -guarantee:
Definition 1 (ℓ1 -(ǫ, φ)-Heavy Hitters Problem) In the (ǫ, φ)-Heavy Hitters Problem, we are
given parameters 0 < ǫ < φ < 1, as well as a stream a1 , . . . , am of items aj ∈ {1, 2, . . . , n}. Let fi
denote the number of occurrences of item i, i.e., its frequency. The algorithm should make one pass
over the stream and at the end of the stream output a set S ⊆ {1, 2, . . . , n} for which if fi ≥ φm,
then i ∈ S, while if fi ≤ (φ − ǫ)m, then i ∈
/ S. Further, for each item i ∈ S, the algorithm should
output an estimate f˜i of the frequency fi which satisfies |fi − f˜i | ≤ ǫm.
We are interested in algorithms which use as little space (i.e., memory) in bits as possible to
solve the ℓ1 -(ǫ, φ)-Heavy Hitters Problem. We allow the algorithm to be randomized and to
succeed with probability at least 1 − δ, for 0 < δ < 1. We do not make any assumption on the
∗
A preliminary version of this paper appeared as an invited paper in ICDT 2016.
1
ordering of the stream a1 , . . . , am . This is desirable, as often in applications one cannot assume a
best-case or even a random order. We will assume m is known in advance, though many of the
algorithms below (including ours) can deal with unknown m.
The first algorithm for the ℓ1 -(ǫ, φ)-Heavy Hitters Problem was given by Misra and Gries
[39], who achieved O(ǫ−1 log n) bits of space for any φ > 2ǫ. This algorithm was rediscovered
by Demaine et al. [20], and again by Karp et al. [32]. Other than these algorithms, which are
deterministic, there are a number of randomized algorithms, such as the CountSketch [15], CountMin sketch [19], sticky sampling [35], lossy counting [35], space-saving [37], sample and hold [21],
multi-stage bloom filters [13], and sketch-guided sampling [33]. Berinde et al. [6] show that using
O(kǫ−1 log(mn)) bits of space, one can achieve the stronger guarantee of reporting, for each item
res(k)
res(k)
, where F1
< m denotes the sum of frequencies of items
i ∈ S, f˜i with |f˜i − fi | ≤ kǫ F1
in {1, 2, . . . , n} excluding the frequencies of the k most frequent items. This is particularly useful
res(k)
will depend only on the
when there are only a few large frequencies, since then the error kǫ F1
remaining small frequencies.
While the ℓ1 -heavy hitters have a number of applications, there is also a sometimes stronger
notion known as the ℓ2 -heavy hitters, which we now define.
Definition 2 (ℓ2 -(ǫ, φ)-Heavy Hitters Problem) In the (ǫ, φ)-Heavy Hitters Problem, we are
given parameters 0 < ǫ < φ < 1, as well as a stream a1 , . . . , am of items aP
j ∈ {1, 2, . . . , n}. Let fi
denote the number of occurrences of item i, i.e., its frequency. Let F2 = ni=1 fi2 . The algorithm
should make one pass over the stream and at the end of the stream output a set S ⊆ {1, 2, . . . , n}
for which if fi2 ≥ φF2 , then i ∈ S, while if fi2 ≤ (φ − ǫ)F2 , then i ∈
/ S. Further, for each item
√ i ∈ S,
the algorithm should output an estimate f˜i of the frequency fi which satisfies |fi − f˜i | ≤ ǫ F2 .
One of the algorithms for ℓ1 -heavy hitters mentioned above, the CountSketch [16], refined in
[46], actually solves the ℓ2 -(ǫ, φ)-Heavy Hitters Problem. Notice that this guarantee can be
significantly stronger than the aforementioned ℓ1 -guarantee that fi ≥ ǫm. Indeed, if fi ≥ φm, then
fi2 ≥ φ2 m2 ≥ φ2 F2 . So, an algorithm for finding the ℓ2 -heavy hitters, with φ replaced by φ2 , will
find all items satisfying the ℓ1 -guarantee with parameter φ. On the other hand, given a stream
√
of n distinct items in which fi∗ = n for an i∗ ∈ [n] = {1, 2, 3, . . . , n}, yet fi = 1 for all i 6= i∗ ,
an algorithm satisfying the ℓ2 -heavy hitters guarantee will identify item i with constant φ, but an
√
√
algorithm which only has the ℓ1 -guarantee would need to set φ = 1/ n, therefore using Ω( n) bits
of space. In fact, ℓ2 -heavy hitters are in some sense the best one can hope for with a small amount
of space in a data stream, as it is known for p > 2 that finding those i for which fip ≥ φFp requires
n1−2/p bits of space even for constant φ [5, 14].
The ℓ2 -heavy hitter algorithms of [16, 46] have broad applications in compressed sensing [23,
38, 43] and numerical linear algebra [10, 17, 36, 41], and are often used as a subroutine in other
data stream algorithms, such as ℓp -sampling [3, 30, 40], cascaded aggregates [29], and frequency
moments [9, 28].
Given the many applications of heavy hitters, it is natural to ask what the best space complexity
for them is. For simplicity of presentation, we make the common assumption that the stream length
m is polynomially related to the universe size n.
It is clear that for constant ǫ and φ, that there is an Ω(log n) bit lower bound, as this is just
the number of bits needed to specify the identity of the heavy hitter.
For constant ǫ, given the aforementioned results, this is actually tight for the ℓ1 -(ǫ, φ)-Heavy
Hitters Problem. The main focus then, for the ℓ1 -(ǫ, φ)-Heavy Hitters Problem is on obtaining
2
tight bounds as a function of ǫ and φ.
On the other hand, for the ℓ2 -(ǫ, φ)-Heavy Hitters Problem, even for constant ǫ and φ, the
best previous algorithms of [16] and the followup [46] achieve Θ(log2 n) bits of space. It is known
that if one allows deletions in the stream, in addition to insertions, then Θ(log2 n) bits of space is
optimal [4, 30]. However, in many cases we just have a stream of insertions, such as in the model
studied in the seminal paper of Alon, Matias, and Szegedy [2]. Thus, for the ℓ2 -(ǫ, φ)-Heavy
Hitters Problem, our focus will be on the regime of constant ǫ and φ and on understanding the
dependence on n.
There are a number of other desirable properties one would want out of a heavy hitters algorithm. For instance, one is often also interested in minimizing the update time and reporting time
of such algorithms. Here, the update time is defined to be the time the algorithm needs to update
its data structure when processing a stream insertion. The reporting time is the time the algorithm
needs to report the answer after having processed the stream. In this article we will focus primarily
on the space complexity. For other very interesting recent work on improving the reporting time
in a stream of insertions and deletions, see [34]. The results in this survey are focused on a stream
of insertions only (for which, as mentioned above, smaller space bounds are possible).
2
Our Recent Results
In several recent works [8, 11, 12], we significantly improve known algorithms for finding both
ℓ1 -heavy hitters as well as ℓ2 -heavy hitters. For many settings of parameters, our algorithms are
optimal.
2.1
ℓ1 -Heavy Hitters
In joint work with Bhattacharyya and Dey [8], we improve upon the basic algorithm of Misra and
Gries [39] for the ℓ1 -(ǫ, φ)-Heavy Hitters Problem, the latter achieving O(ǫ−1 log n) bits of space
for any φ ≥ 2ǫ. There are two algorithms of [8], the first a bit simpler and already achieving a large
improvement over [39], and the second an optimal algorithm. We first discuss the first algorithm.
We first recall the algorithm of Misra and Gries. That algorithm initializes a table of 1/ǫ + 1
pairs of (v, c) to (⊥, 0), where v is an element in the universe {1, 2, . . . , n} ∪ ⊥, and c is a nonnegative integer. When receiving a new stream insertion ai , the algorithm checks if v = ai for some
(v, c) pair in the table. If so, it replaces (v, c) with (v, c + 1). Otherwise, if there is a (v, c) in the
table with v = ⊥, then the algorithm replaces that (v, c) pair with (ai , 1). If neither of the previous
two cases hold, the algorithm takes each (v, c) pair in the table, and replaces it with (v, c − 1). If
c − 1 = 0, then the corresponding v is replaced with ⊥.
Note that the algorithm, as described in the previous paragraph, naturally can be implemented
using O(ǫ−1 log n) bits of space (recall we assume the stream length m and the universe size n
are polynomially related, so log m = Θ(log n)). Moreover, a nice property is that the algorithm is
deterministic.
For the correctness, note that if an item i occurs fi ≥ 2ǫm times, then it will appear in the
table at the end of the stream. Indeed, notice that for each occurrence of i in the stream, if it is
not included in the table via the operation of replacing a pair (i, c) with (i, c + 1) for some value
of c, or replacing a pair (⊥, 0) with (i, 1), then this means that there were at least 1/ǫ + 1 stream
updates that were removed from the table upon seeing this occurrence of i, since each counter c
3
for each (v, c) pair in the table is decremented by 1. We can therefore charge those stream updates
to this occurrence of i. Moreover, if (i, c) is in the table for some value of c and is replaced with
(i, c − 1) or (⊥, 0), this means we can charge at least 1/ǫ stream updates to items not equal to i
to this occurrence of i. Since we are charging distinct stream updates for each occurrence of i, we
have the relationship that fi · (1/ǫ) ≤ m, which is a contradiction to fi ≥ 2ǫm. Therefore, i will
occur in a pair in the table at the end of the stream. The same analysis in fact implies that at most
ǫm occurrences of i will not be accounted for in the table at the end of the stream, which means
that for the (i, c) pair in the table, we have fi ≥ c ≥ fi − ǫm. This latter guarantee enables us to
solve the ℓ1 -(ǫ, φ)-Heavy Hitters Problem for any φ ≥ 2ǫ.
One shortcoming of the algorithm above is that if φ is much larger than ǫ, say φ is constant,
then the above algorithm still requires O(ǫ−1 log n) bits of space, that is, it is insensitive to the
value of φ. Consider for instance, the case when ǫ = 1/ log n and φ = 1/10, so one wants a very
high accuracy estimate to each of the item frequencies for items occurring at least 10% of the time.
The above algorithm would use O(log2 n) bits of space for this problem. In this case, the only
known lower bound is Ω(log n) bits, which just follows from the need to return the identities of the
heavy hitters. Is it possible to improve this O(log2 n) bits of space upper bound?
This is precisely what we show in [8]. Here we sketch how to achieve a bound of O((1/φ) log n +
(1/ǫ) log(1/ǫ)) bits of space and refer to [8] for further optimizations as well as extensions to related
problems. Note that this translates to a space bound of O(log n log log n) bits for the above setting
of parameters.
The first observation is that if we randomly sample r = Θ(1/ǫ2 ) stream updates, then with
probability 99%, simultaneously for every universe item i, if we let fˆi denote its frequency among
the samples, and fi its frequency in the original stream, then we have
fi
ǫ
fˆi
−
≤ .
r
m
2
This follows by Chebyshev’s inequality and a union bound. Indeed, consider a given i ∈ [n] with
frequency fi and suppose we sample each of its occurrences pairwise-independently with probability
r/m, for a parameter r. Recall that pairwise independence here implies that any single occurrence
is sampled with probability r/m and any two occurrences are jointly sampled with probability
exactly r 2 /m2 , though we do not impose any constraints on the joint distribution of any three or
more samples. Also, a pairwise independent hash function can be represented with only O(log n)
bits of space. Then the expected number E[fˆi ] of sampled occurrences is fi · r/m and the variance
Var[fˆi ] is fi · r/m(1 − r/m) ≤ fi r/m (here we use pairwise independence to conclude the same
variance bound as if the samples were fully independent). Applying Chebyshev’s inequality,
h
4fi r
rǫ i Var[fˆi ]
≤
≤
.
Pr fˆi − E[fˆi ] ≥
2
(rǫ/2)2
mr 2 ǫ2
4fi
. By the union bound, if we
Setting r = ǫC2 for a constant C > 0 makes this probability at most Cm
r
sample each element in the stream independently with probability m
, then the probability there
Pn 4fi
4
1
exists an i for which |fˆi − E[fˆi ]| ≥ rǫ
i=1 Cm ≤ C , which for C ≥ 400 is at most 100 ,
2 is at most
as desired.
After sampling so that the stream length is reduced to O(1/ǫ2 ), it follows that the number
of distinct items in the stream is also O(1/ǫ2 ), and therefore if we hash the item identifiers to a
4
universe of size O(1/ǫ4 ), by standard arguments with probability 99% the items will be perfectly
hashed, that is, there will be no collisions. This follows even with a pairwise-independent hash
function h. The high level idea then is to run the algorithm of Misra and Gries, but the pairs (v, c)
correspond to the hashed item identity and the count in the sampled stream, respectively. Notice
that it takes only O(log(1/ǫ)) bits to represent such pairs and so the algorithm of Misra and Gries
would take O(ǫ−1 log(1/ǫ)) bits of space.
However, we still want to return the actual item identifiers! To do this, we maintain a parallel
data structure containing actual item identifiers in [n], but the data structure only contains O(1/φ)
items. In particular, these item identities correspond to the items v for which (h(v), c) is stored in
the algorithm of Misra and Gries, for which the c values are largest. Namely, the items with top 1/φ
c-values have their actual identities stored. This can be maintained under stream insertions since
given a new stream update, one has the actual identity in hand, and therefore can appropriately
update the identities of the items with top O(1/φ) counts. Moreover, when we subtract one from
all counters in the algorithm of Misra and Gries, the only thing that changes in the top O(1/φ)
identities is that some of them may now have zero frequency, and so can be thrown out. Thus, we
can always maintain the actual top O(1/φ) identities in the original (before hashing) universe.
The second algorithm of [8] achieves an optimal O(ǫ−1 log φ−1 + φ−1 log n + log log m) bits of
space. The algorithm can be seen as an extension of our first algorithm. The idea of the optimal
algorithm, as in our first algorithm, is to have a list of the top O(1/φ)-heavy hitters with exact
identities, and to use a separate data structure to approximate their individual frequencies up to
ǫ · m. In the earlier algorithm, this was an accompanying Misra-Gries data structure on the hashed
universe identities and sample count values; in the new one we optimize this data structure to use
O(ǫ−1 log φ−1 ) bits instead of the earlier O(ǫ−1 log ǫ−1 ) bits. We have O(1/ǫ) counts, as before,
but now in each count we spend O(1) bits on average. We also eliminate the need to maintain
hashed identities in the earlier algorithm by partitioning the items into O(1/ǫ) buckets using a hash
function and maintaining the approximate sum in each bucket. We note that the counts need to
be randomized, but in a different sense than probabilistic counters since we want them to achieve
additive error O(1/ǫ) rather than the relative error guarantee of probabilistic counters. We call
these accelerated counters since their relative error improves as the count gets larger. We are able to
compress the counts since they sum up to O(1/ǫ2 ), which is the length of the sampled stream. Each
count is individually only correct with constant probability, so we have O(log(1/φ)) repetitions and
take a median across the repetitions to get a correct count for each of the O(1/φ) heavy hitters in
our list.
We refer the reader to [8] for further details about both algorithms.
2.2
ℓ2 -Heavy Hitters
In joint work with Braverman, Chestnut, and Ivkin [12], we improve upon the CountSketch data
structure [16] for the ℓ2 -(ǫ, φ)-Heavy Hitters Problem. To illustrate the algorithm of [12], we
consider ǫ and φ to be constants in what follows, and further, we suppose there is only a single
i∗ ∈ [n] for which fi2∗ ≥ φF2 and there is no i for which (φ − ǫ)F2 ≤ fi2 < φF2 . It is not hard to
reduce to this case by first hashing into O(1) buckets (recall φ, ǫ are constants for this discussion),
since the O(1/φ) heavy hitters will go to separate buckets with large constant probability (if, say,
we have Ω(1/φ2 ) buckets). Thus, we focus on this case. In this case the CountSketch algorithm
would use Θ(log2 n) bits of space, whereas in [12] we achieve O(log n log log n) bits of space, nearly
matching the trivial Ω(log n) bit lower bound.
5
We first explain the CountSketch data structure. The idea is to assign each item i ∈ [n] a random
sign σ(i) ∈ {−1, 1}. WePalso randomly partition [n] into B buckets via a hash function h and
maintain a counter cj = i|h(i)=j σ(i) · fi in the j-th bucket. Then, to estimate any given frequency
P
fi , we estimate it as σ(i) · ch(i) . Note that E[σ(i) · ch(i) ] = E[σ(i)2 fi + j6=i,h(j)=h(i) fj σ(j)σ(i)] =
fi , using that E[σ(i)σ(j)] = 0 for i 6= j. Moreover, by computing the variance and applying
Chebyshev’s inequality, one has that
p
|σ(i) · ch(i) − fi | = O( F2 /B)
with probability at least 9/10. The intuitive explanation is that due to the random sign combination
of remaining items in the same hash bucket as i, the absolute value of this linear combination
concentrates to the Euclidean norm of the frequency vector of these items. The idea then is to
repeat this independently O(log n) times in parallel. Then we estimate fi by taking the median
of the estimates across each of the O(log n) repetitions. By Chernoff bounds,
p we have that with
2
probability 1 − 1/n , say, the resulting estimate is within an additive O( F2 /B) of the true
frequency fi . This then holds for every i ∈ [n] simultaneously by a union bound, at which point
one can then find the ℓ2 -heavy hitters, if say, one sets B = Θ(1/ǫ2 ).
Notice that it is easy to maintain the CountSketch data structure in a data stream since we
just need to hash the new item i to the appropriate bucket and add σ(i) to the counter in that
bucket, once for each of the O(log n) repetitions. The total space complexity of the CountSketch
algorithm is O(B · log2 n), where the “B” is the number of hash buckets, one log n factor is to
store the counter in each bucket, and the other log n factor is for the number of repetitions. For
constant ǫ and B = Θ(1/ǫ2 ) this gives O(log2 n) bits of space. It is also not hard to see that the
CountSketch data structure can be maintained in a stream with deletions as well as insertions,
since given a deletion to item i, this just corresponds to subtracting σ(i) from the bucket i hashes
to in each repetition. Moreover, as mentioned earlier, this O(log2 n) space bound is optimal for
streams with deletions.
To give some intuition for our new algorithm, let i∗ ∈ [n] be the identity of the single ℓ2 -heavy
√
hitter that we wish to find. Suppose first that fi∗ ≥ n log n and that fi ∈ {0, 1} for all i ∈ [n]\{i∗ }.
For the moment, we are also going to ignore the issue of storing random bits, so assume we can
store poly(n) random bits for free (which can be indexed into using O(log n) bits of space). We will
later sketch how to remove this assumption. As in the CountSketch algorithm, we again assign a
random sign σ(i) to each item i ∈ [n]. Suppose we randomly partition [n] into twoPbuckets using a
hash function
two counters c1 = i|h(i)=1 σ(i) · fi
P h : [n] → {1, 2}, and correspondingly maintain
∗
and c2 = i|h(i)=2 σ(i) · fi . Suppose for discussion that h(i ) = 1. A natural question is what the
values c1 and c2 look like as we see more updates in the stream.
Consider the values c1 − σ(i∗ ) · fi∗ and c2 . Then, since all frequencies other than i∗ are assumed
to be 0 or 1, and since the signs σ(j) are independent, these two quantities evolve as random walks
starting at 0 and incrementing by +1 with probability 1/2, and by −1 with probability 1/2, at each
step of the walk. By standard theory of random walks (e.g., Levy’s theorem), there is a constant
C > 0 so that with probability at least 9/10, simultaneously at all times during the stream we
√
have that |c1 − σ(i∗ ) · fi∗ | and |c2 | are upper bounded by C n. The constant of 9/10, like typical
constants in this paper, is somewhat arbitrary. This suggests the following approach to learning
√
√
i∗ : at some point in the stream we will have that fi∗ > 2C n, and at that point |c1 | > C n, but
then we know that i∗ occurs in the first bucket. This is assuming that the above event holds for
the random walks. Since we split [n] randomly into two pieces, this gives us 1 bit of information
6
about the identity of i∗ . If we were to repeat this O(log n) times in parallel, we would get exactly
the CountSketch data structure, which would use Θ(log2 n) bits of space. Instead, we get much
better space by repeating Θ(log n) times sequentially!
To repeat this sequentially, we simply wait until either |c1 | or |c2 | exceeds Cn1/2 , at which point
we learn one bit of information about i∗ . Then, we reset the two counters to 0 and perform the
√
procedure again. Assuming fi∗ = Ω( n log n), we will have Ω(log n) repetitions of this procedure,
each one succeeding independently with probability 9/10. By Chernoff bounds, there will only be
a single index i ∈ [n] which match a 2/3 fraction of these repetitions, and necessarily i = i∗ .
2.2.1
Gaussian Processes
√
In general we do not have fi∗ = Ω( n log n), nor do we have that fi ∈ {0, 1} for all i ∈ [n] \ {i∗ }.
We fix both problems using the theory of Gaussian processes.
Definition 3 A Gaussian process is a collection {Xt }t∈T of random variables, for an index set
T , for which every finite linear combination of the random variables is Gaussian.
We assume E[Xt ] = 0 for all t, as this will suffice for our application. It then follows that the
Gaussian process is entirely determined by its covariances E[Xs Xt ]. This fact is related to the
fact that a Gaussian distribution is determined by its mean and covariance. The distance function
d(s, t) = (E[(Xs − Xt )2 ])1/2 is then a pseudo-metric on T (the only property it lacks of a metric is
that d(s, t) may equal 0 if s 6= t).
The connection to data streams is the following. Suppose we replace the signs σ(i) with standard
normal random variables
g(i) in our counters above, and consider a counter c at time t, denoted
P
c(t), of the form i g(i) · fi (t). Here fi (t) is the frequency of item i after processing t stream
insertions. The main point is that c(t) is a Gaussian process! Indeed, any linear combination of
the c(t) values for different t is again Gaussian since the sum of normal random variables is again
a normal random variable.
The reason we wish to make such a connection to Gaussian processes is the following powerful
inequality called the “chaining inequality”.
Theorem 4 (Talagrand [45]) Let {Xt }t∈T be a Gaussian process and let T0 ⊆ T1 ⊆ T2 ⊆ · · · ⊆ T
i
be such that |T0 | = 1 and |Ti | ≤ 22 for i ≥ 1. Then,
X
E sup Xt ≤ O(1) · sup
2i/2 d(t, Ti ),
t∈T
t∈T i≥0
where d(t, Ti ) = mins∈Ti d(t, s).
We wish to apply Theorem 4 to the problem of finding ℓ2 -heavy hitters. Let F2 (t) be the value of
the second moment F2 after seeing t stream insertions. We now describe how to choose the sets Ti
in order to apply the chaining inequality; the intuition is that we recursively partition the stream
based on its F2 value.
Let at be the first stream update for which F2 (m)/2 ≤ F2 (t). Then T0 = {t}. We then let Ti be
i
the set of 22 times t1 , t2 , . . . , t22i in the stream for which tj is the first point in the stream for which
i
j · F2 (m)/22 ≤ F2 (tj ). Then, we have created a nested sequence of subsets T0 ⊆ T1 ⊆ T2 ⊆ · · · ⊆ T
i
with |T0 | = 1 and |Ti | ≤ 22 for i ≥ 1.
7
We are now in position to apply Theorem 4. A straightforward computation based on our
recursive partitioning of the stream around where F2 changes (see [12] for details) shows that for
any stream position t and set Ti we have created,
1/2
F2
2 1/2
.
d(t, Ti ) = min E[|c(t) − c(s)| ]
=O
s∈Ti
22i
Applying Theorem 4, we have
E[sup Xt ] ≤ O(1) sup
t∈T
X
i/2
2
t∈T i≥0
F2
22i
1/2
1/2
= O(F2 ).
This P
is exactly the same bound that the theory for random walks gave us earlier! (recall in that
case i6=i∗ fi2 < n).
Using Gaussian processes has therefore allowed us to remove our earlier assumption that fi ∈
{0, 1} for all i ∈ [n] \ {i∗ }. The same
random walk based algorithm will now work; however, we
√
still need to assume the fi∗ = Ω( F2 log n) in order to learn log n bits of information
to identify
√
F
i∗ , as before. This is not satisfactory, as an ℓ2 -heavy hitter only satisfies
f
=
Ω(
)
(recall we
i
2
√
have assumed φ and ǫ are constants), which is weaker than the fi∗ = Ω( F2 log n) that the above
analysis requires.
2.2.2
Amplification
√
To remove the assumption that fi∗ = Ω( F2 log n), our work [12] designs what we call an “amplification” procedure. This involves for j = 1, 2, . . . , O(log log n), independently choosing a pairwise
independent
hash function hj : [n]P→ {1, 2}. For each j, we as before maintain two counters
P
j
c1 = i|hj (i)=1 gj (i) · fi and cj2 = i|hj (i)=2 gj (i) · fi , where the gj (i) are independent standard
normal random variables.
Applying the chaining inequality to each of the O(log log n) counters created, we have that with
j
j
large constant probability,
√ both counters c1 and c2
√ in a constant fraction of the O(log log n) pairs,
will be bounded by O( F2 ) in magnitude. It follows that if fi∗ ≥ C F2 for a sufficiently large
constant C > 0 (which we can assume by first hashing the universe into O(1) buckets before the
streaming algorithm begins), then in say, a 9/10 fraction of pairs j, the counter cjk , k ∈ {1, 2},
of larger magnitude will contain i∗ . Moreover, by Chernoff bounds, only a log1c n fraction of other
i ∈ [n] will hash to the larger counter in at least a 9/10 fraction of such pairs, where c > 0 is a
constant that can be made arbitrarily large by increasing the constant in the number O(log log n)
of pairs of counters created. Now the idea is to effectively run our previous algorithm only on items
which hash to the heavier counter in at least a 9/10 fraction of pairs. By definition, this will contain
i∗ , and now the expected second moment of the√other items for which we run the algorithm on will
be F2 / logc n, which effectively makes fi∗ = Ω( F2 log n), where F2 is now measured with respect
to the items for which we run the algorithm on. Now we can sequentially learn O(log n) bits of
information about i∗ in our algorithm, as before.
One thing√to note about this approach is that after seeing a sufficiently large number of insertions
∗
of i , i.e., Θ( F2 ) such insertions, then most of the pairs of counters will have the property that
the larger counter (in absolute value) stays larger forever. This is due to the chaining inequality.
This can be used to fix the itemset for which we run the algorithm on. In fact, this is precisely
why this does not result in a 2-pass algorithm, which one might expect since one does not know the
8
itemset to run our algorithm on in advance. However, we always run the algorithm on whichever
current itemset agrees with at least a 9/10 fraction of the larger counters, and just accept the fact
that in the beginning of the stream the bits we learn about i∗ are nonsense; however, after enough
updates to i∗ have occurred in the stream then the counters “fix” themselves in the sense that the
larger counter does not change. At this point the bits we learn about i∗ in our algorithm are the
actual bits that we desire. At the end of the stream, we only look at a suffix of these bits to figure
out i∗ , thereby ignoring the nonsensical bits at the beginning of the stream. We refer the reader to
[12] for more details.
2.2.3
Derandomization
The final piece of the algorithm is to account for the randomness used by the algorithm. We need
to derandomize the counters, which use the theory of Gaussian processes to argue their correctness.
We also cannot afford to maintain all of the hash functions that were used to learn specific bits of
i∗ (which we need ad the end of the stream to figure out what i∗ is).
To derandomize the Gaussian processes, we use a derandomized Johnson Lindenstrauss transform of Kane, Meka, and Nelson [31]. The rough idea is to first apply a Johnson-Lindenstrauss
transform to the frequency vectors for which we take inner products with independent Gaussian
random variables in our counters. This will reduce the dimension from n to O(log n), for which we
can then afford to take an inner product with fully independent Gaussian random variables. The
nice thing about Johnson-Lindenstrauss transforms is that they preserve all the covariances up to
a constant factor in our specific Gaussian process, and therefore we can use Slepian’s Lemma (see
[12] for details) to argue that the Gaussian process is roughly the same as before, since it is entirely
determined by its covariances. Here the derandomized Johnson-Lindenstrauss transform of [31] can
be represented using only O(log n log log n) bits of space. Also, instead of using Gaussian random
variables, which require truncation, we can directly use sign random variables (+1 with probability
1/2, −1 with probability 1/2), which results in what are called Bernoulli processes, together with
a comparison theorem for Bernoulli processes and Gaussian processes. This enables us to avoid
arguments about truncating Gaussians.
To derandomize the hash functions, we use Nisan’s pseudorandom generator in a similar way
that Indyk uses it for derandomizing his algorithms for norm estimation [27, 42]. Please see [12]
for further details.
2.3
Followup Work
Very recently, in followup work by Braverman et al. [11], we improved the space bound further
to the optimal O(log n) bits of space (for constant ǫ, φ). The high level idea of using Gaussian or
Bernoulli processes is the same, but several additional insights were needed. This involves both a
new algorithm which we call BPTree, which avoids the amplification step and Nisan’s pseudorandom
generator described above, as well as a better derandomization of the Bernoulli processes using
O(1)-wise independence. We refer the reader to that work for further details.
3
Conclusions
We presented new algorithms for finding ℓ1 -heavy hitters and ℓ2 -heavy hitters in a data stream.
We refer the reader to the original papers cited above for further details. As these problems are
9
inspired from applications in practice, it is very interesting to see how the improved theoretical
algorithms perform in practice. In ongoing work we are testing these algorithms in practice on real
datasets.
Another interesting aspect is that the technique of using Gaussian processes in the ℓ2 -heavy
hitters algorithm has led to a number of other improvements to data stream algorithms, including
for example the ability to estimate the second moment F2 at all times in a stream of insertions.
Previously, given a stream of length n and a universe of size n, to estimate F2 at all points in a
stream up to a constant factor would require Θ(log2 n) bits of space, since it takes Θ(log n log(1/δ))
bits to estimate it at a single point with failure probability δ, and one needs to union bound over
n stream positions. Using Gaussian processes, [12] achieves only O(log n log log n) bits of space for
this task. It would be interesting to see if Gaussian processes are useful for other problems in data
streams.
References
[1] Rakesh Agrawal and Ramakrishnan Srikant. Fast algorithms for mining association rules in
large databases. In VLDB’94, Proceedings of 20th International Conference on Very Large
Data Bases, September 12-15, 1994, Santiago de Chile, Chile, pages 487–499, 1994.
[2] Noga Alon, Yossi Matias, and Mario Szegedy. The space complexity of approximating the
frequency moments. J. Comput. Syst. Sci., 58(1):137–147, 1999.
[3] Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Streaming algorithms via precision sampling. In IEEE 52nd Annual Symposium on Foundations of Computer Science, FOCS
2011, Palm Springs, CA, USA, October 22-25, 2011, pages 363–372, 2011.
[4] Khanh Do Ba, Piotr Indyk, Eric Price, and David P. Woodruff. Lower bounds for sparse
recovery. CoRR, abs/1106.0365, 2011.
[5] Ziv Bar-Yossef, T. S. Jayram, Ravi Kumar, and D. Sivakumar. An information statistics
approach to data stream and communication complexity. J. Comput. Syst. Sci., 68(4):702–
732, 2004.
[6] Radu Berinde, Piotr Indyk, Graham Cormode, and Martin J. Strauss. Space-optimal heavy
hitters with strong error bounds. ACM Trans. Database Syst., 35(4):26, 2010.
[7] Kevin S. Beyer and Raghu Ramakrishnan. Bottom-up computation of sparse and iceberg cubes.
In SIGMOD 1999, Proceedings ACM SIGMOD International Conference on Management of
Data, June 1-3, 1999, Philadelphia, Pennsylvania, USA., pages 359–370, 1999.
[8] Arnab Bhattacharyya, Palash Dey, and David P. Woodruff. An optimal algorithm for l1-heavy
hitters in insertion streams and related problems. PODS, 2016.
[9] Lakshminath Bhuvanagiri, Sumit Ganguly, Deepanjan Kesh, and Chandan Saha. Simpler
algorithm for estimating frequency moments of data streams. In Proceedings of the Seventeenth
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2006, Miami, Florida, USA,
January 22-26, 2006, pages 708–713, 2006.
10
[10] Jean Bourgain and Jelani Nelson. Toward a unified theory of sparse dimensionality reduction
in euclidean space. CoRR, abs/1311.2542, 2013.
[11] Vladimir Braverman, Stephen R. Chestnut, Nikita Ivkin, Jelani Nelson, Zhengyu Wang, and
David P. Woodruff. BPTree: an ℓ2 heavy hitters algorithm using constant memory. CoRR,
abs/1603.00759, 2016.
[12] Vladimir Braverman, Stephen R. Chestnut, Nikita Ivkin, and David P. Woodruff. Beating
countsketch for heavy hitters in insertion streams. STOC, 2016.
[13] Yousra Chabchoub, Christine Fricker, and Hanene Mohamed. Analysis of a bloom filter algorithm via the supermarket model. In 21st International Teletraffic Congress, ITC 2009, Paris,
France, September 15-17, 2009, pages 1–8, 2009.
[14] Amit Chakrabarti, Subhash Khot, and Xiaodong Sun. Near-optimal lower bounds on the
multi-party communication complexity of set disjointness. In 18th Annual IEEE Conference
on Computational Complexity (Complexity 2003), 7-10 July 2003, Aarhus, Denmark, pages
107–117, 2003.
[15] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in data
streams. Theoretical Computer Science, 312(1):3–15, 2004.
[16] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in data
streams. Theor. Comput. Sci., 312(1):3–15, 2004.
[17] Kenneth L. Clarkson and David P. Woodruff. Low rank approximation and regression in input
sparsity time. In Symposium on Theory of Computing Conference, STOC’13, Palo Alto, CA,
USA, June 1-4, 2013, pages 81–90, 2013.
[18] Graham Cormode and Marios Hadjieleftheriou. Finding frequent items in data streams. Proceedings of the VLDB Endowment, 1(2):1530–1541, 2008.
[19] Graham Cormode and S Muthukrishnan. An improved data stream summary: the count-min
sketch and its applications. Journal of Algorithms, 55(1):58–75, 2005.
[20] Erik D Demaine, Alejandro López-Ortiz, and J Ian Munro. Frequency estimation of internet
packet streams with limited space. In AlgorithmsESA 2002, pages 348–360. Springer, 2002.
[21] Cristian Estan and George Varghese. New directions in traffic measurement and accounting:
Focusing on the elephants, ignoring the mice. ACM Trans. Comput. Syst., 21(3):270–313,
2003.
[22] Min Fang, Narayanan Shivakumar, Hector Garcia-Molina, Rajeev Motwani, and Jeffrey D.
Ullman. Computing iceberg queries efficiently. In VLDB’98, Proceedings of 24rd International
Conference on Very Large Data Bases, August 24-27, 1998, New York City, New York, USA,
pages 299–310, 1998.
[23] Anna C. Gilbert, Yi Li, Ely Porat, and Martin J. Strauss. Approximate sparse recovery:
optimizing time and measurements. In Proceedings of the 42nd ACM Symposium on Theory
of Computing, STOC 2010, Cambridge, Massachusetts, USA, 5-8 June 2010, pages 475–484,
2010.
11
[24] Jiawei Han, Jian Pei, Guozhu Dong, and Ke Wang. Efficient computation of iceberg cubes
with complex measures. In Proceedings of the 2001 ACM SIGMOD international conference
on Management of data, Santa Barbara, CA, USA, May 21-24, 2001, pages 1–12, 2001.
[25] Jiawei Han, Jian Pei, and Yiwen Yin. Mining frequent patterns without candidate generation.
In Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data,
May 16-18, 2000, Dallas, Texas, USA., pages 1–12, 2000.
[26] Christian Hidber. Online association rule mining. In SIGMOD 1999, Proceedings ACM SIGMOD International Conference on Management of Data, June 1-3, 1999, Philadelphia, Pennsylvania, USA., pages 145–156, 1999.
[27] Piotr Indyk. Stable distributions, pseudorandom generators, embeddings, and data stream
computation. J. ACM, 53(3):307–323, 2006.
[28] Piotr Indyk and David P. Woodruff. Optimal approximations of the frequency moments of
data streams. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing,
Baltimore, MD, USA, May 22-24, 2005, pages 202–208, 2005.
[29] T. S. Jayram and David P. Woodruff. The data stream space complexity of cascaded norms.
In 50th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2009, October
25-27, 2009, Atlanta, Georgia, USA, pages 765–774, 2009.
[30] Hossein Jowhari, Mert Saglam, and Gábor Tardos. Tight bounds for lp samplers, finding
duplicates in streams, and related problems. In Proceedings of the 30th ACM SIGMODSIGACT-SIGART Symposium on Principles of Database Systems, PODS 2011, June 12-16,
2011, Athens, Greece, pages 49–58, 2011.
[31] Daniel M. Kane, Raghu Meka, and Jelani Nelson. Almost optimal explicit johnsonlindenstrauss families. In Approximation, Randomization, and Combinatorial Optimization.
Algorithms and Techniques - 14th International Workshop, APPROX 2011, and 15th International Workshop, RANDOM 2011, Princeton, NJ, USA, August 17-19, 2011. Proceedings,
pages 628–639, 2011.
[32] Richard M Karp, Scott Shenker, and Christos H Papadimitriou. A simple algorithm for finding
frequent elements in streams and bags. ACM Transactions on Database Systems (TODS),
28(1):51–55, 2003.
[33] Abhishek Kumar and Jun (Jim) Xu. Sketch guided sampling - using on-line estimates of flow
size for adaptive data collection. In INFOCOM 2006. 25th IEEE International Conference on
Computer Communications, Joint Conference of the IEEE Computer and Communications
Societies, 23-29 April 2006, Barcelona, Catalunya, Spain, 2006.
[34] Kasper Green Larsen, Jelani Nelson, Huy Le Nguyen, and Mikkel Thorup. Optimal space
heavy hitters with fast update and query time. 2016.
[35] Gurmeet Singh Manku and Rajeev Motwani. Approximate frequency counts over data streams.
In Proceedings of the 28th international conference on Very Large Data Bases, pages 346–357.
VLDB Endowment, 2002.
12
[36] Xiangrui Meng and Michael W. Mahoney. Low-distortion subspace embeddings in inputsparsity time and applications to robust linear regression. In Symposium on Theory of Computing Conference, STOC’13, Palo Alto, CA, USA, June 1-4, 2013, pages 91–100, 2013.
[37] Ahmed Metwally, Divyakant Agrawal, and Amr El Abbadi. Efficient computation of frequent
and top-k elements in data streams. In Proceedings of the 10th International Conference on
Database Theory, ICDT’05, pages 398–412, Berlin, Heidelberg, 2005. Springer-Verlag.
[38] Gregory T. Minton and Eric Price. Improved concentration bounds for count-sketch. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2014, Portland, Oregon, USA, January 5-7, 2014, pages 669–686, 2014.
[39] Jayadev Misra and David Gries. Finding repeated elements. Sci. Comput. Program., 2(2):143–
152, 1982.
[40] Morteza Monemizadeh and David P. Woodruff. 1-pass relative-error lp -sampling with applications. In Proceedings of the Twenty-First Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2010, Austin, Texas, USA, January 17-19, 2010, pages 1143–1160, 2010.
[41] Jelani Nelson and Huy L. Nguyen. OSNAP: faster numerical linear algebra algorithms via
sparser subspace embeddings. In 54th Annual IEEE Symposium on Foundations of Computer
Science, FOCS 2013, 26-29 October, 2013, Berkeley, CA, USA, pages 117–126, 2013.
[42] Noam Nisan. Pseudorandom generators for space-bounded computation. Combinatorica,
12(4):449–461, 1992.
[43] Eric Price. Efficient sketches for the set query problem. In Proceedings of the Twenty-Second
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2011, San Francisco, California, USA, January 23-25, 2011, pages 41–56, 2011.
[44] Ashok Savasere, Edward Omiecinski, and Shamkant B. Navathe. An efficient algorithm for
mining association rules in large databases. In VLDB’95, Proceedings of 21th International
Conference on Very Large Data Bases, September 11-15, 1995, Zurich, Switzerland., pages
432–444, 1995.
[45] Michel Talagrand. Majorizing measures: The generic chaining. The Annals of Probability,
24(3), 1996.
[46] Mikkel Thorup and Yin Zhang. Tabulation-based 5-independent hashing with applications to
linear probing and second moment estimation. SIAM J. Comput., 41(2):293–331, 2012.
[47] Hannu Toivonen. Sampling large databases for association rules. In VLDB’96, Proceedings
of 22th International Conference on Very Large Data Bases, September 3-6, 1996, Mumbai
(Bombay), India, pages 134–145, 1996.
13
| 8cs.DS
|
Text-based Adventures of the Golovin AI Agent
Bartosz Kostka, Jarosław Kwiecień, Jakub Kowalski, Paweł Rychlikowski
arXiv:1705.05637v1 [cs.AI] 16 May 2017
Institute of Computer Science
University of Wrocław, Poland
Email: {bartosz.kostka,jaroslaw.kwiecien}@stud.cs.uni.wroc.pl, {jko,prych}@cs.uni.wroc.pl
Abstract—The domain of text-based adventure games has been
recently established as a new challenge of creating the agent that
is both able to understand natural language, and acts intelligently
in text-described environments.
In this paper, we present our approach to tackle the problem.
Our agent, named Golovin, takes advantage of the limited game
domain. We use genre-related corpora (including fantasy books
and decompiled games) to create language models suitable to
this domain. Moreover, we embed mechanisms that allow us
to specify, and separately handle, important tasks as fighting
opponents, managing inventory, and navigating on the game map.
We validated usefulness of these mechanisms, measuring
agent’s performance on the set of 50 interactive fiction games.
Finally, we show that our agent plays on a level comparable to the
winner of the last year Text-Based Adventure AI Competition.
I. I NTRODUCTION
The standard approach to develop an agent playing a given
game is to analyze the game rules, choose an appropriate AI
technique, and incrementally increase the agent’s performance
by exploiting these rules, utilizing domain-dependent features
and fixing unwanted behaviors. This strategy allowed to beat
the single games which were set as the milestones for the AI
development: Chess [1] and Go [2].
An alternative approach called General Game Playing
(GGP), operating on a higher level of abstraction has recently
gained in popularity. Its goal is to develop an agent that can
play any previously unseen game without human intervention.
Equivalently, we can say that the game is one of the agent’s
inputs [3].
Currently, there are two main, well-established GGP domains providing their own game specification languages and
competitions [4]. The first one is the Stanford’s GGP, emerged
in 2005 and it is based on the Game Description Language
(GDL), which can describe all finite, turn-based, deterministic
games with full information [5], and its extensions (GDL-II
[6] and rtGDL [7]).
The second one is the General Video Game AI framework
(GVGAI) from 2014, which focuses on arcade video games
[8]. In contrast to Stanford’s GGP agents are provided with the
forward game model instead of the game rules. The domain is
more restrictive but the associated competition provides multiple tracks, including procedural content generation challenges
[9], [10].
What the above-mentioned approaches have in common is
usually a well-defined game state the agent is dealing with.
It contains the available information about the state (which
may be partially-observable), legal moves, and some kind of
scoring function (at least for the endgame states). Even in a
GGP case, the set of available moves is known to the agent,
and the state is provided using some higher-level structure
(logic predicates or state observations).
In contrast, the recently proposed Text-Based Adventure AI
Competition, held during the IEEE Conference on Computational Intelligence and Games (CIG) in 2016, provides a new
kind of challenge by putting more emphasis on interaction
with the game environment. The agent has access only to the
natural language description about his surroundings and effects
of his actions.
Thus, to play successfully, it has to analyze the given
descriptions and extract high-level features of the game state
by itself. Moreover, the set of possible actions, which are also
expected to be in the natural language, is not available and an
agent has to deduct it from his knowledge about the game’s
world and the current state.
In some sense, this approach is coherent with the experiments on learning Atari 2600 games using the Arcade Learning Environment (ALE), where the agent’s inputs were only
raw screen capture and a score counter [11], [12]. Although
in that scenario the set of possible commands is known in
advance.
The bar for text-based adventure games challenge is set
high – agents should be able to play any interactive fiction
(IF) game, developed by humans for the humans. Such environment, at least in theory, requires to actually understand the
text in order to act, so completing this task in its full spectrum,
means building a strong AI.
Although some approaches tackling similar problems exist
since early 2000s ([13], [14]), we are still at the entry point
for this kind of problems, which are closely related to the
general problem solving. However, recent successes of the
machine learning techniques combined with the power of
modern computers, give hope that some vital progress in the
domain can be achieved.
We pick up the gauntlet, and in this work we present our
autonomous agent that can successfully play many interactive
fiction games. We took advantage of the specific game domain,
and trained agent using matching sources: fantasy books and
texts from decompiled IF games. Moreover, we embed some
rpg-game-based mechanisms, that allow us to improve fighting
opponents, managing hero’s inventory, and navigating in the
maze of games’ locations.
We evaluated our agent on a set of 50 games, testing
the influence of each specific component on the final score.
Also, we tested our agent against the winner of the last year
competition [15]. The achieved results are comparable. Our
agent scored better in 12 games and worse in 11 games.
The paper is organized as follows. Section II provides background for the domain of interactive fiction, Natural Language
Processing (NLP), Text-Based Adventure AI Competition,
and the related work. In Section III, we presented detailed
description of our agent. Section IV contains the results of the
performed experiments. Finally, in Section V, we conclude and
give perspective of the future research.
II. BACKGROUND
A. Interactive Fiction
Interactive Fiction (IF), emerged in 1970s, is a domain of
text-based adventure or role playing games, where the player
uses text commands to control characters and influence the
environment. One of the most famous example is the Zork
series developed by the Infocom company. From the formal
point of view, they are single player, non-deterministic games
with imperfect information. IF genre is closely related to
MUDs (Multi-User Dungeons), but (being single-player) more
focused on plot and puzzles than fighting and interacting
with other players. IF was popular up to late 1980s, where
the graphical interfaces become available and, as much user
friendlier, more popular. Nevertheless, new IF games are still
created, and there are annual competitions for game authors
(such as The Interactive Fiction Competition).
IF games usually (but not always) take place in some fantasy
worlds. The space the character is traversing has a form of
labyrinth consisting of so called rooms (which despite the
name can be open areas like forest). Entering the room, the
game displays its description, and the player can interact with
the objects and game characters it contains, or try to leave
the room moving to some direction. However, reversing a
movement direction (e.g. go south ↔ go north) not necessarily
returns the character to the previous room.
As a standard, the player character can collect objects from
the world, store them in his equipment, and combine with
other objects to achieve some effects on the environment (e.g.
put the lamp and sword in the case). Thus, many games
require solving some kind of logical puzzle to push the action
forward. After performing an action, the game describes its
effect. Many available actions are viable, i.e. game engine
understands them, but they are not required to solve the game,
or even serve only for the player amusement.
Some of the games provide score to evaluate the player’s
progress, however the change in the score is often the result of
a complex series of moves rather than quick “frame to frame”
decisions, or the score is given only after the game’s end.
Other games do not contain any scoring function and the only
output is win or lose.
B. Playing Text-Based Games
Although the challenge of playing text-based games was
not take on often, there are several attempts described in the
literature, mostly based on the MUD games rather than the
classic IF games.
Adventure games has been carefully revised as the field of
study for the cognitive robotics in [14]. First, the authors identify the features of “tradition adventure game environment” to
point-out the specifics of the domain. Second, they enumerate
and discuss existing challenges, including e.g. the need for
commonsense knowledge (its learning, revising, organization,
and using), gradually revealing state space and action space,
vague goal specification and reward specification.
In [13], the agent able to live and survive in an existing
MUD game have been described. The authors used layered
architecture: high level planning system consisting of reasoning engine based on hand-crafted logic trees, and a low level
system responsible for sensing the environment, executing
commands to fulfill the global plan, detecting and reacting
in emergency situations.
While not directly-related to playing algorithms, it is worth
to note the usage of computational linguistics and theorem
proving to build an engine for playing text-based adventure
games [16]. Some of the challenges are similar for both
tasks, as generating engine requires e.g. object identification
(given user input and a state description) and understanding
dependencies between the objects.
The approach focused on tracking the state of the world
in text-based games, and translating it into the first-order
logic, has been presented in [17]. Proposed solution was able
to efficiently update agent’s belief state from a sequence of
actions and observations.
The extension of the above approach, presents the agent that
can solve puzzle-like tasks in partially observable domain that
is not known in advance, assuming actions are deterministic
and without conditional effects [18]. It generates solutions by
interleaving planning (based on the traditional logic-based approach) and execution phases. The correctness of the algorithm
is formally proved.
Recently, an advanced MUD playing agent has been described in [19]. Its architecture consists of two modules. First,
responsible for converting textual descriptions to state representation is based on the Long Short-term Memory (LSTM)
networks [20]. Second, uses Deep Q-Networks [11] to learn
approximated evaluations for each action in a given state.
Provided results show that the agent is able to to successfully
complete quests in small, and even medium size, games.
C. Natural Language Processing
Natural Language Processing is present in the history of
computers almost from the very beginning. Alan Turing in
his famous paper [21] state (approximately) that “exhibit
intelligent behavior” means “understand natural language and
use it properly in conversations with human being”. So, since
1950 Turing test is the way of checking whether computer has
reached strong AI capability.
First natural language processing systems were rule based.
Thanks to the growing amount of text data and increase of
the computer power, during last decades one can observe the
shift towards the data driven approaches (statistical or machine
learning). Nowadays, NLP very often is done “almost from
scratch”, as it was done if [22] where the authors have used
neural network in order to solve many NLP tasks, including
part-of-speech tagging, named entity recognition and semantic
role labeling. The base for this was the neural language model.
Moreover, this systems produced (as a side effect) for every
word in a vocabulary a dense vector which reflected word
properties. This vectors are called word embeddings and can
be obtained in many ways. One of the most popular is the one
proposed in [23] that uses very simple, linear language model
and is suitable to large collections of texts.
Language models allow to compute probability of the sentence treated as a sequence of items (characters, morphemes
or words). This task was traditionally done using Markov
models (with some smoothing procedures, see [24]). Since
predicting current words is often dependent on the long part
of history, Markov models (which, by definition, looks only
small numbers of words behind) are outperformed by the
modern methods that can model long distance dependencies.
This methods use recursive (deep) neural networks, often
augmented with some memory.
We will use both word embeddings (to model words similarity) and LSTM neural networks [25] with attention mechanism
(see [26] and [27]). We are particularly interested in the
information given from the attention mechanism, which allows
us to estimate how important is each word, when we try to
predict the next word in the text.
D. The Text-Based Adventure AI Competition
The first Text-Based Adventure AI Competition1 , organized
by the group from the University of York, has been announced
in May 2016 and took place at the 2016 IEEE CIG conference
in September. The second, will be held this year, also colocated with CIG.
The purpose of the competition is to stimulate research
towards the transcendent goal of creating General Problem
Solver, the task stated nearly six decades ago [28]. The organizers plan to gradually increase the level of given challenges,
with time providing more complex games that require more
sophisticated approaches from the competitors. Thus, finally
force them to develop agents that can autonomously acquire
knowledge required to solve given tasks from the restricted
domain of text-based games.
The domain of the competition is specified as any game that
can be described by the Z-machine, the classic text adventuring
engine. Interactive Fiction games are distributed as compiled
byte code, requiring the special interpreter to run them. The
first Z-machine has been developed in 1979 by Infocom,
and supports games developed using a LISP-like programming language named Infocom’s ZIL (Zork Implementation
1 http://atkrye.github.io/IEEE-CIG-Text-Adventurer-Competition/.
Language). The Text-based AI Competition uses Frotz2 , the
modern version of the Z-machine, compatible with the original
interpreter.
The competition organizers provide a Java package managing the communication between a game file and an agent
process. Also, example random agents in Java and Python 3 are
available. The interpreter is extended by the three additional
commends, allowing players to quit the game, restart it with
a new instance of the agent, and restart without modifying
the agent. Given that, the text-based AI framework supports
learning and simulation-based approaches.
Little details about the competition insides are available. In
particular, the number of participants is unknown, and the test
environment game used to evaluate agents remained hidden,
as it is likely to be used again this year. The game has been
developed especially for the purpose of the competition and
supports graduated scale of scoring points, depending on the
quality of agent’s solution.
The winner of the first edition was the BYU-Agent3 from
the Perception Control and Cognition lab at Brigham Young
University, which achieved a score 18 out of 100. The idea
behind the agent has been described in [15]. It uses Q-learning
[29] to estimate the utility of an action in a given game state,
identified as the hash of its textual description.
The main contribution concerns affordance detection, used
to generating reasonable set of actions. Based on the word2vec
[30], an algorithm mapping words into a vector representations
based on their contextual similarities, and the Wikipedia as the
word corpus, the verb-noun affordances are generated. Thus,
the algorithm is able to detect, for an in-game object, words
with a similar meaning, and provide a set of actions that are
possible to undertake with that object.
Provided results, based on the IF games compatible with Zmachine interpreter, shows the overall ability of the algorithm
to successfully play text-based games. Usually, the learning
process results in increasing score, and requires playing a game
at least 200 times to reach the peek. However, there are some
games that achieve that point much slower, or even the score
drops as the learning continues.
III. T HE G AME P LAYING AGENT
A. Overview
Our agent is characterized by the following features.:
•
•
•
it uses a huge set of predefined command patterns,
obtained by analyzing various domain-related sources; the
actual commands are obtained by suitable replacements;
it uses language models based on selection of fantasy
books;
it takes advantage of the game-specific behaviors, natural
for adventure games, like fight mode, equipment management, movement strategy;
2 http://frotz.sourceforge.net.
3 The agent is open source and available at https://github.com/danielricks/
BYU-Agent-2016.
it memorizes and uses some aspects of the current play
history;
• it tries to imitate human behavior: after playing several
games and exploring the game universe it repeats the most
promising sequence of commands. We treat the result
reached in this final trial as the agent’s result in this game.
The agent was named “Golovin”, as one of the first answers
it gives after asking Hey bot, what is your name?, was your
name is Golovin, a phrase from the game Doomlords.
•
B. Preprocessing
1) Language Models: We used language models for two
purposes. First, they allow us to define words similarity (which
in turns gives us opportunity to replace some words in commands with their synonyms). For this task we use word2vec
[30] (and its implementation in TensorFlow [31]). Secondly,
we use neural network language models to determine which
words in the scene description plays more important role than
other (and so are better candidates to be a part of the next
command). We use the LSTM neural networks operating on
words [25], augmented by the attention mechanism ([26] and
[27]). This combination was previously tested in [32].
Since the action of many games is situated in fantasy
universe, we decided to train our models on the collection
of 3000 fantasy books from bookrix.com (instead of using
Wikipedia, as in [15]).
2) Commands: In order to secure out agent against overfitting, we fix the set of games used in tests (the same 50 games
as in [15]). No data somehow related to this games were used
in any stage of preprocessing.
We considered three methods to gather commands:
• walkthroughs – for several games, sequence of commands from winning path can be found in the Internet.
This source provides raw valid commands, that are useful
in some games.
• tutorials – on the other hand, some games have tutorials
written in natural language. Analyzing such tutorials4
seemed to be a promising way of getting valid command.
Concept of reading manuals has been successfully used
to learn how to play Civilization II game [33].
• games – at the end, there are many games that don’t have
tutorials nor walkthroughs. We downloaded a big collection of games, decompiled their codes, and extracted all
texts from them.
The last two sources required slightly more complicated
preprocessing. After splitting texts into sentences, we parsed
them using PCFG parser from NLTK package [34]. Since
commands are (in general) verb phrases, we found all minimal
VP phrases from parse trees. After reviewing some of them,
we decided not to take every found phrase, but manually
create the list of conditions which characterizes ’verb phrases
useful in games’. In this way we obtained the collections of
4 This tutorials were downloaded from the following sites: http://www.
ifarchive.org/, https://solutionarchive.com/, http://www.gameboomers.com/,
http://www.plover.net/~davidw/sol/
approximately 250,000 commands (or, to be more precisely,
command patterns). We also remember the count of every
command (i.e. the number of parse tree it occurs in).
Some of the commands have special tag: “useful in the
battle”. We have manually chosen five verbs, as the most
commonly related to fighting: attack, kill, fight, shoot, and
punch. Approximately 70 most frequent commands containing
one of these verbs received this tag.
The commands used by our agent were created from these
patterns by replacing (some) nouns by nouns taken from the
game texts.
C. Playing Algorithm
The algorithm uses 5 types of command generators: battle
mode, gathering items, inventory commands, general actions
(interacting with environment), and movement. The generators
are fired in the given order, until a non-empty set of commands
is proposed.
There are multiple reasons why some generator may not
produce any results, e.g. the agent reaches the predefined limit
of making actions of that type, all the candidates are blacklisted, or simply we cannot find any appropriate command in
the database. We will describe other special cases later.
When the description of the area changes, all the command
lists are regenerated.
1) Generating Commands: Our general method to compute
available commands and choose the one which is carried out,
looks as follows:
1) Find all nouns in the state description and agent’s
equipment. (We accept all type of nouns classified by
the nltk.pos_tag function.)
2) Determine their synonyms, based on the cosine similarity between word vectors. (We use n-best approach with
n being subject to Spearmint optimization; see IV-A.)
3) Find the commands containing nouns from the above
described set. If a command contains a synonym, it is
replaced by the word originally found in the description.
4) Score each command taking into account:
• cosine similarity between used synonyms and the
original words
• uniqueness of the words in command, computed as
the inverse of number of occurrences in the corpora.
• the weight given by the neural network model
• the number of words occurring both in the description and in the command
The score is computed as the popularity of the command
(number of occurrences in the command database) multiplied by the product of the above. The formula uses
some additional constants influencing the weights of the
components.
5) Then, using the score as the command’s weight, randomly pick one command using the roulette wheel
selection.
2) Battle Mode: The battle mode has been introduced to
improve the agent’s ability to survive. It prevents from what
has been the main cause of agent’s death before – careless
walking into an enemy or spending too much time searching
for the proper, battle-oriented and life-saving, action.
The agent starts working in battle mode after choosing
one of the “fight commands”. Being in this mode, the agent
strongly prefers using battle command, moreover it repeats
the same command several times (even if it fails), because
in many games the opponent has to be attacked multiple
times to be defeated. Therefore, between the consecutive
fighting actions we prevent using standard commands (like
look, examine), as wasting precious turns usually gives the
opponent an advantage.
3) Inventory Management (gathering and using items):
In every new area, the algorithm searches its description for
interesting objects. It creates a list of nouns ordered by the
weight given by the neural network model and their rarity.
Then, the agent tries take them.
If it succeeds (the content of the inventory has changed), a
new list of commands using only the newly acquired item is
generated (using the general method). The constant number of
highest scored commands is immediately executed.
4) Exploration: The task of building an IF game map is
difficult for two reasons. One, because a single area can be
presented using multiple descriptions and they may change as
the game proceeds. Two, because there may be different areas
described by the same text. Our map building algorithm tries
to handle these problems.
We have found that usually the first sentence of the area
description remains unchanged, so we use it to label the
nodes of the graph (we have tried other heuristics as well
but they performed worse). This heuristic divides all visited
nodes into the classes suggesting that corresponding areas may
be equivalent. The edges of the graph are labeled by the move
commands used to translocate between the nodes (we assume
that movement is deterministic).
We initialize the map graph using the paths corresponding
to the past movements of the agent. Then, the algorithm takes
all pairs of nodes with the same label and considers them in
a specific, heuristic-based, order. For every pair, the MergeNodes procedure (Listing 1) is fired. The procedure merges two
states joining their outcoming edges and recursively proceeds
to the pairs of states that are reachable using the same move
command. If the procedure succeeds, we replaces current map
graph with the minimized one, otherwise the changes are
withdrawn.
We use a small fixed set of movement commands (south,
northwest, up, left, etc.) to reveal new areas and improve the
knowledge about the game layout. When the agent decides
to leave the area, it tries a random direction, unless it already discovered all outgoing edges – then it uses map to
find a promising destination. We evaluate destination nodes
minimizing the distance to that node plus the number of
tested commands divided by the node’s curiosity (depending
on scores of available commands and untested movement
commands). Then, the agent follows the shortest path to the
best scored destination.
Algorithm 1 MergeNodes(A, B)
1: if A = B then return True end if
2: if label(A)6=label(B) then return False end if
3: mergelist ← {}
4: for all m ∈ M oveCommands do
5:
if A.moveby(m)6=None ∧ B.moveby(m)6=None then
6:
mergelist.append((A.moveby(m), B.moveby(m)))
7:
end if
8: end for
9: JoinIncomingAndOutgoingEdges(A, B)
10: for all (A0 , B 0 ) ∈ mergelist do
11:
if ¬ MergeNodes(A0 , B 0 ) then return False end if
12: end for
13: return True
5) Failing Commands: When, after executing a command,
the game state (description) remains unchanged, we assume
the command failed. Thus, the algorithm puts it on a blacklist
assigned to the current location. The command on the location’s black list is skipped by the command generators.
After any change in the agent’s inventory, all blacklists are
cleared.
6) Restarts: The Frotz environment used for the contest
allows to restart the game, i.e. abandon current play and start
again from the beginning.
Me make use of this possibility in a commonsense imitating
of the human behavior. When the agent dies, it restarts the
game and, to minimize the chance of the further deaths, it
avoids repeating the last commands of his previous lives. The
agent also remembers the sequence of moves that lead to the
best score and eventually repeats it. The final trial’s result is
used as the agent’s result in the game.
IV. E XPERIMENTS
Our experiments had two main objectives: creating the most
effective agent, and analyze how some parameters influence
the agents performance.
The most natural way to measure the agent performance is
to use the score given by the game (divided by the maximum
score, when we want to compare different games). However,
there are many games in which our agent (as well as BYUAgent) has problems with receiving non zero points. So, we
have decided to reward any positive score and add to the
positive agent result arbitrarily chosen constant 0.2. Therefore,
optimal (hypothetical) agent would get 1.2 in every game.
We selected 20 games for the training purposes, for all
of them the maximum score is known. The performance of
the agent is an averaged (modified) score computed on these
games.
A. Creating The Best Agent
The agent’s play is determined by some parameters, for
instance:
• the set of command patterns,
• the number of synonyms taken from word2vec,
the number of items, we are trying to gather, after visiting
new place,
• the number of standard command, tried after gathering
phase,
• how to reward the commands containing many words
from description (the actual reward is bk , where k is the
number of common words, and b is a positive constant),
• how to punish the commands containing words with no
good reason to be used (neither in state description nor in
generated synonyms), the score is divided by pk , where
k is the number of such words, and p is a constant,
• how many command should be done before trying movement command.
Furthermore we wanted to check, whether using battle mode
or a map has an observable effect on agent performance. The
number of parameter combinations was too large for grid
search, so we decided to use Spearmint5 .
We started this optimization process with (total) score equal
to 0.02, and after some hours of computation we end with
0.08 (which means that the score has been multiplied 4 times).
From now all parameters (if not stated otherwise) will be taken
from the final Spearmint result.
•
optimal configuration uses only two sources: T and W6 . We,
however, still believe that decompiled games can be a useful
source for game commands. But they cannot be found in
descriptions, but in command interpreter – which requires
more advanced automated code analysis. We left it as a future
work.
Fig. 2. Comparison of agent using different sources of commands. Best
variant scaled to 100%.
B. Evaluation of Domain-based Mechanisms
D. Gameplay Examples
We wanted to check whether some more advanced features
of our agent give observable influence on agent performance.
We checked the following 4 configurations with battle-mode
and map turned on or off. The results are presented in Figure
1. One can see that map is useful (but only to some extent),
and battle mode is undoubtedly useful.
While playing detective, our agent finds himself in a
closet. We get the following state description:
Game: You are in a closet. There is a gun on the
floor. Better get it. To exit, go east.
Our agent determines items: closet, gun, floor, exit. Our agent
is choosing from the commands listed in Table I. We see that
what is the most important for the agent is gun and how to
take it, which is reasonable and even suggested by the game.
Moreover, the agent also tries to search using synonyms of
the word gun in order to find proper commands (for instance:
we can see that he recognizes gun as some kind of weapon,
and some weapons, as for example knife, can be sharpened).
Fortunately, after using a command get gun, Golovin obtains
a black little pistol.
TABLE I
B EST 10 ( OUT OF 25) COMMANDS PROPOSED BY OUR AGENT FOR THE
SITUATION DESCRIBED IN THE D E T E C T I V E EXAMPLE (S ECTION IV-D)
Fig. 1. Comparison of agent version with and without map and battle mode.
Best variant scaled to 100%.
C. Evaluation of Language Model Sources
Commands were taken from 3 sources: tutorials (T), walkthroughs (W), and state description from games (G). We
compared the agents used command from all combination
of these sources. The results are presented in Figure 2. The
5 Spearmint
is a package which performs Bayesian optimization of hyperparameters. It allows to treat the optimized function as a black-box, and tries
to choose the parameters for the next run considering the knowledge gathered
during previous runs. See [35].
action
get gun
drop gun
take gun
open closet
examine gun
fire gun
load gun
examine closet
buy gun
sharp gun
score
0.1736
0.1129
0.0887
0.0557
0.0309
0.0252
0.0237
0.0128
0.0042
0.0006
Another example comes from zork3. We start with the
following location description:
6 The difference between T+W and G+T+W is not very big. In the previous
version of this experiment the winner was G+T+W.
Game: You are in a shadowy land of low, rolling
hills stretching out to the west and south. The
land is bordered to the north by a massive stone
wall. Ancient and weathered, the wall has crumbled
enough at one point to permit passage. A cloaked
and hooded person, carrying a sword not unlike your
own, is standing blocking the way to the northeast.
The hooded figure has a great deal of strength,
perhaps matching your own.
Golovin: examine figure
Game: There is nothing special to be seen. Your
quick reflexes cannot stop the hooded figure’s stroke!
You are hit!
In this very moment, in the agent’s inventory are: a sword,
a lamp. In the command list we can find for instance kill
man with sword. As the word man is close enough to figure,
and Golovin also have sword, he chooses to use command
kill figure with sword. Because the fighting command was
chosen, the battle mode is turned on, and the game continues
as follows:
Golovin: kill figure with sword
Game: The hooded figure is hit with a quick slash!
The figure has a light wound which hasn’t affected
its seemingly great strength.
the agent’s score improves and the fight continues. . .
Golovin: kill figure with sword
Game: You charge, but the hooded figure jumps
nimbly aside. The hooded figure stabs nonchalantly
with its sword and misses.
E. The Comparison with BYU-Agent
Finally, we validate our approach by comparing it with the
BYU-Agent. We were using the same set of 50 Z-machine
games7 as in [15].
The results of the comparison are presented in Table II.
The BYU agent was trained for 1000 epochs (each epoch
containing 1000 game steps), and its score was noted after
each epoch. Because the learning curves vary depending on
the game, including degeneration of the results (see [15,
Figure 5]), as the main measure we took the maximum score
achieved over all epochs.
As for the Golovin, we restricted his playing time to
1000 steps (i.e. an equivalent of one epoch) and use our
commonsense restarting mechanism.
The BYU-Agent results are obtained using the verb and
action space reduction algorithm, except the games marked
with an asterisk, where the verb space reduction experienced
errors, so we present scores obtained by the action space
reduction variant instead.
Eventually, there are 24 games, out of 50, where some of
the agents received any positive reward. Golovin scored better
in 12 games, including 7 games where BYU-Agent received
no reward. BYU-Agent scored better in 11 games, including
7 The game set is available at https://github.com/danielricks/textplayer/tree/
master/games.
6 games where Golovin scored no points. One game is a nonzero tie.
Thus, despite significantly shorter learning time (i.e. available number of steps), our agent is able to outperform BYUAgent on a larger number of games than he is outperformed on.
On the other hand, BYU-Agent gains in the games where the
Q-learning is effective and gradually increases score through
the epochs, e.g. curses, detective or Parc.
Last observation concerns the number of games where only
one of the agents scored 0, which is surprisingly large. This
may suggest that the two compared approaches are effective
on a different types of games, and may, in some sense,
complement each other.
TABLE II
AVERAGE SCORES FOR 10 RUNS OF EACH GAME . F OR BYU-AGENT WE
TOOK THE MAXIMUM ACHIEVED SCORE DURING THE 1000 EPOCHS
TRAINING . G OLOVIN PLAYS FOR ONE EPOCH . I N THE GAMES THAT ARE
NOT LISTED BOTH AGENTS GAIN NO REWARD . T HE ASTERISK MARKS
GAMES THAT USES OTHER VERSION OF BYU-AGENT
game
balances
break-in
bunny
candy
cavetrip
curses
deephome
detective
gold
library
mansion
Murdac
night
omniquest
parallel
Parc
reverb
spirit
tryst205
zenon
zork1
zork2
zork3
ztuu
better in:
Golovin
9.0
0
2.7
10.0
15.0
0.4
1.0
71.0
0.3
5.0
0.1
10.0
0.8
7,5
0
1.6
0
3.2
0.2
0
13.5
-0.1
0.7
0
12
BYU-Agent
0
0.3
2.0
10.0
10.5
1.9
0
213.0
0
0
2.2
0
0
5.0
5.0
5.0
1.8
2.0
2.0
2.8
*8.8
*3.3
*0
0.5
11
max score
51
150
60
41
500
550
300
360
100
30
68
250
10
50
150
50
50
250
350
20
350
400
7
100
games
V. C ONCLUSIONS AND F UTURE W ORK
We have presented an agent able to play any interactive fiction game created for human players, on the level comparable
to the winner of the last year Text-Based Adventure AI Competition. Due to the number of domain-based mechanisms, our
agent can successfully handle the game in a limited number
of available steps. The results of the presented experiments
show that the mechanisms we embed (battle mode, mapping)
and a choice of learning sources, indeed improves the agent’s
performance.
Although the results are promising, we are still at the
beginning of the path towards creating the agent that can
really understand the natural language descriptions in order
to efficiently play the text-based adventure games.
There are multiple future work directions we would like to
point out. First, and one of the most important, is to embed
a learning mechanisms: the in-game learning, that uses restart
functionality to improve player efficiency in one particular
game; and preliminary learning, that is able to gain useful
knowledge from playing entire set of games. Also, we plan to
take a closer look at the decompiled game codes, as we believe
that analyzing them may provide very useful knowledge.
We would like to improve the battle mode behavior, especially mitigate the agent and make it more sensitive to the
particular situation. We hope that the mapping mechanism can
be further extended to allow the casual approach, where the
agent travels to distant locations for some specific reason (e.g.
item usage), instead of a simple reactive system that we have
now.
Lastly, we would like to continue the domain-based approach, and so focus our efforts on discovering the subgames
(like we did with fighting and exploring) that we are able to
properly detect, and handle significantly better than the general
case.
ACKNOWLEDGMENTS
The authors would like to thank Szymon Malik for his
valuable contribution in the early stage of developing Golovin.
We would also like to thank Nancy Fulda for helpful
answers to our questions and providing up-to date results of
the BYU-Agent.
R EFERENCES
[1] M. Campbell, A. J. Hoane, and F. Hsu, “Deep Blue,” Artificial intelligence, vol. 134, no. 1, pp. 57–83, 2002.
[2] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den
Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever,
T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis,
“Mastering the game of Go with deep neural networks and tree search,”
Nature, vol. 529, pp. 484–503, 2016.
[3] M. Genesereth and M. Thielscher, General Game Playing. Morgan &
Claypool, 2014.
[4] M. Genesereth, N. Love, and B. Pell, “General Game Playing: Overview
of the AAAI Competition,” AI Magazine, vol. 26, pp. 62–72, 2005.
[5] N. Love, T. Hinrichs, D. Haley, E. Schkufza, and M. Genesereth,
“General Game Playing: Game Description Language Specification,”
Stanford Logic Group, Tech. Rep., 2006.
[6] M. Thielscher, “A General Game Description Language for Incomplete
Information Games,” in AAAI Conference on Artificial Intelligence,
2010, pp. 994–999.
[7] J. Kowalski and A. Kisielewicz, “Towards a Real-time Game Description Language,” in International Conference on Agents and Artificial
Intelligence, vol. 2, 2016, pp. 494–499.
[8] D. Perez, S. Samothrakis, J. Togelius, T. Schaul, S. Lucas, A. Couëtoux,
J. Lee, C. Lim, and T. Thompson, “The 2014 General Video Game
Playing Competition,” IEEE Transactions on Computational Intelligence
and AI in Games, vol. 8, no. 3, pp. 229–243, 2015.
[9] D. Perez, S. Samothrakis, J. Togelius, T. Schaul, and S. M. Lucas,
“General Video Game AI: Competition, Challenges and Opportunities,”
in AAAI Conference on Artificial Intelligence, 2016, pp. 4335–4337.
[10] A. Khalifa, D. Perez, S. Lucas, and J. Togelius, “General Video
Game Level Generation,” in Genetic and Evolutionary Computation
Conference, 2016, pp. 253–259.
[11] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.
Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski
et al., “Human-level control through deep reinforcement learning,”
Nature, vol. 518, no. 7540, pp. 529–533, 2015.
[12] S. Kelly and M. I. Heywood, “Emergent Tangled Graph Representations
for Atari Game Playing Agents,” in EuroGP 2017: Genetic Programming, ser. LNCS, 2017, vol. 10196, pp. 64–79.
[13] M. A. DePristo and R. Zubek, “being-in-the-world,” in Proceedings
of the 2001 AAAI Spring Symposium on Artificial Intelligence and
Interactive Entertainment, 2001, pp. 31–34.
[14] E. Amir and P. Doyle, “Adventure games: A challenge for cognitive
robotics,” in Proc. Int. Cognitive Robotics Workshop, 2002, pp. 148–
155.
[15] N. Fulda, D. Ricks, B. Murdoch, and D. Wingate, “What can you do with
a rock? Affordance extraction via word embeddings,” in International
Joint Conference on Artificial Intelligence, 2017, (to appear).
[16] A. Koller, R. Debusmann, M. Gabsdil, and K. Striegnitz, “Put my
galakmid coin into the dispenser and kick it: Computational linguistics
and theorem proving in a computer game,” Journal of Logic, Language
and Information, vol. 13, no. 2, pp. 187–206, 2004.
[17] B. Hlubocky and E. Amir, “Knowledge-gathering agents in adventure
games,” in AAAI-04 workshop on Challenges in Game AI, 2004.
[18] A. Chang and E. Amir, “Goal Achievement in Partially Known, Partially
Observable Domains,” in Proceedings of the Sixteenth International
Conference on International Conference on Automated Planning and
Scheduling, 2006, pp. 203–211.
[19] K. Narasimhan, T. Kulkarni, and R. Barzilay, “Language Understanding
for Text-based Games using Deep Reinforcement Learning,” in Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing, 2015, pp. 1–11.
[20] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural
Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
[21] A. M. Turing, “Computing machinery and intelligence,” Mind,
vol. 59, no. 236, pp. 433–460, 1950. [Online]. Available: http:
//www.jstor.org/stable/2251299
[22] R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, and P. P.
Kuksa, “Natural language processing (almost) from scratch,” CoRR, vol.
abs/1103.0398, 2011. [Online]. Available: http://arxiv.org/abs/1103.0398
[23] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation
of word representations in vector space,” CoRR, vol. abs/1301.3781,
2013. [Online]. Available: http://arxiv.org/abs/1301.3781
[24] S. F. Chen and J. Goodman, “An empirical study of smoothing
techniques for language modeling,” in Proceedings of the 34th
Annual Meeting on Association for Computational Linguistics, ser.
ACL ’96. Stroudsburg, PA, USA: Association for Computational
Linguistics, 1996, pp. 310–318. [Online]. Available: http://dx.doi.org/
10.3115/981863.981904
[25] Y. Bengio, R. Ducharme, P. Vincent, and C. Jauvin, “A neural probabilistic language model,” JOURNAL OF MACHINE LEARNING RESEARCH, vol. 3, pp. 1137–1155, 2003.
[26] C. Raffel and D. P. W. Ellis, “Feed-forward networks with attention can
solve some long-term memory problems,” CoRR, vol. abs/1512.08756,
2015. [Online]. Available: http://arxiv.org/abs/1512.08756
[27] J. Cheng, L. Dong, and M. Lapata, “Long short-term memory-networks
for machine reading,” CoRR, vol. abs/1601.06733, 2016. [Online].
Available: http://arxiv.org/abs/1601.06733
[28] A. Newell, J. C. Shaw, and H. A. Simon, “Report on a general problem
solving program,” in Proceedings of the International Conference on
Information Processing, 1959, pp. 256–264.
[29] C. J. Watkins and P. Dayan, “Q-learning,” Machine learning, vol. 8, no.
3-4, pp. 279–292, 1992.
[30] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of
word representations in vector space,” 2013, arXiv:1301.3781 [cs.CL].
[31] M. A. et al., “TensorFlow: Large-scale machine learning on
heterogeneous systems,” 2015, software available from tensorflow.org.
[Online]. Available: http://tensorflow.org/
[32] S. Malik, “Application of artificial neural networks with attention mechanism for discovering distant dependencies in time series,” Bachelor
Thesis, University of Wrocław, 2016.
[33] S. Branavan, D. Silver, and R. Barzilay, “Learning to Win by Reading
Manuals in a Monte-Carlo Framework,” Journal of Artificial Intelligence
Research, vol. 43, pp. 661–704, 2012.
[34] S. Bird, E. Klein, and E. Loper, Natural Language Processing with
Python, 1st ed. O’Reilly Media, Inc., 2009.
[35] J. Snoek, H. Larochelle, and R. P. Adams, “Practical bayesian optimization of machine learning algorithms,” in Advances in Neural Information
Processing Systems 25, F. Pereira, C. J. C. Burges, L. Bottou, and K. Q.
Weinberger, Eds. Curran Associates, Inc., 2012, pp. 2951–2959.
| 2cs.AI
|
Reduction of some homological conjectures to excellent unique factorization domains
arXiv:1607.00025v2 [math.AC] 28 Jul 2016
E HSAN TAVANFAR∗
I N THE MEMORY OF MY FATHER , M ANOUCHEHR TAVANFAR ,
WHO PASSED AWAY AT THE TIME OF WRITING THIS PAPER
A BSTRACT. In this article, applying the quasi-Gorenstein analogous of the Ulrich’s
deformation of certain Gorenstein rings we show that some homological conjectures,
including the Monomial Conjecture, Big Cohen-Macaulay Algebra Conjecture as well as
the Small Cohen-Macaulay Conjecture reduce to the excellent unique factorization domains. Some other reductions of the Monomial Conjecture are also proved. We, moreover, show that certain almost complete intersections adhere the Monomial Conjecture.
1 Introduction
Throughout this article, (R, m) denotes a commutative Noetherian local ring of dimension d
with identity where m denotes the unique maximal ideal of R.
In the early seventies, in [29], Hochster proposed the following two equivalent conjectures and proved them for the equicharacteristic rings.
Monomial Conjecture: Suppose that x1 , . . . , xd is a system of parameters for R. Then,
t
x1 · · · xdt ∉ (x1t+1 , . . . , xdt+1 ) for each t ≥ 1.
Direct Summand Conjecture: Suppose that A is a regular local ring and R is a module
finite extension of A. Then the inclusion A → R splits, in the category of A-modules.
The foregoing conjectures have been around for decades in the mixed characteristic case.
In the sequel we succinctly review certain substantial advances in the study of the aforementioned conjectures. In the eighties, in his extraordinary paper [28], among other things,
Hochster introduced the following further equivalent forms of the Monomial Conjecture.
Canonical Element Conjecture: Suppose that (F• , ∂• ) is a free resolution of R/m and consider the natural projection π : Fd → syzdR (R/m) which defines an element,
∗
¢
¡
ǫR := [π] ∈ ExtdR R/m, syzdR (R/m) .
This research was in part supported by a grant from IPM and supported in part by NSF grant DMS 1162585.
2010 Mathematics Subject Classification. 13D22, 13H10.
Key words and phrases. Almost complete intersections, big Cohen-Macaulay algebra, homological conjectures, linkage, Monomial Conjecture, quasi-Gorenstein rings, unique factorization domains.
1
2
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
Set, ηR := φ(ǫR ), wherein,
¡
¢
¡
¢
¡ d
¢
d
n
d
d
φ : ExtdR R/m, syzdR (R/m) → lim
Ext
R/
m
,
syz
(R/
m
)
=
H
syz
(R/
m
)
,
m
R
R
R
−→
n∈N
is the natural map. Then ηR 6= 0.
CE Property, also known as the Canonical Element Conjecture: Assume that F• is minimal resolution of R/m. Consider the Koszul complex K• (x, R) of a system of parameters x of R.
Then for any lifting, φ : K• (x, R) → F• , of the natural surjection R/(x) → R/m, we have φd 6= 0.
In [28], Hochster, moreover shows that the mentioned equivalent forms of the Monomial
Conjecture imply the following conjecture:
Improved New Intersection Conjecture: Suppose
¡
¢that F• : 0 → Fs → Fs−1 → · · · → F0 → 0
is a complex of finite free modules such that ℓ Hi (F• ) < ∞ for each i > 1 and H0 (F• ) has a
minimal generator whose generated R-module has finite length. Then s ≥ d .
The important point here is that a slightly stronger version of the Improved New Intersection Conjecture, wherein H0 (F0 ) has finite length as well, has been settled in full generality
by Roberts (see,[47]). The validity of this stronger form, which is called the New Intersection
Theorem, itself established a bunch of other old conjectures, most notably the Bass’s question regarding the Cohen-Macaulayness of rings having a non-zero finite module of finite
injective dimension (see, [10, Chapter 9.] and [44]). Originally the Improved New Intersect
Conjecture is, implicitly, stated in [16] among the proof of [16, Theorem 1.1]. Subsequently,
miraculously, in [12] Dutta showed that the improved new intersection is, in fact, an another
equivalent form of the Monomial Conjecture. Recently, in [14] Dutta also proved that all of
the equivalent forms of the Monomial Conjecture are in addition equivalent to the following
conjecture which has its origin, again, in [16] (see, [16, Proposition 1.6.]).
Order Ideal Conjecture: Assume that R is a regular ring and M is an R-module. Suppose
that x is a minimal generator of the i th-syzygy of M for some i > 0. Then the ideal,
¡
¢
O syzi (M) (x) = { f (x) : f ∈ HomR syzi (M), R },
has height greater than or equal to i .
An R-module M (algebra A) is said to be a balanced big Cohen-Macaulay module ( big
Cohen-Macaulay algebra) if mM 6= M (mA 6= A) and any system of parameters x of R is a regular sequence on M (on A). The existence of balanced big Cohen-Macaulay module (big
Cohen-Macaulay algebra), which is a long standing conjecture, settles the Monomial Conjecture affirmatively. In [32] Hochster showed that any equicharacteristic ring has a balanced
big Cohen-Macaualy module. Moreover, in [35] Hochster and Huneke proved that if R is an
excellent domain of prime characteristic, then, R+ , the integral closure of R in the algebraic
closure of its fraction field is a big Cohen-Macaulay algebra. Unfortunately, using the trace
map, we can deduce that the analogous statement about R+ in equal characteristic zero does
not hold provided dim(R) ≥ 3. In spite of this obstruction, in [33] Hochster and Huneke established the existence of big Cohen-Macaulay algebras in the equicharacteristic case.
Although the Monomial Conjecture is quite easy in dimension less than or equal to two,
but not in dimension three. In [24] Heitmann provided a very complicated proof of the validity of Monomial Conjecture in dimension three of mixed characteristic. In fact he proved
more, that is, if R is a mixed characteristic excellent normal domain of dimension three then
E HSAN TAVANFAR
3
H2m (R+ ) is almost zero, i.e. R+ is an almost Cohen-Macaulay R-algebra1. Quite obviously, the
existence of almost Cohen-Macaulay algebra is stronger than the existence of big CohenMacaulay algebras. In general the Monomial Conjecture follows from the existence of almost Cohen-Macaulay algebra(see, e.g. [49, Proposition 1.3]). In [27] when R has dimension three and mixed characteristic Hochster constructed a big Cohen-Macaulay algebra using the Heitmann’s result regarding the almost Cohen-Macaulayness of R+ . Recently, in [8],
Bhattacharyya showed that the Big Cohen-Macaulay Algebra Conjecture follows from the existence of almost big Cohen-Macaulay algebras. The question whether R+ is always almost
Cohen-Macaulay in mixed characteristic is open in dimension greater than or equal to 4. Although the Monomial Conjecture, has been proved in equal characteristic zero but the question whether R+ is almost Cohen-Macaulay in equal characteristic zero is of self interest and
is open in dimension greater than or equal to 3. But the three dimensional equal characteristic zero case for the Segre product of Cohen-Macaulay N-graded rings has been investigated
in [49], where the authors apply the Albanese varieties to deduce that the image of non-top
local cohomologies of R are almost zero in R+ . Recently, in [48] Roberts showed that the image of non-top local cohomologies of R in R+ are almost zero provided R is a Segre product
of N-graded Cohen-Macaulay rings of mixed characteristic. Moreover, recently, applying the
Almost Purity Theorem, in [52] Shimomoto proved that R has a big Cohen-Macaulay algebra
provided R has mixed characteristic p > 0 such that R[1/p] is an étale extension of A[1/p]
wherein A is a Noether normalization of R.
A finitely generated big Cohen-Macaulay R-module is said to be a small Cohen-Macaulay
module or maximal Cohen-Macaulay module. Hochster’s Small Cohen-Macaulay Conjecture states that every complete ring has a small Cohen-Macaulay Module. This conjecture is
even open in three dimensional equicharacteristic rings.
In Section 2 we prove that for the validity of the Monomial Conjecture it suffices to take
into account only systems of parameters of R which form a part of a minimal basis of m.
This fact will be used in Section 4 to deduce the main result of that section. Here, it is worth
to point out that, since for proving this reduction we descend to a finite free extension of
R with complete intersection fibers, we can merge some of other known reductions of the
Monomial Conjecture with our reduction. For instance, it suffices to confine our attention
only on such system of parameters x := x1 , . . . , xd for almost complete intersections. In view
of [13] the Monomial Conjecture reduces to the almost complete intersections.
In Section 3 we show that the aforementioned homological conjectures reduce to the
class of excellent (homomorphic image of regular) unique factorization domains. The main
idea of the proof is the fact that the Ulrich’s deformation of certain Gorenstein rings to the
class of unique factorization domains, developed in [55], has a quasi-Gorenstein counterpart. But prior to applying this reduction, for a normal domain R with canonical ideal ωR , we
L
endow R ωR with a ring structure such that it is a quasi-Gorenstein domain which is a complete intersection in codimension 1 (recall that the trivial extension is never a domain). For
the Monomial Conjecture with the aid of Hochster’s first general grade reduction technique
developed in [31], we deduce that the reduction to excellent unique factorization domains
is dimensionwise in the sense that the Monomial Conjecture for d -dimensional local rings
1
Here, almost zero means that each element of H2m (R+ ) is killed by elements of arbitrary small valuation,
with respect to a fixed rank one valuation which is non-negative on R and positive on m.
4
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
descends to d -dimensional excellent unique factorization local domains. It is noteworthy to
mention that in the light of Heitmann’s remarkable paper, [23], loosely speaking, any complete ring with depth at least 2 is a completion of a unique factorization domain! But, unfortunately, Heitmann’s unique factorization domains are neither excellent nor catenary 2 .
For instance, his unique factorization domains are not supposed to have a maximal CohenMacaulay module or even a canonical module. So the excellence property can be important
from this point of view. Besides the known good properties of unique factorization domains,
they may have even more tacit important properties as we will point out them in the sequel.
The first example of a non-Cohen-Macaulay unique factorization domain was constructed
by virtue of the invariant theory, in Bertin’s paper [7]. In fact at that time, Bertin’s counterexample settled negatively Samuel’s conjecture regarding the Cohen-Macaulayness of unique
factorization domains. In spite of the existence of this counterexample, complete unique
factorization domains at least in equal characteristic zero, have very good properties. For
instance, as stated in [42, Page 539] (see, also [21]), if R is a complete equicharacteristic zero
unique factorization domain of depth ≥ 3, then R satisfies S 3 . Moreover, if the residue field
is algebraically closed (char 0), then the depth condition is superfluous. Furthermore, in
view of [42, Page 540], any complete unique factorization domain with algebraically closed
residue field of characteristic zero is Cohen-Macaulay provided dim(R) ≤ 4. So it might be,
fairly, more helpful, if we can take more steps and reduce any of the homological conjectures
to the class of complete unique factorization domains.
In Section 4, we investigate the homological conjectures for almost complete intersections satisfying m2 ⊆ (x) for some system of parameters x of R. In mixed characteristic we
furthermore assume that x1 = p, the residual characteristic of R. We will see that in this
case R has a balanced big Cohen-Macaulay algebra. In the sequel, we explain our motivation behind this investigation. As we mentioned before, the Monomial Conjecture is proved
for equicharacteristic rings of arbitrary dimension and for (at most) 3-dimensional rings of
any characteristic. Due to the Goto’s paper [20], another subclass of rings which adhere the
Monomial Conjecture ¡is Buchsbaum
rings. In fact any system
of parameters x of¢ a Buchs¢
S ¡ t+1
lim
lim
(x1 , . . . , xdt+1 ) : x1t · · · xdt . This inbaum ring satisfies, m {x}R ⊆ (x), wherein {x}R =
t∈N
clusion immediately settles the Monomial Conjecture. Now set A to be the class of rings
consisting of rings R satisfying,
¡
¢
m2 {x}lim
⊆ (x),
R
for some system of parameters x of R. Note that A contains, properly, the class of Buchsbaum rings (see, [54, Theorem 5.12(ii)]). In order to see why this inclusion is proper, note
that in the light of [19], there exists a non-Buchsbaum quasi-Buchsbaum ring R such that
there are exactly two non-zero non-top local cohomologies of R which both of them are R/mvector spaces. Then applying the argument of the proof of [46, Proposition 1.] we can conclude that R belongs to the class A . Note also that by [54, Theorem 5.12.(i)], roughly speaking, A is a subclass of generalized Cohen-Macaulay rings. Quite obviously, if the Monomial Conjecture is valid for any system of parameters x of any ring R with m2 ⊆ (x), then
2
In fact Heitmann’s construction of such unique factorization domains provided us with an example of a
non-catenary local unique factorization domain. Prior to this counterexample, it was conjectured that any
local unique factorization domain is catenary.
5
E HSAN TAVANFAR
any member of A satisfies the Monomial Conjecture. Although, at this time, we could not
prove the validity of the Monomial Conjecture in this general setting, but we succeeded to
deduce our desired statement for almost complete intersections satisfying m2 ⊆ (x) with the
extra assumption that x1 = p in the mixed characteristic p > 0. For justifying our investigation in this restricted case, from another point of view, recall that in the light of Dutta’s
result [13, 1.2 Proposition] in conjunction with [28, (6.1) Theorem] Monomial Conjecture reduces to the case where R is an almost complete intersection of mixed characteristic p > 0
and x1 , . . . , xd is a system of parameters for R with x1 = p. The key ingredient of our proof is
the fact that those almost complete intersections have multiplicities up to two in the nonCohen-Macaulay case. We give an example of a non-Cohen-Macaulay ring R ∈ A such that
its multiplicity is strictly greater than two (so, R is not an almost complete intersection).
In view of the Monomial Conjecture, we end the paper with a question regarding existence of a particular non-commutative R-algebra.
2 Reduction of the Monomial Conjecture to systems of parameters which
are part of a minimal basis of m
This section is devoted to prove that for the validity of the Monomial Conjecture it suffices
only to check systems of parameters of the ring which can be extended to a minimal set of
generators of the maximal ideal. This result will be used in the Section 4 to prove that certain
almost complete intersections satisfy the Monomial Conjecture.
L
Remark 2.1. Let a ∈ R. Then the free R-module R R acquires a ring structure via the following rule,
(r, s)(r ′ , s ′ ) = (r r ′ + ss ′ a, r s ′ + r ′ s).
L
We use the notation R(a 1/2 ) to denote the forgoing ring structure of R R. In fact it is easily
seen that the map, R(a 1/2 ) → R[X]/(X 2 − a), which takes (r, s) to (sX + r ) + (X 2 − a)R[X] is an
isomorphism of R-algebras. We are given the extension map R → R(a 1/2 ) by the rule r 7→
(r, 0) which turns R(a 1/2 ) into a free R-module with the basis {(1, 0), (0, 1)}. Consequently this
extension is an integral extension of R and it is subject to the following properties which all
are easy to verify.
¡
¢
(i) dim(R) = dim R(a 1/2 ) and a has a square root in R(a 1/2 ), namely (0, 1).
(ii) If a ∈ m then R(a 1/2 ) is a local ring with unique maximal ideal m
L
R.
(iii) If a, x2 , . . . , xd is a system of parameters of R then a 1/2 , x2 , . . . , xd is a system of parameters for R(a 1/2 ).
(iv) If ϕ : R → S is an R-algebra wherein, ϕ(a) has a square root in S, then there exists a
natural induced R-algebra homomorphism Ψ : R(a 1/2 ) → S defined by (r, s) 7→ ϕ(r ) +
ϕ(s)ϕ(a)1/2 , extending ϕ. Moreover, assuming that R and S are domains, Ψ is injective
if and only if ϕ is injective and a does not have a square root in Frac(R).
6
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
(v) If R is a domain, then R(a 1/2 ) is a domain if and only if a does not have a square
root in Frac(R). For one implication we note that if a does not have a square root in
K := Frac(R), then K[x]/(x 2 −a) is a field extension of R. Therefore, by applying the preceding part we get an embedding R(a 1/2 ) → K[x]/(x 2 − a) which shows that R(a 1/2 )
is an integral domain. For the reverse implication note that if a has a square root
r /s ∈ Frac(R) then the identity (r, s)(r, −s) = 0 shows that R(a 1/2 ) is not a domain.
The following remark is in fact the proof of Corollary 2.3(i).
Remark 2.2. Let x1 , . . . , xl be a sequence of elements of R contained in the maximal ideal of
R. We, inductively, construct the local ring (Ri , mi ) by taking a square root of xi in Ri −1 . Then
in Rl we have,
),
xi1/2 = (
0
, 0, . . . , 0, 1, 0, . . .,
0
|{z}
|{z}
0-th coordinate
(2l −1)-th coordinate
whose 2(i −1) -th coordinate is 1 and others are zero.
(i) Let 1 ≤ j ≤ l and 0 ≤ k ≤ 2l − 1. We denote the element (0, . . . , 0,
1
|{z}
, 0, . . . , 0) of Rl
k−th coordinate
by e k . Then we have,
e k x 1/2
j =
(
e k+2 j −1 ,
( j − 1)- th digit of k in base 2 is 0
x j e k−2 j −1 , ( j − 1)- th digit of k in base 2 is 1.
In order to see why this is the case we induct on the least natural number s ≥ j such that
k ≤ 2s − 1. In the case where s = j it is easily seen that the ( j − 1)-th digit of k in its 2-th
base representation is 0 (is 1) if and only if k ≤ 2 j −1 − 1 (k ≥ 2 j −1 ). So an easy use of the
L
multiplication rule of the ring R j := R j −1 R j −1 proves the claim (Recall that R j is subring of
Rl ). Now assume that s > j . Then we have,
³
¢ 1/2 ´
¡
,
0,
.
.
.
,
0
xj .
e k x 1/2
=
0,
.
.
.
,
0
,
0,
.
.
.
,
0,
1
j
|{z}
|{z}
2s−1 −1-th coordinate
(k−2s−1 )-th coordinate
Now set k ′ := k − 2s−1 . Note that the ( j − 1)-th digit of the base 2 representation of k and
k are equal. Consequently the statement follows from our inductive hypothesis.
′
(ii) We are going to show that for each 1 ≤ i ≤ l the projection map τ2(i −1) : m2l → R, which is
the projection to the (2(i −1) )-th coordinate, is not surjective. In the case where l = 1 we have
L
m2l = (m2 + x1 R) m. So, we assume that l ≥ 2 and the statement is true for smaller values of
l . Then,
¡
¢M
m2l = m2l −1 + xl Rl −1
ml −1 .
(1)
Now, if i = l then 2l −1 -th coordinate of m2l is just the first coordinate of,
M M M
ml −1 = m R · · · R.
Hence, clearly, τ2(l −1) is not surjective. On the other hand if i l then by our inductive hypothesis τ2(i −1) : m2l −1 → R is not surjective which, in the light of the equality (1), implies the
statement immediately.
7
E HSAN TAVANFAR
(iii) In continuation of our investigation in the previous part, we need to show, also, that the
projection map τ2(i −1) : x 1/2
Rl −1 → R is not surjective unless i = j (1 ≤ i ≤ l − 1 and 1 ≤ j ≤
j
l − 1). Let (r k )0≤k≤2l −1 −1 ∈ Rl −1 . Then we have,
(r k )0≤k≤2l −1 −1 x 1/2
j
2l −1
P−1
k=0
( j −1)−th digit of k in base 2 is 0
r k e k+2 j −1 +
=
2l −1
P−1
k=0
r k e k x 1/2
=
j
2l −1
P−1
k=0
( j −1)−th digit of k in base 2 is 1
r k x j e k−2 j −1 .
Thus if i < j then evidently τ2i −1 is not surjective. On the other hand if i > j and there
exits 0 ≤ k ≤ 2l −1 − 1 with k + 2 j −1 = 2i −1 then k = 2i −1 − 2 j −1 which after a straightforward
computation shows that the ( j −1)-th digit of k in base 2 is 1. This proves the non-surjectivity
of τ2i −1 .
(iv) By means of the arguments of the forgoing part we can, directly, conclude that,
1/2
xi1/2 ∉ m2l −1 + (x11/2 , . . . , xd
, . . . , xl1/2
−1 , x l )Rl −1 , (i
i
l)
otherwise we must have τ2(i −1) : (x 1/2
)Rl −1 → R for some 1 ≤ j ≤ l − 1 and j 6= i or τ2(i −1) :
j
m2l −1 → R is surjective.
The above remark yields the following corollary, as claimed before. In the second part
of the subsequent corollary the important point is the equality embdim(R) = d + u which
equals to the embedding dimension of the regular local ring V[[X 2 , . . . , X d , Y1 , . . . , Yu ]].
Corollary 2.3.
(i) The Monomial Conjecture holds if and only if every system of parameters
x1 , . . . , xd which is a part of a minimal basis for the maximal ideal satisfies the Monomial
Conjecture.
(ii) For the Monomial Conjecture, without loss of generality, we can assume that,
³
´
1/2
R = (V, p )[[X 2 , . . . , X d , Y1 , . . . , Yu ]] /I,
and the system of parameters x1 , . . . , xd is the image of p 1/2 , X 2 , . . . , X d in R where the
discrete valuation ring (V, p 1/2 ) is a subring of R and embdim(R) = d + u.
Proof. (i) Let x1 , . . . , xd be a system of parameters for R. Consider the ring Rd as in Remark
2.2. It suffices to show that x11/2 , . . . , xd1/2 is a part of a minimal basis for the unique maximal
L
ideal md of Rd . Let (α1 , β1 ), . . . , (αd , βd ) ∈ Rd = Rd−1 Rd−1 be such that,
¢M
¡
md−1 .
(αk , βk )xk1/2 + (αd , βd )(0Rd −1 , 1Rd −1 ) ∈ m2d = m2d−1 + (xd )Rd−1
{z
}
|
k=1
d−1
X
=xd1/2
Then by a simple computation we get
d−1
X
k=1
αk xk1/2 + βd xd ∈ m2d−1 + xd Rd−1 ,
(2)
8
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
and,
d−1
X
k=1
βk xk1/2 + αd ∈ md−1 .
(3)
So the identity (3) yields αd ∈ md−1 and thence (αd , βd ) ∈ md . Moreover, for each 1 ≤ k ≤
L
d − 1 we must have (αk , βk ) ∈ md = md−1 Rd−1 , otherwise we get αi ∉ md−1 for some 1 ≤
1/2
i ≤ d − 1 which in view of the identity (2) yields x 1/2 ∈ m2 + (x 1/2 , . . . , xd
, . . . , x 1/2 , x )R
i
d−1
1
i
d−1
d
d−1
violating Remark 2.2 (iv). Consequently, (αk , βk ) ∈ md for each 1 ≤ k ≤ d . This implies that
x11/2 + m2d , . . . , xd1/2 + m2d is a linearly independent subset of md /m2d over Rd /md and thence
x11/2 , . . . , xd1/2 is part of a minimal basis for md .
(ii) By [28, (6.1) Theorem], for the validity of the Monomial Conjecture it suffices to verify
it only for systems of parameters x1 , . . . , xd of R wherein x1 = p is the residual characteristic
and (V, pV) is a coefficient ring of the complete local ring R. So the statement follows from
the argument in the proof of the preceding part.
3 Reduction to excellent unique factorization domains
In this section we show that some homological conjectures reduce to excellent unique factorization domains. Here, by homological conjectures we mean one of the Small CohenMacaulay Conjecture, Big-Cohen-Macaulay Algebra Conjecture as well as the Monomial Conjecture.
Lemma 3.1. Suppose that R is a Noetherian local domain which is not a field. Then there
exists a ∈ m such that a does not have a square root in F := Frac(R).
Proof. We first claim that there exists a ∈ R without a square root in Frac(R). Assume to the
contrary that each element of R has a square root in F. Then F is square root closed. Let
p be a height one prime ideal of the integral closure R of R which is a Krull domain (see,
[39, Theorem 4.10.5]). Since Rp is an integrally closed domain, so the discrete valuation ring
(Rp , xRp ) is square root closed in Frac(R). But then x has a square root in Rp . This observation
violates x ∉ p2 Rp .
Hence there exists an element a ∈ R without a square root in Frac(R). If a ∈ m then we are
done, so suppose that a is an invertible element of R. Then for any 0 6= t ∈ m we have, t , or,
t a, does not have a square root in Frac(R).
Remark 3.2. Suppose that R is a domain. Since the trivial extension R ⋉ ωR R by its canonical module (even the amalgamated duplication ring of R with its canonical ideal3) is not a
L
domain we will use a different multiplication on R ωR . To be more precise, since R is a
domain so [4, (3.1)] implies that ωR is an ideal of R. By applying the preceding lemma we can
L
choose a ∈ m such that a does not have a square root in Frac(R). We may endow S := R ωR
with a ring structure such that,
(r, x)(r ′ , x ′ ) = (r r ′ + xx ′ a, r x ′ + r ′ x),
3
See, [6].
9
E HSAN TAVANFAR
L
for each (r, x), (r ′ , x ′ ) ∈ R ωR . In particular, S is a subring of R(a 1/2 ) and it is a domain by
Remark 2.1(v). Furthermore, ηR : R → S which maps an element r ∈ R to (r, 0) is an R-algebra
L
homomorphism. Note that S is a local ring with the maximal ideal n := m ωR and η is a local
monomorphism. As we will see in the following remark, S endowed with the aforementioned
multiplication is an integral domain which is a complete intersection in low dimension, provided R is a normal ring. The fact that it is a complete intersection in the low dimension is
essential for reducing the homological conjectures to the subclass of unique factorization
domains.
Lemma 3.3. Suppose that R is a normal domain and that ωR is the canonical ideal of R. Let
a ∈ Frac(R) be an element of m without a square root in Frac(R). Then the following statements
hold.
(i) R(a 1/2 ) is locally complete intersection at codimension ≤ 1.
L
(ii) The subring R ωR of R(a 1/2 ) is locally complete intersection at codimension ≤ 1.
¡
¢
T
Proof. (i) Let p ∈ Spec R(a 1/2 ) be a height one prime ideal and set q := p R. Firstly note
that ht¡R (q) = htR(a
¢ 1/2 ) p, as we have going down, going up and incomparability. In particular,
1/2
p ∈ ass qR(a ) . Since R is normal, Rq is a regular local ring. Consequently, Rq [x] is a regular domain. Hence Rq [x]/(x 2 − a)Rq [x] ∼
= Rq (a 1/2 ) is locally complete intersection. An easy
verification shows that,
l∈ass
Note that qR(a 1/2 ) = q
L
¢
l = {(r, s) ∈ R(a 1/2 ) : r 2 − s 2 a ∈ q}.
(4)
qR(a 1/2 )
q. Set, S := R(a 1/2 )\
l∈ass
¡
S
¢ l. Then there exists the natural em-
qR(a 1/2 )
¢
bedding ϕ : Rq → S R(a ) defined by a/s → (a, 0)/(s, 0). On the other hand a does not
have a square root in Frac(Rq ) = Frac(R). Thus by Remark 2.1(iv) ϕ induces the injection
¡
¢
Ψ : Rq (a 1/2 ) → S −1 R(a 1/2 ) ,
−1
¡
¡
[
1/2
such that maps any
(a/b, a ′ /b ′ ) ∈ Rq (a 1/2 ) to (ab ′ , a ′ b)/(bb ′ , 0). For each element
¡ element
¢
′ ′
−1
1/2
(r, s)/(r , s ) ∈ S R(a ) we have, by(4) , r ′2 − s ′2 a ∉ q and,
¡
¢
Ψ (r r ′ − ss ′ a)/(r ′2 − s ′2 a), (r ′ s − r s ′ )/(r ′2 − s ′2 a) = (r, s)/(r ′ , s ′ ).
¡
¢
−1
1/2
Hence, Ψ is an isomorphism.
So,
S
R(a
)
, is¡ locally ¢complete intersection. Conse¡
¢
quently, as p ∈ ass qR(a 1/2 ) , we have R(a 1/2 )p ∼
= S −1 R(a 1/2 ) S −1 p is a complete intersection.
L
(ii) Let p′ ∈ Spec(R ωR ) be a height¡ one prime
ideal. Since R(a 1/2 ) is an integral exten¢
L
sion of R ωR , so there exists p ∈ Spec R(a 1/2 ) lying over p′ . Let q be the contraction of p′
in R. Note that htR (q) = htR(a 1/2 ) (p), because R(a 1/2 ) is an integral flat extension of R. On
L
the other hand, because R is integrally closed so R → R ωR has the going down property.
Hence, htR(a 1/2 ) (p) = htR (q) ≤ htR L ωR (p′ ). This shows that htR(a 1/2 ) (p) = 1, as well.
L
If ωR * q, then it is easily verified that the natural map (R ωR )p′ → R(a 1/2 )p is an isomorphism.
10
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
L
L
Now, we deal with the case where ωR ⊆ q. Note that (R ωR )p′ is a localization of (R ωR )q .
L
So, it suffices to show that the latter is (locally) complete intersection. Moreover, (R ωR )q is
L
an Rq -algebra by the map r /t 7→ (r, 0)/t and this underlying Rq -module structure of (R ωR )q
L
L
L
is the same Rq module (R ωR )q ∼
= Rq Rq . In particular, (R ωR )q is a flat Rq -algebra. So it
L
is sufficient to show that the closed fiber of Rq → (R ωR )q is a complete intersection. Since
(ωR )q ⊆ qRq so it is easily seen that there exists a natural ring isomorphism,
M
M
(R
ωR )q /(q
qωR )q → Rq /qRq ⋉ (ωR )q /q(ωR )q ,
such that maps an element (r, s)/t to (r /t , s/t). Now from the fact that (ωR )q ∼
= Rq , we can
deduce that Rq /qRq ⋉ (ωR )q /q(ωR )q is isomorphic to to the complete intersection ring,
(Rq /qRq )[X]/(X 2 ).
In [55] the author proves that any Gorenstein ring which is a homomorphic image of
a regular ring and it is a complete intersection at codimension ≤ 1 is a specialization of a
unique factorization domain. The first part of the proof of the following proposition is more
or less repeating the proof of the Ulrich’s result, and is stated here for the convenience of the
reader and also for the sake of completeness.
Proposition 3.4. Suppose that P is a regular ring and R := P/a is a quasi-Gorenstein ring.
Assume, furthermore, that R is locally complete intersection at codimension ≤ 1. Then there
exists a unique factorization domain S (which is of finite type over P) and a regular sequence
y of S such that R ∼
= S/(y).
Proof. The general idea, here, is similar as given in [55, Proposition 1]. Firstly, we are going
to use the idea of generic linkage presented in [40] to be sure that the almost complete intersection linked to a is a prime ideal which is a complete intersection at codimension ≤ 1. To
be more precise, let a1 , . . . , an be a system of generators for a and assume that grade(a) = g .
n
P
Yi ,j a j . Then, in the light of, [40, ProposiSet, Q = P[Yi ,j : 1 ≤ i ≤ g , 1 ≤ j ≤ n] and ci =
j =1
tion 2.9.(b)] and [40, Proposition 2.6], the linked ideal b = (c1 , . . . , c g ) :Q aQ, to aQ, is a prime
almost complete intersection which is a complete intersection at codimension ≤ 1. In particular, in the light of [38, (5), page 268] one can deduce that b is generated by a d-sequence
and thence [38, Theorem 3.1.] implies that sym(b) ∼
= R(b) wherein R(b) denotes the Rees algebra of b and sym(b) stands for the symmetric algebra of b. Suppose that, b = (c1 , . . . , c g , h).
Now, in the light of [37, Theorem 2.2] in conjunction with [30, Thoerem 1], the extended Rees
algebra Q[t −1 , bt ] is a unique factorization domain and whence so is
U := Q[t −1 , bt ]ht = Q[bt ]ht = sym(b)ht = (Q[Z 0 , . . . , Z g ]/d)Z0 ,
wherein, d =< {
g
P
i =0
r i Z i |r 0 h +
g
P
i =1
r i ci = 0} >. Now, as stated in [55, Proposition 1], U has an
Z-grading structure such that,
V := Q[Z 0 , Z 1 , . . . , Z g ]/(d + (Z 0 − 1)),
11
E HSAN TAVANFAR
is its degree zero subring and since the unique factorization domain U has an invertible
element of degree one it is easily seen that V is also a unique factorization domain.
The reminder of the proof differs with [55, Proposition 1]. Note that if we localize V at its
maximal ideal n := (m, Y1 , . . . , Yng , Z 1 , . . . , Z g , Z 0 −1) then R = Vn /(Y1 , . . . , Yng , Z 1 , . . . , Z g ). Hence
in order to deduce the statement it suffices to show that Z 1, . . . , Z g forms a regular sequence of
V, i.e. Z 0 −1, Z 1 , . . . , Z g is a regular sequence on sym(b). We achieve this with the aid of theory
of Z -complexes which is introduced and investigated in [26]. We use the notation Z i to denote the i -th cycles of the Koszul complex of Q with respect to the sequence c := h, c1 , . . . , c g .
Let S = Q[Z 0 , . . . , Z g ], be the polynomial ring over Q. Then the Z -complex associated to the
sequence c is the complex,
Z := 0 → Z g
O
S(−g ) → · · · → Z 2
Q
O
Q
whose differential is defined by the rule,
S(−2) → Z 1
O
S(−1) → S → 0,
Q
i
O
O ¢ X ¡X
¡X
¢
∂ ( q j e j1 ∧ · · · ∧ e ji )
Zk f .
f = qj
(−1)k e j 1 ∧ · · · ∧ ec
jk ∧ · · · ∧ e ji
j
j
k=1
It is easily seen that H0 (Z ) = sym(b) = S/d. Moreover, Z is acyclic by virtue of [26, Theorem
12.5] because any d -sequence is a proper sequence(see, [26, Definition 6.1.]). Using these
two facts and by considering the spectral sequence arising from the double complex,
O
Z
K• (Z 0 − 1, Z 1 , . . . , Z g ; S),
S
we can deduce that Z 0 − 1, Z 1 , . . . , Z g is a regular sequence on sym(b) if and only if,
O
S/(Z 0 − 1, Z 1 , . . . , Z g ),
Z
S
remains acyclic. Let e 0 denotes the basis of K1 (h, c1 , . . . , c g ; Q) = Qg +1 corresponding to the
element h. Then each element ζi of Z i can be represented as ζi = z i + z i′ ∧ e 0 such that z i ∈
L
L
′
j 1 6=0 Qe j 1 ∧···∧e j i −1 ( Ki −1 (h, c 1 , . . . , c g ; Q). Note
j 1 6=0 Qe j 1 ∧···∧e j i ( Ki (h, c 1 , . . . , c g ; Q) and z i ∈
that,
O¡
¢
Z
S/(Z 0 − 1, Z 1 , . . . , Z g ) =
S
0 → Zg
O
Q
→ Z1
O
Q[Z 0 ]/(Z 0 − 1) → · · · → Z 2
O
Q[Z 0 ]/(Z 0 − 1) →
Q[Z 0 ]/(Z 0 − 1) → Q[Z 0 ]/(Z 0 − 1) → 0,
Q
¡
N ¢
N
whose differentials satisfy, ∂ (z i + z i′ ∧ e 0 ) f = (−1)i −1 z i′ f ,4 , for each z i + z i′ ∧ e 0 ∈ Z i .
N
Let, α, be a cycle in Z i . Without loss of generality we can say, α = (z i + z i′ ∧ e 0 ) 1. Thus,
N
More precisely, note that Z S S/(Z 0 − 1, Z 1 , . . . , Z g ) is a subcomplex of the complex,
O
O
O
K := 0 → Kg +1
S/(Z 0 − 1, Z 1 , . . . , Z g ) → Kg
S/(Z 0 − 1, Z 1 , . . . , Z g ) → · · · → K0
S/(Z 0 − 1, Z 1 , . . . , Z g ) → 0,
4
Q
Q
Q
12
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
N
N
z i′ 1 = 0 which implies that, z i′ = 0. Hence, α = z i 1. In particular, z i ∈ Z i (c1 , . . . , c g ; Q).
Now the exactness of K• (c1 , . . . , c g ; Q) implies that there exists ζi ∈ Ki +1 (c1 , . . . , c g ; Q) such that
∂(ζi ) = z i . In particular, −hζi + (−1)i z i ∧ e 0 ∈ Z i +1 . Now using the algebra structure we get,
α = zi
O
³¡
¢O ´
1 = ∂ − hζi + (−1)i z i ∧ e 0
1 .
This observation concludes the proof.
Recall that by homological conjectures we mean one of the Small Cohen-Macaulay Conjecture, Big-Cohen-Macaulay Algebra Conjecture as well as the Monomial Conjecture. The
first part of the subsequent theorem reduces the homological conjectures to the class of
quasi-Gorenstein domains. As pointed out in the introduction, as the main result of this
section, the second part of the subsequent theorem gives the promised descent to excellent unique factorization domains. The integral closure of a complete quasi-Gorenstein ring
is not necessarily quasi-Gorenstein. Therefore, we could not reduce to the normal complete quasi-Gorenstein domains directly by taking integral closure in the first part. But by
virtue of the second part this is achievable which is stated separately in the third part. Note
that, as stated in [55], the completion of the unique factorization domains in [55] are not
unique factorization domain necessarily. However any possible reduction of the homological conjectures to the complete unique factorization domains would be fairly helpful as the
complete unique factorization domains have even more interesting properties at least in
the equal characteristic zero. More precisely, any equal characteristic zero complete unique
factorization domain with algebraically closed residue field of dimension up to 4 is CohenMacaulay. Furthermore, with mild conditions, any complete unique factorization domain of
equal characteristic zero with algebraically closed residue field satisfies the Serre-condition
S 3 (see, [42, page 540] or [21] for both of the aforementioned results).
Theorem 3.5. The following statements holds.
(i) If each complete quasi-Gorenstein domain, which is locally complete intersection at
codimension ≤ 1, satisfies the homological conjectures then, each local ring satisfies homological conjectures. This reduction is dimensionwise, in the sense that homological
conjectures for d -dimensional local rings reduce to the d -dimensional complete quasiGorenstein domains.
(ii) The homological conjectures reduce to the excellent (and homomorphic image of regular) unique factorization domains. This reduction is dimensionwise for the Monomial
Conjecture.
wherein Ki denotes the i -th component of the Koszul complex K• (h, c 1 , . . . , c g ; Q). The differential of K is dei
N
P
N
fined by, ∂(e j 1 ∧· · ·∧e j i f ) =
(−1)k e j 1 ∧· · ·∧ ec
Z k f . The complex K has a DG-algebra structure.
j k ∧· · ·∧e j i
k=1
Using this DG-algebra structure of K the identity can be easily verified because in view of the definition of ∂
in conjunction with the fact Z i has zero image in S/(Z 0 − 1, Z 1 , . . . , Z g ) for each 1 ≤ i ≤ g we can conclude that
∂(zi ) = ∂(zi′ ) = 0. This DG-algebra technique is used in the proof of [22, Lemma 4.3].
13
E HSAN TAVANFAR
(iii) The homological conjectures reduce to quasi-Gorenstein normal complete domains. This
reduction is dimensionwise for the Monomial Conjecture.
(iv) For the Canonical Element Conjecture it suffices only to consider mixed characteristic
complete normal quasi-Gorenstein rings of minimum possible depth, i.e. depth 2.
Proof. (i) We may suppose that R is a complete normal domain. In particular R satisfies the
L
S 2 -condition. Let ωR be the canonical ideal of R and set S := R ωR as stated in the Remark
3.2. Note that S is a complete local ring. By [4, Proposition 1.2] we have,
M d
¡ d
¢
¡
¢
Hm (R), E(R/m)
HomR Hdn (S), E(R/m) ∼
= HomR Hm (ωR )
O
¡ d
¡
¢
¢M
∼
ωR , E(R/m)
HomR Hd (R), E(R/m)
= HomR H (R)
m
m
R
∼
= HomR (ωR , ωR )
M
∼
ωR = S.
=R
M
ωR
¡
¢
Hence we get an isomorphism, S → HomR Hdn (S), E(R/m) , of R-modules which is an Sisomorphism too5 . In particular, Hdn (S) ∼
= HomR (S, E(R/m)), as S-modules because on the
category of S-modules two functors HomR (−, E(R/m)) and HomS (−, E(S/n)) are naturally
equivalent . So,
³
O
¡
¢
¡
¢´
¢
¡
HomS S/n, Hdn (S) ∼
S, E(R/m)
= HomS S/n, HomR S, E(R/m) ∼
= HomR S/n
S
¡
¢
∼
= HomR R/m, E(R/m)
∼
= R/m
∼
= S/n.
This observation implies that Hdn (S) has one dimensional socle. Since a does not have a
square root in Frac(R), so Remark 2.1 (v) implies that R(a 1/2 ) and whence its subring S is an
integral domain. Thus, by [54, Theorem 5.7], S is a quasi-Gorenstein domain. Since S is a
finitely generated (as R-module) complete quasi-Gorenstein extension domain of R so the
statement follows.
(ii) Using the previous part in conjunction with Lemma 3.3 and Proposition 3.4 we may
assume that R is a complete quasi-Gorenstein domain such that R = S/yS wherein (S, n) is a
5
In order to see why this is also an S-isomorphism, we fix an R-isomorphism,
¡
¢
ϕ : ωR → HomR Hdm (R), E(R/m) .
Then ¡the composition
of the above isomorphisms which is the mentioned R-isomorphism S →
¢
L
HomR Hdn (S), E(R/m) maps an element (r, α) ∈ R ωR = S to the R-homomorphism defined by the rule,
[(r ′ , α′ ) + (xn )] ∈ Hdn (R
wherein x is a system of parameters for R.
homomorphism.
M
¡
¢¡
¢
ωR ) 7→ ϕ(r α′ + r ′ α) [1 + (xn )] ,
Then it is easily verified that the map is, moreover, an S-
14
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
unique factorization domain and y := y 1 , . . . , y n is a regular sequence contained in S. Clearly,
if S satisfies the homological conjectures then so does R. This proves the first claim. For
proving the second claim we invoke the Hochster’s first general grade reduction technique
developed in [31]. Let a1 , . . . , ad be a system of parameters for R. Assume that the Monomial
Conjecture is valid for every d -dimensional unique factorization domain but, to the contrary,
a1t · · · adt ∈ (a1t+1 , . . . , adt+1 ),
for some t ∈ N. So we get the system of parameters y 1 , . . . , y n , y n+1 , . . . , y n+d for S such that
y i +n is a lift of ai in S for each 1 ≤ i ≤ d and
t
t
t+1
t+1
y n+1
· · · y n+d
∈ (y 1 , . . . , y n , y n+1
, . . . , y n+d
).
t+1
t+1
Since R satisfies the S 2 -condition so y 1 , . . . , y n , y n+1
, y n+2
is a regular sequence of S. As n +2 ≥
3 so [31, Theorem. c)] implies that the first general grade reduction,
T = S[X 1 , . . . , X n+2 ]/(
n
X
t+1
t+1
y k X k + y n+1
X n+1 + y n+2
X n+2 ),
k=1
of S is again a (non-local) unique factorization domain and so is its localization at nT. We
have, dim(TnT ) = dim(S)−1. Furthermore, as X 1 is invertible in TnT so y 2 , . . . , y n+d is a system
t+1
t+1
of parameters for TnT , y 2 , . . . , y n , y n+1
, y n+2
is a regular sequence of TnT and
t
t
t+1
t+1
y n+1
· · · y n+d
∈ (y 2 , . . . , y n , y n+1
, . . . , y n+d
)TnT .
Proceeding in this way, we get a unique factorization domain U with dim(U) = dim(R) = d
such that the image of y n+1 , . . . , y n+d in U is a system of parameters for U which does not
satisfy the Monomial Conjecture. This is a contradiction.
(iii) This is followed by taking the completion of the unique factorizations of the second
part, because by [17, Lemma (2.4)] any unique factorization domain with a canonical module is quasi-Gorenstein. Note also that the completion remains normal, as the formal fibers
of an excellent ring are regular.
(iv) In order to see why this is the case firstly note that by [28] the Canonical Element
Conjecture for all local rings is valid if and only if all local rings satisfies the Monomial Conjecture. Therefore in view of Theorem 3.5 (iii) the Canonical Element Conjecture reduces
to the mixed characteristic normal quasi-Gorenstein complete domains. So let R be such
a quasi-Gorenstein normal domain. Without loss of generality we may assume that R has
infinite residue field. If Depth(R) 2 then applying [43, Theorem 4.4(Local Bertini Theorem)] there exist x ∈ R such that R/xR is a normal domain. In particular R/xR satisfies the
S 2 -condition and thence R/xR is a quasi-Gorenstein normal domain in view of [54, CorolS
lary 3.4.(i)]. Consequently, [54, Proposition 3.1.] implies that x ∉
¡
¢ p. In particular,
p∈Att Hdm−1 (R)
if R/xR satisfies the Canonical Element Conjecture then so does R in the light of [28, (5.3)
Theorem]. Proceeding in this fashion the claim follows.
The theory of generic linkage had an important role in the reduction of the homological
conjectures to unique factorization domains. In the second part of the subsequent remark
15
E HSAN TAVANFAR
we, again, observe that how the linkage theory is related to the homological conjectures. In
the first part we show that the validity of the Canonical Element Conjecture in an open subset
implies the validity in general.
Remark 3.6. (i) The following approach to the Monomial Conjecture might be helpful: The
Canonical Element Conjecture is valid if it is valid, in some sense, in an open dense subset. More precisely, in view of the first part of Theorem 3.5(iii), we can assume that R is a
complete normal quasi-Gorenstein domain of mixed characteristic with Depth(R) ≥ 3. Let
(V, πV , k) be a coefficient ring of R. Let x0 , . . . , xn be a minimal generating set of m. If there exists a dense Zariski open subset O ⊆ Pd (k) such that for any a = (a0 : · · · : an ) in the preimage
n
P
ai xi satisfies the Canonical
of O in Pn (V), as described in [43, Definition 2.1.], we have R/
i =0
Element Conjecture then the Canonical Element Conjecture holds. Indeed, if this is the case
then we can apply [43, Theorem 4.4(Local Bertini Theorem)] to find an element a ∈ R such
that R/aR is a normal domain satisfying the Canonical Element Conjecture. Then [54, Theorem 3.4(i)] and [54, Propositino 3.1.] in conjunction with [28, (5.3) Theorem] imply that R
satisfies the Canonical Element Conjecture.
(ii) In [53, page 158] the following question is proposed. Suppose a and b are ideals in
a d -dimensional Gorenstein ring R which are geometrically linked w.r.t. the empty regular
sequence. Suppose R possess a maximal Cohen-Macaulay module M with 0 :R M = a. Does R
possess a maximal Cohen-Macaulay module N with 0 :R N = b? We show that if this question
has an affirmative answer then Small Cohen-Macaulay Conjecture is true in dimension 3.
Furthermore if Monomial Conjecture analogous of this question has an affirmative answer
then the Monomial Conjecture holds in any dimension.
By Theorem3.5(i) the Small Cohen-Macaulay Conjecture reduces, dimensionwise, to the
class of complete quasi-Gorenstein domains. Let P be a regular local ring and a be a quasiGorenstein prime ideal of P. As in the proof of [13, 1.2. Proposition] Choose a maximal regular sequence x1 , . . . , xh contained in a such that aPa = (x1 , . . . , xh )Pa . Then, a and (x1 , . . . , xh ) :P
a are geometrically linked and since a is a quasi-Gorenstein ideal so b := (x1 , . . . , xh ) :P a is
an unmixed almost complete intersection. Therefore P/b has positive depth and whence
Lemma 4.1(ii) implies that Depth(ωP/b ) ≥ 3. In particular, in the case of dimension three we
have ωP/b is a maximal Cohen-Macaulay P/b-module and therefore an affirmative answer
to the mentioned question implies the validity of the Small Cohen-Macaulay Conjecture in
dimension 3.
For the Monomial Conjecture analogous of the Strooker and Stückrad’s question assume
in addition that a is a normal quasi-Gorenstein prime ideal of P. Then the claim follows by
[15, step 2., page 238]. In spite of this observation, it is by no means clear for us that how the
theory of linkage can be helpful for this type of problems, in this point of view.
Any successful proof of the Monomial Conjecture for the class of generalized CohenMacaulay domains will settle the Monomial Conjecture in dimension three (by means of
a new proof). Although we do not know how to prove the Monomial Conjecture for this
type of rings but we show that again the problem reduces to the class of quasi-Gorenstein
generalized Cohen-Macaulay rings.
Proposition 3.7. The following statements hold.
16
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
(i) If every complete quasi-Gorenstein generalized Cohen-Macaulay ring satisfies the Monomial Conjecture then so does every generalized Cohen-Macaulay ring.
(ii) Monomial Conjecture holds for every complete generalized Cohen-Macaulay domain if
and only if every complete generalized Cohen-Macaualy quasi-Gorenstein domain satisfies Monomial Conjecture.
Proof. (i) Let R be a complete generalized Cohen-Macaulay ring. Denote by S the S 2-ification
of R. By [3, Corollary 4.3] ωRp , for each p ∈ Spec(R), is either the canonical module of the
∼
Cohen-Macaulay
ring
¡
¢ Rp or is zero. Hence, for every p ∈ Spec(R)\m either S p is zero or S p =
∼
∼
HomRp (ωR )p , (ωR )p = HomRp (ωRp , ωRp ) = Rp . On the other hand we have,
\
¡
¢
AssR (S) = AssR HomR (ωR , ωR ) = SuppR (ωR ) AssR (ωR ) = AsshR (R).
Therefore, by [9, 9.5.7 Exercise] S is a generalized Cohen-Macaulay R-module. Let n ∈ Max(S)
such that dim(S) = dim(S n ). Now, we show that S n is a generalized Cohen-Macaulay ring.
Denote the natural map R → S by η. We know that S is a finitely generated R-module. So
S
p and whence
dim(R/ker η) = dim(S). By [28, (4.2) Lemma] we have a := ker(η) ⊆
p
∈Assht(R)
T
dim(S) = dim(R). Since S is integral over R/a so m/a = n R/a and thereby mS n is an nS n primary ideal by the incomparability property of R/a → S. There exists l ≥ 0 such that
ml Him (S) = 0 (0 ≤ i ≤ dim(S) − 1). So that ml HimS n (S n ) = 0 (0 ≤ i ≤ dim(S) − 1) for some l ≥ 0.
Thus, (nS n )k HinS n (S n ) = 0 (0 ≤ i ≤ dim(S n ) − 1). Hence S n is a generalized Cohen-Macaulay
ring. We have S n is S 2 . Hence by passing to the completion if necessary, we can assume that
R is an S 2 -complete generalized Cohen-Macaulay ring. Therefore, in view of [3, Theorem
2.11.] the trivial extension T of R by its canonical module is a complete quasi-Gorenstein
ring. We prove that T is a generalized Cohen-Macaulay ring. Clearly if T satisfies the Monomial Conjecture then so does R.
Let q ∈ Spec(T)\{m ⋉ ω}. Then there exists p ∈ Spec(R)\{m} such that q = p ⋉ ωR . So
∼
Tq = (R ⋉ ωR )p⋉ωR ∼
= Rp ⋉ (ωR )p ∼
= Rp ⋉ ωRp . Hence, Tq is a Cohen-Macaulay ring. As T is
equidimensional, [9, 9.5.7 Exercise] implies that T is generalized Cohen-Macaulay.
(ii) Let R be a complete generalized Cohen-Macaulay domain. Set, S := HomR (ωR , ωR ).
Then, by virtue of [34, (3.6)] S is local6 . Also we know that S is complete and domain. By
a similar argument as in the preceding part, S is generalized Cohen-Macaulay. Hence, as
L
stated in the proof of Theorem 3.5(i), T = S ωS ⊆ S(a 1/2 ), is a local quasi-Gorenstein complete domain wherein a is an element of S without a square root in Frac(S). It is enough to
show that T is a generalized Cohen-Macaulay ring.
L
L
L
Let p ∈ Spec(S)\{m}. Then (S ωS )p ∼
= S p (ωS )p ∼
= S p ωS p , is a Cohen-Macaulay S p L
S
module. On the other hand, AssS (S ωS ) = AssS (S) AsshS (ωS ) = AssS (S) = {0}. Hence T is
a generalized Cohen-Macaulay S-module. Now, a similar argument as in the preceding part
shows that T is a generalized Cohen-Macaulay domain.
6
Since S is local so we do not need to localize S. This is important because the completion of a domain is
not necessarily a domain.
17
E HSAN TAVANFAR
4 A class of rings satisfying the Monomial Conjecture
As stated in the introduction for the validity of the Monomial Conjecture it suffices only to
check the almost complete intersections and system of parameters of the form p, x2 , . . . , xd
where p is the residual characteristic. This section is devoted to prove either the Small
Cohen-Macaulay Conjecture or the Big Cohen-Macaulay Algebra Conjecture for an almost
complete intersection R satisfying the extra assumption that,
m2 ⊆ (x)R,
(5)
for some system of parameters x1 , . . . , xd of R. In the mixed characteristic case we moreover
assume that x1 = p.
A motivation behind this consideration is that this case seems to be the most simple
non-trivial case of the Monomial Conjecture. For another motivation see the introduction.
But, parallel, to this aim, several other interesting facts about this class of almost complete
intersections are also given. The following lemma is not required for the next theorem which
is the main result of the section but it is needed for its subsequent remark.
Lemma 4.1. Assume that (A, n) is a regular local ring. Let a = (x1 , . . . , xd ) (resp. b = (y 1 , . . . , y s ))
be a complete intersection (resp. an almost complete intersection) ideal of A. Suppose furthermore that, x1 , . . . , xd , y 1 , . . . , y s−1 , forms a regular sequence of A and that ht(a) + ht(b) =
d + s − 1 = dim(A). Set, R := A/b. Then the following statements hold.
(i) There exists an exact sequence,
0 → H2 (x, R) → ωR /aωR → ωR/a → H1 (x, R) → 0.
(ii) We have Hi (x, R) ∼
= Hi −2 (x, ωR ) for each i ≥ 3. In particular,
(
Depth(ωR ) = Depth(R) + 2, Depth(R) ≤ d − 2
Depth(ωR ) = d ,
Depth(R) ≥ d − 1.
(iii) The system of parameters x of R satisfies the Monomial Conjecture if and only if,
¡
¢
¡
¢
ℓ H2 (x, R)
ℓ ωR /aωR .
Definitely, this is equivalent to say that, ωR → ωR/a , is non-zero in the above exact sequence.
Proof. (i) and (ii): We have,
´
³ ¡
¢O
H1 (x1 , . . . , xd ; R) ∼
(A/y s A) .
= Hi K• x1 , . . . , xd ; A/(y 1 , . . . , y s−1 )
A
¡
¢N
¡
¢
′
This encouraged us to consider the spectral sequence, K• x; A/(y′ )
A K• y s ; A , where y denotes
y 1 , . . . , y s−1 . So we take into account the bicomplex Mp,q :=
¡ the ′truncated
¢N
¡ sequence
¢
Kp x; A/(y )
A Kq y s ; A in which here, as usual, p stands for the column p. Note that,
( ¡
¢
¡
¢
¡
¢
Hi Tot(M) ∼
i ≥ 2.
= Hi x, y s ; A/(y′ ) ∼
= Hi y s ; A/(x, y′ ) = 0,
¡
¢
¡
¢
¡
¢
(6)
′ ∼
′ ∼
∼
Hi Tot(M) = Hi x, y s ; A/(y ) = Hi y s ; A/(x, y ) = ωR/a , i = 1.
18
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
Furthermore we have,
II 2
Ep,q
Hq (x; R),
p =0
µ ³¡
´¶
¢
= Hq x; (y′ ) : y s /(y′ ) = Hq (x; ωR ), p = 1
0,
p 6= 0, 1.
Now the desired exact sequence is just the five term exact sequence of this spectral sequence
(see, [50, Theorem 10.31 (Homology of Five-Term Exact Sequence)]).
For the second part note that according to the vanishings of (6) for i ≥ 2, all of the maps,
d 2 : Hi +2 (x; R) → Hi (x; ωR ), (i ≥ 1)
arising from the second page of the spectral sequence are isomorphisms.
¡
¢
(iii): By [13, 1.3. Proposition] our statement is equivalent to the assertion that ℓ H1 (x; R)
ℓ(R/a). Hence in the light of the exact sequence of the first part we are done once we can
show that,
³¡
´
¢
ℓ(R/a) = ℓ(ωR/a ) = ℓ (x, y′ ) : y s /(x, y′ ) .
But this is clear from the following exact sequence,
¡
¢
ys
0 → (x, y′ ) : y s /(x, y′ ) → R/(x, y′ ) → R/(x, y′ ) → R/a → 0.
In the following example we consider a subclass of almost complete intersections satisfying (5) and we show that they satisfy the Small Cohen-Macaulay Conjecture. The main idea
behind this observation is that this class of almost complete intersections are the same free
square root extensions R(a 1/2 ) of Remark 2.1. This is not the case for the more general setting
of Theorem 4.4. However, in Theorem 4.4, with aid of the multiplicity theory we will again
descend to a similar case of quadratic extensions.
Example 4.2. Suppose that R = (V, pV)[[X 2 , . . . , X d , Z 1 , Z 2 ]]/I, such that I = (g 1 −Z21 , g 2 −Z22 , g 3 −
Z 1 Z 2 ) wherein g 1 , g 2 , g 3 ∈ V[[X 2 , . . . , X d ]] are non-zero. Set, (A, n) := V[[X 2 , . . . , X d , Z 1 , Z 2 ]]. Presume that R is an almost complete intersection, i.e. ht(I) = 2. Note that,
p, x2 := X 2 + I, . . . , xd := X d + I,
forms a system of parameters for R satisfying (5). In this example we show that R has a maximal Cohen-Macaulay module. Setting,
S := V[[X 2 , . . . , X d , Z 1 ]]/(g 1 − Z21 ),
then S is a free extension of B := V[[X 2 , . . . , X d ]] wherein Z 1 is the square root of g 1 , i.e. S =
B(g 11/2 ). If g 1 has a square root in B then the polynomials g 1 −Z21 is reducible and therefore any
p ∈ assht(I) contains an element a ∈ n\n2 . In particular, then, p/aA is a height one prime ideal
of the regular local ring A/aA which implies that p is generated by two elements and thence
19
E HSAN TAVANFAR
A/p is a complete intersection and a maximal Cohen-Macaulay R-module. So without loss
of generality we can assume that g 1 does not have a square root in B and whence in Frac(B).
Consequently, S is a domain hypersurface by Remark 2.1(v). Next set,
T := V[[X 2 , . . . , X d , Z 1 , Z 2 ]]/(g 1 − Z21 , g 2 − Z22 ).
Therefore, T is the free extension of S wherein Z 2 is a square root of g 2 , i.e. T = S(g 21/2 ). In
particular, T is a domain if and only if g 2 does not have a square root Frac(S). But since our
almost complete intersection I is an ideal of height two generated by three elements so indeed
g 2 has a square root in Frac(S). Therefore by Remark 2.1(iv) there exists an induced map from T
to the integral closure S of S, denoted by Ψ. Since the kernel of this map is an associated prime
of the defining ideal of T so, as will be seen later, without loss of generality we can presume
that the kernel of Ψ contains the ideal, I, and thence we get an induced map R → S. Hence it is
enough to prove the statement for S (recall that S is finitely generated as S-module). Note that,
[Frac(S) : Frac(B)] = [Frac(S) : Frac(B)] = 2.
Here the residual characteristic of B is either two or p > 2. In the first case [41, Theorem 3.8.]
implies that R has a maximal Cohen-Macaulay module, while in the second case S is a CohenMacaulay ring in the light of [45]7 .
It remains to show that why we can assume that, I ⊆ ker Ψ. We claim that T has at most
two associated primes. In order to see why this is the case note that, in the light of [18, TheN
orem 1.], E S T is an injective T-module provided E is an injective S-module. In particular,
N
E(S) S T, is an injective T-module which contains T and thence contains ET (T). Therefore,
L
ET (T), as S-module, is an injective submodule of E(S) E(S). Since S is a domain so E(S) is an
indecomposable injective module. This observation shows that ET (T) has at most two direct
summands and, as an immediate consequence, T has at most two associated primes. Now
if α1 and α2 are distinct square roots of g 2 in Frac(S)8 then it is easily verified that the rules
(r, s) 7→ r + sαi , (1 ≤ i ≤ 2) define distinct ring homomorphisms, Ψi : T → S, with distinct
kernels, p1 , p2 ∈ Ass(T). Now the claim, easily, follows from the fact that T has at most two
associated primes.
Example 4.3. It is worth to give a concrete and explicit non-Cohen-Macaulay example of the
class of the rings of the previous example. We do this, but in the equal characteristic case,
where we can use the Macaulay2 system for computations. Set,
R = K[[Y1 , . . . , Y6 , Z 1 , Z 2 ]]/I,
such that,
I = (Y26 Y35 + Z22 , Y33 Y48 + Z21 , Y23 Y34 Y44 + Z 1 Z 2 ).
Then we have, Depth(R) = 4 and dim(R) = 6. In particular, Lemma 4.1 implies that ωR is, also,
a maximal Cohen-Macaulay R-module.
7
By an argument as in the proof of Theorem 4.4 we can deduce that S is necessarily Cohen-Macaulay regardless of the residual characteristic, because we have a quadratic extension. This fact is asserted in Theorem
4.5.
8
Recall that S has mixed characteristic and therefore Frac(S) has characteristic zero. So we have two distinct
roots.
20
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
Theorem 4.4. Suppose that R is an almost complete intersection satisfying (5). Then R has a
b has a maximal Cohen-Macaulay module which is
big Cohen-Macaulay algebra. A fortiori, R
b
an R-algebra.
Proof. If R is a Cohen-Macaulay then there is nothing to prove, so assume that R is not
Cohen-Macaulay (and is complete). The multiplicity theory plays a pivotal role in the proof.
The same idea of the proof of Corollary 2.3 can be applied to impose the extra assumption
that x is a part of minimal
generating
set¡ of m9¢. Consider a minimal Cohen-presentation
¡
¢
R = A/I of R so embdim (A, n) = embdim (R, m) and I ⊆ n2 . Firstly, we have,
³ ¡
¡
¢
¢´
ℓ R/(x) = ℓ A/ (x) + I) = ℓ(A/I),
wherein the notation, , means modulo (x). But, I = n2 , as n2 ⊆ I + (x) and I ⊆ n2 simultaneously. It turns out that,
¡
¢
ℓ R/(x) = ℓ(A/n2 ) = embdim(A) + 1 = embdim(R) − dim(R) + 1.
We claim that, embdim(R) − dim(R) ≤ 2, but before proving this claim let us conclude
the
¡
¢
statement from it. The above inequality in conjunction with our claim implies that ℓ R/(x) ≤
3. Now, in the light of [51, Theorem 1, page. 57] together with [51, Corollary , page 90] we
can deduce a very strong implication that e(x, R) ≤ 3. But if e(x, R) = 3 then R is CohenMacaulay by [25, (1.5) Proposition]. Thus we have even stronger condition that e(x, R) ≤ 2.
In the case where e(x, R) = 1 using [25, (1.10.1)] and [25, (6.8) Theorem] there exists a prime
ideal p ∈ Assht(R) such that R/p is regular, hence is s a maximal Cohen-Macaulay R-module.
It remains to deal with the case where e(x, R) = 2. In this case by a similar argument as above
without loss of generality we can assume that there exists p ∈ Assht(R) for which R/p satisfies
e(x, R/p) = 2. There exists a regular local subring B of R/p, containing a coefficient field (in
the mixed characteristic case, containing a complete discrete valuation ring whose maximal
ideal is generated by the element p 1/2 of R/p and with the same residue field of R), wherein
x forms a minimal basis for its maximal ideal and R/p is module finite over B. Therefore,
[Frac(R/p) : Frac(B)] = 2, in the light of [25, (6.5) Corollary]. Let S be the integral closure
of B in Frac(R/p). Since B is a complete local ring so by [39, Theorem 4.3.4] S is finitely
generated over B. Moreover, by virtue of, [36] the inclusion B → S splits10. Consequently, we
L
may assume that S = B I for some finitely generated B-module I. Since we have quadratic
extension of fraction fields so we can deduce that I has rank 1. In particular, we may presume
that, I is an ideal of B (since S is a domain so I is torsion-free). Using the S 2 -property of R for
each prime ideal p of B we conclude that Ip ∼
= Bp provided Depth(Bp ) = dim(Bp ) ≤ 1 and
Depth(Ip ) ≥ 2 provided Depth(Bp ) = dim(Bp ) ≥ 2. Consequently, I is a reflexive ideal of B in
the light of [10, Proposition 1.4.1]. Now the result follows from the fact that reflexive ideals of
unique factorization domains are principal.
9
Here it is important to point out that the extensions used in the proof of Corollary 2.3 preserves the condition (5) as well as the almost complete intersection property.
10
Hence, the Monomial Conjecture for the system or parameters x of R follows from [36].
21
E HSAN TAVANFAR
Now we are going to prove the above claim, i.e., embdim(R)−dim(R) ≤ 2. Let us use a presentation p 1/2 , X 2 , . . . , X d for the sequence x in R where R = (V, p 1/2 )[[X 2 , . . . , X d , Z 1 , . . . , Z u ]]/I11
is a homomorphic image of the regular local ring (A, n) = (V, p 1/2 )[[X 2 , . . . , X d , Z 1 , . . . , Z u ]]. So,
n2 ⊆ (p 1/2 , Y2 , . . . , Yd ) + I. Set, I = ( f 1 , . . . , f l ) where l = µ(I). We denote by f iX the sum of those
monomials of f i whose power of X i is non-zero for some 2 ≤ i ≤ d . Subsequently, we set,
p 1/2
fi
, to be the sum of those monomials of f i − f iX whose coefficients are multiple of p 1/2 . It
p 1/2
follows that, f iZ := f i − f iX − f i
∈ V[[Z 1 , . . . , Z u ]], and that the coefficients of the monomials
Z
of f i are all invertible. Now, set ≥3fiZ (resp., ≤2fiZ ) to be the sum of the monomials of f iZ of
total degree greater than or equal to 3 (resp., less than or equal to 2). Since, I ⊆ n2 , so it turns
out that, ≤2fiZ , is an V-linear combination of the elements of the form {Z i Z i : 1 ≤ i , j ≤ u}
with invertible coefficients in V. Otherwise, ≤2fiZ and thence f i , would have a summand of
the form kZαj where, α ∈ {0, 1}, 1 ≤ j ≤ u and k ∈ V\p 1/2 V. But this contradicts with f i ∈ n2 . In
particular, ≤2fiZ ∈ (Z 1 , . . . , Z u )2 .
On the other hand the fact that, Z i Z j ∈ (p 1/2 , X 2 , . . . , X d ) + ( f 1 , . . . , f l ) yields,
Z k Z s = g 1 p 1/2 +
d
X
i =2
g i Xi +
l
X
i =1
hi
≥3 Z
fi +
l
X
i =1
hi
≤2 Z
fi ,
for each 1 ≤ k, s ≤ u and power series h i , g i . Thus an elementary, quite easy, computation
shows that (Z 1 , . . . , Z u )2 ⊆ ( ≤2f1Z , . . . , ≤2flZ ). This in conjunction with the concluding assertion
of the preceding paragraph yields (Z 1 , . . . , Z u )2 = ( ≤2f1Z , . . . , ≤2flZ ).
Since R is an almost complete intersection, so we have, l = µ(I) = ht(I) + 1 = dim(A) −
dim(R) + 1 = u + 1. Consequently, we get,
³X
´
l
¡
¢
u +1 = l ≥ µ
( ≤2fiZ )V[[X 2 , . . . , X d , Z 1 , . . . , Z u ]] = µ (Z 1 , . . . , Z u )2 = u(u + 1)/2,
i =1
i.e. u ≤ 2.
The proof of the above theorem implies also the following result which is worth stating it
separately as a theorem. Recall that by [45] the integral closure of a regular local ring A in a
finite Galois extension L of K := Frac(A) with Abelian Galois group GalL/K is Cohen-Macaulay
provided the order of GalL/K is not divisible by the characteristic of the residue field of A. In
view of the following theorem for the quadratic extension case there is no restriction on the
characteristic of the residue field, provided A is a complete regular local ring.
Theorem 4.5. Suppose that A is a complete regular local ring and L is a quadratic field extension of Frac(A). Then the integral closure of A in L is Cohen-Macaulay.
11
We choose a presentation of A in the mixed characteristic, because the Monomial Conjecture is open in the
mixed characteristic. But the proof is characteristic free and someone can use presentation of A in equicharacteristic zero. Furthermore, we write p 1/2 instead of p, because we applied our square root technique, developed
in the proof of Corollary 2.3, to have the extra assumption that x1 , . . . , xd is a part of a minimal generating set of
m.
22
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
Remark 4.6. In this remark we will state some nice properties as well as several questions
about the class of rings of Theorem 4.4. So we are in the case of the Theorem 4.4 with an
extra assumption that the system of parameters x of (5) is a part of a minimal basis of m
whence embdim(R) − dim(R) ≤ 2.
(i) All of the examples of such almost complete intersections that we have computed in
Macaulay2 system had Depth ≥ d − 2. So we guess Depth(R) ≥ d − 2 whence ωR is a
maximal Cohen-Macaulay module by Lemma 4.1 (ii). This observation leads us to the
following question.
Question 4.7. Suppose that U is an almost complete intersection. Then do we have,
Depth(U) ≥ dim(U) − e(U)?
(7)
Question 4.7 is certainly true in dimension ≤ 1. But even in dimension two we need to
use the validity of the Monomial Conjecture for the confirmation of the inequality (7).
Namely, in dimension two we only need to show that if Depth(U) = 0 then e(U) ≥ 2.
But if Depth(U) = 0 then we have H2 (y, U) 6= 0, wherein y is a system of parameters for
U. Hence,
¡
¢ ¡
¢ ¡
¢
e(U) = ℓ U/(y) − ℓ H1 (y, U) + ℓ H2 (y, U) ≥ 2,
¡
¢ ¡
¢
as the validity of the Monomial Conjecture implies that, ℓ U/(y) − ℓ H1 (y, U)
0 by
[13, 1.3. Proposition].
For another case where inequality (7) holds, consider an almost complete intersection
ideal a of an equicharacteristic complete regular local ring A with infinite residue field
such that Assh(A/a) = {p/a}. Set, U := A/a. Suppose, furthermore, that dim(U) = 3
and S := A/p is an integrally closed domain12. If the inequality (7) fails for U then we
must have Depth(U) = 0 and e(U) ≤ 2. It is easily seen that ωU = ωS . On the other
hand, Lemma 4.1(ii) implies that Depth(ωU ) = 2 and thence ωU = ωS is not a maximal
Cohen-Macaulay module. In particular, S is not a Cohen-Macaulay ring. By [25, (4.15)
Remark.] there exists a system of parameters x of U such that e(x, U) = 2. Since we
are in the equicharacteristic case so there exists a complete regular local subring B of
S with the same residue field as S such that x forms a minimal basis of the maximal
ideal of B. Consequently, in view of [25, (1.10.1.)], e(x, S) ≤ 2. If e(x, S) = 1 then [25, (6.8)
Theorem] implies that S is regular, a contradiction. On the other hand if e(x, S) = 2,
then [25, (6.5) Corollary] implies that [Frac(S) : Frac(B)] = 2. Now S is Cohen-Macaulay
by Theorem 4.5, again, a contradiction.
Note that the inequality (7) does not hold if we relax the almost complete intersection
condition. More precisely, taking into account the Abhyankar’s local domains [1, (3)],
we can construct local domains with multiplicity two and depth 1 of arbitrary dimension.
12
In each prime ideal q of A we can find an almost complete intersection ideal a ⊆ q such that Assh(A/a) =
{q/a} (see, the proof of [13, 1.2 Proposition]). So, indeed, there exists concrete examples of such almost complete intersections.
23
E HSAN TAVANFAR
(ii) In view of [13, 1.3. Proposition] the Monomial Conjecture for the system of parameters
x := p 1/2 , X 2 , . . . , X d of R is equivalent to the assertion that,
¡
¢
¡
¢
¡
¢
ℓ H1 (x, R)
ℓ R/(x) = ℓ K[[Z 1 , Z 2 ]]/(Z21 , Z22 , Z 1 Z 2 ) = 3.
On the other hand in the light of Lemma 4.1(i) the Monomial Conjecture holds for x if
and only if H1 (x, R) is a quotient of (Z 1 , Z 2 )/(Z21 , Z22 ) by a non-zero submodule. It is easily seen that this is equivalent to the assertion that mH1 (x, R) = 0. Accordingly, in the
light
4.4.], there exists a complex, denoted by, 0 Z •+ (x, Z 1 ), such that
¡ of+[22, Theorem
¢
H0 0 Z • (x, Z 1 ) = R/m and H1 (0 Z •+ (x, Z 1 )) = 0 if and only if the Monomial Conjecture
holds for R. Consequently, using Theorem
¡ +4.4, we¢ have H1 (x, R) is a non-zero at most
two dimensional vector space and H1 0 Z • (x, Z 1 ) = 0. Here it is noteworthy to mention that, we strongly, guess more, i.e. Z •+ (x, Z 1 ) is an acyclic finite complex consisting
of Koszul cycles of x1 , . . . , xd , Z 1 which resolves the residue field of R (This is equivalent
to say that mHi (x, R) = 0 for each i ≥ 1).
(iii) We claim that the parameter ideal a = (p 1/2 , x2 , . . . , xd ) of R is a Lech ideal in the sense
that a/a2 is a free R/a-module. Before proving, we wish to clear why this claim sounds
interesting to us. This fact in conjunction with the exact sequence in the paragraph
prior to [11, Proposition 2.6] shows that H1 (x, R) ∼
= δ1 (a) wherein,
\
¡
¢
¡
¢
δ1 (a) = Z1 (x, R) (aRd ) /B1 (x, R) ∼
= ker sym2 (a) → a2 ,
by [26, Corollary 2.6.]. So [2, XV., Proposition 12] implies that H1 (x, R) is the second
André-Quillen homology D2(R, R/a, R/a). Hence, in view of the preceding part, it would
be highly desirable to give a proof of the validity of the Monomial Conjecture for the
sequence x of R with aid of the André-Quillen homology avoiding, [36]. The fact that a
is Lech ideal follows from the following two steps.
Step 1: Bearing in mind the notations of the proof of Theorem 4.4, this step is devoted
to prove that, without loss of generality, we can assume that f 1Z = Z21 , f 2Z = Z22 and f 3Z =
Z 1 Z 2 . Recall that R is a homomorphic image of (A, n) = (V, p 1/2 V)[[X 2 , . . . , X d , Z 1 , Z 2 ]], by
the ideal ( f 1 , f 2 , f 3 ). As we have seen in the proof of Theorem 4.4, ( f 1Z , . . . , f 3Z ) ⊆ (Z 1 , Z 2 )2 .
d
3
P
P
On the other hand, by our hypothesis, Z i Z j =
X k g k + p 1/2 h +
f kZ l k , for some,
k=2
k=1
g 2 , . . . , g d , h, l 1 , l 2 , l 3 ∈ A.
We use a similar method as in the case of f i and write h = h X +h Z,p
for each 1 ≤ s ≤ 3, so that h Z,p
Z i Z j − p 1/2 h Z,p
1/2
1/2
−
Z,p 1/2
, l1
3
X
k=1
which implies that,
Z i Z j = p 1/2 h Z,p
Z,p 1/2
, . . . , l3
Z,p 1/2
lk
1/2
+
3
X
k=1
f kZ =
Z,p 1/2
and l s = l sX +l s
∈ V[[Z 1 , Z 2 ]]. Therefore,
d
X
k=2
Z,p 1/2
lk
1/2
X k g k + p 1/2 h X +
3
X
k=1
f kZ ∈ ( f 1Z , f 2Z , f 3Z , p 1/2 ).
f kZ l kX ,
24
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
Set, B := A/p 1/2 = k[[X 2 , . . . , X d , Z 1 , Z u ]] where k is the residue field of V. By the above
argument we have ( f 1Z , f 2Z , f 3Z ) = (Z 1 , Z 2 )2 in B. Thus there exists a square matrix H =
[h st ] such that H[ f 1Z , f 2Z , f 3Z ]τ = [Z21 , Z22 , Z 1 Z 2 ]τ , in B. Let us to denote the maximal ideal
of B by nB . In particular, H := [h st ], is the change of basis matrix with respect to the two
basis { f 1Z , f 2Z , f 3Z } and {Z21 , Z22 , Z 1 Z 2 } of the K-vector space (Z 1 , Z 2 )2 /nB (Z 1 , Z 2 )2 . Whence,
det(H) 6= 0, in K. Let Y = [y i j ] be a square matrix which lifts the matrix H. Then det(Y) ∉
n. So Y is invertible. Hence, if we put, Y[ f 1 , f 2 , f 3 ]τ = [ f 1′ , f 2′ , f 3′ ]τ , then I = ( f 1′ , f 2′ , f 3′ ).
Moreover for each 1 ≤ i ≤ 3 we have the identity,
f i′ =
3
X
j =1
y i j f jX +
3
X
j =1
p 1/2
yi j f j
+
3
X
j =1
y i j f jZ .
(8)
On the other hand, since H[ f 1Z , f 2Z , f 3Z ]τ = [Z21 , Z22 , Z 1 Z 2 ] in B, so we get
3
X
j =1
y 1j f jZ ≡ Z21 ,
3
X
j =1
y 2j f jZ ≡ Z22 and
3
X
j =1
y 3j f jZ ≡ Z 1 Z 2 , mod p 1/2 A.
(9)
So our claim follows from (8) and (9).
Step 2: In this step we assume that
d
P
i =1
r i xi = 0 ∈ R and we prove that,
r i ∈ a = (p 1/2 , x2 , . . . , xd ),
for each 1 ≤ i ≤ d (x1 = p 1/2 )13 . If so, then there are s 1 , . . . , s d ∈ (A, n) such that,
d
X
(r i + s i )X i = Z21 h 1 + Z22 h 2 + Z 1 Z 2 h 3 ,
(10)
i =1
because f 1Z = Z21 , f 2Z = Z22 , and f 3Z = Z 1 Z 2 . Clearly, we must have, h i ∈ n, for each 1 ≤
p 1/2
i ≤ 3. Hence, s i ∈ n2 , for every 1 ≤ i ≤ d , because f jX + f j
other hand, (10) shows that,
d
P
i =1
∈ n2 (1 ≤ j ≤ 3). On the
(r i + s i )X i + (Z21 , Z22 , Z 1 Z 2 ) = 0 in the Cohen-Macaulay
ring A/(Z21 , Z22 , Z 1 Z 2 ) whence r i + s i ∈ (X 1 , . . . , X d , Z21 , Z22 , Z 1 Z 2 ), for each 1 ≤ i ≤ d . Consequently,
r i ∈ (X 1 , . . . , X d , Z21 , Z22 , Z 1 Z 2 ) = (X 1 , . . . , X d , f 1 , f 2 , f 3 ),
for each 1 ≤ i ≤ d , because s i ∈ n2 .
Example 4.8. We give an example to show that, in general, the inclusion m2 ⊆ (x) in conjunction with the non-Cohen-Macaulayness does not imply that e(x, R) = 2 (without assuming that
R is an almost complete intersection). Set,
R = Q[X 1 , . . . , X 4 , Z 1 , Z 2 , Z 3 ],
13
This, immediately, shows that a is a Lech ideal.
25
E HSAN TAVANFAR
wherein X 1 , X 2 , X 3 have degree 1 and X 4 , Z 1 , Z 2 , Z 3 have degree 2. Let,
I := (Z21 + X 23 Z 2 + X 4 Z 2 , Z22 − X 1 X 2 Z 3 + X 4 Z 3 , Z23 , Z 1 Z 2 , Z 1 Z 3 , Z 2 Z 3 ).
Then, in S = R/I the image of the sequence x := X 1 , X 2 , X 3 , X 4 forms a system of parameters
satisfying (5) while e(x, R) = 3. Note that Depth(R) = 3 and R is not Cohen-Macaulay.
5 A question
We end the paper with the following question.
Question 5.1. In prime characteristic, Hdm (R) = lim R/(xn ), is a cyclic module over the Frobe−→
n∈N
nius Skew Polynomial ring R[x; f ], which is generated by [1/(x)]. This shows, in particular,
that, [1/(x)] 6= 0, by the Grothendieck’s non-vanishing theorem, i.e. the Monomial Conjecture
holds. In characteristic zero also the Frobenius-like endomorphism exists in a faithfully flat
extension with aid of the Schoutens’s Lefschetz extensions (see, [5]). Hence probably the same
strategy works in the equal characteristic zero. Here it seems to be natural to ask whether in
the mixed characteristic there exists an R-algebra S such that Hdm (R) is a cyclic S-module generated by [1/(x)]? It is worth to point out that if this question has an affirmative answer then
S is necessarily a non-commutative ring. To be more precise suppose that there exists such
an R-algebra S, but to the contrary S is commutative. Then since Hdm (R) is a cyclic S-module
by our assumption so Hdm (R) carries a commutative R-algebra structure. In particular there
N d
Hdm (R) induced by the multiplication. But on the
exists a surjective map Hdm (R) R H
m (R) →
¢
¡
N
other hand Hdm (R) R Hdm (R) ∼
= Hdm Hdm (R) = 0 which contradicts with the Grothendieck’s nonvanishing theorem (here, we assume that d > 0).
Acknowledgement
We would like to express our deepest gratitude to Massoud Tousi, Anurag K. Singh, Linquan
Ma, S. H. Hassanzadeh as well as Kazuma Shimomoto for their helpful comments and valuable suggestions. We also would like to thank the University of Utah for its hospitality during
our visit in the academic year 2015-2016 as well as for its support for this research.
References
[1] S. S. Abhyankar, Local rings of high embedding dimension, Amer. J. Math. 89 (1967), 1073–1077.
[2] M. André, Homologie des algèbres commutatives, Springer-Verlag, Berlin-New York, 1974. Die Grundlehren
der mathematischen Wissenschaften, Band 206.
[3] Y. Aoyama, Some basic results on canonical modules, J. Math. Kyoto Univ. 23 (1983), no. 1, 85–94.
[4] Y. Aoyama and S. Goto, On the endomorphism ring of the canonical module, J. Math. Kyoto Univ. 25 (1985),
no. 1, 21–30.
26
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
[5] M. Aschenbrenner and H. Schoutens, Lefschetz extensions, tight closure and big Cohen-Macaulay algebras,
Israel J. Math. 161 (2007), 221–310.
[6] A. Bagheri, M. Salimi, E. Tavasoli and S. Yassemi, A construction of quasi-Gorenstein rings, J. Algebra Appl.,
11, (2012), no. 1, 9 pages.
[7] M. -J. Bertin, Anneaux d’invariants d’anneaux de polynomes, en caractéristique p, C. R. Acad. Sci. Paris Sér.
A-B 264 (1967), A653–A656.
[8] R. Bhattacharyya, Existence of almost Cohen-Macaulay algebras implies the existence of big CohenMacaulay algebras., J. Algebra 457 (2016), 1–6.
[9] M. P. Brodmann and R. Y. Sharp, Local cohomology, Second, Cambridge Studies in Advanced Mathematics,
vol. 136, Cambridge University Press, Cambridge, 2013. An algebraic introduction with geometric applications.
[10] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39,
Cambridge University Press, Cambridge, 1993.
[11] A. Corso, C. Huneke, D. Katz and W. V. Vasconcelos, Integral closure of ideals and annihilators of homology.,
Boca Raton, FL: Chapman & Hall/CRC, 2006.
[12] S.P. Dutta, On the canonical element conjecture, Trans. Amer. Math. Soc. 299 (1987), no. 2, 803–811.
[13]
, The monomial conjecture and order ideals, J. Algebra 383 (2013), 232–241.
[14]
, The monomial conjecture and order ideals II, J. Algebra 454 (2016), 123–138.
[15] S.P. Dutta and P. Griffith, Intersection multiplicity, canonical element conjecture and the syzygy problem.,
Mich. Math. J. 57 (2008), 227–247.
[16] E. G. Evans and P. Griffith, The syzygy problem, Ann. of Math. (2) 114 (1981), no. 2, 323–333.
[17] R. Fossum, H. -B. Foxby, P. Griffith and I. Reiten, Minimal injective resolutions with applications to dualizing modules and Gorenstein modules., Publ. Math., Inst. Hautes Étud. Sci. 45 (1975), 193–215.
[18] H. -B. Foxby, Injective modules under flat base change., Proc. Am. Math. Soc. 50 (1975), 23–27.
[19] S. Goto, A note on quasi-Buchsbaum rings, Proc. Amer. Math. Soc. 90 (1984), no. 4, 511–516.
[20]
, On the associated graded rings of parameter ideals in Buchsbaum rings, J. Algebra 85 (1983), no. 2,
490–534.
[21] R. Hartshorne and A. Ogus, On the factoriality of local rings of small embedding codimension, Comm.
Algebra 1 (1974), 415–437.
[22] S. H. Hassanzade and J. Naéliton, Residual Intersections and the Annihilator of Koszul Homologies, Algebra
Number Theory 10 (2016), no. 4, 737–770.
[23] R. C. Heitmann, Characterization of completions of unique factorization domains, Trans. Amer. Math. Soc.
337 (1993), no. 1, 379–387.
[24]
, The direct summand conjecture in dimension three, Ann. of Math. (2) 156 (2002), no. 2, 695–712.
[25] M. Herrmann, S. Ikeda and U. Orbanz, Equimultiplicity and blowing up, Springer-Verlag, Berlin, 1988. An
algebraic study, With an appendix by B. Moonen.
[26] J. Herzog, A. Simis and W. V. Vasconcelos, Koszul homology and blowing-up rings, Commutative algebra
(Trento, 1981), pp. 79–169.
[27] M. Hochster, Big Cohen-Macaulay algebras in dimension three via Heitmann’s theorem, J. Algebra 254
(2002), no. 2, 395–408.
[28]
, Canonical elements in local cohomology modules and the direct summand conjecture, J. Algebra 84
(1983), no. 2, 503–553.
E HSAN TAVANFAR
[29]
, Contracted ideals from integral extensions of regular rings, Nagoya Math. J. 51 (1973), 25–43.
[30]
, Criteria for equality of ordinary and symbolic powers of primes, Math. Z. 133 (1973), 53–65.
[31]
27
, Properties of Noetherian rings stable under general grade reduction., Arch. Math. 24 (1973), 393–
396.
[32]
, The equicharacteristic case of some homological conjectures on local rings, Bull. Amer. Math. Soc.
80 (1974), 683–686.
[33] M. Hochster and C. Huneke, Applications of the existence of big Cohen-Macaulay algebras, Adv. Math. 113
(1995), no. 1, 45–117.
[34]
, Indecomposable canonical modules and connectedness, Contemp. Math., vol. 159, Amer. Math.
Soc., Providence, RI, 1994.
[35]
, Infinite integral extensions and big Cohen-Macaulay algebras, Ann. of Math. (2) 135 (1992), no. 1,
53–89.
[36] M. Hochster and J. E. McLaughlin, Splitting theorems for quadratic ring extensions, Illinois J. Math. 27
(1983), no. 1, 94–103.
[37] C. Huneke, Almost complete intersections and factorial rings, J. Algebra 71 (1981), no. 1, 179–188.
[38]
, On the symmetric and Rees algebra of an ideal generated by a d-sequence., J. Algebra 62 (1980),
268–275.
[39] C. Huneke and I. Swanson, Integral closure of ideals, rings, and modules. (2006), xiv + 431.
[40] C. Huneke and B. Ulrich, Divisor class groups and deformations, Amer. J. Math. 107 (1985), no. 6, 1265–1303
(1986).
[41] D. Katz, On the existence of maximal Cohen-Macaulay modules over pth root extensions, Proc. Amer. Math.
Soc. 127 (1999), no. 9, 2601–2609.
[42] J. Lipman, Unique factorization in complete local rings, Algebraic geometry (Proc. Sympos. Pure Math.,
Vol. 29, Humboldt State Univ., Arcata, Calif., 1974), 1975, pp. 531–546.
[43] T. Ochiai and K. Shimomoto, Bertini theorem for normality on local rings in mixed characteristic (applications to characteristic ideals)., Nagoya Math. J. 218 (2015), 125–173.
[44] C. Peskine and L. Szpiro, Dimension projective finie et cohomologie locale. Applications à la démonstration
de conjectures de M. Auslander, H. Bass et A. Grothendieck, Inst. Hautes Études Sci. Publ. Math. 42 (1973),
47–119.
[45] P. C. Roberts, Abelian extensions of regular local rings, Proc. Amer. Math. Soc. 78 (1980), no. 3, 307–310.
[46]
, Fontaine rings and local cohomology, J. Algebra 323 (2010), no. 8, 2257–2269.
[47]
, Le théorème d’intersection, C. R. Acad. Sci. Paris Sér. I Math. 304 (1987), no. 7, 177–180.
[48]
, Local cohomology of Segre product type rings, J. Pure Appl. Algebra 219 (2015), no. 3, 652–665.
[49] P. C. Roberts, A. K. Singh and V. Srinivas, Annihilators of local cohomology in characteristic zero, Illinois J.
Math. 51 (2007), no. 1, 237–254 (electronic).
[50] J. J. Rotman, An introduction to homological algebra, Second, Universitext, Springer, New York, 2009.
[51] J. -P. Serre, Local algebra, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2000. Translated
from the French by CheeWhye Chin and revised by the author.
[52] K. Shimomoto, An application of the almost purity theorem to the homological conjectures, J. Pure Appl.
Algebra 220 (2016), no. 2, 621–632.
[53] J. R. Strooker and J. Stückrad, Monomial conjecture and complete intersections., Manuscr. Math. 79 (1993),
no. 2, 153–159.
28
R EDUCTION
OF SOME HOMOLOGICAL CONJECTURES TO
...
[54] E. Tavanfar and M. Tousi, A Study of Quasi-Gorenstein Rings, arXiv:1508.04597 [math.AC].
[55] B. Ulrich, Gorenstein rings as specializations of unique factorization domains, J. Algebra 86 (1984), no. 1,
129–140.
E HSAN TAVNAFAR , D EPARTMENT OF M ATHEMATICS , S HAHID B EHESHTI U NIVERSITY, G.C., T EHRAN ,
I RAN .
E-mail address: [email protected]
| 0math.AC
|
Fair comparison of skin detection approaches on publicly available datasets
Alessandra Luminia and Loris Nannib
a.
b
DISI, Università di Bologna, Via Sacchi 3, 47521 Cesena, Italy.
DEI - University of Padova, Via Gradenigo, 6 - 35131- Padova, Italy.
Abstract
Skin detection is the process of discriminating skin and non-skin regions in a digital image and it is widely used in several applications
ranging from hand gesture analysis to tracking body parts and face detection. Skin detection is a challenging problem which has drawn
extensive attention from the research community, nevertheless a fair comparison among approaches is very difficult due to the lack of a
common benchmark and a unified testing protocol. In this work, we investigate the most recent research in this field and we propose a
fair comparison among approaches using several different datasets.
The major contributions of this work is a framework to evaluate and combine different skin detector approaches, whose source code
will be made freely available for future research, and an extensive experimental comparison among several recent methods which have
also been used to define an ensemble that works well in many different problems.
Experiments are carried out in 10 different datasets including more than 10000 labelled images: experimental results confirm that the
ensemble here proposed obtains a very good performance with respect to other stand-alone approaches, without requiring ad hoc
parameter tuning.
A MATLAB version of the framework for testing and ensemble proposed in this paper will be freely available from
(https://www.dei.unipd.it/node/2357 + Pattern Recognition and Ensemble Classifiers).
Keywords: skin classification, skin detection, skin segmentation, skin database.
1. Introduction
Skin texture and color are important signs that people use to understand variety of culture-related aspects about each other, as: health,
ethnicity, age, beauty, wealth and so on. The presence of skin color in images and videos is a signal of the presence of humans in such
media. Therefore in the last two decades extensive research has focused on skin detection in video and images. Skin detection is the
process of discriminating skin and non-skin regions in a digital image and consists in performing a binary classification of pixels between
the classes “skin”/ “not skin” and in executing a fine segmentation to define the boundaries of the skin regions.
Skin detection is used within many application domains: it is used as a preliminary step for face detection and tracking [1], body
tracking [2], hand gesture recognition, biometric authentication (i.e. palm print recognition) [3], objectionable content filtering [4],
medical imaging.
A useful feature for the discrimination of skin and non-skin pixels is the pixel color; nevertheless, obtaining skin color consistency
across variations in illumination, diverse ethnicity and different acquisition devices is a very challenging task. Moreover, skin detection
when used as preliminary step of other applications is required to be computationally efficient, invariant against geometrics
transformations, partial occlusions or changes of posture/facial expression, insensitive to complex or pseudo-skin background, robust
against the quality of the acquisition device. The factor that worst influence skin detection is color constancy problem [5]: i.e. the
dependency of pixel intensity on both reflection and illumination which have a nonlinear and unpredictable behavior. To be effective
when the illumination conditions vary rapidly some skin detection approaches use image preprocessing strategies based on color
constancy (i.e. a color correction method based on an estimate of the illuminant color) and/or dynamic adaptation techniques (i.e. the
transformation of a skin-color model according to the changing illumination conditions). Static skin color approaches that rely on image
preprocessing can only partially solve this problem and their performance strongly degrades in real-world application. A feasible solution
consists in considering additional data acquired out the visual spectrum (i.e. infrared images [6] or spectral imaging [7]), however the use
of such sensors is not appropriate for all applications and requires higher acquisition costs which limit their use to specific problems.
Skin detection is a challenging problem and has been extensive studied from the research community. Despite the large number of
methods, there are only few surveys in this topic: the works in [8][9] are quite old and cover only the methods proposed before 2005, the
survey in [10][11] are more recent and contain a good investigation of methods, benchmarking datasets and performance related to a
period of about two decades. The aim of the present work is not limited to survey the most recent research in this field (which is now
enriched of methods based on deep learning [12][13][14]), but also, and above all, to propose a framework for a fair comparison among
approaches.
In this research, a novel framework is proposed that integrates different skin color classification approaches and compare their
performance and their combination on several publicly available datasets. The major contributions of this research work is:
A framework to evaluate and combine different skin detector approaches is presented. The source code of the framework and many
of the tested methods will be made freely available for future research and comparisons. The system can be tuned according to the
target application: on the basis of the application requirement, the acceptance threshold can be tuned to prune a large percentage of
false accepts at a small cost of reduction in genuine accepts or vice versa a larger number of false accepts can be admitted to maximize
the number of genuine accepts. The framework include training and testing protocols for most used benchmark datasets in this field.
A fair comparison among the most recent research and methods in the skin detection field is carried out, using the same testing
protocols, benchmark datasets and performance indicators. A discussion about performance can help researchers and practitioners in
evaluating the approaches most suited to their requirements according to computational complexity, memory requirements, detection
rate and sensitivity.
The arrangement of this paper is as follows. In section 2 related works in skin detection are presented, including a discussion about
taxonomy of existing approaches and a detailed description of the approaches tested in this work. In section 3 the evaluation problem is
treated, the most known datasets used for performance evaluation are listed and commented, testing protocols and performance indicators
used in our experiment are discussed. In section 4 the experiments conducted using the proposed framework are reported and discussed.
Finally, section 5 includes the conclusions and some future research directions.
2. Skin detection approaches
Several skin detection methods are based on the assumption that skin color can be recognized from background colors according to
some clustering rule in a specific color space. Even if this assumption can be valid in a constrained environment where both ethnicity of
the people and background colors are known, it is a very challenging task in complex images captured under unconstrained conditions
and when individuals show a large spectrum of human skin coloration [9]. There are a lot of challenging factors that influence the
performance of a skin detector:
Human characteristics as ethnicity and age: skin color ranges from white to dark brown among human racial groups, the transition
from fresh skin to dry skin related to the age determines a strong variation of tones.
Acquisition conditions: factors such as camera characteristics or illumination variations strongly influence the skin appearance.
Generally, a variation in the illumination level or in the light source distribution determines the presence of shadows and changes in
the color of the skin.
Skin painting: the presence of makeup or tattoo coverage affects the appearance of skin.
Complex background: sometimes the skin detector can be deceived by the presence of objects that have skin-like color in the
background.
Skin detection approaches can be classified according to several taxonomy schemes which evaluates different aspects of the methods:
1. considering the presence or not of preprocessing steps such as color correction and illumination cancelation [15] or dynamic adaptation
[16] designed to reduce the effects of different acquisition conditions;
2. considering the color space used for pixel classification [9]. Most of research papers dealing with the skin detection have to face the
problem of selecting the most appropriate skin color model. The performance of several different color models are compared in [9]:
basic models (i.e. RGB, normalized RGB), perceptual models (i.e. HIS, HSV) perceptual uniform models (i.e. CIE-Lab, CIE-Luv) and
orthogonal models (i.e. YCbCr, YIQ) with the finding that orthogonal model are characterized by a reduced redundancy/correlation
among channels, therefore they are the most suited for skin detection.
3. examining the problem formulation, which can be based on the segmentation of an image in the regions where human skin is present
(segmentation based approaches) or on treating each pixel as skin or non-skin without considering its neighbors (pixel-based
approaches). Despite the presence of a huge number of pixel-based approaches [9], the region-based skin color detection techniques
are very few [17][18][19][20]. Some recent methods [13][14] based on convolutional neural network can be included in this category.
4. distinguish among methods for performing pixel classification (i.e. explicitly defined skin region, parametric approaches, nonparametric approaches) [21]. The first category, also named rule based, defines explicit rules to define the skin color (in an opportune
color space), The other categories includes machine learning approaches that makes use non-parametric or parametric approaches to
estimate the color distribution from a training set (returning a lookup table or a parametric model, respectively).
5. considering the type of classifier used for machine learning approaches: in [11] a taxonomy of 8 not exclusive groups is proposed
which extends the simple division in parametric and non-parametric approaches. Statistical methods includes parametric methods
based on Bayesian rule o Mixture models [22], neural networks models [23][24] are used for the segmentation of color images based
on both color and texture information, diffusion based methods [25][26] extends analysis on neighboring pixels to improve the
classification performance: after an initial pixel based extraction a seed growing method is applied to include neighborhood in the skin
region. Adaptive methods [27] are based on the tuning of models in order to adapt to specific conditions (i.e. lighting, skin tone,
background); the model calibration often grants performance advantages, but at the cost of increased computation time. Hyperspectral
models [28] are based on acquisition devices with hyperspectral capabilities: despite the advantage due to the presence of spectral
information, these approaches will be not considered in this work since they are only applicable to specifically collected datasets. SVM
based systems are parametric models based on SVM classifier: this class also overlaps adaptive methods if the SVM classifier is trained
using active learning [29]. The last class includes the mixture techniques which are the mixture of different methods [30][31].
In the last few years the research in skin detection has taken two main directions, according to the following consideration: even if in
the most general applications nothing should be assumed about the background and the acquisition conditions, in many applications the
difference between skin and background is large, the acquisition conditions are controlled and the skin region is quite easy to detect. For
example, in many gesture recognition applications the hand images are acquired using flatbed scanners and have dark unsaturated
backgrounds [32]. For this reason, in addition to many approaches that adopt sophisticated and computationally expensive techniques,
several simple rule-based methods have been proposed, which are preferred in some applications since they are easier to understand,
implement and reuse, more efficient, while at the same time, adequately effective. Usually simple rule-based methods are not even tested
on benchmarks for pure skin detection but as a step of a more complex task (i.e. face detection, hand gesture recognition). A recent
example of method belonging to this class is the work in [32], which performs a study on different color models, drawing the conclusion
that there are no apparent advantages of using a perceptual uniform color space: therefore they propose an approach based on a simple
RGB lookup table. The work in [33] proposes a dynamic but very straightforward method based on parametric modeling the Cr or Cb
channel from YCbCr color space through a single Gaussian: the final approach requires a limited amount of storage space, is fast and can
be trained using a small training set with small time delay. In [34] a simple skin detector based on RGB histogram thresholding is
employed as a preliminary fast step for motion-based skin region of interest detection. A dynamic skin color model method based on
space switching is proposed in [35], where in order to handle with natural changes in illumination, a system with three robust algorithms
has been built based on different color spaces which are switched according to the statistical mean of value of the skin pixels in the image.
In [36] a new 3D hybrid color space named SKN is proposed, obtained by using Principal Component Analysis and a Genetic Algorithm
to discover the optimal representation of skin color over 17 color spaces, then a pixel-wise skin classification is performed employing a
general purpose trained classifier (i.e. Naïve Bayes, Random Forest, Multilayer Perceptron and Support Vector Machines).
The first class of methods, based on sophisticated and computationally expensive techniques, include recently proposed approaches
based on deep learning [12][13][14]. Convolutional neural networks have recently achieved remarkable results for a variety of computer
vision tasks, including several applications based on pixel-wise prediction (i.e. scene labelling, semantic image segmentation). In [12] a
patch-wise skin segmentation approach is described based on deep neural networks, which uses image patches as processing units instead
of pixels; a dataset of image patches have been appositely collected for training purposes and the trained DSMs are integrated into a
sliding window framework to detect skin regions of the human body parts achieving competitive performance on pixel-wise skin
detection. In [14] an integration of some recurrent neural networks layers into the fully convolutional neural networks is proposed as a
solution to develop an end-to-end network for human skin detection, while in [13] the authors propose a inception-like convolutional
network-in-network structure, which consists of convolution and rectified linear unit layers only (without pooling and subsampling layers).
All the proposed architectures report comparable or better performances than other the state-of-the-art methods, anyway a fair comparison
with traditional existing approaches is always difficult due to the different testing protocols.
In this work, we evaluate and combine several skin detectors with the aim of comparing their performance and we propose an ensemble
able of maximizing their classification performance. The base approaches used to create the ensemble have been selected according to
their effectiveness, their availability to research scopes and their efficiency:
GMM [22] is a simple and efficient Gaussian mixture model for skin detection trained to classify non-skin vs. skin pixels in the RGB
space.
Bayes [22] is a fast and effective method based on a Bayesian classifier. In this work we used a classifier trained in the RGB color
space using 2000 images from the ECU data set.
SPL(τ) [37]1 is a pixel-based skin detection approach which determines skin probability in the RGB domain using a look up table
(LUT). For a testing image the probability of being skin for each pixel x is calculated, then a threshold τ is applied to decide skin/nonskin.
Cheddad(τ) [38] is a pixel-based and real-time approach which reduces the RGB color space to a 1D space derived from
differentiating the grayscale map and the non-red encoded grayscale version. The classification is performed using a skin probability
which delimits the lower and upper bounds of the skin cluster and the final decision depends on a classification threshold τ;
Chen [39] is a statistical skin color model, which is specifically suited to be implemented on hardware. The 3D skin cube is
represented as three 2D sides calculated as the difference of two color channels: sR=R-G, sG=G-B, sB=R-B. The skin cluster region
is delineated in the transformed space, fixing the boundaries to the following ranges: −142<sR<18, −48< sG <92, and −32< sB <192.
SA1(τ) [40] is a method for skin detection based on spatial analysis. Starting from a skin probability map obtained using a color pixelwise detector, the starting step to the spatial analysis is the proper selection of the high probability pixels as “skin seeds”. The second
step consists in finding the shortest routes from each seed to every single pixel, in order to propagate the “skinness”. All the pixels
not adjoined during the propagation process are labeled as non-skin. The performance depends on the threshold τ used to classify.
SA2(τ) [41] is another recent method based on spatial analysis which uses both color and textural features to determine the presence
of skin. The basic idea is to extract the textural features from the skin probability maps rather than from the luminance channel:
therefore simple textural statistics are computed from each pixel’s neighborhood in the probability map using kernels of different
sizes (i.e the median, the minimal values, the difference between the maximum and minimum and the standard deviation). Then skin
and non-skin pixels are transformed into two classes of feature vectors whose size is reduced by Linear Discriminant Analysis (LDA)
to increase their discriminating power. Finally the spatial analysis method proposed in [40] and described above is used for seed
extraction and propagation using the distance transform. The classification depends on a threshold τ used to classify pixels in the
distance domain.
SA3(τ) [27] is a self-adaptive method that consists in combining a local skin color model created using a probability map and spatial
analysis to fix the boundaries of the skin regions. It is an evolution of approach proposed in [41] based on spatial analysis, skin color
model adaptation and textural features. The main difference from previous approach is a new technique for extracting adaptive seeds,
based on the analysis of the skin probability map calculated from the input color image. The performance depends on a classification
threshold τ.
DYC [42]2 is a skin detection approach which works in the YCbCr color space and takes into account the lighting conditions. The
method is based on the dynamic generation of the skin cluster range both in the YCb and YCr subspaces of YCbCr and on the
definition of correlation rules between the skin color clusters.
A rough classification of the method tested in this work according to the criteria reported in section 2 is reported in Table 1
1
2
http://clickdamage.com/sourcecode/index.php
https://github.com/nadiabrancati/skin_detection/
Table 1. Rough classification of the tested approaches
GMM
Bayes
SPL
Cheddad
Chen
x
x
x
x
x
SA1
SA2
SA3
DYC
x
x
x
x
x
x
x
Preprocessing steps
None
Illumination cancellation
Dynamic adaptation
Color consistency
Color space
Basic color spaces
x
x
x
Perceptual color spaces
Orthogonal color spaces
x
Perceptually uniform color spaces
Other (e.g. Color ratio)
x
x
x
x
x
x
x
x
Problem formulation
Segmentation based
x
Pixel based
x
x
x
x
x
x
x
Type of pixel classification
Rule based
Machine learning: parametric
Machine learning: non-parametric
x
Type of classifier
Statistical
x
Mixture techniques
x
x
Adaptive methods
x
x
x
Other
3. Skin detection evaluation: Datasets and performance indicators
To assist research in the area of skin detection, there are some well-known color image datasets provided with ground truth. The use
of a standard and representative benchmark is essential to execute a fair empirical evaluation of skin detection techniques.
3.1. Datasets
In Table 2 some of the most used datasets are summarized and in this section a brief description of each of them is given.
Table 2. Some of the most used datasets per skin detection
Name
Compaq
Ref
[22]
Images
4675
TDSD
UChile
[43]
[44]
555
103
ECU
Schmugge
[45]
[46]
4000
845
Feeval
[47]
8991
MCG
Pratheepan
VDM
SFA
[48]
[49]
[50]
[51]
1000
78
285
1118
HGR
SDD
[27]
[52]
1558
21000
Ground truth
Semisupervised
Imprecise
Medium
Precision
Precise
Precise (3
classes)
Low quality,
imprecise
Imprecise
Precise
Precise
Medium
Precision
Precise
Precise
Download
ask to the authors
Year
2002
http://lbmedia.ece.ucsb.edu/research/skin/skin.htm
http://agami.die.uchile.cl/skindiff/
2004
2004
http://www.uow.edu.au/~phung/download.html (currently not available)
https://www.researchgate.net/publication/257620282_skin_image_Data_se
t_with_ground_truth
http://www.feeval.org/Data-sets/Skin_Colors.html
2005
2007
http://mcg.ict.ac.cn/result_data_02mcg_skin.html (ask to authors )
http://web.fsktm.um.edu.my/~cschan/downloads_skin_dataset.html
http://www-vpu.eps.uam.es/publications/SkinDetDM/
http://www1.sel.eesc.usp.br/sfa/
2011
2012
2013
2013
http://sun.aei.polsl.pl/~mkawulok/gestures/
Not available yet
2014
2015
2009
Compaq [22] is one of the first and most used large skin datasets; it consists of images collected by crawling Web: 9731 images
containing skin pixels (but only 4675 skin images have been segmented and included in the ground truth) and 8965 images with no skin
pixels. This dataset have been extensity used for testing and comparing methods, but without using a standard testing protocol, therefore
comparisons using this dataset are not always fair. Moreover, the ground truth for this dataset has been obtained using an automatic
software tool leading to imprecise results.
TDSD [43] is an old dataset containing 555 images with very imprecise annotations (automatic labeling).
The UChile [44] dbskin2 complete set includes 103 images acquired in different lighting conditions and with complex background.
The ground truth has been manually annotated with medium precision (in some images the boundaries between skin and non-skin pixels
are not marked precisely).
ECU [45] skin and face datasets are a collection of about 4000 color images annotated with relatively accurate ground truth. ECU
dataset is quite challenging, since it includes diversity in terms of the lighting conditions, background scenes, and skin types.
The skin dataset named Schmugge [46] is a collection of 845 images taken from different face datasets (i.e. the UOPB dataset, the AR
face dataset, and University of Chile database). The ground truth is very accurate since all images are labelled in 3 classes: skin/ notskin/don’t care.
Feeval [47] is a dataset based on 8991 frames extracted from 25 online videos of low quality. The ground truth is not precisely
annotated. Here, the performance indicator is calculated for each of the 25 videos then the results are averaged.
MCG skin database [29] contains 1000 images selected from internet in order to include confusing backgrounds, variable ambient
lights and diversity of human races. The ground-truth is obtained through manually labeling, but it is not very precise since eyes, eyebrows,
and even bracelets are sometimes marked as skin.
VMD [50] is obtained selecting 285 images from several public datasets for human activity recognition (i.e. LIRIS, EDds, UT, AMI
and SSG). The dataset is already divided in subsets for training and testing, anyway in this work we used all the images for testing. The
images cover a wide range of illumination levels and situations.
SFA dataset [51] includes images from FERET (876 images) and AR (242 images) face databases manually labelled (with medium
precision). SFA includes an internal organization in folders in order to separate 1118 original images (ORI), 1118 ground truths (GT)
masks, 3354 samples of skin (SKIN) and 5590 samples of nonskin (NS) ranging differently from 1 to 35×35 dimensions. We have used
ORI/GT for assessing the performance.
Pratheepan [49] is a small dataset which includes 78 images downloaded randomly from Google; the dataset is divided in two subsets:
FacePhoto including 32 single subject images with simple background and FamilyPhoto including 46 images with multiple subjects and
complex background.
HGR [27] is a dataset for gesture recognition which contains also ground truth binary skin presence masks; the dataset includes 1558
images representing gestures from Polish and American sign language with controlled and uncontrolled background divided in 3 subsets
(HGR1, HGR2A, HGR2B). In our experiments the size of the images of HGR2A and HGR2B has been reduced of a factor 0.3.
SDD [52] is a dataset of 21000 images acquired using different imaging devices, in a variety of illumination conditions and including
different skin colors from people all around the world. The dataset is composed from some images extracted from videos and others
belonging to popular face datasets. Images are provided in four sets: a set including mainly single face images, made for training purposes,
and other three sets to be used for testing.
FvNF [53] (Face vs. NonFace) is not a real skin dataset, it is composed by 800 face and 770 non-face images, extracted from Caltech
dataset [54]. This dataset has been collected and used in [53] to evaluate the capability of a skin detector method to detect the presence of
a face, based on the number of pixels classified as skin. The average precision (AP) is used as performance indicator, AP[0,100]. The
AP summarizes the shape of the precision/recall curve, since it is the area under the precision-recall curve.
3.2. Performance indicators
Skin detection is a two-class classification problem, therefore standard measures for general classification problems [55] can be used
for performance evaluation: including Accuracy, Precision, Recall, F-measure, Kappa, ROC-Curve, Area Under Curve, and others.
Anyway, due to the particular nature of this problem which is based on pixel-level classification and on unbalanced distribution, the
following measures are usually considered for performance evaluation: the confusion matrix, the F-measure, the True Positive Rate (TPR)
and the False Positive Rate (FPR).
The confusion matrix is obtained comparing results with the ground-truth data to determine the number of true negatives (tn), false
negatives (fn), true positives (tp) and false positives (fp). Several useful measures can be obtained from the confusion matrix, including
precision=tp/(tp+fp), that is the percentage of correctly classified pixels out of all the pixels classified as skin and recall =tp/(fn+tp), that
is the percentage of the ground-truth skin pixels correctly classified as skin.
The F1 measure is the harmonic mean of precision and recall and it is calculated according to the following formula F1 = 𝑡𝑝/(2𝑡𝑝 +
𝑓𝑛 + 𝑓𝑝) , F1[0,1]. According to other works in the literature F1 is averaged at pixel level not at image level; in such a way the final
indicator is not dependent on the image size in the different databases. Besides F1 several papers use the True Positive Rate (TPR=
recall=tp/(fn+tp)) and the False Positive Rate (FPR= fp/(fp+tn)).
Even if the F1 measure is able to evaluate algorithms only at a fixed operating threshold, and therefore it is a worse way to quantify
the accuracy of an algorithm evaluation with respect to ROC curve or precision-recall curve, we use F1 is used in this work since it is
widely used in the literature for skin classification and allow a better comparison. Moreover, several methods evaluated here work at fixed
threshold and do not allow tuning for true positives and false positives rates.
4. A fair experimental comparison
A fair comparison among different approaches is very difficult due to the lack of a universal standard in evaluation: most of published
works are tested on self-collected datasets which often are not available for further comparison; in many cases the testing protocol is not
clearly explained, many datasets are not of high quality and the precision of the ground truth is questionable since sometimes lips, mouth,
rings and bracelets have been labelled as skin. In this section, we carry out a comparison of some well-known approaches whose source
code is freely available. In order to accomplish a fair comparison we do not perform re-training of methods in each dataset, conversely,
we use the knowledge provided by the original authors, therefore all the datasets have been used only for testing purposes. When the
performance depends on a classification threshold, we tested several values in order to select the one that performs best in all the datasets.
In Table 3 we report the performance in 10 datasets of the nine stand-alone approaches described in section 2 and four ensembles
combined by the vote rule. Both the stand-alone approaches and the ensembles have been tested using different thresholds (only the best
one is reported in Table 3 for sake of space, the complete version of the table is included as supplementary material):
GMM [22], Bayes [22] (τ[50,70,90,110,140]), SPL [37] (τ[-2.5,-2,-1.5,-1,-0.5]), Cheddad [38], Chen [39], SA1 [40]
(τ[100,150,175,200,225]), SA2 [41] (τ[30,40,50,85,120]), SA3 [27] (τ[25,50,75,100,125]), DYC [42]
OldVote is the best ensemble based on vote rule proposed in [56], it is based on the weighted vote rule among SPL (weight 0.1),
GMM (weight 0.9) and SA2(weight 1). After weighting the responses from different classifiers, a pixel is classified as skin if it
receives more than one vote for the class Skin. In [56] we have tested different rules for the fusion of classifiers, finding that the vote
rule is the best suited for this particular problem.
Vote1, Vote2 and Vote3 are the ensembles proposed in this work and based on the weighed fusion of the following methods using
the vote rule: w1×SA1(175) + w2× SA2(50) + w3×SA3(50) + w4× Cheddad(125) + w5×DYC + w6×Bayes(110) > (w1+ w2+ w3+ w4+
w5+ w6)/ wτ, where w=(w1, w2, w3, w4, w5, w6) is the weight vector and wτ[1.25,1.5,1.75,2] is a threshold parameter used to tune the
sensitivity of the method. The weight vector used for the three ensemble are, respectively:
w1=(0.5, 1.5, 1, 1.5, 0.5, 1),
w2=(0.5, 1.5, 1, 1.5, 0, 1),
w3=(0, 1, 1, 1, 0, 1);
For each dataset the best result is highlighted and in the last column the rank (calculated as the rank of the average rank) is reported.
Table 3. Performance in 10 different datasets
Dataset
Perf. indicator
Method
τ/ wτ
GMM
110
Bayes
-2
SPL
Cheddad 125
Chen
175
SA1
50
SA2
50
SA3
DYC
OldVote
1.25
Vote1
1.25
Vote2
1.5
Vote3
F1 measure
Feeval Pratheepan
UChile
Compaq
SFA
HGR
Schmugge
VMD
AP
FnF
Rank
MCG
0.513
0.581
0.688
0.615
0.600
0.789
0.658
0.595
0.130
91.590
10
0.565
0.631
0.694
0.661
0.599
0.760
0.871
0.569
0.252
92.380
7
0.544
0.551
0.621
0.568
0.494
0.700
0.845
0.490
0.321
89.380
11.5
0.596
0.597
0.667
0.649
0.588
0.683
0.767
0.571
0.261
89.200
9
0.287
0.540
0.656
0.598
0.549
0.791
0.732
0.571
0.165
87.200
13
0.558
0.613
0.664
0.567
0.593
0.788
0.768
0.482
0.199
86.770
11.5
0.515
0.693
0.755
0.663
0.645
0.771
0.806
0.594
0.156
91.910
6
0.539
0.709
0.762
0.625
0.647
0.863
0.877
0.586
0.147
89.800
5
0.588
0.599
0.680
0.663
0.618
0.569
0.616
0.613
0.275
88.710
8
0.552
0.701
0.755
0.664
0.666
0.757
0.725
0.590
0.259
94.680
4
0.594
0.717
0.754
0.670
0.666
0.737
0.849
0.625
0.269
92.980
1
0.575
0.714
0.754
0.676
0.667
0.731
0.800
0.616
0.262
93.250
2
0.571
0.711
0.753
0.671
0.658
0.739
0.802
0.616
0.265
93.240
3
Maybe some approaches can be advantaged in some datasets, in case that they used some images of the dataset for training (i.e. LUT
creation), anyway considering the results on all the 10 datasets, the comparison is quite fair.
From the results in Table 3 it is clear that no approach overcomes the others in all the problems, on the contrary it is very difficult to
find a solution that works well in all the datasets without retraining and performing a parameter tuning. The design of an ensemble method
able to exploit the good performance of each of its composing approaches is a feasible solution to this problem. Our three proposed
ensembles based on the vote rule grant good performance on average in almost all the datasets and reach the first 3 positions in the rank
column. This is a valuable result considering that our experiments compares 13 approaches on 10 different datasets.
From a comparison among stand-alone approaches in Table 3, the following conclusions can be drawn:
There is a noticeable difference among performance of different approaches in each dataset. The best stand-alone method is different
for each dataset, anyway SA2 and SA3 work better than others, this is the reason why we gave them the higher weights in our vote
rule.
In particular SA3 reaches the best performance in three datasets: MCG, SFA, HGR. In our opinion, it is a very valuable method in
particular for problems where the background is quite simple (e.g. gesture recognition).
The method Bayes perform very well (it is ranked third among stand-alone methods) even though it is a very simple approach with
low computational requirements.
On the contrary the statistical color model proposed by Chen is very efficient but at the expense of performance.
The methods SPL is the best in VMD but performs very poorly in Compact.
For many approaches, the selection of an appropriate threshold is crucial for performance: the most appropriate choice depends of the
specific application, since a strict value decreases the number of false positives despite of the true positive rate, vice versa a low threshold
increases the true positive rate but also the number of false positives.
The ensemble that we propose as the best solution is Vote1(1.25); besides F1 we also report TPR and FPR as performance indicators
for Vote1(1.25) in the MGC and UChile datasets, for an easier comparison with several methods in the literature that use these two
indicators in these datasets [48]:
MGC (TPR=0.772, FPR=0.091),
UChile (TPR=0.7231, FPR=0.0915).
5. Conclusion and future research directions
In this work, a framework to evaluate and combine different skin detector approaches is presented and an extensive evaluation of
several approaches is carried out on 10 different datasets including more than 10000 labelled images. A survey of most recent existing
approaches is carried out and a new ensemble based on the combination of six methods is proposed and evaluated. Experimental results
confirm that our proposed ensemble obtains a very good performance with respect to other stand-alone approaches. The proposed
ensemble can be used in a wide range of skin detection applications such as body tracking, face detection, detection of objectionable
content and hand gesture analysis. The performance obtained on 10 different datasets confirms that the proposed ensemble is well suited
for detecting skin in a wide range of images without requiring re-training and ad hoc parameter optimization.
In conclusion, we show that skin detection is a very difficult problem, which can be hardly solved by a stand-alone approach. The
performance of any stand-alone method for skin detection depends on many factors, such as the color space used, the parameters used,
the nature of data, the image characteristics, the shape of the distribution, the size of samples for training, the presence of data noise, etc.
Our experiments show that our ensemble suffers less of these problems, since it is able to exploit the advantages of each composing
method and it is a feasible solution for problems where training is impracticable.
Future work for designing an effective skin-color detector can be done by adding a preprocessing module able to deal with the color
constancy problem in order to improve the robustness to illumination changes. Another idea to improve the precision of the detection is
the use of morphological operators as a post-processing step. Finally, in order to exploit the diversity in stand-alone approaches to be
fused in an ensemble, recent deep learning methods can be considered in the fusion, as soon as freely available method is provided.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
S. Kawato, J. Ohya, Automatic skin-color distribution extraction for face detection and tracking, WCC 2000 - ICSP 2000. 2000 5th Int. Conf. Signal
Process. Proceedings. 16th World Comput. Congr. 2000. 2 (2000). doi:10.1109/ICOSP.2000.891809.
A. Argyros, M. Lourakis, Real-time tracking of multiple skin-colored objects with a possibly moving camera, Eccv 2004. vol3 (2004) 368–379.
doi:10.1007/978-3-540-24672-5_29.
H. Sang, Y. Ma, J. Huang, Robust Palmprint Recognition Base on Touch-Less Color Palmprint Images Acquired, J. Signal Inf. Process. 4 (2013) 134–
139. doi:10.4236/jsip.2013.42019.
J.-S. Lee, Y.-M. Kuo, P.-C. Chung, E.-L. Chen, Naked image detection based on adaptive and extensible skin color model, Pattern Recognit. 40 (2007)
2261–2270. doi:10.1016/j.patcog.2006.11.016.
R. Khan, A. Hanbury, J. St??ttinger, A. Bais, Color based skin classification, Pattern Recognit. Lett. 33 (2012) 157–163.
doi:10.1016/j.patrec.2011.09.032.
S.G. Kong, J. Heo, B.R. Abidi, J. Paik, M.A. Abidi, Recent advances in visual and infrared face recognition - A review, Comput. Vis. Image Underst.
97 (2005) 103–135. doi:10.1016/j.cviu.2004.04.001.
G. Healey, M. Prasad, B. Tromberg, Face recognition in hyperspectral images, IEEE Trans. Pattern Anal. Mach. Intell. 25 (2003) 1552–1560.
doi:10.1109/TPAMI.2003.1251148.
C. Prema, M. Manimegalai, Survey on Skin Tone Detection using Color Spaces, Int. J. Appl. Inf. Syst. 2 (2012) 18–26.
P. Kakumanu, S. Makrogiannis, N. Bourbakis, A survey of skin-color modeling and detection methods, Pattern Recognit. 40 (2007) 1106–1122.
doi:10.1016/j.patcog.2006.06.010.
W. Chen, K. Wang, H. Jiang, M. Li, Skin color modeling for face detection and segmentation: a review and a new approach, Multimed. Tools Appl. 75
(2016) 839–862.
M.R. Mahmoodi, S.M. Sayedi, A Comprehensive Survey on Human Skin Detection, Int. J. Image, Graph. Signal Process. 8 (2016) 1–35.
doi:10.5815/ijigsp.2016.05.01.
T. Xu, Z. Zhang, Y. Wang, Patch-wise skin segmentation of human body parts via deep neural networks, J. Electron. Imaging. 24 (2015) 43009.
doi:10.1117/1.JEI.24.4.043009.
Y. Kim, I. Hwang, N.I. Cho, A New Convolutional Network-in-Network Structure and Its Applications in Skin Detection, Semantic Segmentation, and
Artifact Reduction, arXiv Prepr. arXiv1701.06190. (2017).
H. Zuo, H. Fan, E. Blasch, H. Ling, Combining Convolutional and Recurrent Neural Networks for Human Skin Detection, IEEE Signal Process. Lett.
24 (2017) 289–293.
B.D. Zarit, B.J. Super, F.K.H. Quek, Comparison of five color models in skin pixel classification, Proc. Int. Work. Recognition, Anal. Track. Faces
Gestures Real-Time Syst. Conjunction with ICCV’99 (Cat. No.PR00378). (1999) 58–63. doi:10.1109/RATFG.1999.799224.
N.B. Ibrahim, M.M. Selim, H.H. Zayed, A Dynamic Skin Detector Based on Face Skin Tone Color, in: 2012 8th Int. Conf. Informatics Syst., 2012: pp.
1–5.
R.P.K. Poudel, J.J. Zhang, D. Liu, H. Nait-Charif, Skin Color Detection Using Region-Based Approach, Int. J. Image Process. 7 (2013) 385.
W.C. Chen, M.S. Wang, Region-based and content adaptive skin detection in color images, Int. J. Pattern Recognit. Artif. Intell. 21 (2007) 831–853.
doi:10.1142/s0218001407005715.
H. Kruppa, M.A. Bauer, B. Schiele, Skin Patch Detection in Real-World Images, in: Annu. Symp. Pattern Recognit. DAGM 2002, 2002: p. 109f.
doi:10.1007/3-540-45783-6.
N. Sebe, I. Cohen, T.S. Huang, T. Gevers, Skin detection: a Bayesian network approach, Proc. 17th Int. Conf. Pattern Recognit. 2004 ICPR 2004. 2
(2004) 2–5. doi:10.1109/ICPR.2004.1334405.
A. Kumar, S. Malhotra, Pixel-Based Skin Color Classifier: A Review, Int. J. Signal Process. Image Process. Pattern Recognit. 8 (2015) 283–290.
M.J. Jones, J.M. Rehg, Statistical color models with application to skin detection, Int. J. Comput. Vis. 46 (2002) 81–96.
doi:10.1023/A:1013200319198.
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
QingXiang Wu, R. Cai, L. Fan, C. Ruan, G. Leng, Skin detection using color processing mechanism inspired by the visual system, in: IET Conf. Image
Process. (IPR 2012), 2012: pp. B10–B10. doi:10.1049/cp.2012.0459.
L. Chen, J. Zhou, Z. Liu, W. Chen, G. Xiong, A skin detector based on neural network, in: Commun. Circuits Syst. West Sino Expo. IEEE 2002 Int.
Conf., 2002: pp. 615–619 vol.1. doi:10.1109/ICCCAS.2002.1180694.
M.R. Mahmoodi, S.M. Sayedi, Leveraging spatial analysis on homogonous regions of color images for skin classification, in: Comput. Knowl. Eng.
(ICCKE), 2014 4th Int. eConference, 2014: pp. 209–214.
R. Nidhu, M.G. Thomas, Real Time Segmentation Algorithm for Complex Outdoor Conditions, Int. J. Sci. Technoledge. 2 (2014) 71.
M. Kawulok, J. Kawulok, J. Nalepa, B. Smolka, Self-adaptive algorithm for segmenting skin regions, EURASIP J. Adv. Signal Process. 2014 (2014)
1–22. doi:10.1186/1687-6180-2014-170.
A.S. Nunez, M.J. Mendenhall, Detection of Human Skin in Near Infrared Hyperspectral Imagery, Int. Geosci. Remote Sens. Symp. 2 (2008) 621–624.
doi:10.1109/IGARSS.2008.4779069.
J. Han, G.M. Award, A. Sutherland, H. Wu, Automatic skin segmentation for gesture recognition combining region and support vector machine active
learning, in: FGR 2006 Proc. 7th Int. Conf. Autom. Face Gesture Recognit., 2006: pp. 237–242. doi:10.1109/FGR.2006.27.
Z. Jiang, M. Yao, W. Jiang, Skin Detection Using Color, Texture and Space Information, Fourth Int. Conf. Fuzzy Syst. Knowl. Discov. (FSKD 2007).
(2007) 366–370. doi:10.1109/FSKD.2007.518.
H.K. Al-Mohair, J. Mohamad Saleh, S.A. Suandi, Hybrid Human Skin Detection Using Neural Network and K-Means Clustering Technique, Appl.
Soft Comput. 33 (2015) 337–347. doi:10.1016/j.asoc.2015.04.046.
F.E. Sandnes, L. Neyse, Y.-P. Huang, Simple and practical skin detection with static RGB-color lookup tables: A visualization-based study, in: Syst.
Man, Cybern. (SMC), 2016 IEEE Int. Conf., 2016: pp. 2370–2375.
S. Jairath, S. Bharadwaj, M. Vatsa, R. Singh, Adaptive Skin Color Model to Improve Video Face Detection, in: Mach. Intell. Signal Process., Springer,
2016: pp. 131–142.
W. Song, D. Wu, Y. Xi, Y.W. Park, K. Cho, Motion-based skin region of interest detection with a real-time connected component labeling algorithm,
Multimed. Tools Appl. (2016) 1–16.
A. Gupta, A. Chaudhary, Robust skin segmentation using color space switching, Pattern Recognit. Image Anal. 26 (2016) 61–68.
M.M. Oghaz, M.A. Maarof, A. Zainal, M.F. Rohani, S.H. Yaghoubyan, A hybrid Color space for skin detection using genetic algorithm heuristic
search and principal component analysis technique, PLoS One. 10 (2015). doi:10.1371/journal.pone.0134828.
C.Ó. Conaire, N.E. O’Connor, A.F. Smeaton, Detector adaptation by maximising agreement between independent data sources, in: Proc. IEEE
Comput. Soc. Conf. Comput. Vis. Pattern Recognit., 2007. doi:10.1109/CVPR.2007.383448.
A. Cheddad, J. Condell, K. Curran, P. Mc Kevitt, A skin tone detection algorithm for an adaptive approach to steganography, Signal Processing. 89
(2009) 2465–2478. doi:10.1016/j.sigpro.2009.04.022.
Y.H. Chen, K.T. Hu, S.J. Ruan, Statistical skin color detection method without color transformation for real-time surveillance systems, Eng. Appl.
Artif. Intell. 25 (2012) 1331–1337. doi:10.1016/j.engappai.2012.02.019.
M. Kawulok, Fast propagation-based skin regions segmentation in color images, in: 2013 10th IEEE Int. Conf. Work. Autom. Face Gesture
Recognition, FG 2013, 2013. doi:10.1109/FG.2013.6553733.
M. Kawulok, J. Kawulok, J. Nalepa, Spatial-based skin detection using discriminative skin-presence features, Pattern Recognit. Lett. 41 (2014) 3–13.
doi:10.1016/j.patrec.2013.08.028.
N. Brancati, G. De Pietro, M. Frucci, L. Gallo, Human skin detection through correlation rules between the YCb and YCr subspaces based on dynamic
color clustering, Comput. Vis. Image Underst. 155 (2017) 33–42. doi:10.1016/j.cviu.2016.12.001.
Q. Zhu, C.-T. Wu, K. Cheng, Y. Wu, An adaptive skin model and its application to objectionable image filtering, Proc. 12th Annu. ACM Int. Conf.
Multimed. - Multimed. ’04. (2004) 56. doi:10.1145/1027527.1027538.
J. Ruiz-Del-Solar, R. Verschae, Skin detection using neighborhood information, in: Proc. - Sixth IEEE Int. Conf. Autom. Face Gesture Recognit.,
2004: pp. 463–468. doi:10.1109/AFGR.2004.1301576.
S.L. Phung, A. Bouzerdoum, D. Chai, Skin segmentation using color pixel classification: Analysis and comparison, IEEE Trans. Pattern Anal. Mach.
Intell. 27 (2005) 148–154. doi:10.1109/TPAMI.2005.17.
S.J. Schmugge, S. Jayaram, M.C. Shin, L. V. Tsap, Objective evaluation of approaches of skin detection using ROC analysis, Comput. Vis. Image
Underst. 108 (2007) 41–51. doi:10.1016/j.cviu.2006.10.009.
J. Stöttinger, A. Hanbury, C. Liensberger, R. Khan, Skin paths for contextual flagging adult videos, in: Lect. Notes Comput. Sci. (Including Subser.
Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), 2009: pp. 303–314. doi:10.1007/978-3-642-10520-3_28.
L. Huang, T. Xia, Y. Zhang, S. Lin, Human skin detection in images by MSER analysis, 2011 18th IEEE Int. Conf. Image Process. (2011) 1257–1260.
doi:10.1109/ICIP.2011.6115661.
W.R. Tan, C.S. Chan, P. Yogarajah, J. Condell, A Fusion Approach for Efficient Human Skin Detection, Ind. Informatics, IEEE Trans. 8 (2012) 138–
147. doi:10.1109/TII.2011.2172451.
J.C. Sanmiguel, S. Suja, Skin detection by dual maximization of detectors agreement for video monitoring, Pattern Recognit. Lett. 34 (2013) 2102–
2109. doi:10.1016/j.patrec.2013.07.016.
J.P.B. Casati, D.R. Moraes, E.L.L. Rodrigues, SFA: A human skin image database based on FERET and AR facial images, in: IX Work. Visao
Comput. Rio Janeiro, 2013.
M.R. Mahmoodi, S.M. Sayedi, F. Karimi, Z. Fahimi, V. Rezai, Z. Mannani, SDD: A skin detection dataset for training and assessment of human skin
classifiers, in: Knowledge-Based Eng. Innov. (KBEI), 2015 2nd Int. Conf., 2015: pp. 71–77.
L. Nanni, A. Lumini, Skin Detection for Reducing False Positive in Face Detection, in: V.M. Petrova (Ed.), Adv. Eng. Res. vol.16, Nov. Publ. 2017,
Nova Publisher, 2017. https://www.novapublishers.com/catalog/product_info.php?products_id=60326&osCsid=.
A. Angelova, Y. Abu-Mostafa, P. Perona, Pruning training sets for learning of object categories, in: Proc. IEEE Comput. Soc. Conf. Comput. Vis.
Pattern Recognit., 2005: pp. 494–501. doi:10.1109/CVPR.2005.283.
D.M.W. POWERS, Evaluation: From Precision, Recall and F-Measure To Roc, Informedness, Markedness & Correlation, J. Mach. Learn. Technol. 2
(2011) 37–63. doi:10.1.1.214.9232.
L. Nanni, A. Lumini, M. Migliardi, Learning-based Skin Classification, Autom. Identif. Technol. (2016).
Extended version of Table 2 (including results for different thresholds)
Dataset
Perf. indicator
Method
τ/ wτ
GMM
SPL
-2.5
-2
-1.5
-1
-0.5
Bayes
50
70
90
110
140
Cheddad
75
100
125
150
175
Chen
SA1
225
200
175
150
100
SA2
30
40
50
85
120
SA3
125
100
75
50
25
DYC
OldVote
Vote1
1.25
1.5
1.75
2
Vote2
1.25
1.5
1.75
2
Vote3
1.25
1.5
1.75
2
F1 measure
Feeval Prath.
0.513
0.581
0.549
0.537
0.544
0.551
0.544
0.567
0.529
0.578
0.512
0.582
0.552
0.672
0.564
0.663
0.567
0.649
0.565
0.631
0.547
0.599
0.561
0.633
0.619
0.641
0.596
0.597
0.518
0.513
0.439
0.448
0.287
0.540
0.486
0.579
0.535
0.603
0.558
0.613
0.575
0.615
0.560
0.576
0.524
0.683
0.519
0.690
0.515
0.693
0.513
0.665
0.497
0.626
0.506
0.638
0.521
0.674
0.536
0.699
0.539
0.709
0.544
0.707
0.588
0.599
0.552
0.701
0.594
0.717
0.594
0.717
0.593
0.684
0.593
0.684
0.575
0.714
0.576
0.711
0.576
0.694
0.575
0.664
0.567
0.713
MCG
0.688
0.615
0.621
0.627
0.624
0.615
0.702
0.711
0.706
0.694
0.667
0.668
0.704
0.667
0.585
0.526
0.656
0.639
0.657
0.664
0.668
0.645
0.739
0.749
0.755
0.749
0.714
0.715
0.738
0.755
0.762
0.752
0.680
0.755
0.754
0.754
0.733
0.733
0.754
0.751
0.740
0.722
0.750
UChile
0.615
0.562
0.568
0.565
0.554
0.543
0.605
0.643
0.660
0.661
0.642
0.563
0.637
0.649
0.567
0.484
0.598
0.511
0.553
0.567
0.571
0.543
0.656
0.661
0.663
0.647
0.623
0.575
0.592
0.608
0.625
0.631
0.663
0.664
0.670
0.670
0.684
0.684
0.676
0.655
0.655
0.669
0.636
Compaq
0.600
0.492
0.494
0.497
0.494
0.485
0.613
0.617
0.609
0.599
0.574
0.615
0.608
0.588
0.567
0.540
0.549
0.536
0.580
0.593
0.599
0.588
0.641
0.644
0.645
0.627
0.591
0.591
0.613
0.633
0.647
0.647
0.618
0.666
0.666
0.666
0.635
0.635
0.667
0.656
0.641
0.620
0.664
SFA
0.789
0.726
0.700
0.628
0.583
0.521
0.473
0.593
0.686
0.760
0.830
0.281
0.468
0.683
0.550
0.581
0.791
0.809
0.823
0.788
0.730
0.572
0.758
0.760
0.771
0.799
0.801
0.897
0.896
0.886
0.863
0.817
0.569
0.757
0.737
0.737
0.792
0.792
0.731
0.793
0.793
0.820
0.662
HGR
0.658
0.845
0.845
0.845
0.828
0.801
0.711
0.793
0.840
0.871
0.885
0.357
0.596
0.767
0.850
0.812
0.732
0.752
0.787
0.768
0.732
0.614
0.842
0.831
0.806
0.508
0.354
0.498
0.667
0.785
0.877
0.893
0.616
0.725
0.849
0.849
0.872
0.872
0.800
0.837
0.837
0.852
0.729
Schmugge
0.595
0.501
0.490
0.482
0.462
0.435
0.508
0.547
0.565
0.569
0.555
0.571
0.601
0.571
0.384
0.349
0.571
0.453
0.479
0.482
0.471
0.440
0.580
0.590
0.594
0.584
0.553
0.511
0.535
0.566
0.586
0.591
0.613
0.590
0.625
0.625
0.611
0.611
0.616
0.611
0.611
0.597
0.569
VMD
0.130
0.302
0.321
0.341
0.354
0.367
0.264
0.272
0.268
0.252
0.221
0.209
0.274
0.261
0.151
0.067
0.165
0.087
0.153
0.199
0.234
0.276
0.169
0.163
0.156
0.130
0.108
0.101
0.114
0.129
0.147
0.172
0.275
0.259
0.269
0.269
0.250
0.250
0.262
0.245
0.245
0.219
0.209
AP
FnF
91.590
87.580
89.380
90.290
91.180
92.140
95.710
94.680
93.100
92.380
90.510
93.710
92.860
89.200
80.700
75.110
87.200
76.980
83.860
86.770
88.500
90.780
91.660
91.770
91.910
92.100
91.480
86.640
87.930
89.150
89.800
90.980
88.710
94.680
92.980
92.980
92.980
92.980
93.250
93.250
93.250
93.250
93.240
0.571
0.711
0.753
0.671
0.658
0.739
0.802
0.616
0.265
93.240
0.571
0.700
0.745
0.671
0.648
0.739
0.802
0.616
0.265
93.240
0.571
0.678
0.732
0.671
0.632
0.739
0.802
0.616
0.265
93.240
Rank
10
11.5
7
9
13
11.5
6
5
8
4
1
2
3
| 1cs.CV
|
Discrete Extremes
arXiv:1707.05033v1 [math.ST] 17 Jul 2017
Adrien Hitz, Richard Davis and Gennady Samorodnitsky
Our contribution is to widen the scope of extreme value analysis applied to discretevalued data. Extreme values of a random variable X are commonly modeled using
the generalized Pareto distribution, a method that often gives good results in practice.
When X is discrete, we propose two other methods using a discrete generalized Pareto
and a generalized Zipf distribution respectively. Both are theoretically motivated and
we show that they perform well in estimating rare events in several simulated and real
data cases such as word frequency, tornado outbreaks and multiple births.
1
Introduction
Extreme quantile estimation is an important but difficult problem in statistics, especially
when the quantile is beyond the range of the data. In the univariate case, extreme value
theory motivates the choice of a parametric family called the generalized Pareto distribution
(GPD) which is used to model the tail and estimate the probability of rare events (Pickands
[1975]). Let X be a random variable taking values in [0, xF ) with survival function F̄X ,
where xF ∈ R+ ∪ {∞} and R+ = (0, ∞). Suppose that there exists a strictly positive
sequence au such that
d
a−1
u (X − u) | X ≥ u → Z,
(1)
as u → xF , for some Z following a non-degenerate probability distribution on [0, ∞), where
d
→ denotes weak convergence. A stunning result is that this assumption is sufficient to
characterize the limiting distribution: Z follows a generalized Pareto distribution (GPD),
defined by its survival function
x −1/ξ
1{x<τ } ,
1 − FGPD (x; σ, ξ) = F̄GPD (x; σ, ξ) = 1 + ξ
σ
x ≥ 0,
with τ = ∞ if ξ ≥ 0, τ = σ/|ξ| if ξ < 0, where 1{x<τ } is 1 if x < τ and 0 otherwise, and
σ > 0. We use the convention that if ξ = 0, then (1 + ξx)1/ξ = ex . Condition (1), written
as X ∈ MDAξ , means that X is in the maximum domain of attraction of an extreme value
distribution with shape parameter ξ (see e.g. Resnick [1987]). Sometimes one says that the
Adrien Hitz, University of Oxford, 24-29 St Giles, Oxford OX1 3LB, UK.
Gennady Samorodnitsky, Cornell University, 220 Rhodes Hall, Ithaca, NY.
Richard Davis, Columbia University, 1255 Amsterdam Avenue, New York, NY.
law F of X is in MDAξ . In this case, the sequence of cumulative distribution functions of
a−1
u (X − u) | X ≥ u converges uniformly to FGPD on [0, ∞). Thus, for large u,
−1
P (X − u > x | X ≥ u) = P {a−1
u (X − u) > au x | X ≥ u}
is often approximated by
≈ F̄GPD (x; σau , ξ) ,
(2)
motivating the approximation of the distribution of exceedances of X above a large threshold u by a GPD (Davison and Smith [1990]). In the continuous case, where X is assumed
to have a continuous distribution, most common distributions belong to some maximum
domain of attraction, and the GPD approximation to the tail of the distribution can be
applied. This approximation often works well in practice, although it can be poor when u
is not large enough, for instance if X is normal and u is around the 90th percentile of its
distribution.
Two issues are apparent with applying the GPD method of approximating the distribution tail to a discrete distribution. First, a necessary condition for a discrete distribution FX to be in some maximum domain of attraction in the case xF = ∞ is that FX
is long-tailed, i.e., F̄X (u + 1)/F̄X (u) → 1 as u → ∞, see Shimura [2012] and Anderson
[1970, 1980]. Without being in a maximum domain of attraction the GPD approximation
(2) does not necessarily apply. Note that many common discrete distributions, including
geometric, Poisson and negative binomial distributions, are not long-tailed. The second
issue in approximating a discrete distribution by a GPD, a continuous distribution, is that
ties are not allowed. To overcome these limitations, we suggest two alternative methods
of modeling the tails of discrete observations, each one relying on a specific assumption
on the underlying distribution. For this work, we will only consider approximations to
distributions with infinite support.
We say that a discrete random variable X with non-negative values is in the discrete
maximum domain of attraction, which we write as X ∈ D-MDAξ , if there exists a random
variable Y ∈ MDAξ with ξ ≥ 0 such that Pr(X ≥ k) = Pr(Y ≥ k) for k = 0, 1, 2, . . . , i.e.,
d
the equality in distribution, X = bY c, holds. We call Y an extension of X and such an
extension is not unique. Shimura [2012] showed that X ∈ MDAξ for some ξ ≥ 0 if and only
if X ∈ D-MDAξ and X is long-tailed. It was also shown by Shimura [2012] that geometric,
Poisson and negative binomial distributions are in D-MDA. This set is, therefore, strictly
larger than the set of all discrete distributions in MDAξ .
d
Let X ∈ D-MDAξ and Y ∈ MDAξ be the corresponding extension satisfying X = bY c.
Then, for large integers u,
P (X − u = k | X ≥ u) = P (Y − u ≥ k | Y ≥ u) − P (Y − u ≥ k + 1 | Y ≥ u)
2
which we will approximate as in (2) by
≈ pD-GPD (k; σau , ξ),
(3)
where pD-GPD is the probability mass function of the discrete generalized Pareto distribution (D-GPD) defined by
pD-GPD (k; σ, ξ) = F̄GPD (k; σ, ξ) − F̄GPD (k + 1; σ, ξ),
(4)
for k = 0, 1, 2, . . .. The discrete generalized Pareto distribution has been used by Prieto
et al. [2014] to model road accidents, while various aspects of discrete Pareto-type distributions were studied in Krishna and Pundir [2009], Buddana and Kozubowski [2014], and
Kozubowski et al. [2015]. Notice that the scale parameter in (3) is undetermined since the
extension Y is non-unique.
We now discuss an alternative assumption on the distribution of the discrete random
variable X that will allow us to construct another approximation of its tail. Let pX be
the probability mass function of X and suppose that there exists a non-negative random
variable Y ∈ MDAξ/(1+ξ) with ξ ≥ 0 such that pX (k) = c F̄Y (k) for k = d, d + 1, d + 2, . . . ,
for some c > 0 and d ∈ N0 = {0, 1, . . .}. We denote this condition by pX ∈ D-MDAξ/(1+ξ)
and call F̄Y an extension of pX . Then for large integers u,
pX (u + k)/pX (u)
P (X − u = k | X ≥ u) = P∞
i=0 pX (u + i)/pX (u)
P (Y > u + k)/P (Y > u)
= P∞
i=0 P (Y > u + i)/P (Y > u)
which can be approximated as in (2),
≈ pGZD k; (1 + ξ)σau , ξ ,
(5)
where
1 + ξ σk
pGZD (k; σ, ξ) = P
∞
i=0
−1/ξ−1
1 + ξ σi
−1/ξ−1 ,
k = 0, 1, 2, . . . .
(6)
We call the probability mass function in (6) generalized Zipf distribution (GZD). In the
case ξ > 0, the GZD is a Zipf–Mandelbrot distribution whose probability mass function is
usually written in the form p(k) = (k + q)−s /Hs,q , for k = 0, 1, 2, . . . , s > 1 and q > 0,
where Hs,q is the Hurwitz-Zeta function (Mandelbrot [1953]). The GZD is of this form
with s = 1 + 1/ξ and q = σ/ξ. When q = 1 the distribution is a Zipf law which is
sometimes presented as the counterpart of the Pareto distribution because its probability
mass function, after a shift, can be written in a homogeneous form (Arnold [1983]). Zipftype families of discrete distributions have been fitted to various data sets such as word
3
frequencies (Booth [1967]), city sizes (Gabaix [1999]), company sizes (Axtell [2001]) and
numbers of website hits (Clauset et al. [2009]). In the case ξ = 0, the generalized Zipf
distribution is a geometric distribution with probability of success p = 1 − e−1/σ (and so
is the discrete generalized Pareto distribution).
We will show that the geometric, Poisson and negative binomial probability mass functions belong to D-MDA0 , so the tail approximation (5) makes sense in these cases. In
addition, we will see that pX ∈ D-MDAξ/(1+ξ) for ξ ≥ 0 implies X ∈ MDAξ (under an
additional condition in the case ξ = 0).
Equations (3) and (5) motivate the approximation of the tails of some discrete distributions by a D-GPD and GZD, respectively. If ξ ≥ 0, then
sup
k=0,1,2...
fGPD (k; σ, ξ)
− 1 −→ 0,
σ→∞
q(k; σ, ξ)
for either q = pD-GPD or q = pGZD . One thus expects the GPD, D-GPD and GZD approximations to give similar results when the scale parameter σ is large.
In Section 2, we justify theoretically the three approximations in the case ξ > 0 by
showing that if pX ∈ MDAξ/(1+ξ) , then for any sequence ku ⊂ N0 such that supu ku /u < ∞,
P (X = ku + u | X ≥ u)
→ 1,
q(ku ; ξu, ξ)
k = 0, 1, 2, . . . ,
where q ≡ fGPD , pD-GPD and pGZD (see Theorem 2.3 and Proposition 2.4). A similar
justification is provided in the case ξ = 0 for the D-GPD and GZD approximations (Theorem 2.5). We further present an important invariance property satisfied by the D-GPD
(Proposition 2.7).
In Section 3, we simulate data from different discrete distributions and compare the
ability of the three approximations for estimating the probability of regions far from the
origin. The D-GPD and GZD approximations outperform the GPD when there are many
tied observations, otherwise the results are similar. As opposed to the GZD, the D-GPD
has a closed-form survival and probability mass function allowing for an exact likelihood
based inference.
In Section 4, we study a data set counting the occurrence of British words in a corpus
and show that the three distributions are appropriate to describe the tails of these word
frequencies. We also use the D-GPD and GZD to model the length of French words, the
number of tornado outbreaks in the United States over 50 years, and the number of births
in the United States in the last 20 years, illustrating how they offer potential methods for
estimating the probability of rare events when the data are discrete.
4
2
Theoretical Results
In this section we describe a number of properties of the approximation procedures introduced in Section 1.
Proposition 2.1. If X ∈ D-MDAξ , then there exists a positive sequence (au , u = 1, 2, . . .)
such that
lim
sup
u∈N0 , u→∞ k=0,1,2,...
P (X = u + k | X ≥ u) − pD-GPD (k; au , ξ) = 0 ,
(7)
where pD-GPD is defined in (4).
We remark that if au → ∞, then the result is less interesting because the two terms in
(7) converge to 0.
Proof. By assumption, there exists a random variable Y ∈ MDAξ for ξ ≥ 0 and a positive
d
function (ãu , u > 0) such that X = bY c and the sequence of functions P ã−1
u (Y − u) ≥ x |
Y ≥ u , x ≥ 0, converges uniformly, as u → ∞, to the function F̄GPD (x; σ, ξ), x ≥ 0, for
some σ > 0 and ξ ≥ 0. For a positive integer u we let au = ãu σ. Then
| P (X = u + k | X ≥ u) − pD-GPD (k; au , ξ) |
sup
k=0,1,2,...
=
sup
k=0,1,2,...
−1
−1
−1
P ã−1
u (Y − u) ≥ ãu k | Y ≥ u − P ãu (Y − u) ≥ ãu (k + 1) | Y ≥ u
− F̄GPD (k; au , ξ) + F̄GPD (k + 1; au , ξ)
≤ 2 sup P ã−1
u (Y − u) ≥ x | Y ≥ u − F̄GPD (x; σ, ξ) → 0
x≥0
as u → ∞ over the integers.
The following auxiliary lemma is elementary (as the sum can be sandwiched between
two integrals).
Lemma 2.2. If ξ > 0, then, as u → ∞,
u1/ξ H1+1/ξ,u → ξ,
where Hs,q =
P∞
i=0 (q
+ i)−s is the Hurwitz-Zeta function.
Recall that a positive and measurable function f on [1, ∞) is regularly varying if there
exists a positive function ` such that
lim
u→∞
f (ux)
→ `(x),
f (u)
5
x ≥ 1.
In this case, there exists α ∈ R such that `(x) = xα and we write f ∈ RVα (see e.g.
Bingham et al. [1989]). If f ∈ RV−α for α ≥ 0, then
lim sup
u→∞ x∈[1,b]
f (ux)
− x−α → 0,
f (u)
(8)
for b = ∞ if α > 0, and for any b < ∞ if α = 0. If f ∈ RV−α for α > 0, then by Potter’s
bounds (see e.g. Resnick [1987]) for any > 0 there is u ∈ (0, ∞) such that
e− x−α− ≤
f (ux)
≤ e x−α+ ,
f (u)
x ≥ 1,
(9)
for u ≥ u . We say that X is regularly varying if F̄X ∈ RV−α for some α > 0, a necessary
and sufficient condition for X ∈ MDA1/α .
The following result sheds some light on the approximation suggested in (5).
Theorem 2.3. If pX ∈ D-MDAξ/(1+ξ) for ξ > 0, then X ∈ MDAξ and for any sequence
of nonnegative integers (ku , u = 1, 2, . . .) such that supu ku /u < ∞,
lim
u∈N0 , u→∞
P (X = ku + u | X ≥ u)
= 1.
pGZD (ku ; ξu, ξ)
(10)
Proof. By assumption, there exists a survival function F̄ such that F̄ (k) = c pX (k) for c >
0, k large enough and F̄ ∈ RV−1/ξ−1 . The last condition is equivalent to F̄ (b·c) ∈ RV−1/ξ−1
(Shimura [2012]). Therefore,
P∞
R duxe
R∞
F̄ (byc)dy − ux F̄ (byc)dy
Pr(X ≥ ux)
i=duxe pX (i)
ux
R∞
=
= P∞
Pr(X ≥ u)
i=u pX (i)
u F̄ (byc)dy
R∞
u x F̄ (buzc)/F̄ (buc)dz − (duxe − ux)F̄ (buxc)/F̄ (buc)
R∞
=
→ x−1/ξ , x ≥ 1,
u 1 F̄ (buzc)/F̄ (buc)dz
applying (9) and dominated convergence. Thus, F̄X ∈ RV−1/ξ and X ∈ MDAξ .
For the second part of the theorem, we have
P
−1/ξ−1
F̄ (u + ku )/F̄ (u) ∞
P (X = ku + u | X ≥ u)
i=0 (1 + i/u)
P
=
.
pGZD (ku ; ξu, ξ)
(1 + ku /u)−1/ξ−1 ∞
i=0 F̄ (u + i)/F̄ (u)
By the uniform convergence (8) and the the fact that ku grows at most linearly fast, we
conclude that
F̄ (u + ku )/F̄ (u)
→ 1,
(1 + ku /u)−1/ξ
as u → ∞ over the integers. Second, Lemma 2.2 yields
u−1
∞
X
(1 + i/u)−1/ξ−1 → ξ.
i=0
6
Third, it follows from (9) that for ∈ (0, 1/ξ), there exists u > 0 such that for u ≥ u ,
∞
∞
X
X
ξe
i −1−1/ξ+
→
,
u−1
F̄ (u + i)/F̄ (u) ≤ u−1 e
1+
u
1 − ξ
i=0
i=0
using again Lemma 2.2. A lower bound is found in the same manner and we let → 0 to
conclude.
We now present a tail equivalence property between the probability mass and density
functions of the GZD, D-GPD and GPD. A direct consequence is that the denominator
pGZD in (10) can be replaced either by pD-GPD or by fGPD .
Proposition 2.4. If ξ ≥ 0, then
lim
sup
σ→∞ k=0,1,2,...
pD-GPD (k; σ, ξ)
pD-GPD (k; σ, ξ)
− 1 = lim sup
− 1 = 0.
σ→∞
pGZD (k; σ, ξ)
k=0,1,2,... fGPD (k; σ, ξ)
Proof. Suppose first that ξ > 0. Then
−1/ξ
−1/ξ
1 + ξ σk
− 1 + ξ k+1
pD-GPD (k; σ, ξ)
σ
=
1
k −1/ξ−1
fGPD (k; σ, ξ)
1
+
ξ
σ
σ
(
−1/ξ )
ξ
= 1− 1+
(σ + ξk) → 1,
σ + ξk
uniformly in k = 0, 1, 2, . . .. Furthermore,
∞
X
fGPD (k; σ, ξ)
= σ −1
(1 + ξi/σ)−1/ξ−1 → 1
k=0,1,2,... pGZD (k; σ, ξ)
sup
i=0
by Lemma 2.2. In the case ξ = 0,
pD-GPD (k; σ, 0)/fGPD (k; σ, 0) = pGZD (k; σ, 0)/fGPD (k; σ, 0) = σ(1 − e−1/σ ) → 1
as σ → ∞.
Next we extend Theorem 2.3 to the case ξ = 0. Recall that a distribution F is in MDA0
if and only if the survival function has a representation
Z x
1
F̄ (x) = c(x) exp −
dy , −∞ < x < xF ,
(11)
0 a(y)
where c(·) is a positive function with c(x) → c > 0 as x ↑ xF , and a(·) is a positive,
differentiable function a(·) with limx↑xF a0 (x) = 0. If c(x) = c on (d, xF ) for some d < xF ,
then we say that the distribution F satisfies the von Mises condition. The function a(·)
in (11) is sometimes called the auxiliary function. Note, however, that it is only uniquely
defined (on (d, xF )) under the von Mises condition; see Embrechts et al. [2013]. Recall that
in the sequel we only consider the case of unbounded support, i.e. xF = ∞.
7
Theorem 2.5. Suppose that pX ∈ D-MDA0 and, moreover, that a distribution F such that
pX (k) = F̄ (k) has the property that an auxiliary function of F̄ satisfies limx→∞ a(x) = σ ∈
[0, ∞]. Then
lim
u∈N0 , u→∞
P (X = k + u | X ≥ u) = pGe (k; σ),
k = 0, 1, 2, . . . ,
(12)
where pGe (k; σ) = (1 − e−1/σ ) e−k/σ is the probability mass function of a geometric distribution if 0 < σ < ∞, and pGe (k; ∞) = pGe (k; 0) = 0. Furthermore, if σ ∈ [0, ∞), then
X ∈ D-MDA0 .
Proof. Note that for large integers u,
F̄ (k + u)/F̄ (u)
.
P (X = k + u | X ≥ u) = P∞
i=0 F̄ (i + u)/F̄ (u)
(13)
We have for every i = 0, 1, 2, . . .,
Z i
c(i + u)
exp −
1/a(u + y)dy → e−i/σ
F̄ (i + u)/F̄ (u) =
c(u)
0
as u → ∞. If 0 < σ < ∞, then the dominated convergence theorem gives us
∞
X
F̄ (i + u)/F̄ (u) →
∞
X
e−i/σ = 1/(1 − e−1/σ ) ,
i=0
i=0
and (12) follows. If σ = ∞,
∞
X
F̄ (i + u)/F̄ (u) → ∞
i=0
by Fatou’s lemma, and (12), once again, follows. If σ = 0, the claim follows from the fact
that the denominator in (13) cannot be smaller than 1.
For the second part of the proposition, it follows from
Z n
1
pX (n) = c(n) exp −
dy
0 a(y)
for all n and a(y) → σ ∈ [0, ∞) that
pX (n)
= 1 − e−1/σ ∈ (0, ∞) ,
n→∞ P (X ≥ n)
lim
which immediately implies that X ∈ D-MDA0 as well.
To summarize, for a discrete random variable X, the conditions X ∈ MDA, X ∈
D-MDA and pX ∈ D-MDA are related to each other as follows. If ξ ≥ 0, then X ∈ MDAξ
if and only if X ∈ D-MDAξ and X is long-tailed. If ξ > 0, then pX ∈ D-MDAξ/(1+ξ) implies
8
X ∈ D-MDAξ ; the same implication holds in the case ξ = 0 if the auxiliary function of the
extension of pX satisfies a(x) → σ ∈ (0, ∞) as x → ∞.
The condition pX ∈ D-MDA is satisfied, among others, by the Zipf–Mandelbrot, geometric, Poisson and negative binomial distributions, as shown in the next example.
Example 2.6. The probability mass function of a Zipf–Mandelbrot distribution with parameters s > 1 and q > 0 is in D-MDA1/s because it is regularly varying of order −s.
The probability mass function of a geometric distribution belongs to D-MDA0 as it
coincides up to a constant with the survival function of an exponential distribution. The
latter distribution clearly satisfies the von Mises condition and thus is a member of MDA0 .
The auxiliary function is, in fact, equal (eventually) to 1/λ, where λ is the rate of the
exponential distribution.
The probability mass function p of a Poisson distribution with rate λ > 0 coincides on
k = 0, 1, 2, . . . with the function
λx e−λ
,
g(x) =
Γ(x + 1)
a continuous function on R+ satisfying limx→∞ g(x) = 0. Moreover,
d
log g(x) = −ψ0 (x + 1) + log λ ,
dx
where ψ0 is the polygamma function of order 0. Since ψ0 (x) → ∞ as x → ∞, we see that
g 0 (x) < 0 for x sufficiently large. Therefore, F̄Y (x) = g(x)/g(d) is a survival function on
[d, ∞) for some d ≥ 0. Furthermore,
1
ψ1 (x + 1)
d
− 0
=−
,
dx
g (x)
(ψ0 (x + 1) − log λ)2
where ψ1 = ψ00 is is the polygamma function of order 1. Since ψ1 (x) → 0 as x → ∞,
we conclude that F satisfies the von Mises condition, with the auxiliary function a(x) =
(ψ0 (x + 1) − log λ)−1 → 0 as x → ∞. Therefore, the Poisson probability mass function is
in D-MDA0 .
Similarly, the probability mass function of the negative binomial distribution with probability of success p ∈ (0, 1) and number of successes r > 0 is also in D-MDA0 because it
coincides on {0, 1, 2, . . .} with the function
g(x) =
pr Γ(x + r)
(1 − p)x ,
Γ(r) Γ(x + 1)
a continuous function on R+ . It is simple to check that limx→∞ g(x) = 0, and g 0 (x) < 0 for
x large enough, so that F̄Y (x) = g(x)/g(d) is a survival function on [d, ∞) for some d ≥ 0.
9
Furthermore, g(x) ∼ cxr−1 (1 − p)x for large x, where c is a positive constant. Therefore,
F̄Y is of the form (11) with the auxiliary function
a(x) =
1
, x large,
− log(1 − p) − (r − 1)/x
and so it converges to −1/ log(1 − p) as x → ∞.
We conclude this section with a discussion designed to provide some intuition on how
the approximation methods suggested above differ, assuming that both apply in a given
situation. First of all, one would expect the D-GPD and GZD approximations to perform
similarly when ξ is close to zero because both approximating distributions coincide with a
geometric distribution when ξ = 0. Second, Proposition 2.4 suggests that, regardless of the
underlying justification, if one uses either pD-GPD (k; σ, ξ), fGPD (k; σ, ξ) or pGZD (k; σ, ξ) as
an approximation to P (X − u = k | X ≥ u), one should not expect major differences as
long as one chooses the scale parameter σ to be large. This would always be the case if X
is long-tailed, since the scale parameter is chosen to be proportional to the normalization
sequence au defined in (1), which grows to infinity if and only if X is long-tailed.
When using a continuous distribution, such as the generalized Pareto distribution, to
approximate the probabilities related to a discrete distribution, it is also common to use a
“continuity correction” and shift the argument in the continuous approximation by some
δ ∈ [0, 1). In our situation this amounts to replacing fGPD (k; σ, ξ) by fGPD (k + δ; σ, ξ),
some δ ∈ [0, 1). When ξ > 0, it is elementary to check that, as σ → ∞,
pD-GPD(σ,ξ) (k)
(1 + ξ)(2δ − 1)
=1+
+ O(σ −2 )
fGPD(σ,ξ) (k + δ)
2σ
for every k = 0, 1, 2, . . .. Therefore, the approximations by pD-GPD(σ,ξ) (k) and fGPD (k +
δ; σ, ξ) with large σ are most similar when δ = 1/2, a property that will be illustrated in
the empirical part. Similarly, in the case ξ = 0, as σ → ∞,
pD-GPD(σ,ξ) (k)
2δ − 1
= σeδ/σ (1 − e−1/σ ) = 1 +
+ O(σ −2 )
fGPD(σ,ξ) (k + δ)
2σ
for every k = 0, 1, 2, . . ., and the fastest convergence to the unity is again found when
δ = 1/2.
The final result of this section accomplishes two things. Its first part is related to the
well-known invariance property of the generalized Pareto distribution: its residual lifetime
is again generalized Pareto distributed. More precisely, if Y has the generalized Pareto
distribution with scale parameter σ and shape parameter ξ ≥ 0, then the exceedance Y − u
has, given Y ≥ u, the generalized Pareto distribution with scale parameter σ+ξu and shape
parameter ξ, for all u ≥ 0. This invariance property is important when approximating
10
the exceedance distribution using generalized Pareto distributions because changing the
threshold does not alter the distributional assumptions used in the approximation. It is
easily checked that the D-GPD has an analogous property. Moreover, the D-GPD also
exhibits the property that discretizing a GPD using different types of rounding does not
affect the fact that a D-GPD is obtained, and the shape parameter remains invariant.
Proposition 2.7. Let Y follow a GPD with scale parameter σ > 0 and shape parameter
ξ ≥ 0. Let 0 < h ≤ 1. If X = bλY + 1 − hc, then for any integer u ≥ 1 − h, the distribution
of X −u | X ≥ u is a D-GPD with shape parameter λσ +ξ(u+h−1) and scale parameter ξ.
3
Empirical Results
0.0
0.5
1.0
1.5
2.0
2.5
3.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
●
Fitted Quantiles
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
2.0
1.5
1.0
0.0
0.5
Fitted Quantiles
2.5
3.0
We now assess the performance of the generalized Pareto distribution (GPD), the discrete
generalized Pareto distribution (D-GPD) and the generalized Zipf distribution (GZD) approximations for estimating extreme quantiles from several simulated and real data sets,
starting with a simple example to illustrate the methods.
●
Empirical Quantiles
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.5
1.0
1.5
2.0
2.5
3.0
Empirical Quantiles
Figure 1: Quantile-quantile plots for the fit of a GPD (left) and D-GPD (right) to X − 3 | X ≥ 3,
where X is Poisson distributed with rate 1 and the sample size is 5000. A good fit occurs on the
right because the points are contained between the two lines, but not on the left.
3.1
Simulated Data
Let X be Poisson distributed with rate λ = 1 and consider an i.i.d. sample of size n = 5000.
We are interested in inferring the distribution of the exceedances X − u | X ≥ u above a
11
large threshold u, say u = 3 which is the 95th empirical percentile of the data in this
case. Since X ∈ D-MDA0 , we can approximate the distribution of X − u | X ≥ u
by a D-GPD as explained in Section 1. Moreover, pX , the probability mass function
of X, also belongs to D-MDA0 and satisfies the additional assumption in Theorem 2.5 as
shown in Example 2.6, motivating the D-GPD and GZD approximations. However, X is
not in MDA and thus the GPD approximation does not necessarily apply. In order to
compare these three approximations, we fit a GPD, D-GPD and GZD to the observations
above u = 3, estimating their two parameters σ and ξ by maximum likelihood. Figure 1
compares the fitted quantiles of the GPD (left) and D-GPD (right) relative to the empirical
quantiles. Since the data are discrete, we applied a slightly different graphical method
−1
than the standard quantile-quantile (QQ) plots: FGPD
{i/(n + 1); σ, ξ}, for i = 1, . . . , n,
were plotted against empirical quantiles, where (σ, ξ) are the estimated parameters. A
good fit occurs when the points accumulate between the two diagonal lines in Figure 1
and touch the bottom line, as would be the case if the quantiles of a continuous random
variable Y were plotted against those of bY c. One clearly sees that the D-GPD fits well the
observations compared to the GPD which delivers a poor fit. We mention that the D-GPD
and GZD approximations produce very similar estimates in this case; the QQ-plot of the
GZD (not displayed here) looks visually identical to the one of the D-GPD. The estimated
scale parameters σ for the GPD, D-GPD and GZD are 1.91, 0.71 and 0.69 respectively.
Increasing λ would increase these estimates and render the three methods indistinguishable
as expected from Proposition 2.4.
We now compare the performances of the GPD, D-GPD and GZD approximations in
estimating the probability of a rare event in the following simulated case. Let
Y ∼ IG(2, 1),
X = bY c,
(14)
where IG(α, β) denotes an inverse-gamma distribution with probability density function
f (x) = Γ(α)−1 β α x−α−1 exp(−β/x), x > 0. We repeat 500 times the experiment described
below. An i.i.d. sample of size 8000 is drawn from the distribution of X. From these
observations, the goal is to estimate the probability of the extreme region
pe = P (X ≥ bqe c),
(15)
where qe is the 99.99 percentile of Y, i.e., the value exceeded once every 10 000 times
on average. The strategy pursued is to select an integer threshold u as the 95th empirical
percentile of the sample, fit a parametric distribution to the exceedances X −u | X ≥ u, and
use it to extrapolate the tail and estimate pe . It holds X ∈ D-MDA1/2 and X ∈ MDA1/2 ,
which motivates the choice of a GPD and D-GPD as seen in Section 1. We implement two
variants of the GPD approximation: the first has no continuity correction and the second
12
250
150
100
50
Frequency
200
100 150 200 250 300
0
0
50
Frequency
0
20
40
60
80
100
100
1000
10000
Word Frequency
Figure 2: On the left: frequency plot of a sample of X − 2 | X ≥ 2 of size 701 for X defined in
(14). On the right: frequency plot of the 5588 most frequent words in a British corpus (x axis on
log-scale).
shifts the observations by −δ = − 21 . It is not immediate if the probability mass function
of X is in D-MDA and we thus apply the GZD approximation heuristically.
As a benchmark, we will also estimate pe from a sample of the continuous variable Y
(as opposed to its discretization X). In this context, the GPD approximation is motivated
by the fact that Y ∈ MDA1/2 , and we thus fit a GPD to Y − u | Y ≥ u.
A frequency plot of a sample of X − u | X ≥ u is displayed in Figure 2 on the lefthand side. For each model, we compute maximum likelihood estimators for σ and ξ by
performing a two dimensional maximization using the function optim of R (R Core Team
[2015]) with starting values (1, 1). We then compute pe and approximate 90% confidence
intervals under asymptotic normality of the estimators. Table 1 displays: the average
parameters pe , ξ and σ over the 500 experiments, the average length of the confidence
intervals, the true length and their coverage. True length is the length the intervals should
have had to contain the estimates across the 500 experiments 90% of the time. Coverage
indicates the proportion of time the truth lies in the confidence interval.
It appears that the D-GPD and GZD approximations accurately estimate pe from the
discretized data with a coverage close to the correct one of 90%, and that their performance
is good relative to the situation of full information where the continuous data are available.
On the other hand, the GPD approximation with δ = 21 is inaccurate and the GPD
with δ = 0 performs very poorly (misleadingly, the latter has a larger likelihood at the
13
pe × 103
truth
Y −u|Y ≥u
GPD
X −u|X ≥u
D-GPD
GZD
GPDδ= 21
GPDδ=0
ξ
σ
mean (cov, len)
0.10
true length
mean (cov, len)
0.5
mean (len)
0.10 (87%, 0.16)
0.16
0.49 (95%, 0.22)
1.19 (0.30)
0.10 (86%, 0.17)
0.11 (88%, 0.17)
0.04 (20%, 0.07)
7.93 (83%, 23.97)
0.16
0.17
0.08
11.25
0.49 (95%, 0.23)
0.50 (95%, 0.24)
0.37 (22%, 0.18)
8.27 (0%, 1.24)
1.19 (0.33)
1.39 (0.29)
1.43 (0.32)
0.00 (0.00)
Table 1: Performance of several methods in estimating the probability pe of a rare event defined
in (15). A generalized Pareto distribution (GPD) is fitted to the observations of Y exceeding
u = 2, and the following distributions are fitted to exceedances of X = bY c: a discrete generalized
Pareto distribution (D-GPD), a generalized Zipf distribution (GZD) and a GPD with continuity
correction δ. The table displays average maximum likelihood estimators for pe , ξ and σ accross
500 experiments. Coverage (cov) and average length (len) of 90% confidence intervals are shown
between brackets. In each experiment, about 700 observations exceeded the threshold.
maximum likelihood estimate than the former). This illustrates why the D-GPD or GZD
approximations should be preferred to the GPD approximation in situations similar to this
simulated case.
We know that the D-GPD and GZD coincide with a geometric distribution when ξ = 0
and that they are asymptotically equivalent when ξ > 0 and σ → ∞ (Proposition 2.4).
Although σ is not particularly large in the above simulated case, the D-GPD and GZD
approximations deliver very similar results. We point out that they outperform the Poisson
and negative binomial distributions which would estimate pe very poorly.
We remark few differences between the GPD fitted to exceedances of Y and the D-GPD
fitted to exceedances of its discretization X, except that the former method yields slightly
shorter confidence intervals. As a consequence of the invariance property in Proposition 2.7,
the two methods would still give similar estimates of ξ if X would be defined by other forms
of rounding than X = bY c, such as using rounding or ceiling function.
In conclusion, we showed that the D-GPD and GZD approximations were able to estimate the probability of rare events in this example. These findings are supported by
two complementary simulated cases with Y ∈ MDAξ for ξ = 0 and ξ < 0 (Hitz [2016],
Chapter 2). The D-GPD provides a more efficient method because its probability mass,
survival and quantile functions are closed-form, but both distributions are useful to describe extreme values of discrete random variables, which we will further illustrate on real
data sets.
14
600
800
●
●
0
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
400
Fitted Quantiles
10000
●
200
50000
30000
●
●
0
Fitted Quantiles
●
0 10000
30000
50000
●●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
Empirical Quantiles
200
400
600
800
Empirical Quantiles
Figure 3: QQ-plots for a D-GPD fitted to the frequencies of the 5588 most frequent words in a
British corpus. On the right, only percentiles below 99% are plotted. Dashed lines denote pointwise
90% confidence intervals.
3.2
Word Frequencies and Word Lengths
We consider a data set based on the British National corpus counting written and spoken
English words (British National Corpus, [2007]) and are interested in modeling the frequency of the most popular words. The six most frequent words in the corpus are: “the”,
“of”, “and”, “a”, “in” and “to.” Let X be the frequency at which a word occurs. We select
a large threshold u and fit a GPD, a D-GPD and a GZD to X − u | X ≥ u by maximum likelihood. The GPD is implemented with continuity correction (δ = 12 ) and without
(δ = 0). Selecting an appropriate threshold is crucial when estimating high quantiles and
can be based on techniques such as mean residual plots (see e.g. Davison and Smith [1990]).
As our focus here is rather on describing the tail distribution of these word frequencies,
we choose a relatively low threshold exceeded by 5588 observations. A frequency plot of
X | X ≥ u is displayed on the right-hand side in Figure 2.
The D-GPD delivers a good fit as revealed by the QQ-plot in Figure 3; indeed, most
observations lie within the pointwise 90% confidence intervals obtained by simulating 2000
times from the fitted model and computing empirical quantiles at each simulation. Table 2
presents maximum likelihood estimators for σ and ξ with 90% confidence intervals. Apart
from the case of the GPD with δ = 0, all models provide similar results as expected from
Proposition 2.4 when σ is large. A Kolmogorov–Smirnov test for discrete data (Arnold
and Emerson [2011]) leads to the same conclusion (in order to perform the test for the
15
0
●
●
●
●
●
●
●
●
●
10
8
6
2
Fitted Quantiles
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
2
●
4
12
8
6
4
0
2
Fitted Quantiles
10
●
●
●
●
●
●
4
6
8
10
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
Empirical Quantiles
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
4
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
6
●
●
●
●
●
●
●
●
●
●
8
Empirical Quantiles
Figure 4: QQ-plots for a D-GPD (left) and GPD with δ = 12 (right) fitted to the length of the
2875 French words consisting of 15 letters or more. A good fit occurs when the lowest part of each
accumulation of points is close to the line.
GPD, which is a continuous distribution, we assumed that data were rounded realizations
of the fitted model). The analysis of frequencies of French words in a collection of books
and movie subtitles leads to analogous results (Hitz [2016], Chapter 2). Since the GZD
coincides with a Zipf–Mandelbrot distribution when ξ > 0, the above results are consistent
with a common hypothesis in linguistic that word frequencies follow a Zipf-type law (see
e.g. Booth [1967]).
We now consider the set of 150 000 words in the French lexical (New et al. [2004]) and
denote by X the length of a word. The longest French word, for instance, is “anticonstitutionnellement”, consisting of 25 letters. Contrary to the word frequencies, this data
set contains many tied observations and we want to see if this translates into a marked
difference between the methods. We thus fit the models to X − u | X ≥ u with u = 15
(the 98% empirical percentile of the data), leaving 2875 exceedances. The D-GPD and
GZD deliver a good fit and similar estimations between each other, and this time they
clearly outperform the GPD approximation with δ = 21 as shown by QQ-plots in Figure 4
and discrete Kolmogorov–Smirnov tests in Table 2 (p-values are here computed by Monte
Carlo simulation). Notice that the negative binomial also fits well the observations in this
case.
To summarize, we have illustrated the adequacy of the D-GPD and GZD in describing
the frequencies of the most common and longest words from large corpora. This data
analysis supports the conclusion drawn earlier from the simulated case: the D-GPD and
16
Model
p-value
Word Frequency
British
D-GPD
GZD
GPDδ= 12
GPDδ=0
Negative binomial
0.56
0.56
0.55
0.00
0.00
Word length
French
D-GPD
GZD
GPDδ= 12
Negative binomial
0.85
0.84
0.00
0.75
NLL
ξ
σ
27896.6
27896.6
0.88[0.84,0.93]
0.88[0.84,0.93]
0.88[0.84,0.93]
0.93[0.89,0.97]
22.38[21.42,23.34]
22.82[21.87,23.77]
22.39[21.43,23.34]
20.89[19.96,21.81]
0.02[−0.01,0.06]
0.02[−0.01,0.06]
−0.04[−0.06,−0.01]
1.36[1.30,1.43]
1.37[1.32,1.43]
1.51[1.45,1.57]
29663.2
3894.0
3894.0
3893.9
Table 2: Fit of a GPD, D-GPD and GZD to frequencies of the most frequent words in a British corpus and length of the longest French words. The table displays the p-value of discrete Kolmogorov–
Smirnov tests, negative log-likelihood (NLL) and maximum likelihood estimators with 90% confidence intervals. The sample sizes are 5588 for word frequencies and 2875 for word lengths.
GZD are preferred over the GPD to model extremes of discrete data when tied observations
are frequent.
3.3
Tornadoes
Accurately assessing the risk of environmental hazards is crucial for insurance companies in
particular, and we illustrate here how the D-GPD and GZD approximations can be useful
techniques for this purpose.
We consider the data set studied in Tippett et al. [2016] which reports the number of
extreme tornadoes per outbreak in the United States between 1965 and 2015. They defined
an “extreme outbreak” as a sequence of twelve or more tornadoes occurring close to one
another in time and that are rated F1 and greater on the Fujita scale, or EF1 and greater
on the enhanced Fujita scale [Fuhrmann et al., 2014].
Let X be the number of such tornadoes per extreme outbreak. The authors found that
observations from X − u | X ≥ u for u = 12 were well modeled by a GPD with linear
temporal trend in the scale parameter and continuity correction δ = 12 . The GPD, however,
is not a discrete distribution and the D-GPD and GZD seem appropriate choices for this
type of data. We fit these three distributions and find that they all achieve comparable
quality of fit as shown in Table 3; linear trend in σ leads to significative improvements in
17
each case.
Model
D-GPD
GPDδ= 12
NLL
1439.92
1439.93
ξ
0.27[0.16,0.37]
0.26[0.16,0.37]
σ0
4.81[3.64,5.99]
4.86[3.68,6.04]
σt
6.11[3.74,8.48]
6.13[3.75,8.50]
Table 3: Fit of a D-GPD and GPD (with linear trend in the scale parameter) to the number of
tornadoes per extreme tornado outbreak in the United States. Only the 435 outbreaks with more
than 12 tornadoes were considered. The table displays negative log-likelihood (of the discretized
model in the case of the GPD) and maximum likelihood estimates with 90% confidence intervals.
3.4
Multiple Births
We now turn our attention to a data set consisting of very small integer values to see if
the D-GPD and GZD can describe tail distributions in this special case. We examine data
counting multiple births in the United States from 1995 to 2014 (Hamilton et al. [2015]);
its frequency table reads
single
78 178 588
twin
2 500 340
triplet
117 603
quadruplet
8 108
quint. or more
1 353.
Let X be the number of children at birth. We fit a right-censored D-GPD, GZD,
negative binomial and Poisson distribution to X C − u | X C ≥ u for u = 1, where X C =
min(X, 5). As shown in the table below, which displays Bayesian information criteria (BIC)
and maximum likelihood estimates, the D-GPD and GZD outperform the Poisson and
negative binomial distributions, and are useful methods if one must estimate, for example,
the probability that an American women delivers sextuplets.
D-GPD
GZD
Negative binomial
Poisson
BIC
546 441.2
546 440.6
546 547
552 284
ξ
0.06[0.06,0.06]
0.06[0.06,0.07]
σ
0.30[0.30,0.30]
0.32[0.32,0.32]
The applicability of the D-GPD and GZD approximations for estimating tail distributions when the selected threshold u is a small integer should be more rigorously explored.
Future work could also assess the validity of the approximations in the case ξ < 0, and
compare them to a broader class of discrete distributions such as discrete compound Poisson distributions. Since the D-GPD and GZD delivered similar performances in the data
18
analysis carried out in this article, it would be interesting to further understand how they
relate to one another.
Acknowledgements The first author is grateful to the Berrow Foundation for financial support and would like to thank Robin Evans, Gesine Reinert, David Steinsaltz and
Jonathan Tawn for valuable advice and encouragement. This research was partially supported by the ARO grant W911NF-12-10385.
References
C. W. Anderson. Extreme value theory for a class of discrete distributions with applications to
some stochastic processes. Journal of Applied Probability, 7:99–113, 1970.
C. W. Anderson. Local limit theorems for the maxima of discrete random variables. Mathematical
Proceedings of the Cambridge Philosophical Society, 88(1):161–165, 1980.
B. C. Arnold. Pareto distribution. International Cooperative Publishing House, Maryland, 1983.
T. B. Arnold and J. W. Emerson. Nonparametric goodness-of-fit tests for discrete null distributions.
The R Journal, 3(2):34–39, 2011.
R. L. Axtell. Zipf distribution of US firm sizes. Science, 293:1818–1820, 2001.
N. H. Bingham, C. M. Goldie, and J. L. Teugles. Regular Variation. Cambridge University Press,
1989.
A. D. Booth. A law of occurrences for words of low frequency. Information and Control, 10(4):
386–393, 1967.
British National Corpus. Version 3 BNC XML edition, 2007.
A. Buddana and T. J. Kozubowski. Discrete Pareto distributions. Economic Quality Control, 29
(2):143–156, 2014.
A. Clauset, C. R. Shalizi, and M. E. J. Newman. Power-law distributions in empirical data. SIAM
Review, 51(4):661–703, 2009.
A. C. Davison and R. L. Smith. Models for exceedances over high thresholds. Journal of the Royal
Statistical Society. Series B. Methodological, 52(3):393–442, 1990.
P. Embrechts, C. Klüppelberg, and T. Mikosch. Modelling extremal events: for insurance and
finance. Springer, New York, 2013.
C. M. Fuhrmann, C. E. Konrad, M. M. Kovach, J. T. McLeod, W. G. Schmitz, and P. G. Dixon.
Ranking of tornado outbreaks across the United States and their climatological characteristics.
Weather and Forecasting, 29(3):684–701, 2014.
19
X. Gabaix. Zipf’s law and the growth of cities. The American Economic Review, 89(2):129–132,
1999.
B. E. Hamilton, J. A. Martin, M. J. Osterman, S. C. Curtin, and M. TJ. Births: Final data for
2014. National Vital Statistics Reports, 64(1), 2015.
A. S. Hitz. Modelling of Extremes. PhD thesis, University of Oxford, 2016.
T. J. Kozubowski, A. K. Panorska, and M. L. Forister. A discrete truncated Pareto distribution.
Statistical Methodology, 26:135–150, 2015.
H. Krishna and P. S. Pundir. Discrete Burr and discrete Pareto distributions. Statistical Methodology, 6(2):177–188, 2009.
B. Mandelbrot. Contribution à la théorie mathématique des jeux de communication. Publications
de l’Institut de statistique de l’Université de Paris, 1953.
B. New, C. Pallier, M. Brysbaert, and L. Ferrand. Lexique 2: A new French lexical database.
Behavior Research Methods, Instruments, & Computers, 36(3):516–524, 2004.
J. Pickands, III. Statistical inference using extreme order statistics. The Annals of Statistics, 3:
119–131, 1975.
F. Prieto, E. Gómez-Déniz, and J. M. Sarabia. Modelling road accident blackspots data with the
discrete generalized Pareto distribution. Accident Analysis & Prevention, 71:38–49, 2014.
R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for
Statistical Computing, 2015.
S. I. Resnick. Extreme values, regular variation, and point processes. Springer, New York, 1987.
T. Shimura. Discretization of distributions in the maximum domain of attraction. Extremes, 15(3):
299–317, 2012.
M. K. Tippett, C. Lepore, and J. E. Cohen. More tornadoes in the most extreme US tornado
outbreaks. Science, 354(6318):1419–1423, 2016.
20
| 10math.ST
|
Near-Optimal UGC-hardness of Approximating Max k-CSPR
Pasin Manurangsi, Preetum Nakkiran, and Luca Trevisan
arXiv:1511.06558v1 [cs.CC] 20 Nov 2015
University of California, Berkeley
{pasin,preetum,luca}@berkeley.edu
Abstract
In this paper, we prove an almost-optimal hardness for Max k-CSPR based on Khot’s Unique Games
Conjecture (UGC). In Max k-CSPR , we are given a set of predicates each of which depends on exactly
k variables. Each variable can take any value from 1, 2, . . . , R. The goal is to find an assignment to
variables that maximizes the number of satisfied predicates.
Assuming the Unique Games Conjecture, we show that it is NP-hard to approximate Max k-CSPR
to within factor 2O(k log k) (log R)k/2 /Rk−1 for any k, R. To the best of our knowledge, this result improves
on all the known hardness of approximation results when 3 ≤ k = o(log R/ log log R). In this case, the
previous best hardness result was NP-hardness of approximating within a factor O(k/Rk−2 ) by Chan.
When k = 2, our result matches the best known UGC-hardness result of Khot, Kindler, Mossel and
O’Donnell.
In addition, by extending an algorithm for Max 2-CSPR by Kindler, Kolla and Trevisan, we provide
an Ω(log R/Rk−1 )-approximation algorithm for Max k-CSPR . This algorithm implies that our inapproximability result is tight up to a factor of 2O(k log k) (log R)k/2−1 . In comparison, when 3 ≤ k is a constant,
the previously known gap was O(R), which is significantly larger than our gap of O(polylog R).
Finally, we show that we can replace the Unique Games Conjecture assumption with Khot’s d-to-1
Conjecture and still get asymptotically the same hardness of approximation.
1
1
Introduction
Maximum Constraint Satisfaction Problem (Max CSP) is an optimization problem where the inputs are
a set of variables, an alphabet set, and a set of predicates. Each variable can be assigned any alphabet
from the alphabet set and each predicate depends only on the assignment to a subset of variables. The
goal is to find an assignment to the variables that maximizes the number of satisfied predicates.
Many natural optimization problems, such as Max Cut, Max k-CUT and Max k-SAT, can be formulated as Max CSP. In addition, Max CSP has been shown to help approximate other seemingly-unrelated
problems such as Densest k-Subgraph [CHK11]. Due to this, Max CSP has long been researched by
the approximation algorithm community [Tre98, Has05, CMM09, MM14, KKT15, GM15]. Furthermore,
its relation to PCPs ensures that Max CSP is also well-studied on the hardness of approximation
side [ST00, Eng05, ST06, KKMO07, AM08, GR08, EH08, Cha13].
The main focus of this paper is on Max k-CSPR , a family of Max CSP where the alphabet set is of size
R and each predicate depends on only k variables. On the hardness of approximation side, most early
works focused on boolean Max k-CSP. Samorodnitsky
and Trevisan first showed that Max k-CSP2 is
√
NP-hard to approximate to within factor 2O( k) /2k [ST00]. Engebretsen and Holmerin
later improved
√
√
O( k)
k
constant factors in the exponent O( k) but still yielded hardness of a factor 2
/2 [EH08]. To break
this barrier, Samorodnitsky and Trevisan proved a hardness of approximation conditioned on Khot’s
Unique Games Conjecture (UGC), which will be discussed in more detail later; they achieved a ratio of
O(k/2k ) hardness, which is tight up to a constant for the boolean case [ST06]. Chan later showed that
NP-hardness of approximation at this ratio can be achieved unconditionally and, thus, settled down the
approximability of Max k-CSP2 [Cha13].
Unlike the boolean case, the approximability
of Max k-CSPR when R > 2 is still not resolved. In
√
this case, Engebretsen showed RO( k) /Rk NP-hardness of approximation [Eng05]. Under the Unique
Games Conjecture, a hardness of approximation of O(kR/Rk−1 ) factor was proven by Austrin and
Mossel [AM08] and, independently, by Guruswami and Raghavendra [GR08]. For the case k = 2, results
by Khot et al. [KKMO07] implicitly demonstrate UGC-hardness of approximation within O(log R/R),
made explicit in [KKT15]. Moreover, Austrin and Mossel proved UGC-hardness of approximation of
O(k/Rk−1 ) for infinitely many ks [AM08], but in the regime k ≥ R. Recently, Chan was able to remove
the Unique Game Conjecture assumption from these results [Cha13]. More specifically, Chan showed
NP-hardness of approximation of factor O(kR/Rk−1 ) for every k, R and that of factor O(k/Rk−1 ) for
every k ≥ R. Due to an approximation algorithm with matching approximation ratio by Makarychev
and Makarychev [MM14], Chan’s result established tight hardness of approximation for k ≥ R. On the
other hand, when k < R, Chan’s result gives O(kR/Rk−1 ) hardness of approximation whereas the best
known approximation algorithm achieves only Ω(k/Rk−1 ) approximation ratio [MM14, GM15]. In an
attempt to bridge this gap, we prove the following theorem.
Theorem 1 (Main Theorem) Assuming the Unique Games Conjecture, it is NP-hard to approximate
Max k-CSPR to within 2O(k log k) (log R)k/2 /Rk−1 factor, for any k ≥ 2 and any sufficiently large R.
When k = o(log R/ log log R), our result improves upon the previous best known hardness of approximation result in this regime, due to Chan. In particular, when k is constant, our results are tight up to
a factor of O(polylog R). While Chan’s results hold unconditionally, our result, similar to many of the
aforementioned results (e.g. [ST06, AM08, GR08]), rely on the Unique Games Conjecture.
A unique game is a Max 2-CSP instance where each constraint is a permutation. The Unique Games
Conjecture (UGC), first proposed by Khot in his seminal paper in 2002 [Kho02], states that, for any
sufficiently small η, γ > 0, it is NP-hard to distinguish a unique game where at least 1 − η fraction of constraints can be satisfied from a unique game where at most γ fraction of constraints can be satisfied. The
UGC has since made a huge impact in hardness of approximation; numerous hardness of approximation
results not known unconditionally can be derived assuming the UGC. More surprisingly, UGC-hardness
of approximation for various problems, such as Max Cut [KKMO07], Vertex Cover [KR08] and any
2
Range of k, R
k=2
3≤k<R
R≤k
Any k, R
NP-Hardness
√R
O log
R
k
O Rk−2
k
O Rk−1
–
UGC-Hardness
O logRR
–
–
O(k log k)
2
(log R)
Rk−1
k/2
Approximation
Ω logRR
k
Ω Rk−1
k
Ω Rk−1
log R
Ω R
k−1
References
[Cha13, KKMO07, KKT15]
[Cha13, MM14, GM15]
[Cha13, MM14]
this work
Figure 1: Comparison between our work and previous works. We list the previous best known results
alongside our results. From previous works, there is either an NP-hardness or a UGC-hardness result
matching the best known approximation algorithm in every case except when 3 ≤ k < R. Our hardness
result improves on the best known hardness result when k = o(log R/ log log R), and our approximation
algorithm improves on the previously known algorithm when k = o(log R).
Max CSP [Rag08]1 , are known to be tight. For more details on UGC and its implications, we refer
interested readers to Khot’s survey [Kho10] on the topic.
Another related conjecture from [Kho02] is the d-to-1 Conjecture. In the d-to-1 Conjecture, we consider
d-to-1 games instead of unique games. A d-to-1 game is an instance of Max 2-CSP where the constraint
graph is bipartite. Moreover, each constraint must be a d-to-1 function, i.e., for each assignment to a
variable on the right, there exists d assignments to the corresponding variable on the left that satisfy the
constraint. The d-to-1 Conjecture states that, for any sufficiently small γ > 0, it is NP-hard to distinguish
between a fully satisfiable d-to-1 game and a d-to-1 game where at most γ fraction of constraints can be
satisfied. Currently, it is unknown whether the d-to-1 Conjecture implies the Unique Games Conjecture
and vice versa.
While the d-to-1 Conjecture has yet to enjoy the same amount of influence as the UGC, it has been proven
successful in providing a basis for hardness of graph coloring problems [DMR09, DS10, GS11] and for
Max 3-CSP with perfect completeness [OW09b, Tan09]. Here we show that, by assuming the d-to-1
Conjecture instead of UGC, we can get a similar hardness of approximation result for Max k-CSPR as
stated below.
Theorem 2 Assuming the d-to-1 Games Conjecture holds for some d, it is NP-hard to approximate
Max k-CSPR to within 2O(k log k) (log R)k/2 /Rk−1 factor, for any k ≥ 2 and any sufficiently large R.
As mentioned earlier, there has also been a long line of works in approximation algorithms for Max
k-CSPR . In the boolean case, Trevisan first showed a polynomial-time approximation algorithm with
approximation ratio 2/2k [Tre98]. Hast later improved the ratio to Ω(k/(2k log k)) [Has05]. Charikar,
Makarychev and Makarychev then came up with an Ω(k/2k )-approximation algorithm [CMM09]. As
stated when discussing hardness of approximation of Max k-CSP2 , this approximation ratio is tight up
to a constant factor.
For the non-boolean case, Charikar, Makarychev, and Makarychev’s algorithm achieve Ω(k log R/Rk )
ratio for Max k-CSPR . Makarychev, and Makarychev later improved the approximation ratio to
Ω(k/Rk−1 ) when k = Ω(log R) [MM14]. Additionally, the algorithm was extended by Goldshlager and
Moshkovitz to achieve the same approximation ratio for any k, R [GM15]. On this front, we show the
following result.
Theorem 3 There exists a polynomial-time Ω(log R/Rk−1 )-approximation algorithm for Max k-CSPR .
In comparison to the previous algorithms, our algorithm gives better approximation ratio than all the
known algorithms when k = o(log R). We remark that our algorithm is just a simple extension of Kindler,
Kolla and Trevisan’s polynomial-time Ω(log R/R)-approximation algorithm for Max 2-CSPR [KKT15].
1 Raghavendra showed in [Rag08] that it is hard to approximate any Max CSP beyond what a certain type of semidefinite
program can achieve. However, determining the approximation ratio of a semidefinite program is still not an easy task. Typically,
one still needs to find an integrality gap for such a program in order to establish the approximation ratio.
3
1.1
Summary of Techniques
Our reduction from Unique Games to Max k-CSPR follows the reduction of [KKMO07] for Max 2-CSPs.
We construct a k-query PCP using a Unique-Label-Cover “outer verifier”, and then design a k-query Long
Code test as an “inner verifier”. For simplicity, let us think of k as a constant. We essentially construct a
k-query Dictator-vs.-Quasirandom test for functions f : [R]n → [R], with completeness Ω(1/(log R)k/2 )
and soundness O(1/Rk−1 ). Our test is structurally similar to the 2-query “noise stability” tests of
[KKMO07]: first we pick a random z ∈ [R]n , then we pick k weakly-correlated queries x(1) , . . . , x(k) by
choosing each x(i) ∈ [R]n as a noisy copy of z, i.e., each coordinate (x(i) )j is chosen as zj with some
probability ρ or uniformly at random otherwise. We accept iff f (x(1) ) = f (x(2) ) = · · · = f (x(k) ). The
key technical step is our analysis of the soundness of this test. We need to show that if f is a balanced
function with small low-degree influences, then the test passes with probability O(1/Rk−1 ). Intuitively,
we would like to say that for high enough noise, the values f (x(i) ) are roughly independent and uniform,
so the test passes with probability around 1/Rk−1 . To formalize this intuition, we utilize the Invariance
Principle and Hypercontractivity.
More precisely, if we let f i (x) : [R]n → {0, 1} be the indicator function for f (x) = i, then the test accepts
iff f i (x(1) ) = · · · = f i (x(k) ) = 1 for some i ∈ [R]. For each i ∈ [R], this probability can be written as
the expectation of the product: E[f i (x(1) )f i (x(2) ) . . . f i (x(k) )]. Since x(i) ’s are chosen as noisy copies of
z, this expression is related to the k-th norm of a noisy version of f i . Thus, our problem is reduced to
bounding the k-norm of a noisy function fei : [R]n → [0, 1], which has bounded one-norm E[fei ] = 1/R
since f is balanced. To arrive at this bound, we first apply the Invariance Principle, which essentially
states that a low-degree low-influence function on [R]n behaves on random input similarly to its “boolean
analog” over domain {±1}nR . Here “boolean analog” refers to the function over {±1}nR with matching
Fourier coefficients.
Roughly speaking, now that we have transfered to the boolean domain, we are left to bound the k-norm of
a noisy function on {±1}nR based on its one-norm. This follows from Hypercontractivity in the boolean
setting, which bounds a higher norm of any noisy function on boolean domain in terms of a lower norm.
The description above hides several technical complications. For example, when we pass from a function
[R]n → [0, 1] to its “boolean analog” {±1}nR → R, the range of the resulting function is no longer
bounded to [0, 1]. This, along with the necessary degree truncation, means we must be careful when
bounding norms. Further, Hypercontractivity only allows us to pass from k-norms to (1 + ε)-norms for
small ε, so we cannot use the known 1-norm directly. For details on how we handle these issues, see
Section 3. This allows us to prove soundness of our dictator test without passing through results on
Gaussian space, as was done to prove the “Majority is Stablest” conjecture [MOO10] at the core of the
[KKMO07] 2-query dictator test.
To extend our result to work with d-to-1 Games Conjecture in place of UGC, we observe that our proof
goes through even when we assume a conjecture weaker than the UGC, which we name the One-Sided
Unique Games Conjecture. The only difference between the original UGC and the One-Sided UGC is
that the completeness in UGC is allowed to be any constant smaller than one but the completeness is
a fixed constant for the One-Sided UGC. The conjecture is formalized as Conjecture 3. We show that
the d-to-1 Games Conjecture also implies the One-Sided UGC, which means that our inapproximability
result for Max k-CSPR also holds when we change our assumption to d-to-1 Games.
Lastly, for our approximation algorithm, we simply extend the Kindler, Kolla and Trevisan’s algorithm
by first creating an instance of Max 2-CSPR from Max k-CSPR by projecting each constraint to all
possible subsets of two variables. We then use their algorithm to approximate the instance. Finally, we
set our assignment to be the same as that from KKT algorithm with some probability. Otherwise, we pick
the assignment uniformly at random from [R]. As we shall show in Section 4, with the right probability,
this technique can extend not only the KKT algorithm but any algorithm for Max k′ -CSPR to an
algorithm for Max k-CSPR where k > k′ with some loss in the advantage over the naive randomized
algorithm.
4
1.2
Organization of the Paper
In Section 2, we define notations and list background knowledge that will be used throughout the paper.
Next, we prove our hardness of approximation results, i.e., Theorem 1 and Theorem 2, in Section 3. In
Section 4, we show how to extend Kindler et al.’s algorithm to Max k-CSPR and prove Theorem 3. We
also explicitly present the dictator test that is implicit in our hardness proof, in Section 5. Finally, in
Section 6, we discuss interesting open questions and directions for future works.
2
Preliminaries
In this section, we list notations and previous results that will be used to prove our results.
2.1
Max k-CSPR
We start by giving a formal definition of Max k-CSPR , which is the main focus of our paper.
Definition 1 (Max k-CSPR ) An instance (X , C) of (weighted) Max k-CSPR consists of
• A set X = {x1 , . . . , xn } of variables.
• A set C = {C1 , . . . ,P
Cm } of constraints. Each constraint Ci is a triple (Wi , Si , Pi ) of a positive weight
m
Wi > 0 such that
Wi = 1, a subset of variables Si ⊆ X , and a predicate Pi : [R]Si → {0, 1}
i=1
that maps each assignment to variables in Si to {0, 1}. Here we use [R]Si to denote the set of all
functions from Si to [R], i.e., [R]Si = {ψ : Si → [R]}.
For each assignment of variables ϕ
P: mX → [R], we define its value to be the total weights of the predicates
Wi Pi (ϕ |Si ). The goal is to find an assignment ϕ : X → [R] that
satisfied by this assignment, i.e.,
i=1
with maximum value. We sometimes call the optimum the value of (X , C).
Note that, while the standard definition of Max k-CSPR refers to the unweighted version where W1 =
· · · = Wm = 1/m, Crescenzi, Silvestri and Trevisan showed that the approximability of these two cases
are essentially the same [CST01].2 Hence, it is enough for us to consider just the weight version.
2.2
Unique Games and d-to-1 Conjectures
In this subsection, we give formal definitions for unique games, d-to-1 games and Khot’s conjectures
about them. First, we give a formal definition of unique games.
Definition 2 (Unique Game) A unique game (V, W, E, N, {πe }e∈E ) consists of
• A bipartite graph G = (V, W, E).
• Alphabet size N .
• For each edge e ∈ E, a permutation πe : [N ] → [N ].
For an assignment ϕ : V ∪ W → [N ], an edge e = (v, w) is satisfied if πe (ϕ(v)) = ϕ(w). The goal is to
find an assignment that satisfies as many edges as possible. We define the value of an instance to be the
fraction of edges satisfied in the optimum solution.
The Unique Games Conjecture states that it is NP-hard to distinguish an instance of value close one
from that of value almost zero. More formally, it can be stated as follows.
2 More specifically, they proved that, if the weighted version is NP-hard to approximate to within ratio r, then the unweighted
version is also NP-hard to approximate to within r − on (1) where on (1) represents a function such that on (1) → 0 as n → ∞.
5
Conjecture 1 (Unique Games Conjecture) For any sufficiently small η, γ, it is NP-hard to distinguish a unique game of value at least 1 − η from that of value at most γ.
Next, we define d-to-1 games, which is similar to unique games except that each constraint is a d-to-1
function instead of a permutation.
Definition 3 (d-to-1 Game) A d-to-1 game (V, W, E, N, {πe }e∈E ) consists of
• A bipartite graph G = (V, W, E).
• Alphabet size N .
• For each edge e ∈ E, a function πe : [N ] → [N/d] such that |πe−1 (σ)| = d for every σ ∈ [N/d].
Satisfiability of an edge, the goal, and an instance’s value of is defined similar to that of unique games.
In contrast to the UGC, d-to-1 Conjecture requires perfect completeness, i.e., it states that we cannot
distinguish even a full satisfiable d-to-1 game from one with almost zero value.
Conjecture 2 (d-to-1 Conjecture) For any sufficiently small γ, it is NP-hard to distinguish a d-to-1
game of value 1 from that of value at most γ.
2.3
Fourier Expansions
For any function g : [q]n → R over a finite alphabet [q], we define the Fourier expansion of g as follows.
Consider the space of all functions [q] → R, with the inner-product hu, vi := Ex∈[q] [u(x)v(x)], where the
expectation is over a uniform x ∈ [q]. Pick an orthonormal basis l1 , . . . , lq for this space li : Σ → R, such
that l1 is the constant function 1. We can now write g in the tensor-product basis, as
g(x1 , x2 , . . . , xn ) =
X
s∈[q]n
ĝ(s) ·
n
Y
ls(i) (xi ).
(1)
i=1
Since we pick l1 (x) = 1 for all x ∈ [q], we also have Ex∈[q] [li (x)] = hli , l1 i = 0 for every i 6= 1.
Throughout, we use ĝ to refer to the Fourier coefficients of a function g.
For functions g : [q]n → R, the p-norm is defined as
p 1/p
E [|g(x)| ] .
||g||p =
(2)
x∈[q]n
In particular, the squared 2-norm is
||g||22 =
2.3.1
2
E n [g(x) ] =
x∈[q]
X
ĝ(s)2 .
(3)
s∈[q]n
Noise Operators
ρ
For x ∈ [q]n , let y ← x denote that y is a ρ-correlated copy of x. That is, each coordinate yi is
independently chosen to be equal to xi with probability ρ, or chosen uniformly at random otherwise.
Define the noise operator Tρ acting on any function g : [q]n → R as
(Tρ g)(x) = E [g(y)].
(4)
ρ
y ←x
Notice that the noise operator Tρ acts on the Fourier coefficients on this basis as follows.
f (x) = Tρ g(x) =
X
s∈[q]n
ĝ(s) · ρ|s| ·
where |s| is defined as |{i | s(i) 6= 1}|.
6
n
Y
i=1
ls(i) (xi )
(5)
2.3.2
Degree Truncation
For any function g : [q]n → R, let g ≤d denote the (≤ d)-degree part of g, i.e.,
g
≤d
X
(x) =
s∈[q]n ,|s|≤d
ĝ(s) ·
n
Y
ls(i) (xi ),
(6)
i=1
and similarly let g >d : [q]n → R denote the (> k)-degree part of g, i.e.,
X
g >d (x) =
s∈[q]n ,|s|>d
ĝ(s) ·
n
Y
ls(i) (xi ).
(7)
i=1
Notice that degree-truncation commutes with the noise-operator, so writing Tρ g ≤d is unambiguous.
Also, notice that truncating does not increase 2-norm:
||g ≤d ||22 =
X
s∈[q]n ,|s|≤d
ĝ(s)2 ≤
X
s∈[q]n
ĝ(s)2 = ||g||22 .
(8)
We frequently use the fact that noisy functions have small high-degree contributions. That is, for any
function g : [q]n → [0, 1], we have
||Tρ g >d ||22 =
2.3.3
X
s∈[q]n ,|s|>d
ρ2|s| ĝ(s)2 ≤ ρ2d
X
s∈[q]n
ĝ(s)2 = ρ2d ||g||22 ≤ ρ2d .
(9)
Influences
For any function g : [q]n → R, the influence of coordinate i ∈ [n] is defined as
Infi [g] =
E [V arxi ∈[q] [g(x) | {xj }j6=i ]].
x∈[q]n
(10)
This can be expressed in terms of the Fourier coefficients of g as follows:
X
Infi [g] =
ĝ(s)2 .
(11)
s∈[q]n : s(i)6=1
Further, the degree-d influences are defined as
Infi≤d [g] = Infi [g ≤d ] =
X
ĝ(s)2 .
(12)
s∈[q]n :
|s|≤d, s(i)6=1
2.3.4
Binary Functions
The previous discussion of Fourier analysis can be applied to boolean functions, by specializing to q =
2. However, in this case the Fourier expansion can be written in a more convenient form. Let G :
{+1, −1}n → R be a boolean function over n bits. We can choose orthonormal basis functions l1 (y) = 1
and l2 (y) = y, so G can be written as
G(y) =
X
S⊆[n]
for some coefficients Ĝ(S).
7
Ĝ(S)
Y
i∈S
yi
(13)
Degree-truncation then results in
X
G≤d (y) =
Ĝ(S)
Y
yi ,
(14)
i∈S
S⊆[n]:|S|≤d
and the noise-operator acts as follows:
X
Tρ G(y) =
Ĝ(S)ρ|S|
yi .
(15)
i∈S
S⊆[n]
2.3.5
Y
Boolean Analogs
To analyze k-CSPR , we will want to translate between functions over [R]n to functions over {±1}nR .
The following notion of boolean analogs will be useful.
For any function g : [R]n → R with Fourier coefficients ĝ(s), define the boolean analog of g to be a
function {g} : {±1}n×R → R such that
{g}(z) =
Note that
X
s∈[R]n
||g||22 =
and that
ĝ(s) ·
X
s∈[R]n
Y
zi,s(i) .
(16)
i∈[n],s(i)6=1
ĝ(s)2 = ||{g}||22 ,
{g ≤d } = {g}≤d .
(17)
(18)
Moreover, the noise operator acts nicely on {g} as follows:
Tρ {g} = {Tρ g}.
(19)
For simplicity, we use Tρ to refer to both boolean and non-boolean noise operators with correlation ρ.
2.4
Invariance Principle and Mollification Lemma
We start with the Invariance Principle in the form of Theorem 3.18 in [MOO10]:
Theorem 4 (General Invariance Principle [MOO10]) Let f : [R]n → R be any function such that
V ar[f ] ≤ 1, Infi [f ] ≤ δ, and deg(f ) ≤ d. Let F : {±1}nR → R be its boolean analog: F = {f }. Consider
any “test-function” ψ : R → R that is C 3 , with bounded 3rd-derivative |ψ ′′′ | ≤ C everywhere. Then,
E
y∈{±1}nR
[ψ(F (y))] −
√
d d/2
δ.
E n [ψ(f (x))] ≤ C10 R
(20)
x∈[R]
Note that the above version follows directly
from Theorem 3.18 and Hypothesis 3 of [MOO10], and the
√
fact that uniform ±1 bits are (2, 3, 1/ 2)-hypercontractive as described in [MOO10].
As we shall see later, we will want to apply the Invariance Principle for some functions ψ that are not in
C 3 . However, these functions will be Lipschitz-continuous with some constant c ∈ R (or “c-Lipschitz”),
meaning that
|ψ(x + ∆) − ψ(x)| ≤ c|∆| for all x, ∆ ∈ R.
(21)
e that is that is C 3 ,
Therefore, similar to Lemma 3.21 in [MOO10], we can “smooth” it to get a function ψ
and has arbitrarily small pointwise difference to ψ.
Lemma 1 (Mollification Lemma [MOO10]) Let ψ : R → R be any c-Lipschitz function. Then for
e : R → R such that
any ζ > 0, there exists a function ψ
8
e ∈ C3,
• ψ
e − ψ||∞ ≤ ζ, and,
• ||ψ
e′′′ ||∞ ≤ Cc
e 3 /ζ 2 .
• ||ψ
e not depending on ζ, c.
For some universal constant C,
For completeness, the full proof of the lemma can be found in Appendix A.1.
Now we state the following version of the Invariance Principle, which will be more convenient to invoke.
It can be proved simply by just combining the two previous lemmas. We list a full proof in Appendix A.2.
Lemma 2 (Invariance Principle) Let ψ : R → R be one of the following functions:
1. ψ1 (t) := |t|,
2. ψk (t) :=
k
t
0
1
if t ∈ [0, 1],
if t < 0,
if t ≥ 1.
Let f : [R]n → [0, 1] be any function with all Infi≤d [f ] ≤ δ. Let F : {±1}nR → R be its boolean analog:
F = {f }. Let f ≤d : [R]n → R denote f truncated to degree d, and similarly for F ≤d : {±1}nR → R.
Then, for parameters d = 10k log R and δ = 1/(R10+100k log(R) ), we have
E
y∈{±1}nR
2.5
[ψ(F ≤d (y))] −
k
≤d
E [ψ(f (x))] ≤ O(1/R ).
(22)
x∈[R]n
Hypercontractivity Theorem
Another crucial ingredient in our proof is the hypercontractivity lemma, which says that, on {±1}n
domain, the operator Tρ smooths any function so well that the higher norm can be bound by the lower
norm of the original (unsmoothed) function. Here we use the version of the theorem as stated in [O’D14].
Theorem 5 (Hypercontractivity Theorem [O’D14]) Let 1 ≤ p ≤ q ≤ ∞. For any ρ ≤
n
for any function h : {±1} → R, the following inequality holds:
||Tρ h||q ≤ ||h||p .
p
In particular, for choice of parameter ρ = 1/
q
p−1
q−1
and
(23)
(k − 1) log R, we have
||T2ρ h||k ≤ ||h||1+ε .
(24)
where ε = 4/ log(R).
3
Inapproximability of Max k-CSPR
In this section, we prove Theorem 1 and Theorem 2. Before we do so, we first introduce a conjecture,
which we name One-Sided Unique Games Conjecture. The conjecture is similar to UGC except that the
completeness parameter ζ is fixed in contrast to UGC where the completeness can be any close to one.
Conjecture 3 (One-Sided Unique Games Conjecture) There exists ζ < 1 such that, for any sufficiently small γ, it is NP-hard to distinguish a unique game of value at least ζ from that of value at most
γ.
9
It is obvious that the UGC implies One-Sided UGC with ζ = 1 − η for any sufficiently small η. It is
also not hard to see that, by repeating each alphabet on the right d times and spreading each d-to-1
constraint out to be a permutation, d-to-1 Games Conjecture implies One-Sided UGC with ζ = 1/d. A
full proof of this can be found in Appendix B.
Since both UGC and d-to-1 Games Conjecture imply One-Sided UGC, it is enough for us to show the
following theorem, which implies both Theorem 1 and Theorem 2.
Theorem 6 Unless the One-Sided Unique Games Conjecture is false, for any k ≥ 2 and any sufficiently
large R, it is NP-hard to approximate Max k-CSPR to within 2O(k log k) (log R)k/2 /Rk−1 factor.
The theorem will be proved in Subsection 3.3. Before that, we first prove an inequality that is the heart
of our soundness analysis in Subsection 3.2.
3.1
Parameters
We use the following parameters throughout, which we list for convenience here:
p
• Correlation: ρ = 1/ (k − 1) log R
• Degree: d = 10k log R
• Low-degree influences: δ = 1/(R10+100k log(R) )
3.2
Hypercontractivity for Noisy Low-Influence Functions
Here we show a version of hypercontractivity for noisy low-influence functions over large domains. Although hypercontractivity does not hold in general for noisy functions over large domains, it turns out
to hold in our setting of high-noise and low-influences. The main technical idea is to use the Invariance
Principle to reduce functions over larger domains to boolean functions, then apply boolean hypercontractivity.
Lemma 3 (Main Lemma) Let g : [R]n → [0, 1] be any function with Ex∈[R]n [g(x)] = 1/R. Then, for
our choice of parameters ρ, d, δ: If Infi≤d [g] ≤ δ for all i, then
k
O(k)
/Rk .
E [(Tρ g(x)) ] ≤ 2
x∈[R]n
Before we present the full proof, we outline the high-level steps below. Let f = Tρ g, and define boolean
analogs G = {g}, and F = {f }. Let ψk : R → R be defined as in Lemma 2. Then,
k
≤d
E [f (x) ] ≈ E[ψk (f (x))]
x∈[R]n
(Lemma 2: Invariance Principle)
(Definition of ψk )
(Definition of F )
≈
(Invariance, etc.)
(Since E[|g|] = 1/R)
Proof.
10
[ψk (F ≤d (y))]
(26)
≤ ||F ≤d ||kk
(27)
=
(28)
=
(Hypercontractivity, for small ε)
E
y∈{±1}nR
(25)
≤
||Tρ G≤d ||kk
||T2ρ T1/2 G≤d ||kk
||T1/2 G≤d ||k1+ε
O(k)
(29)
(30)
≈2
||g||k1
(31)
=2
/R .
(32)
O(k)
k
To establish line (25), first notice that
ψk (f (x)) = ψk (f ≤d (x) + f >d (x)) ≤ ψk (f ≤d (x)) + k|f >d (x)|
(33)
where the last inequality is because the function ψk is k-Lipschitz.
Moreover, since g(x) ∈ [0, 1], we have f (x) ∈ [0, 1], so
f (x)k = ψk (f (x)).
(34)
Thus,
k
E[f (x) ] = E[ψk (f (x))]
≤ E[ψk (f
≤d
= E[ψk (f
≤d
≤ E[ψk (f
≤d
(35)
(x))] + k E[|f
>d
(x))] + k||f
>d
(x))] + k||f
>d
(x)|]
(36)
||1
(37)
||2 .
(38)
(39)
And we can bound the 2-norm of f >d , since f is noisy, we have
||f >d ||22 = ||Tρ g >d ||22 ≤ ρ2d ≤ O(1/R2k ).
(40)
The last inequality comes from our choice of ρ and d.
So line (25) is established:
k
≤d
k
E[f (x) ] ≤ E[ψk (f (x))] + O(k/R ).
(41)
Line (26) follows directly from our version of the Invariance Principle (Lemma 2), for the function ψk :
≤d
E [ψk (f (x))] ≤
x∈[R]n
E
y∈{±1}nR
[ψk (F ≤d (y))] + O(1/Rk ).
(42)
We can now rewrite Ey∈{±1}nR [ψk (F ≤d (y))] as
E
y∈{±1}nR
[ψk (F ≤d (y))] ≤
E
[|F ≤d (y)|k ]
(43)
y∈{±1}nR
= ||F ≤d ||kk
(44)
=
(45)
=
||Tρ G≤d ||kk
||T2ρ T1/2 G≤d ||kk .
(46)
(47)
Now, from the Hypercontractivity Theorem, Equation (24), we have
||T2ρ T1/2 G≤d ||k ≤ ||T1/2 G≤d ||1+ε
(48)
for ε = 4/ log R. This establishes line (30):
||T2ρ T1/2 G≤d ||kk ≤ ||T1/2 G≤d ||k1+ε = E[|T1/2 G≤d (y)|1+ε ]k/(1+ε) .
(49)
To show the remaining steps, we will apply the Invariance Principle once more. Notice that for all
t ∈ R : |t|1+ε ≤ |t| + t2 . Hence, we can derive the following bound:
≤d
E[|T1/2 G
(y)|1+ε ] ≤ E[|T1/2 G≤d (y)|] + E[(T1/2 G≤d (y))2 ]
(Matching Fourier expansion)
(Lemma 2, Invariance Principle)
≤d
= E[|T1/2 G
≤ E[|T1/2 g
11
≤d
(y)|] + E[(T1/2 g
(x)|] + E[(T1/2 g
≤d
≤d
(50)
2
(y)) ]
2
(51)
k
(x)) ] + O(1/R ).
(52)
Here we applied our Invariance Principle (Lemma 2) for the function ψ1 as defined in Lemma 2. We will
bound each of the expectations on the RHS, using the fact that g is balanced, and T1/2 g is noisy.
First,
≤d
>d
E[|T1/2 g (x)|] = E[|T1/2 g(x) − T1/2 g (x)|]
(Triangle Inequality)
≤ E[|T1/2 g(x)|] + E[|T1/2 g
= ||g||1 + ||T1/2 g
≤ ||g||1 + ||T1/2 g
(By our choice of d)
≤ 1/R + (1/2)
>d
>d
d
>d
(x)|]
(53)
(54)
||1
(55)
||2
(56)
= O(1/R).
(57)
(58)
Second,
E[(T1/2 g
≤d
X
(x))2 ] =
(1/2)2|s| ĝ(s)2
(59)
s∈[R]n ,|s|≤d
≤
(Since g ∈ [0, 1])
X
(1/2)2|s| ĝ(s)2
(60)
s∈[R]n
= E[(T1/2 g(x))2 ]
(61)
≤ E[T1/2 g(x)]
(62)
= E[g(x)] = 1/R.
(63)
Finally, plugging these bounds into (52), we find:
||T1/2 G≤d ||k1+ε = E[|T1/2 G≤d (y)|1+ε ]k/(1+ε)
(64)
≤ (O(1/R))
(65)
O(k)
/R
k/(1+ε)
(66)
≤2
/R
k(1−ε)
(67)
=2
k
/R .
(68)
=2
O(k)
(Recall ε = 4/ log R)
k/(1+ε)
O(k)
This completes the proof of the main lemma.
3.3
Reducing Unique Label Cover to Max k-CSPR
Here we reduce unique games to Max k-CSPR . We will construct a PCP verifier that reads k symbols
of the proof (with an alphabet of size R) with the following properties:
• (Completeness) If the unique game has value at least ζ, then the verifier accepts an honest proof
with probability at least c = 1/((log R)k/2 2O(k log k) ).
• (Soundness) If the unique game has value at most γ = 2O(k) δ 2 /(4dRk ), then the verifier accepts
any (potentially cheating) proof with probability at most s = 2O(k) /Rk−1 .
Since each symbol in the proof can be viewed as a variable and each accepting predicate of the verifier can
be viewed as a constraint of Max k-CSPR , assuming the One-sided UGC, this PCP implies NP-hardness
of approximating Max k-CSPR of factor s/c = 2O(k log k) (log R)k/2 /Rk−1 and, hence, establishes our
Theorem 6.
12
3.3.1
The PCP
Given a unique game (V, W, E, n, {πe }e∈E ), the proof is the truth-table of a function hw : [R]n → [R] for
each vertex w ∈ W . By folding, we can assume hw is balanced, i.e. hw takes on all elements of its range
with equal probability: Prx∈[R]n [hw (x) = i] = 1/R for all i ∈ [R]. 3
Notationally, for x ∈ [R]n , let (x ◦ π) denote permuting the coordinates of x as: (x ◦ π)i = xπ(i) . Also,
−1
for an edge e = (v, w), we write πe = πv,w , and define πw,v = πv,w
.
The verifier picks a uniformly random vertex v ∈ V , and k independent uniformly random neighbors of v:
w1 , w2 , . . . , wk ∈ W . Then pick z ∈ [R]n uniformly at random, and let x(1) , x(2) , . . . , x(k) be independent
ρ-correlated noisy copies of z (each coordinate xi chosen as equal to zi w.p. ρ, or uniformly at random
otherwise). The verifier accepts if and only if
hw1 (x(1) ◦ πw1 ,v ) = hw2 (x(2) ◦ πw2 ,v ) = · · · = hwk (x(k) ◦ πwk ,v ).
p
To achieve the desired hardness result, we pick ρ = 1/
3.3.2
(69)
(k − 1) log R.
Completeness Analysis
First, note that that we can assume without loss of generality that the graph is regular on V side.4 Let
the degree of each vertex in V be d.
Suppose that the original unique game has an assignment of value at least ζ. Let us call this assignment
ϕ. The honest proof defines hw at each vertex w ∈ W as the long code encoding of this assignment, i.e.,
hw (x) = xϕ(w) . We can written the verifier acceptance condition as follows:
The verifier accepts ⇔ hw1 (x(1) ◦ πw1 ,v ) = · · · = hwk (x(k) ◦ πwk ,v )
⇔ (x
⇔ (x
(1)
(1)
◦ πw1 ,v )ϕ(w1 ) = · · · = (x
)πw1 ,v (ϕ(w1 )) = · · · = (x
(k)
(k)
◦ πwk ,v )ϕ(wk )
)πwk ,v (ϕ(wk )) .
(70)
(71)
(72)
Observe that, if the edges (v, w1 ), . . . , (v, wk ) are satisfied by ϕ, then πw1 ,v (ϕ(w1 )) = · · · =
πwk ,v (ϕ(wk )) = ϕ(v). Hence, if the aforementioned edges are satisfied and x(1) , . . . , x(k) are not
perturbed at coordinate ϕ(v), then (x(1) )πw1 ,v (ϕ(w1 )) = · · · = (x(k) )πwk ,v (ϕ(wk )) .
For each u ∈ V , let su be the number of satisfied edges touching u. Since w1 , . . . , wk are chosen from the
neighbors of v independently from each other, the probability that the edges (v, w1 ), (v, w2 ), . . . , (v, wk )
are satisfied can be bounded as follows:
Pr
v,w1 ,...,wk
=
X
u∈V
=
X
u∈V
= E
u∈V
[(v, w1 ), . . . , (v, wk ) are satisfied]
(73)
Pr
(74)
w1 ,...,wk
[(v, w1 ), . . . , (v, wk ) are satisfied | v = u] Pr[v = u]
(su /d)k Pr[v = u]
(su /d)k
≥ E [su /d]k .
(75)
(76)
(77)
u∈V
3 More precisely, if the truth-table provided is of some function e
hw : [R]n → [R], we define the “folded” function hw as
hw (x1 , x2 , x3 , . . . xn ) := e
hw (x − (x1 , x1 , . . . , x1 )) + x1 , where the ± is over mod R. Notice that the folded hw is balanced, and
also that folding does not affect dictator functions. Thus we define our PCP in terms of hw , but simulate queries to hw using
the actual proof e
hw .
4 See, for instance, Lemma 3.4 in [KR08].
13
Notice that Eu∈V [su /d] is exactly the value of ϕ, which is at least ζ. As a result,
Pr
v,w1 ,...,wk
[(v, w1 ), . . . , (v, wk ) are satisfied] ≥ ζ k .
Furthermore, it is obvious that the probability that x1 , . . . , xk are not perturbedp
at the coordinate ϕ(v)
k
k k
is ρ . As a result, the PCP accepts with probability at least ζ ρ . When ρ = 1/ (k − 1) log R and ζ is
a constant not depending on k and R, the completeness is 1/((log R)k/2 2O(k log k) ).
3.3.3
Soundness Analysis
Suppose that the unique game has value at most γ = 2O(k) δ 2 /(4dRk ). We will show that the soundness
is 2O(k) /Rk−1 .
Suppose for the sake of contradiction that the probability that the verifier accepts Pr[accept] > t =
2Ω(k) /Rk−1 where Ω(·) hides some large enough constant.
ρ
Let hiw (x) : [R]n → {0, 1} be the indicator function for hw (x) = i and let x ← z denote that x is a
ρ-correlated copy of z. We have
Pr[accept] = Pr[hw1 (x(1) ◦ πw1 ,v ) = · · · = hwk (x(k) ◦ πwk ,v )]
=
X
Pr[i = hw1 (x
(1)
i∈[R]
=
X
i
E[hw1 (x
i∈[R]
(Since wi ’s are independent given v) =
X
i∈[R]
Now define gvi : [R]n → [0, 1] as
E
(1)
◦ πw1 ,v ) = · · · = hwk (x
(k)
(78)
◦ πwk ,v )]
◦ πw1 ,v ) · · · hiwk (x(k) ◦ πwk ,v )]
i
(1)
E [hw1 (x
w1
◦ πw1 ,v )] · · · E
wk
[hiwk (x(k)
(80)
◦ πwk ,v )] .
gvi (x) = E [hiw (x ◦ πw,v )]
(79)
(81)
(82)
w∼v
where w ∼ v denotes neighbors w of v.
We can rewrite Pr[accept] as follows:
Pr[accept] =
X
i
E[gv (x
i∈[R]
(Since x
(j)
’s are independent given z) =
X
i∈[R]
=
X
i∈[R]
= E
v
14
E
"
(1)
E
ρ
)gvi (x(2) ) · · · gvi (x(k) )]
[gvi (x)]k
x←z
#
i
k
E [(Tρ gv (z)) ]
(83)
(84)
(85)
v,z
X
i∈[R]
i
k
E[(Tρ gv (z)) .
z
(86)
Next, notice that
X
i∈[R]
i
k
E[(Tρ gv (z)) ] = E
z
z
X
(Tρ gvi (z))k
i∈[R]
≤ E
z
X
i∈[R]
= E Tρ
z
k
(87)
Tρ gvi (z)
(88)
(89)
X
gvi (z)
i∈[R]
= E[(Tρ 1)k ] = 1.
k
(90)
z
Therefore, since Pr[accept] > t, by (86), at least t/2 fraction of vertices v ∈ V have
X
i∈[R]
i
k
E[(Tρ gv (z)) ] ≥ t/2.
(91)
z
For these “good” vertices, there must exist some i ∈ [R] for which
i
k
E[(Tρ gv (z)) ] ≥ t/(2R).
(92)
k
Ω(k)
i
/Rk .
E[(Tρ gv (z)) ] > 2
(93)
z
Then for “good” v and i as above,
z
By Lemma 3 (Main Lemma), this means gvi has some coordinate j for which
Infj≤d [gvi ] > δ
(94)
for our choice of d, δ as defined in Subsection 3.1. Pick this j as the label of vertex v ∈ V .
Now to pick the label of a vertex w ∈ W , define the candidate labels as
Cand[w] = {j ∈ [n] : ∃ i ∈ [R] s.t. Infj≤d [hiw ] ≥ δ/2}.
(95)
Notice that
X
j∈[n]
Infj≤d [hiw ] =
X
s∈[R]n : |s|≤d
|s|ĥiw (s)2 ≤ d
X
s:|s|>0
ĥiw (s)2 = d V ar[hiw ] ≤ d.
(96)
So for each i ∈ [R], the projection hiw can have at most 2d/δ coordinates with influence ≥ δ/2. Therefore
the number of candidate labels is bounded:
|Cand[w]| ≤ 2dR/δ.
(97)
Now we argue that picking a random label in Cand[w] for w ∈ W is in expectation a good decoding. We
will show that if we assigned label j to a “good” v ∈ V , then πv,w (j) ∈ Cand[w] for a constant fraction
−1
of neighbors w ∼ v. Note here that πv,w = πw,v
.
First, since gvi (x) = Ew∼v [hiw (x ◦ πw,v )], the Fourier transform of gvi is related to the Fourier transform
of the long code labels hiw as
ĝvi (s) = E [ĥiw (s ◦ πw,v )].
w∼v
15
(98)
Hence, the influence Infj≤d [gvi ] of being large implies the expected influence Inf ≤d
−1
πv,w (j)
[hiw ] of its neighbor
labels w ∼ v is also large as formalized below.
δ < Infj≤k [gvi ] =
X
ĝvi (s)2
(99)
n
s∈[R]
|s|≤k,sj 6=1
=
≤
X
X
i
2
E [ĥw (s ◦ πw,v ) ]
(101)
X
n
s∈[R]
|s|≤k,sj 6=1
= E [
w∼v
(100)
w∼v
= E [
w∼v
i
2
E [ĥw (s ◦ πw,v )]
w∼v
X
ĥiw (s ◦ πw,v )2 ]
ĥiw (s)2 ]
(103)
ĥiw (s)2 ]
(104)
n
s∈[R]
|s|≤k,s −1
πw,v (j)
−1
(Since πv,w = πw,v
)= E [
w∼v
(102)
X
6=1
s∈[R]n
|s|≤k,sπv,w (j) 6=1
= E [Infπ≤d
[hiw ]]
v,w (j)
(105)
w∼v
[hiw ] ≥ δ/2, and so πv,w (j) ∈
Therefore, at least δ/2 fraction of neighbors w ∼ v must have Infπ≤d
v,w (j)
Cand[w] for at least δ/2 fraction of neighbors of “good” vertices v.
Finally, recall that at least (t/2) fraction of vertices v ∈ V are “good”. These vertices have at least (δ/2)
fraction of neighbors w ∈ W with high-influence labels and the matching label w ∈ W is picked with
probability at least δ/(2dR). Moreover, as stated earlier, we can assume that the graph is regular on V
side. Hence, the expected fraction of edges satisfied by this decoding is at least
(t/2)(δ/2)(δ/2dR) = tδ 2 /(4dR) = 2Ω(k) δ 2 /(4dRk ) > γ,
(106)
which contradicts our assumption that the unique game has value at most γ. Hence, we can conclude
that the soundness is at most 2O(k) /Rk−1 as desired.
4
Ω(log R/Rk−1)-Approximation Algorithm for Max k-CSPR
Instead of just extending the KKT algorithm to work with Max k-CSPs, we will show a more generalized statement that any algorithm that approximates Max CSPs with small arity can be extended
′
to approximate Max CSPs with larger arities. In particular, we show how to extend any f (R)/Rk ′
O(min{k′ ,k−k′ })
k
approximation algorithm for Max k -CSPR to an (f (R)/2
)/R -approximation algorithm
for Max k-CSPR where k > k′ .
Since the naive algorithm that assigns every variable randomly has an approximation ratio of 1/Rk , we
think of f (R) as the advantage of algorithm A over the randomized algorithm. From this perspective,
′
′
our extension lemma preserves the advantage up to a factor of 1/2O(min{k ,k−k }) .
The extension lemma and its proof are stated formally below.
Lemma 4 Suppose that there exists a polynomial-time approximation algorithm A for Max k′ -CSPR
′
that outputs an assignment with expected value at least f (R)/Rk times the optimum. For any k > k′ , we
can construct a polynomial-time approximation algorithm B for Max k-CSPR that outputs an assignment
′
′
with expected value at least (f (R)/2O(min{k ,k−k }) )/Rk times the optimum.
16
Proof. The main idea of theproof is simple. We turn an instance of Max k-CSPR to an instance of Max
′
k′ -CSPR by constructing kk′ Rk−k new constraints for each original constraint; each new constraint is
a projection of the original constraint to a subset of variables of size k′ . We then use A to solve the newly
constructed instance. Finally, B simply assigns each variable with the assignment from A with a certain
probability and assign it randomly otherwise.
For convenience, let α be
k−k′
.
k
We define B on input (X , C) as follows:
′
1. Create an instance (X , C ) of Max k′ -CSPR with the same variables and, for each C = (W, S, P ) ∈ C
′
′
and for every subset S ′ of S with |S ′ | = k′ and every τ ∈ [R]S−S , create a constraint C S ,τ =
′
′
′
′
′
′
S′
W
(W , S , P ) in C where W = k k−k′ and P : [R] → {0, 1} is defined by
(k ′ )R
P ′ (ψ) = P (ψ ◦ τ ).
Here ψ ◦ τ is defined as follows:
ψ ◦ τ (x) =
if x ∈ S ′ ,
otherwise.
ψ(x)
τ (x)
2. Run A on input (X , C ′ ). Denote the output of A by ϕA .
3. For each x ∈ X , with probability α, pick ϕB (x) randomly from [R]. Otherwise, let ϕB (x) be ϕA (x).
4. Output ϕB .
We now show that ϕB has expected value at least (f (R)/2O(min{k
′
,k−k′ })
)/Rk times the optimum.
′
First, observe that the optimum of (X , C ′ ) is at least 1/Rk−k times that of (X , C). To see that this is
true, consider any assignment ϕ : X → [R] and any constraint C = (W, S, P ). Its weighted contribution
in (X , C) is W P (ϕ|S ). On the other hand, k Wk−k′ P (ϕ|S ) appears kk′ times in (X , C ′ ), once for each
(k ′ )R
′
subset S ′ ⊆ S of size k′ . Hence, the value of ϕ with respect to (X , C ′ ) is at least 1/Rk−k times its value
with respect to (X , C)
′
Recall that the algorithm A gives an assignment of expected value at least f (R)/Rk times the optimum
of (X , C ′ ). Hence, the expected value of ϕA is at least f (R)/Rk times the optimum of (X , C).
Next, we will compute the expected value of ϕB (with respect to (X , C)). We start by computing the
expected value of ϕB with respect to a fixed constraint C = (W, S, P ) ∈ C, i.e., EϕB [W P (ϕB |S )]. For
each S ′ ⊆ S of size k, we define DS ′ as the event where, in step 3, ϕB (x) is assigned to be ϕA (x) for all
x ∈ S ′ and ϕB (x) is assigned randomly for all x ∈ S − S ′ .
Since DS ′ is disjoint for all S ′ ⊆ S of size k, we have the following inequality.
X
E [W P (ϕB |S )] ≥
ϕB
′
S ′ ⊆S
|S ′ |=k
′
Pr[DS ′ ] E [W P (ϕB |S ) | DS ′ ]
(107)
ϕB
′
(Since Pr[DS ′ ] = αk−k (1 − α)k ) = αk−k (1 − α)k
′
X
S ′ ⊆S
|S ′ |=k
W E [P (ϕB |S ) | DS ′ ]
(108)
ϕB
Moreover, since every vertex in S − S ′ is randomly assigned when DS ′ occurs, E[P (ϕB |S ) | DS ′ ] can be
′
view as the average value of P ((ϕA |S ′ ) ◦ τ ) over all τ ∈ [R]S−S . Hence, we can derive the following:
E [P (ϕB |S ) | DS ′ ] =
ϕB
1
Rk−k′
As a result, we have
k−k′
E [W P (ϕB |S )] ≥
ϕB
α
(1 − α)
Rk−k′
k′
E
ϕA
X
τ ∈[R]
X
E
ϕA
17
S−S ′
P ((ϕA |S ′ ) ◦ τ ) .
X
S ′ ⊆S τ ∈[R]S−S ′
|S ′ |=k
(109)
W P ((ϕA |S ′ ) ◦ τ ) .
(110)
By summing (110) over all constraints C ∈ C, we arrive at the following inequality.
E
ϕB
X
C=(W,S,P )∈C
k−k′
≥
=
=
α
(1 − α)
Rk−k′
W P (ϕB |S )
k′
E
ϕA
(111)
X
C=(W,S,P )∈C
X
′
′
k
αk−k (1 − α)k E
k′
ϕA
S ′ ⊆S τ ∈[R]S−S ′
|S ′ |=k
C=(W,S,P )∈C
X
′
′
k
αk−k (1 − α)k E
k′
ϕA
X
X
X
X
S ′ ⊆S τ ∈[R]S−S ′
|S ′ |=k
C ′ =(W ′ ,S ′ ,P ′ )∈C
k
k′
′
W
P ((ϕA |S ′ ) ◦ τ )
Rk−k′
W P (ϕA |S ′ )
′
(112)
W P ((ϕA |S ′ ) ◦ τ )
(113)
(114)
′
′
The first expression is the expected value of ϕB whereas the last is kk′ αk−k (1 − α)k times the expected
value of ϕA . Since the expected value of ϕA is at least f (R)/Rk times the optimum of (X , C), the
′
′
expected value of ϕB is at least ( kk′ αk−k (1 − α)k )(f (R)/Rk ) times the optimum of (X , C).
Finally, we substitute α =
k−k′
k
in to get
k − k′
k
k−k′
k′
′
′
k
αk−k (1 − α)k =
k′
Let l = min{k′ , k − k′ }. We then have
k
k′
k − k′
k
k
k′
k′
k
k−k′
=
k′
k
k
l
k′
k−l
k
l
.
k−l l
k−l
k
l
k
k−l k
≥
k
≥
l
k
k−l l
= (1 − l/k)2k/l
2l
(From Bernoulli’s inequality and from l ≤ k/2) ≥ 1/2 .
(115)
l
k
2l
(116)
(117)
(118)
(119)
(120)
Hence, ϕB has expected value at least (f (R)/2O(l) )/Rk times the optimum of (X , C), which completes
the proof of this lemma.
Finally, Theorem 3 is an immediate consequence of applying Lemma 4 to the algorithm from [KKT15]
with k′ = 2 and f (R) = Ω(R log R).
5
k-Query Large Alphabet Dictator Test
We remark that the results of Section 3 also implicitly yield a k-query nonadaptive Dictator-vs.Quasirandom test for functions over large alphabet.
A Dictator-vs.-Quasirandom test aims to
distinguish dictator functions from functions with small low-degree influences (“quasirandom”). This
concept was essentially introduced in [Hås96], and we borrow the “quasirandom” terminology from
[OW09a] (adapted here for functions over non-binary alphabets). Specifically, we have the following
test:
18
i
Theorem 7 For any function f : [R]n → [R], and any i ∈ [R], let fp
: [R]n → {0, 1} denote the
indicator function for f (x) = i. For any k, R ≥ 2, set parameters ρ = 1/ (k − 1) log R, d = 10k log R,
and δ = 1/(R10+100k log(R) ). Then there exists a k-query nonadaptive Dictator-vs.Quasirandom test with
the following guarantees:
• (Completeness) If f is a dictator, i.e. f (x) = xj for some coordinate j ∈ [n], then the test passes
with probability at least
ρk = 1/((log R)k/2 2O(k log k) )
• (Soundness) If f has Infj≤d [f i ] ≤ δ for all coordinates j ∈ [n] and all projections i ∈ [R], then
the test passes with probability at most
2O(k) /Rk−1
Notice that if we assume f is balanced, then this theorem is immediately implied by the techniques of
Section 3. However, to extend this to general functions via “folding”, we must technically show that
the operation of folding keeps low-influence functions as low-influence. The full proof can be found in
Appendix C.
6
Conclusions and Open Questions
We conclude by posting interesting open questions regarding the approximability of Max k-CSPR and
providing our opinions on each question. First, as stated earlier, even with our results, current inapproximability results do not match the best known approximation ratio achievable in polynomial time when
3 ≤ k < R. Hence, it is intriguing to ask what the right ratio that Max k-CSPR becomes NP-hard
to approximate is. Since our hardness factor 2O(k log k) (log R)k/2 /Rk−1 does not match Chan’s hardness
factor O(k/Rk−2 ) when k = R, it is likely that there is a k between 3 and R − 1 such that a drastic
change in the hardness factor, and technique that yields that factor, occurs.
Moreover, since our PCP has completeness of 1/(2O(k) (log R)k/2 ), even if one cannot improve on the
inapproximability factor, it is still interesting if one can come up with a hardness result with almost
perfect completeness. √In fact, even for k = 2, there is no known hardness of approximation of factor
better than O(log R/ R) with near perfect completeness whereas the best UGC-hardness known is
O(log R/R).
It is also interesting to try to relax assumptions for other known inapproximability results from UGC
to the One-Sided UGC. Since the One-Sided UGC is implied by d-to-1 Games Conjecture, doing so will
imply inapproximability results based on the d-to-1 Games Conjecture. Moreover, without going into too
much detail, we remark that most attempts to refute the UGC and the d-to-1 Conjecture need the value
of the game to be high [ABS15, CMM06a, CMM06b, GT06, Kho02, Kol11, Tre08]. Hence, these algorithms are not candidates to refute the One-Sided UGC. In addition, Arora, Barak and Steurer’s [ABS15]
subexponential time algorithm for unique games suggest that unique games have intermediate complexity,
meaning that, even if the UGC is true, the UGC-hardness would not imply exponential time lower bounds.
On the other hand, to the best of the authors’ knowledge, the ABS algorithm does not run in subexponential time when the completeness is small. Hence, the One-Sided UGC may require exponential time
to solve, which could give similar running time lower bounds for the resulting hardness of approximation
results. Finally, there are evidences suggesting that relaxing completeness or soundness conditions of a
conjecture can make it easier; the most relevant such result is that from Feige and Reichman who proved
that, if one only cares about the approximation ratio and not completeness and soundness, then unique
game is hard to approximate to within factor ε for any ε > 0 [FR04].
References
[ABS15]
Sanjeev Arora, Boaz Barak, and David Steurer. Subexponential algorithms for unique games
and related problems. J. ACM, 62(5):42, 2015.
19
[AM08]
P. Austrin and E. Mossel. Approximation resistant predicates from pairwise independence.
In Computational Complexity, 2008. CCC ’08. 23rd Annual IEEE Conference on, pages
249–258, June 2008.
[Cha13]
Siu On Chan. Approximation resistance from pairwise independent subgroups. In Proceedings
of the Forty-fifth Annual ACM Symposium on Theory of Computing, STOC ’13, pages 447–
456, New York, NY, USA, 2013. ACM.
[CHK11]
Moses Charikar, MohammadTaghi Hajiaghayi, and Howard Karloff. Improved approximation
algorithms for label cover problems. Algorithmica, 61(1):190–206, 2011.
[CMM06a] Moses Charikar, Konstantin Makarychev, and Yury Makarychev. Near-optimal algorithms
for unique games. In Proceedings of the Thirty-eighth Annual ACM Symposium on Theory
of Computing, STOC ’06, pages 205–214, New York, NY, USA, 2006. ACM.
[CMM06b] Eden Chlamtac, Konstantin Makarychev, and Yury Makarychev. How to play unique games
using embeddings. In 47th Annual IEEE Symposium on Foundations of Computer Science
(FOCS 2006), 21-24 October 2006, Berkeley, California, USA, Proceedings, pages 687–696,
2006.
[CMM09]
Moses Charikar, Konstantin Makarychev, and Yury Makarychev. Near-optimal algorithms
for maximum constraint satisfaction problems. ACM Transactions on Algorithms, 5(3), 2009.
[CST01]
Pierluigi Crescenzi, Riccardo Silvestri, and Luca Trevisan. On weighted vs unweighted versions of combinatorial optimization problems. Information and Computation, 167(1):10–26,
2001.
[DMR09]
Irit Dinur, Elchanan Mossel, and Oded Regev. Conditional hardness for approximate coloring.
SIAM Journal on Computing, 39(3):843–873, 2009.
[DS10]
Irit Dinur and Igor Shinkar. On the conditional hardness of coloring a 4-colorable graph
with super-constant number of colors. In Proceedings of the 13th International Conference on
Approximation, and 14 the International Conference on Randomization, and Combinatorial
Optimization: Algorithms and Techniques, APPROX/RANDOM’10, pages 138–151, Berlin,
Heidelberg, 2010. Springer-Verlag.
[EH08]
Lars Engebretsen and Jonas Holmerin. More efficient queries in pcps for np and improved
approximation hardness of maximum csp. Random Struct. Algorithms, 33(4):497–514, December 2008.
[Eng05]
Lars Engebretsen. The nonapproximability of non-boolean predicates. SIAM J. Discret.
Math., 18(1):114–129, January 2005.
[FR04]
Uriel Feige and Daniel Reichman. On systems of linear equations with two variables per
equation. In Klaus Jansen, Sanjeev Khanna, JoséD.P. Rolim, and Dana Ron, editors, Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques,
volume 3122 of Lecture Notes in Computer Science, pages 117–127. Springer Berlin Heidelberg, 2004.
[GM15]
Gil Goldshlager and Dana Moshkovitz.
Approximating kCSP for large alphabets.
https://people.csail.mit.edu/dmoshkov/papers/Approximating%20MAX%20kCSP.pdf,
2015.
[GR08]
Venkatesan Guruswami and Prasad Raghavendra. Constraint satisfaction over a non-boolean
domain: Approximation algorithms and unique-games hardness. In Ashish Goel, Klaus
Jansen, JoséD.P. Rolim, and Ronitt Rubinfeld, editors, Approximation, Randomization and
Combinatorial Optimization. Algorithms and Techniques, volume 5171 of Lecture Notes in
Computer Science, pages 77–90. Springer Berlin Heidelberg, 2008.
[GS11]
Venkatesan Guruswami and Ali Kemal Sinop. The complexity of finding independent sets
in bounded degree (hyper)graphs of low chromatic number. In Proceedings of the TwentySecond Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2011, San Francisco,
California, USA, January 23-25, 2011, pages 1615–1626, 2011.
[GT06]
Anupam Gupta and Kunal Talwar. Approximating unique games. In Proceedings of the
Seventeenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2006, Miami,
Florida, USA, January 22-26, 2006, pages 99–106, 2006.
20
[Hås96]
Johan Håstad. Clique is hard to approximate within n1−ε . In Foundations of Computer
Science, 1996. Proceedings., 37th Annual Symposium on, pages 627–636. IEEE, 1996.
[Has05]
Gustav Hast. Approximating Max kCSP - outperforming a random assignment with almost
a linear factor. In Proceedings of the 32Nd International Conference on Automata, Languages
and Programming, ICALP’05, pages 956–968, Berlin, Heidelberg, 2005. Springer-Verlag.
[IP01]
Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-sat. J. Comput. Syst.
Sci., 62(2):367–375, 2001.
[Kho02]
Subhash Khot. On the power of unique 2-prover 1-round games. In Proceedings of the Thiryfourth Annual ACM Symposium on Theory of Computing, STOC ’02, pages 767–775, New
York, NY, USA, 2002. ACM.
[Kho10]
Subhash Khot. On the unique games conjecture (invited survey). In Proceedings of the
25th Annual IEEE Conference on Computational Complexity, CCC 2010, Cambridge, Massachusetts, June 9-12, 2010, pages 99–121, 2010.
[KKMO07] Subhash Khot, Guy Kindler, Elchanan Mossel, and Ryan O’Donnell. Optimal inapproximability results for MAX-CUT and other 2-variable csps? SIAM J. Comput., 37(1):319–357,
2007.
[KKT15]
Guy Kindler, Alexandra Kolla, and Luca Trevisan. Approximation of non-boolean 2csp.
CoRR, abs/1504.00681, 2015.
[Kol11]
Alexandra Kolla. Spectral algorithms for unique games.
20(2):177–206, 2011.
[KR08]
Subhash Khot and Oded Regev. Vertex cover might be hard to approximate to within 2-ε.
J. Comput. Syst. Sci., 74(3):335–349, May 2008.
[MM14]
Konstantin Makarychev and Yury Makarychev. Approximation algorithm for non-boolean
max-k-csp. Theory of Computing, 10:341–358, 2014.
[MOO10]
Elchanan Mossel, Ryan O’Donnell, and Krzysztof Oleszkiewicz. Noise stability of functions
with low influences: invariance and optimality. Ann. Math. (2), 171(1):295–341, 2010.
[O’D14]
Ryan O’Donnell. Analysis of Boolean Functions. Cambridge University Press, 2014.
[OW09a]
Ryan O’Donnell and Yi Wu. 3-bit dictator testing: 1 vs. 5/8. In Proceedings of the Twentieth
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2009, New York, NY, USA,
January 4-6, 2009, pages 365–373, 2009.
[OW09b]
Ryan O’Donnell and Yi Wu. Conditional hardness for satisfiable 3-csps. In Proceedings of the
41st Annual ACM Symposium on Theory of Computing, STOC 2009, Bethesda, MD, USA,
May 31 - June 2, 2009, pages 493–502, 2009.
[Rag08]
Prasad Raghavendra. Optimal algorithms and inapproximability results for every csp? In
Proceedings of the Fortieth Annual ACM Symposium on Theory of Computing, STOC ’08,
pages 245–254, New York, NY, USA, 2008. ACM.
[ST00]
Alex Samorodnitsky and Luca Trevisan. A PCP characterization of NP with optimal amortized query complexity. In Proceedings of the Thirty-Second Annual ACM Symposium on
Theory of Computing, May 21-23, 2000, Portland, OR, USA, pages 191–199, 2000.
[ST06]
Alex Samorodnitsky and Luca Trevisan. Gowers uniformity, influence of variables, and pcps.
In Proceedings of the Thirty-eighth Annual ACM Symposium on Theory of Computing, STOC
’06, pages 11–20, New York, NY, USA, 2006. ACM.
[Tan09]
Linqing Tang. Conditional hardness of approximating satisfiable max 3csp-q. In Yingfei
Dong, Ding-Zhu Du, and Oscar Ibarra, editors, Algorithms and Computation, volume 5878
of Lecture Notes in Computer Science, pages 923–932. Springer Berlin Heidelberg, 2009.
[Tre98]
Luca Trevisan. Parallel approximation algorithms by positive linear programming. Algorithmica, 21(1):72–88, 1998.
[Tre08]
Luca Trevisan. Approximation algorithms for unique games. Theory of Computing, 4(1):111–
128, 2008.
21
Computational Complexity,
A
Proofs of Preliminary Results
For completeness, we prove some of the preliminary results, whose formal proofs were not found in the
literature by the authors.
A.1
Mollification Lemma
Below is the proof of the Mollification Lemma. We remark that, while its main idea is explained
in [O’D14], the full proof is not shown there. Hence, we provide the proof here for completeness.
Proof. (of Lemma 1) Let p : R → R be a C 4 function supported only on [−1, +1], such that p(y) forms
2
a probability distribution. (For example, an appropriately normalized version of e−1/(1+y ) for |y| ≤ 1).
Define pλ (y) to be re-scaled to have support [−λ, +λ] for some λ > 0:
pλ (y) := (1/λ)p(y/λ).
(121)
Let Yλ be a random variable with distribution pλ (y), supported on [−λ, +λ]. We will set λ = ζ/c.
Now, define
e := E [ψ(x + Yλ )].
ψ
(122)
Yλ
This is pointwise close to ψ, since ψ is c-Lipschitz:
e − ψ(x)| = | E [ψ(x + Yλ ) − ψ(x)]| ≤ E [|ψ(x + Yλ ) − ψ(x)|] ≤ E [c|Yλ |] ≤ cλ = ζ.
|ψ(x)
(123)
Yλ
Yλ
Yλ
e is C 3 , because ψ(x)
e can be written as a convolution:
Further, ψ
(124)
e′′′ is bounded, for a fixed x ∈ R, define the constant z := ψ(x). Then,
To see that ψ
(125)
e = (ψ ∗ pλ )(x) =⇒ ψe′′′ = (ψ ∗ pλ )′′′ = (ψ ∗ p′′′
ψ(x)
λ ).
e′′′ (x)| = |(ψ ∗ p′′′
|ψ
λ )(x)|
′
(z is constant, so z = 0)
= |(ψ ∗
= |(ψ ∗
p′′′
λ
p′′′
λ
−
−
= |((ψ − z) ∗
=
=
≤
(c-Lipschitz)
Z
Z
Z
′
z ∗ p′′λ )(x)|
z ∗ p′′′
λ )(x)|
′′′
pλ )(x)|
(126)
(127)
(128)
+∞
−∞
+∞
−∞
+λ
−λ
p′′′
λ (y)(ψ(x − y) − z)dy
(129)
p′′′
λ (y)(ψ(x − y) − ψ(x))dy
(130)
|p′′′
λ (y)||ψ(x − y) − ψ(x)|dy
≤ ||p′′′
λ ||∞
Z
+λ
−λ
2
|cy|dy
= ||p′′′
λ ||∞ cλ .
e := ||p′′′ ||∞ . We have
Define the universal constant C
4 e
4 ′′′
′′′
p′′′
λ (y) = (1/λ )p (y/λ) =⇒ ||pλ ||∞ ≤ (1/λ )C.
e′′′ (x)| ≤ Cc
e 3 /ζ 2 , which completes the proof of Lemma 1.
With our choice of λ = ζ/c, this yields |ψ
22
(131)
(132)
(133)
(134)
A.2
Proof of Lemma 2
Below we show the proof of Lemma 2.
e by applying Lemma 1 for
Proof. First, we “mollify” the function ψ to construct a C 3 function ψ,
k
ζ = 1/R . Notice that both choices of ψ are k-Lipschitz. Therefore the Mollification Lemma guarantees
e′′′ (x)| ≤ Ck
e 3 R2k for some universal constant C.
e
that |ψ
e is pointwise close to ψ, with deviation at most 1/Rk , we have
Since ψ
E
y∈{±1}nR
[ψ(F ≤d (y))] −
≤d
E [ψ(f (x))] ≤
x∈[R]n
E
y∈{±1}nR
e ≤d (y))] −
[ψ(F
e ≤d (x))] + O(1/Rk ).
E [ψ(f
x∈[R]n
e we have
Applying the General Invariance Principle (Theorem 4) with the function ψ,
E
y∈{±1}nR
e ≤d (y))] −
[ψ(F
√
e ≤d (x))] ≤ Ck
e 3 R2k 10d Rd/2 δ.
E n [ψ(f
(135)
(136)
x∈[R]
By our choice of parameters d, δ, this is O(1/Rk ).
B d-to-1 Games Conjecture implies One-Sided Unique
Games Conjecture
In this section, we prove that if d-to-1 Games Conjecture is true, then so is One-Sided Unique Games
Conjecture.
Lemma 5 For every d ∈ N, d-to-1 Games Conjecture implies One-Sided UGC.
Proof. Suppose that d-to-1 Games Conjecture is true for some d ∈ N. We will prove One-Sided UGC;
more specifically, ζ in the One-Sided UGC is 1/d. The reduction from a d-to-1 game (V, W, E, N, {πe }e∈E )
to a unique game (V ′ , W ′ , E ′ , N ′ , {πe′ }e∈E ) can be described as follows:
• Let V ′ = V, W ′ = W, E ′ = E, and N ′ = N
• We define πe′ as follows. For each θ ∈ [N/d], let the elements of πe−1 (θ) be σ1 , σ2 , . . . , σd ∈ [N ]. We
then define πe′ (σi ) = d(θ − 1) + i.
Now, we will prove the soundness and completeness of this reduction.
(Completeness) Suppose that the d-to-1 game is satisfiable. Let ϕ : V ∪ W → [N ] be the assignment
that satisfies every constraint in the d-to-1 game. We define ϕ′ : V ′ ∪ W ′ → [N ′ ] by first assign
ϕ′ (v) = ϕ(v) for every v ∈ V . Then, for each w ∈ W , pick ϕ′ (w) to be an assignment that satisfies as
many edges touching w in the unique game as possible, i.e., for a fixed w, ϕ′ (w) is select to maximize
′
|{v ∈ N (w) | π(v,w)
(ϕ(v)) = ϕ′ (w)}| where N (w) is the set of neighbors of w. From how ϕ′ (w) is picked,
we have
d
′
|{v ∈ N (w) | π(v,w)
(ϕ(v)) = ϕ′ (w)}| ≥
1X
′
|{v ∈ N (w) | π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i}|.
d
(137)
i=1
′
′
Let 1[π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i] be the indicating variable whether π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i,
we can rewrite the right hand side as follows:
23
d
1X
′
|{v ∈ N (w) | π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i}|
d
(138)
i=1
=
d
1X X
′
1[π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i]
d
(139)
i=1 v∈N(w)
=
d
1 X X
′
1[π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i].
d
(140)
v∈N(w) i=1
′
′
is defined and since π(v,w) (ϕ(v)) = ϕ(w), there exists i ∈ [d] such that π(v,w)
(ϕ(v)) =
From how π(v,w)
d(ϕ(w) − 1) + i. As a result, we have
d
|N (w)|
1 X X
1 X
′
1=
.
1[π(v,w)
(ϕ(v)) = d(ϕ(w) − 1) + i] ≥
d
d
d
v∈N(w) i=1
(141)
v∈N(w)
In other words, at least 1/d fraction of edges touching w is satisfied in the unique game for every w ∈ W .
Hence, ϕ′ has value at least 1/d, which means that the unique game also has value at least 1/d.
(Soundness) Suppose that the value of the d-to-1 game is at most γ. For any assignment ϕ′ : V ′ ∪ W ′ →
[N ′ ] to the unique game, we can define an assignment ϕ : V ∪ W → [N ] by
ϕ(u) =
ϕ′ (u)
⌊(ϕ′ (u) − 1)/d⌋ + 1
if u ∈ V,
if u ∈ W.
(142)
From how πe′ is defined, it is easy to see that, if πe′ (ϕ′ (v)) = ϕ′ (w), then πe (ϕ(v)) = ϕ(w). In other
words, the value of ϕ′ with respect to the unique game is no more than the value of ϕ with respect to
the d-to-1 game. As a result, the value of the unique game is at most ε.
As a result, if it is NP-hard to distinguish a satisfiable d-to-1 game from one with value at most γ, then
it is also NP-hard to distinguish a unique game of value at least ζ = 1/d from that with value at most
γ, which concludes the proof of this lemma.
C
Proof of Dictator Test
Here we prove our result for the Dictator-vs.-Quasirandom test (Theorem 7).
Proof. (of Theorem 7) For c ∈ [R], define the function
fc (x1 , x2 , . . . , xn ) := f (x1 + c, x2 + c, . . . , xn + c) − c.
(143)
Note that ±c is performed modulo R.
The test works as follows: Pick z ∈ [R]n uniformly at random, and let x(1) , x(2) , . . . , x(k) be independent
ρ-correlated noisy copies of z. Then, pick c1 , c2 , . . . , ck independently uniformly at random, where each
ci ∈ [R]. Accept iff
fc1 (x(1) ) = fc2 (x(2) ) = · · · = fck (x(k) ).
(144)
For completeness, notice that if f is a dictator, then fc = f for all c ∈ [R]. Say f is a dictator on the
j-th coordinate: f (x) = xj . Then the test clearly accepts with probability at least ρk (if none of the
ρ
coordinates j were perturbed in all the noisy copies x(i) ← z).
24
For soundness: For any i ∈ [R], let f i : [R]n → {0, 1} denote the indicator function for f (x) = i, and
similarly for fci : [R]n → {0, 1}. Notice that
fci (x) = f i+c (x + (c, c, . . . , c))
(145)
Then, write the acceptance probability as
Pr[accept] =
Pr
ρ
ci ,z,x(j) ←z
=
X
[fc1 (x(1) ) = fc2 (x(2) ) = · · · = fck (x(k) )]
Pr
ρ
(j) ←z
i∈[R] ci ,z,x
=
(Independence of ci )
=
X
E
ρ
(j) ←z
i∈[R] ci ,z,x
X
E
i∈[R] z,x
ρ
(146)
[i = fc1 (x(1) ) = fc2 (x(2) ) = · · · = fck (x(k) )]
(147)
[fci1 (x(1) )fci2 (x(2) ) . . . fcik (x(k) )]
(148)
[ E [fci1 (x(1) )] E [fci2 (x(2) )] . . . E [fcik (x(k) )]].
(j) ←z c1
c2
(149)
c2
(150)
If we define the function g i : [R]n → [0, 1] as
g i (x) := E[fci (x)].
(151)
c
Then this acceptance probability is
Pr[accept] =
X
[g i (x(1) )g i (x(2) ) . . . g i (x(k) )]
E
ρ
(152)
i∈[R] z,xi ←z
=
X
i∈[R]
i
k
E[(Tρ g (z)) ].
(153)
z
Notice that Ex [g i (x)] = 1/R, because
i
i
i+c
E[g (x)] = E [fc (x)] = E [f (x + (c, c, . . . , c))]
x
x,c
(c, x same joint distribution as i + c, x + c)
(154)
x,c
= E [f c (x)]
(155)
x,c
= E[E[f c (x)]] = E[1/R] = 1/R.
x c
(156)
x
Thus, if the function g i has small low-degree influences, then Lemma 3 (Main Lemma) applied to g i in
line (153) directly implies that this acceptance probability is 2O(k) /Rk−1 . We will now formally show
that the influences of the “expected folded function” g i are bounded by the influences of the original f i .
First, the Fourier coefficients of g i are
ĝ i (s) = E[fˆci (s)].
c
25
(157)
Thus the low-degree influences of g i are bounded as
X
Infj≤d [g i ] =
ĝ i (s)2
(158)
i
2
E[fˆc (s)]
(159)
i
2
E[fˆc (s) ]
(160)
fˆci (s)2 ]
(161)
s∈[R]n
s(j)6=1,|s|≤d
X
=
c
s∈[R]n
s(j)6=1,|s|≤d
≤
X
s∈[R]n
s(j)6=1,|s|≤d
X
= E[
c
c
s∈[R]n
s(j)6=1,|s|≤d
= E[Infj≤d [fci ]].
(162)
c
Finally, we must relate the influences of fci to the influences of f i . For a fixed c ∈ [R], we have
Infj≤d [fci ] = Infj [(fci )≤d ]
=
=
=
E [V
x∈[R]n
(163)
arxj ∈[R] [(fci )≤d ]]
(164)
i+c ≤d
E [V arxj ∈[R] [(f ) (x1 + c, x2 + c, . . . , xn + c)]]
(165)
i+c ≤d
E [V arxj ∈[R] [(f ) (x1 , x2 , . . . , xn )]]
(166)
x∈[R]n
x∈[R]n
= Infj [(f i+c )≤d ]
(167)
Infj≤d [f i+c ].
(168)
=
Therefore, if Infj≤d [f i ] ≤ δ for all coordinates j ∈ [n] and all projections i ∈ [R] (as we assume for
soundness), then from (162) and (168) we have
Infj≤d [g i ] ≤ E[Infj≤d [fci ]] = E[Infj≤d [f i+c ]] ≤ δ.
c
(169)
c
Thus the function g i has small low-degree influences as well.
So we can complete the proof, continuing from line (153) and applying our Main Lemma to g i :
Pr[accept] =
X
i∈[R]
(Lemma 3)
≤
X
i
k
E[(Tρ g (z)) ]
(170)
2O(k) /Rk
(171)
z
i∈[R]
= 2O(k) /Rk−1 .
26
(172)
| 8cs.DS
|
Enhanced q-Least Mean Square
Shujaat Khana , Alishba Sadiqb , Imran Naseemb,c,∗, Roberto Togneric ,
Mohammed Bennamound
a Department
arXiv:1801.00410v1 [math.OC] 1 Jan 2018
of Bio and Brain Engineering, Korea Advanced Institute of Science and
Technology (KAIST), Daejeon, Republic of Korea.
b College of Engineering, Karachi Institute of Economics and Technology, Korangi Creek,
Karachi 75190, Pakistan.
c School of Electrical, Electronic and Computer Engineering, The University of Western
Australia, 35 Stirling Highway, Crawley, Western Australia 6009, Australia.
d School of Computer Science and Software Engineering, The University of Western
Australia, 35 Stirling Highway, Crawley, Western Australia 6009, Australia.
Abstract
In this work, a new class of stochastic gradient algorithm is developed based on
q-calculus. Unlike the existing q-LMS algorithm, the proposed approach fully
utilizes the concept of q-calculus by incorporating time-varying q parameter.
The proposed enhanced q-LMS (Eq-LMS) algorithm utilizes a novel, parameterless concept of error-correlation energy and normalization of signal to ensure
high convergence, stability and low steady-state error. The proposed algorithm
automatically adapts the learning rate with respect to the error. For the evaluation purpose the system identification problem is considered. Extensive experiments show better performance of the proposed Eq-LMS algorithm compared
to the standard q-LMS approach.
Keywords: Adaptive algorithms, Least Mean Squares Algorithm, q-calculus,
Jackson derivative, system identification, q-LMS.
1. Introduction
5
Least square method is considered to be widely used optimization technique.
It has been applied in diversified applications such as plant identification [1],
detection of elastic inclusions [2], noise cancellation [3], echo cancellation [4],
ECG signal analysis [5], elasticity imaging [6], and time series prediction [7], etc.
The least mean square (LMS) is one of the most popular least square algorithms
for adaptive filtering due to its low computational complexity, however, it has
∗ Corresponding
author
Email addresses: [email protected] (Shujaat Khan),
[email protected] (Alishba Sadiq), [email protected] (Imran
Naseem), [email protected] (Roberto Togneri),
[email protected] (Mohammed Bennamoun)
Preprint submitted to Signal Processing - Journal - Elsevier
January 3, 2018
10
15
20
25
30
35
40
45
50
a slow convergence rate due to the dependency on the eigenvalue-spread of
the input correlation matrix [8]. Extensive research has been done towards the
optimization of the LMS algorithm [3, 9, 10, 11, 12]. One of the disadvantages of
the LMS is that it is sensitive to the scaling of its input. In [8, 11], the normalized
LMS (NLMS) and its variants were proposed to solve this problem through
normalization. To improve the convergence rate and steady state performance,
variable step size frameworks were devised in [13, 14]. In [15, 16, 17, 18], different
solutions for complex signal processing were proposed. Similarly, to deal with
non-linear signal processing problem, the concept of kernel function-based LMS
algorithms was proposed in [19, 20, 21].
Beside these variants, various definitions of gradient have also been used
to derive improved LMS algorithms; for instance in [22], a fractional order
calculus (FOC) based least mean square algorithm, named as robust variable
step size fractional least mean square (RVSS-FLMS), is proposed. The algorithm
is derived using Riemann-Liouville fractional derivative for high convergence
performance. In [23, 24], some adaptive schemes were proposed for maintaining
stability through adaptive variable fractional power. The FOC variants are,
however, not stable and diverge if the weights are negative or the input signal
is complex [25].
Recently, the q-LMS algorithm is proposed which utilizes q-gradient from the
Jackson’s derivative so that the secant of the cost function is computed instead of
the tangent [12]. The algorithm takes larger steps towards the optimum solution
and therefore, achieves a higher convergence rate. The q-LMS algorithm has
also been used for various applications, such as adaptive noise cancellation [26],
system identification, and designing of whitening filter [27]. In [28] q-normalized
LMS algorithm is proposed and its convergence performance is analyzed. In
[29], using the same definition of q-calculus, variants of steady state least mean
algorithms are derived.
All the aforementioned variants of the q-LMS algorithm enhance convergence
speed at the cost of increased computational complexity and steady-state error.
In order to improve convergence rate without compromising the steady-state
performance, a time-varying q-LMS is proposed in [30]. However, it requires
the tuning of two additional parameters (β and γ), and the performance of
the time-varying q-LMS [30] is very sensitive to the selection of the tuning
parameters. In this paper, we propose a new variant of the q-LMS by making the
q-parameter time-varying. The proposed enhanced q-LMS (Eq-LMS) utilizes a
novel, parameterless concept of error-correlation energy and normalization to
ensure rapid convergence without compromising stability and low steady-state
error. The proposed algorithm automatically adapts the learning rate with
respect to error. It takes larger steps in case of larger error and reduces the
learning rate with decreased error. Unlike the contemporary methods [12, 13,
14], the proposed method is a parameterless technique and does not require
manual tuning of any parameter. The proposed algorithm is evaluated for the
system identification problem and the results are demonstrated for both the
steady-state performance and the convergence rate. Extensive experiments are
performed to show the superiority of the proposed Eq-LMS algorithm over a
2
55
variety of contemporary methods.
The rest of the paper is structured as follows. An overview of the q-calculus
and q-least mean square algorithm is provided in Section 2. The details of
proposed algorithm are discussed in Section 3, followed by the experimental
results in Section 4. The paper is finally concluded in Section 5.
2. Overview of q -Least Mean Square Algorithm
60
The conventional LMS algorithm is derived using the concept of steepest
descent with the weight- update rule
w(i + 1) = w(i) −
µ
∇w J(w),
2
(1)
where J(w) is the cost function for the LMS algorithm and is defined as
J(w) = E[e2 (i)],
(2)
where E[·] is the expectation operator and e(i) is the estimation error between
the desired response d(i) and the output signal at the ith instant, i.e.,
e(i) = d(i) − x| (i)w(i),
(3)
Here, x(i) is the input signal vector defined as
x(i) = [x1 (i), x2 (i), . . . xM (i)]| ,
(4)
and w(i) is the weight vector defined as:
w(i) = [w1 (i), w2 (i), . . . wM (i)]
(5)
where M is the length of the filter.
65
70
2.1. Overview of q-Calculus
The Quantum calculus or q-calculus is sometimes referred to as the calculus
without a limit [31]. It has been successfully used in various areas such as
number theory, combinatorics, orthogonal polynomials, basic hyper-geometric
functions and other sciences quantum theory, operational theory, mechanics,
and the theory of relativity [32, 33, 34, 35].
In q-calculus, the differential of a function is defined as (See, [36])
dq (f (x)) = f (qx) − f (x).
(6)
The derivative therefore takes the form
Dq (f (x)) =
f (qx) − f (x)
dq (f (x))
=
.
dq (x)
(q − 1)x
3
(7)
When q → 1, the expression becomes the derivative in the classical sense. The
q-derivative of a function of the form xn is
n
q − 1 xn−1 , q 6= 1,
n
q−1
Dq,x x =
(8)
nxn−1 ,
q = 1.
For a function f (x) of n number of variables, x = [x1 , x2 , ....xn ]| , the q-gradient
is defined as
∇q,w f (x) , [Dq1,x1 f (x), Dq2,x2 f (x), ...Dqn,xn f (x)]| ,
(9)
where q = [q1 , q2 , . . . qN ]| .
2.2. q-Least Mean Square ( q-LMS) Algorithm
75
80
The performance of the LMS algorithm depends on the eigenvalue spread of
the input correlation matrix. The LMS is therefore regarded as an inherently
slowly converging approach [14]. In order to resolve this issue the q-LMS has
been proposed in [12]. Instead of the conventional gradient, the q-LMS is derived
using the q-calculus and utilizes the Jackson derivative method [12], it takes
larger steps (for q > 1) in the search direction as it evaluates the secant of
the cost function rather than the tangent [12]. By replacing the conventional
gradient in (1) with the q-gradient, we get
w(i + 1) = w(i) −
µ
∇q,w J(w).
2
(10)
The q-gradient of the cost function J(w) for the k th weight is defined as
∇q,wk J(w) =
∂qk
∂q
∂qk
J(w) k e(i)
y(i).
∂qk e
∂qk y
∂qk wk (i)
(11)
Solving partial derivatives in (11) using the Jackson derivative defined in Section
(2) gives
∂q
q2 − 1
∂q
J(w) =
(E[e2 (i)]) = E[ k
e(i)] = E[(qk + 1)e(i)],
∂q e
∂q e
qk − 1
85
(12)
where J(w) = E[e2 (i)], E[·] is the expectation operator and e(i) = d(i) − y(i).
Similarly
∂qk
y(i) = xk (i),
∂qk wk (i)
(13)
∂qk
e(i) = −1,
∂qk y
(14)
and
4
Substituting equations (12), (13), and (14) in (11) gives
∇q,wk (i)J(w) = −E[(qk + 1)e(i)xk (i)].
(15)
Similarly, for k = 1, 2, . . . , M ,
∇q,w J(w) = −E[(q1 +1)e(i)x1 (i), (q2 +1)e(i)x2 (i), . . . (qM +1)e(i)xM (i)]. (16)
Consequently, Eq. (16) can be written as
∇q,w J(w) = −2E[Gx(i)e(i)],
(17)
where µ is the learning rate (step-size) and G is a diagonal matrix
diag(G) = [(
90
qM + 1 |
q1 + 1 q2 + 1
), (
), .....(
)] .
2
2
2
(18)
Dropping the expectation from the q-gradient in (17) results in
∇q,w J(w) ≈ −2Gx(i)e(i).
(19)
Substituting (19) in (1) renders the weight update rule of the q-LMS algorithm
by
w(i + 1) = w(i) + µGx(i)e(i).
95
(20)
2.2.1. q-LMS as a whitening filter-q-Normalized LMS
The q-normalized least mean square (q-NLMS) algorithm is defined (See
[28])
w(i + 1) = w(i) + µ
Gx(i)e(i)
,
ζ + ||x(i)||2G
(21)
where ζ is a small value added in the denominator to avoid the indeterminate
form, and ||x(i)||2G is the weighted norm of the input vector. By selecting the q
parameter in (20) as q = 1/λ max, we can design a whitening filter and hence
it can remove the dependency on the input correlation [28].
100
2.2.2. Time-varying q-LMS
The time-varying q-LMS algorithm is based on the variable step size (VSS)
method [14] and is given (See [30])
Ψ (i + 1) = βΨ (i) + γe(i)2 ,
(0 < β < 1, γ > 0),
qupper , Ψ (i + 1) > qupper ,
q(i + 1) = 1,
Ψ (i + 1) < 1,
Ψ
otherwise.
5
(22)
(23)
where qupper is so chosen to satisfy the stability bound, (See [12])
qupper =
2
.
µλmax
(24)
The q(i + 1) is updated according to the estimation of the square of the
estimation error. When the estimation error is large, q(i) will approach its
upper bound denoted by qupper , while for smaller values q(i) goes to unity for a
lower steady-state error.
105
3. The Proposed Enhanced q -Least Mean Square (Eq-LMS) Algorithm
The q-LMS algorithm has an extra degree of freedom to control the performance via the diagonal matrix G, which comprises of the q-dependent entries.
The weight-update rule of the q-LMS algorithm can be written as
w(i + 1) = w(i) + µx̂(i)e(i),
110
115
120
125
(25)
where x̂(i) = Gx(i). For the special case of G=I (identity matrix), the q-LMS
algorithm will be transformed into the conventional LMS. Based on the above
discussion, we make the following important observations.
• We argue that the q-gradient with q > 1 enhances the speed of convergence
as it takes the secant of function rather than the tangent [12]. The larger
the value of the q parameter, the faster the convergence of the algorithm.
But this improvement in the rate of convergence comes at the cost of a
degradation in the steady-state performance.
• The time varying q-LMS technique [30] is based on variable step-size
method [14], which requires the tuning of additional parameters such as
β and γ.
• By selecting the q parameter in (20) as q = 1/λ max, we can design a
whitening filter and hence it can remove the dependency on the input
correlation [28]. However, with a large step-size the q-NLMS converges
rapidly with a compromised steady state performance. Similarly, a smaller
step-size results in better steady state performance but with slow convergence. As such, the two important performance parameters cannot be
optimized simultaneously.
3.1. Proposed Improvements
130
To overcome the aforementioned issues, we propose the Eq-LMS algorithm
with the following improvements.
6
135
• To achieve higher convergence rate with lower steady-state error, we propose to incorporate the instantaneous error energy to adapt the q-parameter.
The proposed algorithm automatically takes large steps when the error is
large and reduces the step-size with the decreasing error. Note that, unlike the time-varying q-LMS [30], no additional tuning parameters are
introduced and the proposed approach is completely automatic.
• The whitening factor (q = 1/λ max) is also utilized to set the limits of
adaptive q-vector, this allows the algorithm to operate at a higher convergence rate without worrying about the divergence issues.
140
• To update each q-parameter, the proposed Eq-LMS utilize, a responsible
error for each tap of the filter. With this improvement, the q-variable for
each tap will be updated accordingly, hence, both the steady-state error
and convergence performance can be improved significantly.
3.2. Formulation of the Eq-LMS Algorithm
By replacing the fixed G in (20) with its time-varying form G(i), the weight
update rule of the proposed Eq-LMS is given as
w(i + 1) = w(i) + µe(i)x(i)G(i),
(26)
th
where µ is the learning rate, and e(i) is the error at the i
instant defined by
e(i) = d(i) − y(i),
(27)
where d(i) and y(i) are the desired and estimated output at the ith instant, respectively. Here, G(i) is a diagonal matrix with time-varying diagonal elements,
and is defined as
q1 (i)
q2 (i)
(28)
G(i) =
.
..
.
qM (i)
It can also be written as
G(i) = diag(q1 (i), . . . , qM (i)) = diag(q(i)),
(29)
q(i) = {q1 (i), q2 (i), . . . qM (i)}.
(30)
where,
145
We propose an update rule for vector q defined in the following steps.
• Step1: Initialize vector q with any positive random values.
• Step2: Use instantaneous error to update first entry q1 of q vector, which
is associated with weight of the instant input tap, i.e.,
M
X
1
q1 (i + 1) =
{|e(i)| +
qj (i)},
M +1
j=1
where M is the length of the filter.
7
(31)
• Step3: To avoid divergence while maintaining the higher convergence rate,
the following conditions will be evaluated:
1
1
if |q1 (i + 1)| > λmax
,
λmax
(32)
q(i + 1) =
1
,
q1 (i + 1) if |q1 (i + 1)| < λmax
where λmax is the maximum eigenvalue of the input auto-correlation matrix.
• Step4: Update all entries of vector α except for the first entry, simply by
shifting:
qk+1 (i + 1) = qk (i),
(33)
where 1 < k < M − 1
150
• Step5: For next iterations, repeat steps 2 to 5.
Finally, the weight-update equation of the proposed Eq-LMS can be written
as:
w(i + 1) = x(i) + µe(i)x(i)
where
q(i),
(34)
indicates the element wise multiplication.
4. Experiments
For the evaluation of the proposed algorithm the problem of system identification is used. Adaptive learning methods have been successfully used to
identify the unknown system, with numerous applications, for example, in control engineering, communication systems [11, 37, 38, 39, 40, 41]. Channel estimation, for instance, is a widely used method in communication systems to
estimate the characteristics of an unknown channel. Consider a linear channel
shown in Fig. 1.
Figure 1: Channel estimation using adaptive learning algorithm.
8
y(t) = h1 x(t) + h2 x(t − 1) + h3 x(t − 2) + h4 x(t − 3) + h5 x(t − 4).
155
160
165
Equation (35) shows the mathematical model of the system, where x(t) and y(t)
are the input and output of the system, respectively and d(t) is the disturbance
which is taken to be white Gaussian noise in this case. For this experiment, x(t)
is chosen to be consisting of 1 × 106 randomly generated samples obtained from
Gaussian distribution of mean zero and variance of 1. In (35), the system is
defined by its impulse response h(t) while ŷ(t), ĥ(t), and e(t) are the estimated
output, estimated impulse response, and the error of estimation, respectively.
The simulation parameters selected are as follows: coefficient values of h1 = −2,
h2 = −1, h3 = 0, h4 = 1 and h5 = 2 are selected for the channel, the experiments
are performed on three noise levels with the SNR values of 10dB, 20dB and 30dB.
The weights are initialized to zero for all algorithms. Specifically, the objective
of these simulations is to compare the performance of the proposed enhanced
q−LMS (Eq-LMS) algorithm with the contemporary counterparts, i.e., Least
Mean Square (LMS)/ q-LMS [12] at q = 1, q-LMS [12] at q = 2, time-varying
q-LMS [30] and the normalized LMS (NLMS), for given convergence rate and
given steady state error in three different scenarios.
For the performance evaluation, the normalized weight deviation (NWD) in
the actual and the obtained weights is compared. Specifically, we define
NWD =
170
175
(35)
kh − wk
,
khk
(36)
where h is the actual impulse response of the channel and w is the estimated
weight-vector. The simulations are repeated for 1000 independent runs and
mean results are reported. The simulations are performed primarily to evaluate
the steady-state and convergence performances of the proposed algorithm for
various learning rates. Accordingly, three Evaluation protocols are designed:
1. Evaluation protocol 1: learning rate=1 × 10−1 , SNR={10,20,30} dB.
2. Evaluation protocol 2: learning rate=1 × 10−2 , SNR={10,20,30} dB.
3. Evaluation protocol 3: learning rate=1 × 10−3 , SNR={10,20,30} dB.
4.1. Experiments to Evaluate the Steady-State Performance
180
4.1.1. Evaluation Protocol 1: Fast Convergence
For the comparison of the steady-state performance, all algorithms were
setup for equal convergence and after 10000 iterations, the steady-state value
of the NWD is examined. The learning rate (step size) configurations for equal
convergence rate (Evaluation Protocol 1) is shown in Table 1.
9
Table 1: Evaluation protocol 1: Configuration of learning rates of different approaches for an
equal convergence rate of 1 × 10−1 .
Algorithm
LMS/q-LMS
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
185
190
195
10 dB SNR
4.4 × 10−2
3.5 × 10−1
2.45 × 10−1
1 × 10−1
Learning Rate µ
20 dB SNR 30 dB SNR
2.45 × 10−2
2.7 × 10−5
−1
1.3 × 10
3.2 × 10−5
−1
1.2 × 10
8.7 × 10−5
1 × 10−1
1 × 10−1
The relevant normalized weight difference (NWD) curves with three different
SNR values are depicted in Fig. 2. From the Fig. 2 (a),(b), and (c), it can be
seen that the proposed Eq-LMS produced the best performance under all three
conditions: (1) for the SNR value of 10dB, it outperformed the LMS/q-LMS
at q = 1, the q-LMS at q = 2, the time-varying q-LMS, and the NLMS by the
NWD value of −1.03 dB, −2.95 dB, −8.06 dB, and −2.14 dB, respectively, (2)
for the SNR value of 20 dB, it surpassed the listed algorithms by the NWD
value of −2.36 dB, −4.05 dB, −6.96 dB, and −3.29 dB, respectively, and (3)
for the SNR value of 30 dB, the above mentioned algorithms were outperformed
by a margin of −2.29 dB, −3.86 dB, −7.18, and −3.47 dB, respectively. Note
that the proposed Eq-LMS algorithm showed the lowest steady state error in
all conditions while the LMS, NLMS and the q-LMS at q = 2 show faster
convergence than the time-varying q-LMS but with greater steady state error
than the proposed Eq-LMS. With the above discussed settings, results for the
channel estimation problem are summarized in Table 2.
Table 2: Evaluation protocol 1: Results of various approaches for an equal convergence rate.
Algorithm
LMS/q-LMS (q = 1)
q-LMS (q = 2)
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
200
Steady-state NWD (dB)
10 dB SNR 20 dB SNR 30 dB SNR
-14.63
-21.06
-28.77
-12.71
-19.37
-27.02
-7.60
-16.46
-23.88
-13.52
-20.13
-27.59
-15.66
-23.42
-31.06
4.1.2. Evaluation protocol 2: Medium Convergence
The learning rate (step size) configurations for an equal convergence rate
(Evaluation protocol 2) is shown in Table 3.
10
Table 3: Evaluation protocol 2: Configuration of learning rates of different approaches for an
equal convergence rate of 1 × 10−2 .
Algorithm
LMS/q-LMS
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
205
210
215
10 dB SNR
4 × 10−3
2 × 10−2
2 × 10−2
1 × 10−2
Learning Rate µ
20 dB SNR 30 dB SNR
1.5 × 10−3
5.2 × 10−4
−3
1 × 10
1.8 × 10−3
−3
9 × 10
2.5 × 10−3
1 × 10−2
1 × 10−2
The relevant normalized weight difference (NWD) curves with three different
SNR values are depicted in Fig. 2. From the Fig. 2 (d),(e), and (f), it can be
seen that the proposed Eq-LMS produced the best performance under all three
conditions: (1) for the SNR value of 10 dB, it outperformed the LMS/q-LMS
at q = 1, q-LMS at q = 2, time-varying q-LMS and the NLMS by the NWD
value of −0.68 dB, −2.23 dB, −4.08 dB, and −1.74 dB, respectively, (2) for the
SNR value of 20 dB, it surpassed the above-mentioned algorithms by the NWD
value of −1.22 dB, −2.74 dB, −5.34 dB, and −2.45 dB, respectively, and (3)
for the SNR value of 30 dB, the above mentioned algorithms were outperformed
by a margin of −1.44 dB, −2.94 dB, −4.07 and −2.47 dB, respectively. Note
that the proposed Eq-LMS algorithm showed the lowest steady state error in
all conditions while the LMS, NLMS and the q-LMS at q = 2 show faster
convergence than time-varying q-LMS but with a greater steady state error
than the proposed Eq-LMS. With the above discussed settings, results for the
channel estimation problem are summarized in Table 4.
Table 4: Evaluation protocol 2: Results of various approaches for an equal convergence rate.
Algorithm
LMS/qLMS (q=1)
qLMS (q=2)
Time-varying qLMS
Normalized LMS
Proposed Eq-LMS
Steady-state NWD (dB)
10 dB SNR 20 dB SNR 30 dB SNR
-25.42
-32.70
-40.00
-23.62
-30.86
-38.19
-22.40
-29.40
-37.45
-24.18
-31.45
-38.80
-25.85
-33.60
-41.13
4.1.3. Evaluation protocol 3: Slow Convergence
The learning rate (step size) configurations for equal convergence rate (Evaluation protocol 3) is shown in Table 5.
11
Table 5: Evaluation protocol 3: Configuration of learning rates of different approaches for the
fix convergence rate of 1 × 10−3 .
Algorithm
LMS/qLMS
Time-varying qLMS
Normalized LMS
Proposed Eq-LMS
220
225
230
10 dB SNR
3.6 × 10−4
1.6 × 10−3
1.9 × 10−3
1 × 10−3
Learning Rate µ
20 dB SNR 30 dB SNR
1.3 × 10−4
4.5 × 10−3
−4
6 × 10
1.5 × 10−4
−4
7 × 10
2.3 × 10−4
1 × 10−3
1 × 10−3
The relevant NWD curves with three different SNR values are delineated in
Fig. 2. From the Fig. 2 (g), (h), and (i), it can be seen that the proposed
Eq-LMS produced the best performance under all three conditions: (1) for the
SNR value of 10 dB, it outperformed the LMS/q-LMS at q = 1, q-LMS at
q = 2, Time-varying q-LMS, and the NLMS by the NWD value of −0.43 dB,
−1.92 dB, −3.45 dB, and −1.67 dB, respectively, (2) for the SNR value of 20
dB, it surpassed the above-mentioned algorithms by the NWD value of −0.9
dB, −2.43 dB, −4.2 dB, and −2.15 dB, respectively, and (3) for the SNR value
of 30 dB, the above mentioned algorithms were outperformed by a margin of
−1.13 dB, −2.66 dB, −3.68, and −2.33 dB, respectively. Note that the proposed
Eq-LMS algorithm showed the lowest steady state error in all conditions while
the q-LMS at q = 2 showed faster convergence with greater steady state error
than the proposed Eq-LMS. With the above discussed settings, results for the
channel estimation problem are summarized in Table 6.
Table 6: Evaluation protocol 3: Results of various approaches for an equal convergence rate.
Algorithm
LMS/q-LMS (q = 1)
q-LMS (q = 2)
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
Steady-state NWD (dB)
10 dB SNR 20 dB SNR 30 dB SNR
-20.16
-27.34
-34.67
-18.92
-26.13
-33.45
-16.76
-23.22
-32.04
-19.10
-26.11
-33.64
-20.84
-28.56
-36.11
4.2. Experiments to Evaluate the Convergence Performance
235
4.2.1. Evaluation Protocol 1: Fast Convergence
For the comparison of convergence performances, all algorithms were setup
for equal steady-state error. The learning rate (step size) configurations for
equal steady-state (Evaluation Protocol 1) is shown in Table (7). Learning rate
for proposed Eq-LMS has been set according to three evaluation protocols.
12
Normalized weight difference (dB)
-1
0
200
400
600
800
10-2
10-3
1000
0
1000
Iterations
10 (dB) SNR and 1e -2 Learning rate
Normalized weight difference (dB)
Normalized weight difference (dB)
10-2
2000
4000
6000
8000
10-1
10-3
10000
0
1
Normalized weight difference (dB)
Normalized weight difference (dB)
-3
4
Iterations
g
3
4
6
8
10
104
10
10-3
0
2
-4
1
2
Iterations
h
4
6
8
10
104
Iterations
3
4
30 (dB) SNR and 1e -3 Learning rate
100
10-3
0
10000
f
10-2
10
8000
10-2
10-4
5
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
-1
6000
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
104
20 (dB) SNR and 1e -3 Learning rate
100
10-2
2
4000
30 (dB) SNR and 1e -2 Learning rate
100
e
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
2
Iterations
10 (dB) SNR and 1e -3 Learning rate
0
2000
c
10-2
d
10
0
Iterations
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
Iterations
100
10-3
10-4
5000
20 (dB) SNR and 1e -2 Learning rate
100
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
0
4000
10-2
b
10-1
10-3
3000
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
Iterations
a
100
2000
Normalized weight difference (dB)
10-2
10-1
30 (dB) SNR and 1e -1 Learning rate
100
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
Normalized weight difference (dB)
Normalized weight difference (dB)
10
20 (dB) SNR and 1e -1 Learning rate
100
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
Normalized weight difference (dB)
10 (dB) SNR and 1e -1 Learning rate
100
5
105
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
10-2
10-3
10-4
10-5
0
2
4
Iterations
6
8
10
105
i
Figure 2: NWD curves for the LMS/q-LMS at q = 1, q-LMS at q = 2, time-varying q-LMS,
NLMS, and the Eq-LMS. normalized weight deviation with learning rate and SNR of (a) 1e−1 ,
10 dB, (b) 1e−1 , 20 dB,(c) 1e−1 , 30 dB, (d) 1e−2 , 10 dB, (e) 1e−2 , 20 dB, (f) 1e−2 , 30 dB,
(g) 1e−3 , 10 dB, (h) 1e−3 , 20 dB, and (i) 1e−3 , 30 dB.
13
Table 7: Evaluation protocol 1: Configuration of learning rates of different approaches for an
equal steady-state error.
Algorithm
LMS/q-LMS
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
240
245
250
255
10 dB SNR
3 × 10−2
3.3 × 10−2
1 × 10−1
1 × 10−1
Learning Rate µ
20 dB SNR 30 dB SNR
8.9 × 10−3
2.7 × 10−3
−3
8.8 × 10
3.1 × 10−3
−2
2.8 × 10
8.5 × 10−3
1 × 10−1
1 × 10−1
The relevant normalized weight difference (NWD) curves with three different
SNR values are depicted in Fig. 3. From the Fig. 3 (a), (b), and (c), it can be
seen that the proposed Eq-LMS algorithm produced the best results under all
three conditions: (1) for the SNR value of 10 dB, algorithms are run for 1000
iterations. The convergence point of the proposed Eq-LMS is reached at 120th
iteration, q-LMS at (q = 2) converged on the 80th iteration but its steady state
error is much larger compared to the proposed Eq-LMS, (2) for the SNR value
of 20 dB, algorithms are run for 5000 iterations. Note that the proposed EqLMS algorithm outperformed all competing approaches by converging in only
400 iterations. The q-LMS (q = 2) was unable to reach the given error-floor and
took 400 iterations to reach a much higher error, and (3) for the SNR value of
30 dB, algorithms are run for 10000 iterations, the proposed Eq-LMS algorithm
took the least number of iterations by converging at the 1600th iteration. The
proposed EqLMS shows best performance in terms of steady state error and
convergence rate. Thus, showing the best overall performance. With the above
discussed settings, results for the channel estimation problem are summarized
in Table 8.
Table 8: Evaluation protocol 1: Results of various approaches for an equal steady-state error.
Algorithm
LMS/q-LMS at (q = 1)
q-LMS at (q = 2)
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
260
Convergence point (number of Iterations ×1000))
10 dB SNR 20 dB SNR
30 dB SNR
0.20
0.70
2.70
0.08
0.40
1.55
0.64
2.80
7.20
0.23
1.10
4.40
0.12
0.40
1.60
4.2.2. Evaluation protocol 2: Medium Convergence
The learning rate (step size) configuration for equal steady-state (Evaluation
Protocol 2) is shown in configuration Table (9). Learning rate for proposed EqLMS has been set according to three evaluation protocols.
14
Table 9: Evaluation protocol 2: Configuration of learning rates of different approaches for an
equal steady-state error.
Algorithm
LMS/q-LMS
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
265
270
10 dB SNR
3 × 10−3
3.3 × 10−3
9 × 10−2
1 × 10−2
Learning Rate µ
20 dB SNR 30 dB SNR
8.8 × 10−4
2.7 × 10−4
−4
9.3 × 10
3.1 × 10−4
−3
2.72 × 10
8.5 × 10−4
1 × 10−2
1 × 10−2
The relevant NWD curves with three different SNR values are depicted in
Fig. 3. From the Fig. 3 (d), (e), and (f), it can be seen that the proposed
Eq-LMS algorithm produced the best results under all three conditions: (1) for
the SNR value of 10 dB, algorithms are run for 10000 iterations, the convergence
point of the proposed Eq-LMS is reached at 1500th iteration, the q-LMS at (q =
2) converged on the 1100th iteration but its steady state error is much larger than
the proposed Eq-LMS, (2) for the SNR value of 20 dB, algorithms are run for
50000 iterations, the proposed Eq-LMS algorithm outperformed all competing
approaches in terms of convergence point with least steady state error, and
(3) for the SNR value of 30 dB, algorithms are run for 100000 iterations, the
proposed Eq-LMS algorithm converged on 19000th iteration, it showed best
performance in terms of steady state error and convergence rate. Thus, showing
the best overall performance. With the above discussed settings, results for the
channel estimation problem are summarized in Table 10.
Table 10: Evaluation protocol 2: Results of various approaches for an equal steady-state error.
Algorithm
LMS/q-LMS at (q = 1)
q-LMS at (q = 2)
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
275
Convergence point (number of Iterations ×1000))
10 dB SNR 20 dB SNR
30 dB SNR
2
9
32
1.1
5.9
19
6.2
28
80
3.1
14
52
1.5
6
19
4.2.3. Evaluation protocol 3: Slow Convergence
The learning rate (step size) configuration for equal steady-state (Evaluation
protocol 3) is shown in configuration Table (11). Learning rate for the proposed
Eq-LMS has been set according to three evaluation protocols.
15
Table 11: Evaluation protocol 3: Configuration of learning rates of different approaches for
an equal steady-state error.
Algorithm
LMS/q-LMS
Time-varying q-LMS
Normalized LMS
Proposed Eq-LMS
280
285
290
10 dB SNR
3 × 10−4
3.3 × 10−4
1 × 10−3
1 × 10−3
Learning Rate µ
20 dB SNR 30 dB SNR
8.9 × 10−4
2.7 × 10−5
−5
9 × 10
3.2 × 10−5
−4
2.8 × 10
8.5 × 10−5
1 × 10−3
1 × 10−3
The relevant NWD curves with three different SNR values are depicted in
Fig. 3. From the Fig. 3 (g), (h), and (i), it can be seen that the proposed EqLMS algorithm produced the best results under all three conditions: (1) for the
SNR value of 10 dB, algorithms are run for 100000 iterations, the convergence
point of the proposed Eq-LMS is reached at 21000th iteration, the q-LMS at
(q = 2), converged at the 12000th iteration but its steady state error is much
larger than the proposed Eq-LMS, (2) for the SNR value of 20 dB, the algorithms
are run for 500000 iterations, the proposed Eq-LMS algorithm outperformed all
competing approaches in terms of convergence point with least steady state
error, and (3) for the SNR value of 30 dB, the algorithms are run for 1000000
iterations, the proposed Eq-LMS algorithm converged on 200000th iteration.
The proposed Eq-LMS showed the best performance in terms of steady state
error and convergence rate. With the above discussed settings, results for the
channel estimation problem are summarized in Table 12.
Table 12: Evaluation protocol 3: Results of various approaches for an equal steady-state error.
Algorithm
LMS/qLMS at (q=1)
qLMS at (q=2)
Time-varying qLMS
Normalized LMS
Proposed Eq-LMS
Convergence point (number of Iterations ×1000))
10 dB SNR 20 dB SNR
30 dB SNR
23
100
370
12
60
200
72
28
840
36
320
600
21
70
240
5. Conclusion
295
300
In this work, we proposed a quantum calculus-based steepest descent algorithm called enhanced q-least mean square algorithm (Eq-LMS) using a novel
concept of error correlation energy. The proposed algorithm is a parameterless
method and unlike the contemporary time varyying q-LMS, it does not require
additional tuning. The proposed Eq-LMS was compared with the LMS, q-LMS,
time varying q-LMS, and the NLMS algorithms for a problem of linear channel
estimation. Extensive simulation tests were conducted to analyze the convergence and the steady-state performance at three different SNR levels. For all
16
Normalized weight difference (dB)
10-1
0
200
400
600
800
10-1
10-2
10-3
1000
0
1000
Iterations
10 (dB) SNR and 1e -2 Learning rate
Normalized weight difference (dB)
Normalized weight difference (dB)
10-2
2000
4000
6000
8000
10-1
10-3
10000
0
1
Normalized weight difference (dB)
Normalized weight difference (dB)
10-2
-3
2
4
Iterations
g
3
4
6
8
10
104
10
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-3
0
2
-4
1
2
Iterations
h
4
6
8
10
104
Iterations
3
4
30 (dB) SNR and 1e -3 Learning rate
100
10-3
0
10000
f
10-2
10
8000
10-2
10-4
5
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
-1
6000
10-1
104
20 (dB) SNR and 1e -3 Learning rate
100
4000
30 (dB) SNR and 1e -2 Learning rate
100
e
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
2
Iterations
10 (dB) SNR and 1e -3 Learning rate
0
2000
c
10-2
d
10
0
Iterations
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
Iterations
100
10-3
10-4
5000
20 (dB) SNR and 1e -2 Learning rate
100
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
0
4000
10-2
b
10-1
10-3
3000
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
Iterations
a
100
2000
Normalized weight difference (dB)
10-2
100
30 (dB) SNR and 1e -1 Learning rate
100
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
Normalized weight difference (dB)
Normalized weight difference (dB)
100
20 (dB) SNR and 1e -1 Learning rate
101
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
Normalized weight difference (dB)
10 (dB) SNR and 1e -1 Learning rate
101
5
105
LMS / qLMS at q=1
qLMS at q=2
NLMS
Time varing qLMS
EqLMS (Proposed)
10-1
10-2
10-3
10-4
10-5
0
2
4
Iterations
6
8
10
105
i
Figure 3: NWD curves for the LMS/q-LMS at q = 1, q-LMS at q = 2, time-varying q-LMS,
NLMS, and the Eq-LMS. normalized weight deviation with learning rate and SNR of (a) 1e−1 ,
10 dB, (b) 1e−1 , 20 dB,(c) 1e−1 , 30 dB, (d) 1e−2 , 10 dB, (e) 1e−2 , 20 dB, (f) 1e−2 , 30 dB,
(g) 1e−3 , 10 dB, (h) 1e−3 , 20 dB, and (i) 1e−3 , 30 dB.
17
scenarios, the proposed Eq-LMS algorithm comprehensively outperformed the
contemporary approaches achieving the best performance in terms of steadystate error and convergence.
305
References
[1] Y. Chen, Y. Gu, A. O. Hero, Sparse LMS for system identification, in: IEEE
International Conference on Acoustics, Speech and Signal Processing, 2009.
ICASSP 2009., IEEE, 2009, pp. 3125–3128.
310
315
[2] T. Abbas, S. Khan, M. Sajid, A. Wahab, J. C. Ye, Topological sensitivity based far-field detection of elastic inclusions, Results
in Physics, Volume 8, March 2018, Pages 442-460, ISSN 2211-3797,
https://doi.org/10.1016/j.rinp.2017.12.041.
[3] J. M. Górriz, J. Ramı́rez, S. Cruces-Alvarez, C. G. Puntonet, E. W. Lang,
D. Erdogmus, A novel LMS algorithm applied to adaptive noise cancellation, IEEE Signal Processing Letters 16 (1) (2009) 34–37.
[4] J. Benesty, S. L. Gay, An improved PNLMS algorithm, in: IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP),
2002, Vol. 2, IEEE, 2002, pp. II–1881.
320
[5] N. V. Thakor, Y.-S. Zhu, Applications of adaptive filtering to ECG analysis: noise cancellation and arrhythmia detection, IEEE Transactions on
Biomedical Engineering 38 (8) (1991) 785–794.
[6] H. Ammari, E. Bretin, J. Garnier, H. Kang, H. Lee, A. Wahab, Mathematical methods in elasticity imaging, Princeton University Press, 2015.
325
[7] Y. Zheng, S. Wang, J. Feng, C. K. Tse, A modified quantized kernel least
mean square algorithm for prediction of chaotic time series, Digital Signal
Processing 48 (Supplement C) (2016) 130 – 136. doi:https://doi.org/
10.1016/j.dsp.2015.09.015.
[8] S. C. Douglas, A family of normalized LMS algorithms, IEEE Signal Processing Letters 1 (3) (1994) 49–51.
330
335
[9] W. Wang, H. Zhao, Boxed-constraint least mean square algorithm and its
performance analysis, Signal Processing 144 (Supplement C) (2018) 201 –
213. doi:https://doi.org/10.1016/j.sigpro.2017.10.006.
[10] S. Cheng, Y. Wei, Y. Chen, Y. Li, Y. Wang, An innovative fractional order
LMS based on variable initial value and gradient order, Signal Processing
133 (2017) 260–269.
[11] S. Ciochin, C. Paleologu, J. Benesty, An optimized NLMS algorithm for
system identification, Signal Processing 118 (2016) 115 – 121.
18
340
[12] U. M. Al-Saggaf, M. Moinuddin, M. Arif, A. Zerguine, The q-Least Mean
Squares algorithm, Signal Processing 111 (Supplement C) (2015) 50 – 60.
doi:https://doi.org/10.1016/j.sigpro.2014.11.016.
[13] T. Aboulnasr, K. Mayyas, A robust variable step-size LMS-type algorithm:
analysis and simulations, IEEE Transactions on Signal Processing 45 (3)
(1997) 631–639.
345
[14] R. H. Kwong, E. W. Johnston, A variable step size LMS algorithm, IEEE
Transactions on Signal Processing 40 (7) (1992) 1633–1642. doi:10.1109/
78.143435.
[15] B. Widrow, J. McCool, M. Ball, The complex LMS algorithm, Proceedings
of the IEEE 63 (4) (1975) 719–720. doi:10.1109/PROC.1975.9807.
350
[16] A. Khalili, A. Rastegarnia, W. M. Bazzi, Z. Yang, Derivation and analysis
of incremental augmented complex least mean square algorithm, IET Signal
Processing 9 (4) (2015) 312–319.
[17] A. Khalili, A. Rastegarnia, S. Sanei, Quantized augmented complex leastmean square algorithm: Derivation and performance analysis, Signal Processing 121 (2016) 54–59.
355
360
[18] J. Ahmad, S. Khan, M. Usman, I. Naseem, M. Moinuddin, FCLMS: Fractional Complex LMS Algorithm for Complex System Identification, in: 13th
IEEE Colloquium on Signal Processing and its Applications (CSPA 2017),
IEEE, 2017.
[19] W. Liu, P. P. Pokharel, J. C. Principe, The kernel least-mean-square algorithm, IEEE Transactions on Signal Processing 56 (2) (2008) 543–554.
[20] P. P. Pokharel, W. Liu, J. C. Principe, Kernel least mean square algorithm
with constrained growth, Signal Processing 89 (3) (2009) 257 – 265. doi:
https://doi.org/10.1016/j.sigpro.2008.08.009.
365
370
[21] B. Chen, S. Zhao, P. Zhu, J. C. Principe, Quantized kernel least mean
square algorithm, IEEE Transactions on Neural Networks and Learning
Systems 23 (1) (2012) 22–32.
[22] S. Khan, M. Usman, I. Naseem, R. Togneri, M. Bennamoun, A Robust Variable Step Size Fractional Least Mean Square (RVSS-FLMS) Algorithm, in:
13th IEEE Colloquium on Signal Processing and its Applications (CSPA
2017), IEEE, 2017.
[23] S. Khan, M. Usman, I. Naseem, R. Togneri, M. Bennamoun, VP-FLMS: a
Novel Variable Power Fractional LMS Algorithm, in: 2017 Ninth International Conference on Ubiquitous and Future Networks (ICUFN) (ICUFN
2017), Milan, Italy, 2017.
19
375
380
[24] J. Ahmad, M. Usman, S. Khan, I. Naseem, H. J. Syed, RVP-FLMS : A
Robust Variable Power Fractional LMS Algorithm, in: 2016 IEEE International Conference on Control System, Computing and Engineering (ICCSCE), IEEE, 2016.
[25] N. J. Bershad, F. Wen, H. C. So, Comments on “Fractional LMS algorithm”, Signal Processing 133 (2017) 219–226.
[26] M. Arif, I. Naseem, M. Moinuddin, S. S. Khan, M. M. Ammar, Adaptive noise cancellation using q-LMS, in: 2017 International Conference
on Innovations in Electrical Engineering and Computational Technologies
(ICIEECT), 2017, pp. 1–4. doi:10.1109/ICIEECT.2017.7916527.
385
390
395
[27] U. M. Al-Saggaf, M. Moinuddin, A. Zerguine, An efficient least mean
squares algorithm based on q-gradient, in: 2014 48th Asilomar Conference on Signals, Systems and Computers, 2014, pp. 891–894. doi:
10.1109/ACSSC.2014.7094580.
[28] A. U. Al-Saggaf, M. Arif, U. M. Al-Saggaf, M. Moinuddin, The qnormalized least mean square algorithm, in: 2016 6th International Conference on Intelligent and Advanced Systems (ICIAS), 2016, pp. 1–6.
doi:10.1109/ICIAS.2016.7824098.
[29] A. Ahmed, M. Moinuddin, U. M. Al-Saggaf, q-State Space Least Mean
Family of Algorithms, Circuits, Systems, and Signal Processing. doi:10.
1007/s00034-017-0569-7.
[30] M. Arif, I. Naseem, M. Moinuddin, U. M. Al-Saggaf, Design of an Intelligent
q-LMS Algorithm for Tracking a Non-stationary Channel, Arabian Journal
for Science and Engineering. doi:10.1007/s13369-017-2883-6.
400
[31] T. Ernst, A Comprehensive Treatment of q-Calculus, 1st Edition,
Springer Science & Business Media, Basel, 2012. doi:10.1007/
978-3-0348-0431-8.
[32] G. Bangerezako, Variational q-calculus, Journal of Mathematical Analysis
and Applications 289 (2) (2004) 650 – 665. doi:https://doi.org/10.
1016/j.jmaa.2003.09.004.
405
410
[33] J. Tariboon, S. K. Ntouyas, P. Agarwal, New concepts of fractional quantum calculus and applications to impulsive fractional q-difference equations, Advances in Difference Equations 2015 (1) (2015) 18. doi:10.1186/
s13662-014-0348-8.
[34] J. Tariboon, S. K. Ntouyas, Quantum calculus on finite intervals and applications to impulsive difference equations, Advances in Difference Equations
2013 (1) (2013) 282. doi:10.1186/1687-1847-2013-282.
[35] A. R. A. L. Ali, V. Gupta, R. P. Agarwal, A. Aral, V. Gupta, Applications
of q-Calculus in Operator Theory, Springer New York, 2013.
20
[36] V. Kac, P. Cheung, Quantum Calculus, Springer New York, 2012.
415
420
425
[37] S. Khan, N. Ahmed, M. A. Malik, I. Naseem, R. Togneri, M. Bennamoun,
FLMF: fractional least mean fourth algorithm for channel estimation in
non-gaussian environment, in: International Conference on Information
and Communications Technology Convergence 2017 (ICTC 2017), Jeju Island, Korea, 2017.
[38] P. A. C. Lopes, J. A. B. Gerald, A close to optimal adaptive filter for sudden
system changes, IEEE Signal Processing Letters 24 (11) (2017) 1734–1738.
doi:10.1109/LSP.2017.2757147.
[39] S. Khan, I. Naseem, R. Togneri, M. Bennamoun, A novel adaptive kernel
for the rbf neural networks, Circuits, Systems, and Signal Processing (2016)
1–15.
[40] S. Khan, J. Ahmad, I. Naseem, M. Moinuddin, A novel fractional gradientbased learning algorithm for recurrent neural networks, Circuits, Systems,
and Signal Processing (2017) 1–20.
430
[41] A. I. Sulyman, A. Zerguine, Convergence and steady-state analysis of a
variable step-size nlms algorithm, Signal Processing 83 (6) (2003) 1255–
1273.
21
| 3cs.SY
|
CO-RANK OF WEAKLY PARAFREE 3-MANIFOLD GROUPS
arXiv:1711.10659v1 [math.GT] 29 Nov 2017
SHELLY HARVEY† AND EAMONN TWEEDY
Abstract. Recall that a group is called large if it has a finite index subgroup which surjects onto a
non-abelian free group. By work of Agol and Cooper-Long-Reid, most 3-manifold groups are large;
in particular, the fundamental groups of hyperbolic 3-manifolds are large. In previous work, the
first author gave examples of closed, hyperbolic 3-manifolds with arbitrarily large first homology
rank but whose fundamental groups do not surject onto a non-abelian free group. We call a group
very large if it surjects onto a non-abelian free group. In this paper, we consider the question of
whether the groups of homology handlebodies - which are very close to being free - are very large.
We show that the fundamental group of W. Thurston’s tripus manifold, is not very large; it is known
to be weakly parafree by Stallings’ Theorem and large by the work of Cooper-Long-Reid since the
tripus is a hyperbolic manifold with totally geodesic boundary. It is still unknown if a 3-manifold
group that is weakly parafree of rank at least 3 must be very large. However, we more generally
consider the co-rank of the fundamental group, also known as the cut number of the manifold. For
each integer g ≥ 1 we construct a homology handlebody Yg of genus g whose group has co-rank
equal to r(g), where r(g) = g/2 for g even and r(g) = (g + 1)/2 for g odd. That is, these groups
are weakly parafree of rank g and surject onto a free group of rank roughly half of g but no larger.
1. Introduction
Let M be a compact, connected, orientable 3-dimensional manifold. If M is irreducible and has
non-empty incompressible boundary and M is not covered by a S 1 × S 1 × I, then it was shown by
D. Cooper, D. Long, and A. Reid [CLR97] that π1 (M ) is large. Recall that a group is said to be
large if it has a finite index subgroup that maps onto a non-abelian free group. In addition, using
work of F. Haglund and D. Wise, I. Agol1 [Ago13] showed that if M is closed and irreducible and
π1 (M ) is not finite or solvable then π1 (M ) is large (see also [AFW15, Diagram 4]). As a result,
we see that most 3-manifold groups are large. One could then ask what is the minimal size of the
index one needs or if even if π1 (M ) itself maps onto a non-abelian free group.
Definition 1.1. Let G be a group. We say that G is very large if it admits a surjective homomorphism onto a non-abelian free group.
In 2002, the first author gave examples of closed hyperbolic 3-manifold groups with arbitrarily
large first homology rank but that are not very large. By Agol’s theorem, these examples are all
large.
Theorem ([Har02]). For each n ≥ 1, there is a closed, orientable, hyperbolic 3-dimensional manifold Mn such β1 (Mn ) = n and π1 (Mn ) is not very large.
In fact, it was shown that the groups cannot even map onto F/F4 where F is the free group of
rank 2 and F4 is the 4th term of the lower central series of F [Har02, Proposition 3.3]. See also
work by C. Leinger and A. Reid [LR02], A. Sikora [Sik05]), and I. Gelbukh [Gel15, Gel17].
Key words and phrases. cut number, co-rank, 3-manifold, homology handlebody, fundamental group, large, weakly
parafree.
†
The author was partially supported by National Science Foundation grants DMS-1309070 and DMS-1613279
and by a grant from the Simons Foundation (#304538 to Shelly Harvey).
1Agol proved the case when M is hyperbolic; the other cases were known. See [AFW15] for more details.
1
2
SHELLY HARVEY† AND EAMONN TWEEDY
In this article, we are interested in the fundamental groups of 3-dimensional homology handlebodies. Their fundamental groups are very close to being free; they are weakly parafree, and hence
the question of whether they are very large becomes much more subtle and interesting question.
Recall that M is a homology handlebody of genus g if it has the same homology groups as a
3-dimensional handlebody of genus g; that is, H̃1 (M ) ∼
= Zg and H̃i (M ) = 0 for i 6= 1.
Remark 1.2. If M is a homology handlebody of genus g and G = π1 (M ) then by Stallings’
Theorem [Sta65], G is weakly parafree of rank g; that is,
F (g)/F (g)k ∼
= G/Gk
for all k ≥ 1 where F (g) is the free group of rank g and Gk is the k th term of the lower central
series of G.
Thus, groups of homology handlebodies of genus g look quite similar to free groups of large rank
as g increases. In particular, one cannot hope to use the lower central series quotients to obstruct
being very large.
It was shown in [Ker01, p. 44] (see also Proposition 2.2 in [Har02]) that if the fundamental
group of M is very large then there is an infinite cyclic cover whose first homology has positive
rank as a left Z[t±1 ]-module. The examples in [Har02] were shown to have torsion (as a Z[t±1 ]
module) in first homology for every infinite cyclic cover and hence their groups could not be very
large. However, this obstruction also fails for homology handlebodies. In fact, the rank of the first
homology of any poly-torsion-free-abelian covering space of M is maximal. Recall that a group Γ
is poly-torsion-free-abelian (PTFA) if it admits a normal series {1} = G0 C G1 C · · · C Gn = Γ
such that each of the factors Gi+1 /Gi is torsion-free abelian. If A is a finitely generated left module
over ZΓ with Γ PTFA then it has a well defined rank as a ZΓ-module. Note that any free abelian
group is PTFA.
Remark 1.3. Let Γ be a PTFA group. If M is a 3-dimensional homology handlebody of genus
g, then by [COT03, Lemma 2.12] for any non-trivial homomorphism φ : π1 (M ) → Γ, the covering
space Mφ associated to φ satisfies
rankZΓ H1 (Mφ ) = g − 1.
We show in Lemma 2.2, that if π1 (M ) is very large then there is a Z2 covering space of M ,
Mφ , such that H1 (Mφ ) has a Z[Z2 ] summand. Using this, we show that W. Thurston’s tripus is
a homology handlebody of genus 2 whose group is not very large. It is known to be hyperbolic
manifold manifold and hence is large by Agol’s theorem.
Proposition 2.5. Let T be W. Thurston’s Tripus manifold. Then π1 (T ) is a weakly parafree
group of rank 2 that is large but not very large.
It is unknown if there is a homology handlebody of genus g ≥ 3 whose group is not very large.
All of the examples that the authors have considered with g ≥ 3 have been shown to be very large.
Question 1.4. If Y is a homology handlebody of genus g with g ≥ 3, is π1 (Y ) very large? More
generally, if G is a finitely presented group with H1 (G) ∼
= Zg for g ≥ 3 and H2 (G) = 0, is G very
large?
Often a homology handlebody group is very large. In this case, we ask what is the maximal
rank free group that arises as the quotient of the group. The cut number of M , c (M ), is defined
to be the maximal number of components of a compact, orientable, 2–sided surface F properly
embeddedWin M such that M r F is connected. Hence, for any n ≤ c (M ), we can construct a map
f : M → ni=1 S 1 such that the induced map on π1 is surjective. That is, there exists a surjective
map f∗ : π1 (M ) F (n), where F (n) is the free group of rank n. Conversely,
if we have any
Wn
1
epimorphism φ : π1 (M ) F (n), then we can find a map f : M → i=1 S such that f∗ = φ.
CO-RANK OF WEAKLY PARAFREE 3-MANIFOLD GROUPS
3
After making the f transverse to a non-wedge point xi on each S 1 , Fi = f −1 (xi ) will give n disjoint
surfaces F = ∪Fi with M r F connected. Hence one has the following elementary group-theoretic
characterization of c (M ), as in the closed case.
Theorem ([Jac72, Theorem 2.1]). The cut number c (M ) of M is the maximal integer n such that
there is a surjective homomorphism φ : π1 (M ) F (n) onto the free group of rank n.
The maximal rank of any free quotient of group is called its co-rank. Hence the co-rank of
π1 (M ) is the same as the cut number of M . This is also referred to as the inner rank of π1 (M )
[Lyn59] as well as the non-commutative first Betti number of M [AL86]. For each g, we
construct an example of 3-dimensional handlebody of genus g whose groups maps onto a free group
of rank r(g) (roughly half of g) but no larger.
Theorem 2.1. For each g ≥ 1, there is a compact, connected, orientable 3-dimensional homology
handlebody Yg of genus g with c(Yg ) = r(g) where
(
g
if g is even
2
r(g) = g+1
if g is odd.
2
We expect that this may be optimal.
Question 1.5. If Y is a compact, connected, orientable 3-dimensional homology handlebody of
genus g, is c(Y ) ≥ r(g)?
Acknowledgements. The first author would like to thank Michael Freedman for asking her about
the cut number of homology handlebodies. We would like to thank John Hempel, Neil Fullarton,
and Alan Reid for helpful conversations.
2. Main Theorem
The proof of Theorem 2.1 is almost immediate once we construct examples of homology handlebodies Y2 and Y3 of genus 2 and 3 repectively with c(Y2 ) = 1 and c(Y3 ) = 2. We let Y2 will
be Thurston’s tripus manifold T ; we prove π1 (T ) is not very large in Subsection 2.2. We let
Y3 = Y , a particular genus-3 string link complement; we construct Y and prove that c(Y ) = 2 in
Subsection 2.3.
Theorem 2.1. For each g ≥ 1, there is a compact, connected, orientable 3-dimensional homology
handlebody Yg of genus g with c(Yg ) = r(g) where
(
g
if g is even
2
r(g) = g+1
if g is odd.
2
Proof. For each integer g ≥ 1 we shall construct a homology handlebody Yg of genus g with
c(Yg ) = r(g). Because r(1) = 1, we may choose Y1 to be a handlebody. Let Y2 = T and Y3 = Y ,
where T is Thurston’s Tripus manifold (see Figure 1) and Y is the string link complement from
Figure 2. Then by Propositions 2.5 and 2.6, c(Y2 ) = 1 = r(2) and c(Y3 ) = 2 = r(3). If g ≥ 4, we
define Yg inductively as the boundary connected sum Yg = Yg−2 \Y2 . By Theorem 3.2 of [Jac72],
the cut number is additive under boundary connected sum and so c(Yg ) = c(Yg−2 ) + 1. Since it is
also the case that r(g) = r(g − 2) + 1, it follows by induction that c(Yg ) = r(g) for all g ≥ 0.
2.1. Preliminary notions and lemmas. Before we prove that c(Y2 ) = 1 and c(Y3 ) = 2, we shall
determine some obstructions to a group having a non-abelian free quotient. To do this, we look at
the first homology (and relative first homology) of free abelian covering spaces of the manifolds.
We note that the co-rank is known to be algorithmically computable [Mak82, Raz95] for finitely
presented groups; however, the algorithm seems difficult to use in practice.
SHELLY HARVEY† AND EAMONN TWEEDY
4
Let X be a path connected topological space and φ : π1 (X) → Γ be a surjective group homoπ
morphism. We define Xφ −
→ X to be the covering space of X corresponding to φ. Recall that the
group of deck translations of Xφ → X is identified with Γ making H1 (Xφ ) into a left ZΓ-module.
Ker(φ)
Moreover, for any group G and surjective homomorphism φ : G → Γ, [Ker(φ),Ker(φ)]
has the structure of a left ZΓ-module
where the module action is given as follows. Let γ ∈ Γ and g ∈ Ker(φ).
Then γ [g] := hgh−1 for any h ∈ G such that φ(h) = γ (here [g] denotes the equivalence class
Ker(φ)
as left
of g). In the case that G = π1 (X), we have that H1 (Xφ ) is isomorphic to [Ker(φ),Ker(φ)]
ZΓ-modules.
Lemma 2.2. Let G be a group and Γ = Zm with m ∈ {1, 2}. If G is very large, then there exists
a surjective homomorphism φ : G → Γ such that
Ker(φ)
∼
= ZΓ ⊕ A
[Ker(φ), Ker(φ)]
for some left ZΓ-module A. In particular, if M is a compact, connected, orientable 3-manifold with
π1 (M ) very large, then there is a surjective homomorphism φ : π1 (M ) → Γ such that H1 (Mφ ) has
a ZΓ summand.
Proof. Since G is very large, there is a surjective homomorphism ψ 0 : G F where F is the free
group of rank 2. Let ψ : F Γ be a surjective homomorphism and φ = ψ ◦ ψ 0 . Then ψ induces a
surjective ZΓ-module homomorphism
ψ:
Since
Ker(ψ)
[Ker(ψ),Ker(ψ)]
Ker(φ)
Ker(ψ)
.
[Ker(φ), Ker(φ)]
[Ker(ψ), Ker(ψ)]
∼
= ZΓ is a free ZΓ module, the result follows.
Ker(ψ)
We note that in the previous proof that if m ≥ 3 then [Ker(ψ),Ker(ψ)]
is no longer a free module;
in fact, it is not even projective. Thus, to generalize the previous lemma, we work with relative
homology. Given p ∈ M , let pφ denote the preimage π −1 (p) ⊂ Mφ . The relative homology group
H1 (Mφ , pφ ) has the structure of a left ZΓ-module as before, coming from group of deck translations
acting on the pair (Mφ , pφ ). We have the following generalization of Lemma 2.2, which no longer
restricts us to using free abelian covering spaces.
Lemma 2.3. Let M be a compact, connected, orientable 3-manifold, let p ∈ M , and let Γ be a
quotient of F (n), the non-abelian free group of rank n for some n ≤ c(M ). Then there is a surjective
homomorphism φ : π1 (M ) → Γ such that
H1 (Mφ , pφ ) ∼
= ZΓn ⊕ A
for some left ZΓ-module A.
Proof. Since c(M ) ≥ n, there is an epimorphism ψ 0 : π1 (M, p) F (n), the free group of rank
n. Letting W denote the n-fold wedge of circles and w ∈ W the wedge point, there exists a map
f : M → W such that f (p) = w and such that the induced map on π1 is ψ 0 . Let ψ : π1 (W, w) → Γ
be a surjective map and φ = ψ ◦ ψ 0 . The map f can be lifted to a map g : Mφ → Wψ sending pφ
to wψ .
We examine the following commutative diagram of modules and module homomorphisms over
ZΓ, where the rows are portions of the long exact sequences of the pairs and all of the vertical
maps are the module homomorphisms on homology induced by g.
CO-RANK OF WEAKLY PARAFREE 3-MANIFOLD GROUPS
H1 (Mφ )
g1
H1 (Wψ )
H1 (Mφ , pφ )
g2
H1 (Wψ , wψ )
H0 (pφ )
g3
H0 (wψ )
5
H0 (Mφ )
g4
H0 (Wψ )
Note first that g4 is an isomorphism and g3 is an epimorphism. Also, since ψ 0 is an epimorphism
it follows that g1 is as well. Together these facts imply that g2 is an epimorphism as well. Since
H1 (Wψ , wψ ) is a free module of rank n, the result follows.
The following easy lemma is most likely well-known but we include it for completeness. This will
be used to prove that the modules arising in the proofs Propositions 2.5 and 2.6 are torsion free.
Lemma 2.4. Suppose R is a unique factorization domain and M is a left module over R with a
presentation of the form
hα1 , . . . , αn | p1 α1 + . . . pn αn i
where pk ∈ R for each k. Then:
(i) If pi and pj are relatively prime for some i 6= j, then M is torsion-free.
(ii) If there is a common factor different than 1 which is mutually shared by all of p1 , . . . , pn , then
M is not torsion-free.
In particular, for n = 2, M is torsion free if and only if p1 and p2 are relatively prime.
Proof. Statement (ii) is obvious. We proceed to prove statement (i). Suppose M has a torsion
element, f1 α1 + . . . + fn αn 6= 0 with fk ∈ R for each k. Then there exists r, s ∈ R (with s non-zero
and not a unit) such that s(f1 α1 + . . . + fn αn ) = r(p1 α1 + . . . pn αn ) in the free R module generated
by α and β. Hence sfk = rpk for each k. We can assume that s and r have no common factors,
otherwise, we could reduce them. So s much divide pk for each k. Hence no pair pi , pj can be
relatively prime.
2.2. Thurston’s tripus manifold has cut number equal to 1. Let T be W. Thurston’s tripus
manifold. That is, T is the complement of the three arcs in S 2 × I as shown in Figure 1. Smoothly
embed S 2 × I in S 3 . Then we see that T is the complement of an embedding of a θ graph in S 3
(where the exterior of S 2 × I can be identified with with the two vertices of the spatial graph).
Note that the tripus is a hyperbolic manifold with totally geodesic boundary and is also a genus 2
homology handlebody.
Proposition 2.5. Let T be W. Thurston’s Tripus manifold. Then π1 (T ) is a weakly parafree group
that is large but not very large.
Proof. Since T is a hyperbolic manifold with totally geodesic boundary [Thu97, Chapter 3], it is
irreducible and has incompressible boundary. Thus by [CLR97], π1 (T ) has a finite index subgroup
with a non-abelian free quotient. That is, π1 (T ) is large.
To show that π1 (T ) is not very large, we use Lemma 2.2. We begin by computing the first
homology of the universal abelian cover of T . We can use the Wirtinger presentation to write down
a presentation for π1 (T ):
π1 (T ) ∼
= a, b, c, d, e, f | ace, bdf, (f c)b(f c)−1 a−1 , (be)d(be)−1 c−1 , (da)f (da)−1 e−1 .
Using the relation that e = (ac)−1 and f = (bd)−1 , we get a simplified presentation:
π1 (T ) ∼
= a, b, c, d | (d−1 b−1 c)b(d−1 b−1 c)−1 a−1 , (bc−1 a−1 )d(bc−1 a−1 )−1 c−1 , (da)d−1 b−1 (da)−1 ac .
6
SHELLY HARVEY† AND EAMONN TWEEDY
Figure 1. Thurston’s Tripus, a hyperbolic manifold with totally geodesic boundary
In fact, the third relation above can be derived from the first two, and thus we can use the
presentation
π1 (T ) ∼
= a, b, c, d | (d−1 b−1 c)b(d−1 b−1 c)−1 a−1 , (bc−1 a−1 )d(bc−1 a−1 )−1 c−1 .
We now change the generating set to {a, B, c, D} where B = ba−1 and D = dc−1 . Let φ :
π1 (T ) → H1 (T ) be the abelianization map. We note that the image of two of the generators, a
and c, under φ give a basis for H1 (T ). To make the notation easier, we will call φ(a) (respectively
φ(c)), a (respectively c). Let Θ = aca−1 c−1 . Since B and D are trivial under φ, they lift to Tφ
and the set {B, D, Θ} generates H1 (Tφ ) as a Z[H1 (T )]-module. It is straightforward to write down
a presentation matrix of H1 (Tφ ) as a Z[a±1 , c±1 ]-module (here the rows are the relations and the
columns correspond to B, D, and Θ):
a + c − 1 a(a − 1) a − 1
.
c(1 − c)
1
c−1
Using row and column operations, we can find a simpler
a + c − 1 a(a − 1) a − 1
ac + a − 1
∼
c(1 − c)
1
c−1
0
ac + a − 1
∼
0
ac + a − 1
∼
0
∼ ac + a − 1
presentation:
a2 − a a − 1
1
c−1
a2 − a −a2 c − a2 − ac + a
1
c−1
0 −2a2 c
1 c−1
2a2 c
Since gcd(ac + a − 1, 2a2 c) = 1, Lemma 2.4 implies that H1 (Tφ ) is torsion-free as a H1 (T )-module.
We claim that H1 (Tφ ) is not a free module. To see this, consider the ring homomorphism ξ :
Z[H1 (T )] → Z2 [a±1 , c±1 ] where we reduce the coefficient mod 2. This map endows Z2 [a±1 , c±1 ]
with the structure of a right Z[H1 (T )]-module. If H1 (Tφ ) were free as Z[H1 (T )]-module then the
CO-RANK OF WEAKLY PARAFREE 3-MANIFOLD GROUPS
7
left Z2 [a±1 , c±1 ]-module
Z2 [a±1 , c±1 ] ⊗Z[H1 (T )] H1 (Tφ )
would be free as well. However, this tensor product module is presented by the matrix [ac+a+1, 0],
and is thus not free.
Suppose that π1 (T ) is very large. We note that there is a unique covering space of T with deck
transformation group isomorphic to Z2 , up to covering space isomorphism. Hence by Lemma 2.2,
H1 (Tφ ) ∼
= Z[H1 (T )] ⊕ A. Since the rank of H1 (Tφ ) (as a Z[H1 (T )]-module) in our case is 1, it
follows that A must have rank 0 and hence is a torsion module. However, we know that H1 (Tφ )
is torsion-free so A = 0. This contradicts the fact that H1 (Tφ ) is not a free module and the result
follows.
2.3. A genus-3 homology handlebody with cut number equal to 2.
Proposition 2.6. There is a genus 3 homology handlebody Y with c(Y ) = 2.
x
r
a
y
b
c
d
z
s
w
e
f
u
g
t
Figure 2. The homology handlebody Y is the complement of this pure 3-string
link. Generators of π1 (Y ) are labelled.
Proof. Let Y denote the homology handlebody that is the complement in D2 × I of the threecomponent pure string link show in Figure 2. Removing the middle string leaves of with the trivial
string link whose complement is a handlebody. Hence there is a surjective map π1 (Y ) → F (2),
hence c(Y ) ≥ 2.
A presentation of π1 (Y ) can be computed via the Wirtinger presentation:
π1 (Y ) ∼
=
a, b, c, d, e, f, g, zuz −1 w−1 , eze−1 w−1 , xzx−1 y −1 , bxb−1 y −1 , f sf −1 t−1 , crc−1 s−1 ,
r, s, t, x, y, z, u, w ege−1 f −1 , tet−1 f −1 , wdw−1 e−1 , bdb−1 c−1 , sbs−1 c−1 , yay −1 b−1
First notice that the generators a, g, r, u each appear in only one relation, and can thus be
eliminated along with those relations to produce the following smaller presentation:
b, c, d, e, f, bxb−1 y −1 , xzx−1 y −1 , eze−1 w−1 , tet−1 f −1 ,
∼
π1 (Y ) =
s, t, x, y, z, w sbs−1 c−1 , f sf −1 t−1 , bdb−1 c−1 , wdw−1 e−1
SHELLY HARVEY† AND EAMONN TWEEDY
8
We shall obtain an even simpler presentation by combining relations. The first four relations
imply that
x = b−1 yb = b−1 xzx−1 b = b−1 xe−1 wex−1 b = b−1 xt−1 f −1 twt−1 f tx−1 b
Similarly, the last five relations imply that b = f −1 t−1 f bw−1 t−1 f twb−1 f −1 tf . These observations
lead us to a presentation with seven generators and only two relations:
π1 (Y ) ∼
= b, f, t, x, w | b−1 xt−1 f −1 twt−1 f tx−1 bx−1 , f −1 t−1 f bw−1 t−1 f twb−1 f −1 tf b−1
Consider the abelianization map φ : π1 (Y ) → H1 (Y ). Note that φ(x) = φ(w) and φ(b) = φ(f ),
so that the images of b, t, and x give a basis for H1 (Y ). Considering the universal abelian cover
Yφ , we choose a basepoint p ∈ Y and let pφ = φ−1 (p) ∈ Yφ . Using the Fox differential calculus,
we can obtain a presentation for the structure of H1 (Yφ , pφ ) as a left module over Z[H1 (Y )] ∼
=
Z[b±1 , t±1 , x±1 ]. With respect to the generators b, f, t, x, w, a presentation matrix is
(2.7)
btx − bt
x2 − x
bx2 − bx − x2 + x bt − b2 t − btx
tx
btx − b2 tx − bt2 x bt2 x − t2 x − btx + tx + b2 btx − tx + b3 − b2
0
b3 t − b2 t
We first claim that this module is not free. We use ξ : H1 (Y ) → Z3 [b±1 , t±1 , x±1 ], given by
reducing coefficients modulo 3, to endow Z3 [b±1 , t±1 , x±1 ] with the structure of a right Z[H1 (Y )]module. If H1 (Yφ , pφ ) were a free Z[H1 (Y )]-module, then the left Z3 [b±1 , t±1 , x±1 ]-module
Z3 [b±1 , t±1 , x±1 ] ⊗Z[H1 (Y )] H1 (Yφ , pφ )
would also be free.
Recall that given a rank r module, N , over a multivariable Laurent polynomial ring R, is projective if and only if the rth elementary ideal Er (N ) of N is equal to all of R. We used the
Magma Computational Algebra System [BCP97] to verify that the ideal generated by the 2 × 2
minors of the above matrix, viewed modulo 3, is indeed proper in Z3 [b±1 , t±1 , x±1 ]. Therefore the
Z3 [b±1 , t±1 , x±1 ]-module
Z3 [b±1 , t±1 , x±1 ] ⊗Z[H1 (Y )] H1 (Yφ , pφ )
is not projective and hence not free. It follows that H1 (Yφ , pφ ) is not a free Z[H1 (Y )]-module.
Our Magma source code can be found below. Note that here B = b−1 , T = t−1 , and X = x−1 .
P<x , X, b , B, t , T> := PolynomialRing (GF( 3 ) , 6 ) ;
Q<x , X, b , B, t , T> := quo<P | x∗X−1, t ∗T−1, b∗B−1>;
M := Matrix (Q, 2 , 5 , [ b∗ t ∗ ( x −1) , x ∗ ( x −1) , x ∗ ( b −1)∗(x −1) ,
b∗ t ∗(1−x−b ) , t ∗x , b∗ t ∗x∗(1−b−t ) ,
−t ˆ2∗ x+t ∗x+bˆ2−b∗ t ∗x+b∗ t ˆ2∗x ,
−t ∗x−bˆ2+bˆ3+b∗ t ∗x , 0 , bˆ2∗ t ∗ ( b − 1 ) ] ) ;
Mi := Minors (M, 2 ) ;
I := i d e a l <Q | Mi>;
IsProper ( I )
We now claim that the module H1 (Yφ , pφ ) is torsion-free. After multiplying the second row of
the matrix (2.7) by x, adding b2 − b3 times the first row to the second row, eliminating a generator
and relation, and rescaling several entries by units we obtain a smaller presentation matrix:
CO-RANK OF WEAKLY PARAFREE 3-MANIFOLD GROUPS
9
T
−b3 x + b3 + b2 x − bx2 − tx2 − b2 + x2
−b3 x + bt2 x + b3 + b2 x − btx − t2 x + tx
4
−b x + b4 + 2b3 x − b3 − b2 x + btx − tx
b2 + bx − 2b − x + 1
(2.8)
Now consider the ring homomorphism η : Z[H1 (Y )] → Z[t±1 , x±1 ] generated by the homomorphism H1 (Y ) →< t, x > sending b and t to t and x to x. As before η induces a left Z[t±1 , x±1 ]module structure on
Z[t±1 , x±1 ] ⊗Z[H1 (Y )] H1 (Yφ , pφ )
which has torsion if H1 (Yφ , pφ ) has torsion. We obtain a presentation matrix for the above tensor
module by setting b = t in (2.8) and rescaling several entries by units:
3
T
−t x + t3 + t2 x − 2tx2 − t2 + x2
t2 − tx + x
3
3
2
2
−t x + t + 2t x − t − x
2
t + tx − 2t − x + 1
We verified in Magma that the first two entries of this matrix (in fact, any two entries) are relatively
prime. By Lemma 2.4, the module is torsion-free.
Thus, H1 (Yφ , pφ ) is a torsion-free module of rank 3. Suppose that π1 (Y ) maps onto a free group
of rank 3. Since β1 (Y ) = 3, there is a unique Z3 covering space (up to covering space isomorphism)
so by Lemma 2.3, H1 (Yφ , pφ ) ∼
= Z[H1 (Y )]3 ⊕ A for some A. Since H1 (Yφ , pφ ) is rank 3, it follows
that A must be a torsion module. However H1 (Yφ , pφ ) is torsion free so A = 0. This contradicts
the fact that H1 (Yφ , pφ ) is not free and it follows that c(Y ) = 2.
References
[AFW15] Matthias Aschenbrenner, Stefan Friedl, and Henry Wilton. 3-manifold groups. EMS Series of Lectures in
Mathematics. European Mathematical Society (EMS), Zürich, 2015.
[Ago13] Ian Agol. The virtual Haken conjecture. Doc. Math., 18:1045–1087, 2013. With an appendix by Agol, Daniel
Groves, and Jason Manning.
[AL86]
Pierre Arnoux and Gilbert Levitt. Sur l’unique ergodicité des 1-formes fermées singulières. Invent. Math.,
84(1):141–156, 1986.
[BCP97] Wieb Bosma, John Cannon, and Catherine Playoust. The Magma algebra system. I. The user language. J.
Symbolic Comput., 24(3-4):235–265, 1997. Computational algebra and number theory (London, 1993).
[CLR97] D. Cooper, D. D. Long, and A. W. Reid. Essential closed surfaces in bounded 3-manifolds. J. Amer. Math.
Soc., 10(3):553–563, 1997.
[COT03] Tim D. Cochran, Kent E. Orr, and Peter Teichner. Knot concordance, Whitney towers and L2 -signatures.
Ann. of Math. (2), 157(2):433–519, 2003.
[Gel15] Irina Gelbukh. Co-rank and Betti number of a group. Czechoslovak Math. J., 65(140)(2):565–567, 2015.
[Gel17] Irina Gelbukh. The co-rank of the fundamental group: the direct product, the first Betti number, and the
topology of foliations. Math. Slovaca, 67(3):645–656, 2017.
[Har02] Shelly L. Harvey. On the cut number of a 3-manifold. Geom. Topol., 6:409–424, 2002.
[Jac72]
William Jaco. Geometric realizations for free quotients. J. Austral. Math. Soc., 14:411–418, 1972.
[Ker01] Thomas Kerler. Homology TQFT’s and the Alexander-Reidemeister invariant of 3-manifolds
via Hopf algebras and skein theory. 2001. Version 3 of the arxiv verion of this paper,
https://arxiv.org/pdf/math/0008204v3.pdf.
[LR02]
Christopher J. Leininger and Alan W. Reid. The co-rank conjecture for 3-manifold groups. Algebr. Geom.
Topol., 2:37–50, 2002.
[Lyn59] R. C. Lyndon. The equation a2 b2 = c2 in free groups. Michigan Math. J, 6:89–95, 1959.
[Mak82] G. S. Makanin. Equations in a free group. Izv. Akad. Nauk SSSR Ser. Mat., 46(6):1199–1273, 1344, 1982.
[Raz95] Alexander A. Razborov. On systems of equations in free groups. In Combinatorial and geometric group
theory (Edinburgh, 1993), volume 204 of London Math. Soc. Lecture Note Ser., pages 269–283. Cambridge
Univ. Press, Cambridge, 1995.
[Sik05]
Adam S. Sikora. Cut numbers of 3-manifolds. Trans. Amer. Math. Soc., 357(5):2007–2020, 2005.
SHELLY HARVEY† AND EAMONN TWEEDY
10
[Sta65]
[Thu97]
John Stallings. Homology and central series of groups. J. Algebra, 2:170–181, 1965.
William P. Thurston. Three-dimensional geometry and topology. Vol. 1, volume 35 of Princeton Mathematical Series. Princeton University Press, Princeton, NJ, 1997. Edited by Silvio Levy.
Rice University, Houston, TX, USA
E-mail address: [email protected]
Widener University, Philadelphia, PA, USA
E-mail address: [email protected]
| 4math.GR
|
Safe Non-blocking Synchronization in Ada 202x
Johann Blieberger1 and Bernd Burgstaller2
arXiv:1803.10067v1 [cs.PL] 27 Mar 2018
1
Institute of Computer Engineering, Automation Systems Group, TU Wien, Austria
2
Department of Computer Science, Yonsei University, Korea
Abstract. The mutual-exclusion property of locks stands in the way to
scalability of parallel programs on many-core architectures. Locks do not
allow progress guarantees, because a task may fail inside a critical section
and keep holding a lock that blocks other tasks from accessing shared
data. With non-blocking synchronization, the drawbacks of locks are
avoided by synchronizing access to shared data by atomic read-modifywrite operations.
To incorporate non-blocking synchronization in Ada 202x, programmers
must be able to reason about the behavior and performance of tasks in
the absence of protected objects and rendezvous. We therefore extend
Ada’s memory model by synchronized types, which support the expression of memory ordering operations at a sufficient level of detail. To mitigate the complexity associated with non-blocking synchronization, we
propose concurrent objects as a novel high-level language construct. Entities of a concurrent object execute in parallel, due to a fine-grained, optimistic synchronization mechanism. Synchronization is framed by the semantics of concurrent entry execution. The programmer is only required
to label shared data accesses in the code of concurrent entries. Labels
constitute memory-ordering operations expressed through attributes. To
the best of our knowledge, this is the first approach to provide a nonblocking synchronization construct as a first-class citizen of a high-level
programming language. We illustrate the use of concurrent objects by
several examples.
1
Introduction
Mutual exclusion locks are the most common technique to synchronize multiple tasks to access shared data. Ada’s protected objects (POs) implement the
monitor-lock concept [12]. Method-level locking requires a task to acquire an
exclusive lock to execute a PO’s entry or procedure. (Protected functions allow
concurrent read-access in the style of a readers–writers lock [11].) Entries and
procedures of a PO thus effectively execute one after another, which makes it
straight-forward for programmers to reason about updates to the shared data
encapsulated by a PO. Informally, sequential consistency ensures that method
calls act as if they occurred in a sequential, total order that is consistent with
the program order of each participating task. I.e., for any concurrent execution,
the method calls to POs can be ordered sequentially such that they (1) are consistent with program order, and (2) meet each PO’s specification (pre-condition,
side-effect, post-condition) [11].
2
Although the sequential consistency semantics of mutual exclusion locks facilitate reasoning about programs, they nevertheless introduce potential concurrency bugs such as dead-lock, live-lock and priority inversion. The mutualexclusion property of (highly-contended) locks stands in the way to scalability of
parallel programs on many-core architectures [15]. Locks do not allow progress
guarantees, because a task may fail inside a critical section (e.g., by entering an
endless loop), preventing other tasks from accessing shared data.
Given the disadvantages of mutual exclusion locks, it is thus desirable to
give up on method-level locking and allow method calls to overlap in time. Synchronization is then performed on a finer granularity within a method’s code, via
atomic read-modify-write (RMW) operations. In the absence of mutual exclusion
locks, the possibility of task-failure inside a critical section is eliminated, because
critical sections are reduced to single atomic operations. These atomic operations
are provided either by the CPU’s instruction set architecture (ISA), or the language run-time (with the help of the CPU’s ISA). It thus becomes possible to
provide progress guarantees, which are unattainable with locks. In particular, a
method is non-blocking, if a task’s pending invocation is never required to wait
for another task’s pending invocation to complete [11].
Non-blocking synchronization techniques are notoriously difficult to implement and the design of non-blocking data structures is an area of active research.
To enable non-blocking synchronization, a programming language must provide
a strict memory model. The purpose of a memory model is to define the set of
values a read operation in a program is allowed to return [2].
To motivate the need for a strict memory model, consider the producerconsumer synchronization example in Fig. 1(a) (adopted from [17] and [5]). The
programmer’s intention is to communicate the value of variable Data from Task 1
to Task 2. Without explicitly requesting a sequentially consistent execution, a
compiler or CPU may break the programmer’s intended synchronization via the
Flag variable by re-ordering memory operations that will result in reading R2
= 0 in Line 6 of Task 2. (E.g., a store–store re-ordering of the assignments in
lines 2 and 3 of Task 1 will allow this result.) In Ada 2012, such re-orderings
1
2
3
-- Initial values :
Flag := False ;
Data := 0;
1
1
2
3
2
-- Task 1:
Data := 1;
Flag := True ;
3
4
5
6
(a)
-- Task 2:
loop
R1 := Flag ;
exit when R1 ;
end loop ;
R2 := Data ;
1
2
Data : Integer w i t h Volatile ;
Flag : Boolean w i t h Atomic ;
(b)
Fig. 1: (a) Producer-consumer synchronization in pseudo-code: Task 1 writes the
Data variable and then signals Task 2 by setting the Flag variable. Task 2 is
spinning on the Flag variable (lines 2 to 5) and then reads the Data variable.
(b) Labeling to enforce sequential consistency in Ada 2012.
3
can be ruled out by labeling variables Data and Flag by aspect volatile. The
corresponding variable declarations are depicted in Fig. 1(b). (Note that by [8,
C.6§8/3] aspect atomic implies aspect volatile, but not vice versa.)
The intention for volatile variables in Ada 2012 was to guarantee that all
tasks agree on the same order of updates [8, C.6§16/3]. Updates of volatile
variables are thus required to be sequentially consistent, in the sense of Lamport’s
definition [13]: “With sequential consistency (SC), any execution has a total order
over all memory writes and reads, with each read reading from the most recent
write to the same location”.
However, the Ada 2012 aspect volatile has the following shortcomings:
1. Ensuring SC for multiple tasks without atomic access is impossible. Nonatomic volatile variables therefore should not be provided by the language.
Otherwise, the responsibility shifts from the programming language implementation to the programmer to ensure SC by pairing an atomic (implied
volatile) variable with each non-atomic volatile variable (see, e.g., Fig. 1(b)
and [16] for examples). (Note that a programming language implementation
may ensure atomicity by a mutual exclusion lock if no hardware primitives
for atomic access to a particular type of shared data are available.)
2. Requiring SC on all shared variables is costly in terms of performance on
contemporary multi-core CPUs. In Fig. 1, performance can be improved by
allowing a less strict memory order for variable Data (to be addressed in
Section 2).
3. Although Ada provides the highly abstract PO monitor-construct for blocking synchronization, there is currently no programming primitive available
to match this abstraction level for non-blocking synchronization.
Contemporary CPU architectures relax SC for the sake of performance [3,9,17].
It is a challenge for programming language designers to provide safe, efficient and
user-friendly non-blocking synchronization features. The original memory model
for Java contained problems and had to be revised [14]. It was later found to be
unsound with standard compiler optimizations [18]. The C++11 standard (cf.
[1,19]) has already specified a strict memory model for concurrent and parallel
computing. We think that C++11 was not entirely successful both in terms of
safety and in terms of being user-friendly. In contrast, we are convinced that
these challenges can be met in the upcoming Ada 202x standard.
It has been felt since Ada 95 that it might be advantageous to have language
support for synchronization based on atomic variables. For example, we cite [10,
C.1]: “A need to access specific machine instructions arises sometimes from other
considerations as well. Examples include instructions that perform compound
operations atomically on shared memory, such as test-and-set and compare-andswap, and instructions that provide high-level operations, such as translate-andtest and vector arithmetic.”
Ada is already well-positioned to provide a strict memory model in conjunction with support for non-blocking synchronization, because it provides tasks as
first-class citizens. This rules out inconsistencies that may result from threadfunctionality provided through libraries [6].
4
To provide safe and efficient non-blocking synchronization for Ada 202x, this
paper makes the following contributions:
1. We extend Ada’s memory model by introducing synchronized types, which
allow the expression of memory ordering operations consistently and at a
sufficient level of detail. Memory ordering operations are expressed through
aspects and attributes. Language support for spin loop synchronization via
synchronized variables is proposed.
2. We propose concurrent objects (COs) as a high-level language construct to
express non-blocking synchronization. COs are meant to encapsulate the
intricacies of non-blocking synchronization as POs do for blocking synchronization. Contrary to POs, the entries and procedures of COs execute in
parallel, due to a fine-grained, optimistic synchronization mechanism.
3. We provide an alternative, low-level API on synchronized types, which provides programmers with full control over the implementation of non-blocking
synchronization semantics. Our main purpose with the low-level API is to
provoke a discussion on the trade-off between abstraction versus flexibility.
4. We illustrate the use of concurrent objects and the alternative, low-level API
by several examples.
The remainder of this paper is organized as follows. We summarize the stateof-the-art on memory models and introduce synchronized variables in Sec. 2. We
introduce attributes for specifying memory ordering operations in Sec. 3. We
specify concurrent objects in Sec. 4 and discuss task scheduling in the presence
of COs in Sec. 5. Sec. 6 contains two CO example implementations with varying
memory consistency semantics. We discuss our low-level API in Sec. 7. Sec. 8
contains our conclusions.
Due to space constraints, we have issued a technical report which contains additional programming examples and the rationale for the design of the proposed
non-blocking synchronization mechanisms. Although the paper is intended to
be comprehensive without the appendices, we felt the additional material might
nevertheless be useful. For the same reason of constrained space, the description
of our approach cannot be compared to the specification of Ada on RML-level.
2
The Memory Model
For reasons outlined in Sec. 1, we do not consider the Ada 2012 atomic and
volatile types here. Rather, we introduce synchronized types and variables. Synchronized types provide atomic access. We propose aspects and attributes for
specifying a particular memory model to be employed for reading/writing synchronized variables.
Modern multi-core computer architectures are equipped with a memory hierarchy that consist of main memory, caches and registers. It is important to
2
“Core” refers to the software’s view of a core, which may be an actual core or a
thread context of a multithreaded core [17].
5
distinguish between memory consistency and coherence. We cite from [17]: ‘For
a shared memory machine, the memory consistency model defines the architecturally visible behavior of its memory system. Consistency definitions provide
rules about loads and stores (or memory reads and writes) and how they act
upon memory. As part of supporting a memory consistency model, many machines also provide cache coherence protocols that ensure that multiple cached
copies of data are kept up-to-date.’
The purpose of a memory consistency model (or memory model, for short)
is to define the set of values a read operation is allowed to return [2]. To facilitate programmers’ intuition, it would be ideal if all read/write operations of
a program’s tasks are sequentially consistent. However, the hardware memory
models provided by contemporary CPU architectures relax SC for the sake of
performance [3,9,17]. Enforcing SC on such architectures may incur a noticeable
performance penalty. The workable middle-ground between intuition (SC) and
performance (relaxed hardware memory models) has been established with SC
for data race-free programs (SC-for-DRF) [4]. Informally, a program has a data
race if two tasks access the same memory location, at least one of them is a
write, and there are no intervening synchronization operations that would order
the accesses. “SC-for-DRF” requires programmers to ensure that programs are
free of data races under SC. In turn, the relaxed memory model of a SC-for-DRF
CPU guarantees SC for all executions of such a program.
It has been acknowledged in the literature [2] that Ada 83 was perhaps the
first widely-use high-level programming language to provide first-class support
for shared-memory programming. The approach taken with Ada 83 and later
language revisions was to require legal programs to be without synchronization
errors, which is the approach taken with SC-for-DRF. In contrast, for the Java
memory model it was perceived that even programs with synchronization errors shall have defined semantics for reasons of safety and security of Java’s
sand-boxed execution environment. (We do not consider this approach in the remainder of this paper, because it does not align with Ada’s current approach to
regard the semantics of programs with synchronization errors as undefined, i.e.,
as an erroneous execution, by [8, 9.10§11].) The SC-for-DRF programming model
and two relaxations were formalized for C++11 [7]. They were later adopted for
C11, OpenCL 2.0, and for X10 [21] (without the relaxations).
On the programming language level to guarantee DRF, means for synchronization (ordering operations) have to be provided. Ada’s POs are well-suited for
this purpose. For non-blocking synchronization, atomic operations can be used
to enforce an ordering between the memory accesses of two tasks. It is one goal
of this paper to add language features to Ada such that atomic operations can
be employed with DRF programs. To avoid ambiguity, we propose synchronized
variables and types, which support the expression of memory ordering operations
at a sufficient level of detail (see Sec. 3.1).
The purpose of synchronized variables is that they can be used to safely transfer information (i.e., the value of the variables) from one task to another. ISAs
provide atomic load/store instructions only for a limited set of primitive types.
6
Beyond those, atomicity can only be ensured by locks. Nevertheless, computer
architectures provide memory fences (see e.g., [11]) to provide means for ordering
memory operations. A memory fence requires that all memory operations before
the fence (in program order) must be committed to the memory hierarchy before
any operation after the fence. Then, for data to be transferred from one thread to
another it is not necessary to be atomic anymore. I.e., it is sufficient that (1) the
signaling variable is atomic, and that (2) all write operations are committed to
the memory hierarchy before setting the signaling variable. On the receiver’s
side, it must be ensured that (3) the signaling variable is read atomically, and
that (4) memory loads for the data occur after reading the signaling variable
(Listing 1.2 provides an example.)
In addition to synchronized variables, synchronized types and attribute
Synchronized Components are convenient means for enhancing the usefulness
of synchronized variables.
The general idea of our proposed approach is to define non-blocking concurrent objects similar to protected objects (cf. e.g., [11]). However, entries of
concurrent objects will not block on guards; they will spin loop until the guard
evaluates to true. In addition, functions, procedures, and entries of concurrent
objects are allowed to execute and to modify the encapsulated data in parallel.
Private entries for concurrent objects are also supported. It is their responsibility
that the data provides a consistent view to the users of the concurrent object.
Concurrent objects will use synchronized types for synchronizing data access.
Several memory models are provided for doing this efficiently. It is the responsibility of the programmer to ensure that the entries of a concurrent object are
free from data races (DRF). For such programs, the non-blocking semantics of
a concurrent object will provide SC in the same way as protected objects do for
blocking synchronization.
2.1
Synchronizing memory operations and enforcing ordering
For defining ordering relations on memory operations, it is useful to introduce
some other useful relations.
The synchronizes-with relation can be achieved only by use of atomic types.
Even if monitors or protected objects are used for synchronization, the runtime
implements them employing atomic types. The general idea is to equip read and
write operations on an atomic variable with information that will enforce an
ordering on the read and write operations. Our proposal is to use attributes for
specifying this ordering information. Details can be found below.
The happens-before relation is the basic relation for ordering operations in
programs. In a program consisting of only one single thread, happens-before
is straightforward. For inter-thread happens-before relations the synchronizeswith relation becomes important. If operation X in one thread synchronizes-with
operation Y in another thread, then X happens-before Y. Note that the happensbefore relation is transitive, i.e., if X happens-before Y and Y happens-before Z,
then X happens-before Z. This is true even if X, Y, and Z are part of different
threads.
7
We define different memory models. These memory models originated from
the DRF [4] and properly-labeled [9] hardware memory models. They were formalized for the memory model of C++ [7]. The “sequentially consistent” and
“acquire-release” memory models provide SC for DRF. The models can have
varying costs on different computer architectures. The “acquire-release” memory model is a relaxation of the “sequentially consistent” memory model. As
described in Table 1, it requires concessions from the programmer to weaken SC
in turn for more flexibility for the CPU to re-order memory operations.
Sequentially Consistent Ordering is the most stringent model and the easiest one for programmers to work with. In this case all threads see the same,
total order of operations. This means, a sequentially consistent write to a
synchronized variable synchronizes-with a sequentially-consistent read of the
same variable.
Relaxed Ordering does not obey synchronizes-with relationships, but operations on the same synchronized variable within a single thread still obey
happens-before relationships. This means that although one thread may
write a synchronized variable, at a later point in time another thread may
read an earlier value of this variable.
Acquire-Release Ordering when compared to relaxed ordering introduces
some synchronization. In fact, a read operation on synchronized variables
can then be labeled by acquire, a write operation can be labeled by release.
Synchronization between release and acquire is pairwise between the thread
that issues the release and that acquire operation of a thread that does the
first read-acquire after the release.3 A thread issuing a read-acquire later
may read a different value than that written by the first thread.
It is important to note, that the semantics of the models above have to
be enforced by the compiler (for programs which are DRF). I.e., the compiler
“knows” the relaxed memory model of the hardware and inserts memory fences
in the machine-code such that the memory model of the high-level programming
language is enforced. Compiler optimizations must ensure that reordering of
operations is performed in such a way that the semantics of the memory model
are not violated. The same applies to CPUs, i.e., reordering of instructions is
done with respect to the CPU’s relaxed hardware memory model, constrained
by the ordering semantics of fences inserted by the compiler. The constraints
enforced by the memory model are summarized in Table 1.
3
4
5
In global time!
Memory accesses other than accesses to synchronized variables
Before optimizations performed by the compiler and before reordering done by the
CPU.
8
memory order involved
threads
relaxed
1
release/acquire
2
sequentially
consistent
all
constraints for reordering memory accesses
(for compilers and CPUs)
no inter-thread constraints
(1) ordinary4 stores originally5 before release (in program order) will happen before the release fence (after
compiler optimizations and CPU reordering)
(2) ordinary loads originally after acquire (in program
order) will take place after the acquire fence (after compiler optimizations and CPU reordering)
(1) all memory accesses originally before the sequentially consistent one (in program order) will happen before the fence (after compiler optimizations and CPU reordering)
(2) all memory accesses originally after the sequentially consistent one (in program order) will happen after the fence (after compiler optimizations and CPU reordering)
Table 1: Memory Order and Constraints for Compilers and CPUs
3
3.1
Synchronization primitives
Synchronized Variables
Synchronized variables can be used as atomic variables in Ada 2012, the only
exception being that they are declared inside the lexical scope (data part) of a
concurrent object. In this case aspects and attributes used in the implementation
of the concurrent object’s operations (functions, procedures, and entries) are
employed for specifying behavior according to the memory model. Variables are
labeled by the boolean aspect Synchronized.
Read accesses to synchronized variables in the implementation of the concurrent object’s operations may be labeled with the attribute Concurrent Read,
write accesses with the attribute Concurrent Write. Both attributes have a
parameter Memory Order to specify the memory order of the access. (If the operations are not labeled, the default values given below apply.) In case of read accesses, values allowed for parameter Memory Order are Sequentially Consistent,
Acquire, and Relaxed. The default value is Sequentially Consistent. For
write accesses the values allowed are Sequentially Consistent, Release, and
Relaxed. The default value is again Sequentially Consistent.
For example, assigning the value of synchronized variable Y to synchronized
variable X is given like
X’Concurrent Write(Memory Order => Release) :=
Y’Concurrent Read(Memory Order => relaxed);
In addition we propose aspects for specifying variable specific default values
for the attributes described above. In more detail, when declaring synchronized
variables the default values for read and write accesses can be specified via
aspects Memory Order Read and Memory Order Write. The allowed values are
9
the same as those given above for read and write accesses. If these memory
model aspects are given when declaring a synchronized variable, the attributes
Concurrent Read and Concurrent Write need not be given for actual read and
write accesses of this variable. However, these attributes may be used to temporarily over-write the default values specified for the variable by the aspects.
For example
X: integer with Synchronized, Memory Order Write => Release;
Y: integer with Synchronized, Memory Order Read => Acquire;
...
X := Y;
does the same as the example above but without spoiling the assignment statement.
Aspect Synchronized Components relates to aspect Synchronized in the
same way as Atomic Components relates to Atomic in Ada 2012.
3.2
Read-Modify-Write Variables
If a variable inside the data part of a concurrent object is labeled by the aspect
Read Modify Write, this implies that the variable is synchronized. Write access
to a read-modify-write variable in the implementation of the protected object’s
operations is a read-modify-write access. The read-modify-write access is done
via the attribute Concurrent Exchange. The two parameters of this attribute
are Memory Order Success and Memory Order Failure. The first specifies the
memory order for a successful write, the second one the memory order if the
write access fails (and a new value is assigned to the variable).
Memory Order Success is one of Sequentially Consistent, Acquire,
Release, and Relaxed.
Memory Order Failure may be one of Sequentially Consistent, Acquire,
and Relaxed. The default value for both is Sequentially Consistent. For
the same read-modify-write access the memory order specified for failure must
not be stricter than that specified for success. So, if Memory Order Failure =>
Acquire or Memory Order Failure => Sequentially Consistent is specified,
these have also be given for success.
For read access to a read-modify-write variable, attribute Concurrent Read
has to be used. The parameter Memory Order has to be given. Its value is one of
Sequentially Consistent, Acquire, Relaxed. The default value is
Sequentially Consistent.
Again, aspects for variable specific default values for the attributes described
above may be specified when declaring a read-modify-write variable. The aspects
Memory Order Write Success,
and
are
Memory Order Read,
Memory Order Write Failure with allowed values as above.
3.3
Synchronization Loops
As presented below synchronization by synchronized variables is performed via
spin loops. We call these loops sync loops.
10
4
Concurrent Objects
4.1
Non-Blocking Synchronization
Besides the aspects and attributes proposed in Section 3 that have to be used
for implementing concurrent objects, concurrent objects are different from protected objects in the following way. All operations of concurrent objects can be
executed in parallel. Synchronized variables have to be used for synchronizing
the executing operations. Entries have Boolean-valued guards. The Boolean expressions for such guards may contain only synchronized variables declared in
the data part of the protected object and constants. Calling an entry results
either in immediate execution of the entry’s body if the guard evaluates to true,
or in spin-looping until eventually the guard evaluates to true. We call such a
spin loop sync loop.
4.2
Read-Modify-Write Synchronization
For concurrent objects with read-modify-write variables the attributes proposed
in Section 3 apply. All operations of concurrent objects can be executed in parallel. Read-modify-write variables have to be used for synchronizing the executing
operations. The guards of entries have to be of the form X = X’OLD where X denotes a read-modify-write variable of the concurrent object. The attribute OLD
is well-known from postconditions. An example in our context can be found in
Listing 1.1.
If during the execution of an entry a read-modify-write operation is reached,
that operation might succeed immediately, in which case execution proceeds after
the operation in the normal way. If the operation fails, the whole execution of the
entry is restarted (implicit sync loop). In particular, only the statements being
data-dependent on the read-modify-write variable are re-executed. Statements
not being data-dependent on the read-modify-write variables are executed only
on the first try.6 Precluding non-data-dependent statements from re-execution
is not only a matter of efficiency, it sometimes makes sense semantically, e.g.,
for adding heap management to an implementation.
5
Scheduling and Dispatching
We propose a new state for Ada tasks to facilitate correct scheduling and dispatching for threads synchronizing via synchronized or read-modify-write types.
If a thread is in a sync loop, the thread state changes to “in sync loop”. Note
that sync loops can only happen inside concurrent objects. Thus they can be
spotted easily by the compiler and cannot be confused with “normal” loops.
Note also that for the state change it makes sense not to take place during the
6
For the case that the compiler cannot figure out which statements are datadependent, we propose an additional Boolean aspect only execute on first try
to tag non-data-dependent statements.
11
first iteration of the sync loop, because the synchronization may succeed immediately. For read-modify-write loops, iteration from the third iteration on may
be a good choice; for spin loops, an iteration from the second iteration on may
be a good choice.
In this way the runtime can guarantee that not all available CPUs (cores)
are occupied by threads in state “in sync loop”. Thus we can be sure that at
least one thread makes progress and finally all synchronized or read-modify-write
variables are released (if the program’s synchronization structure is correct and
the program does not deadlock).
After leaving a sync loop, the thread state changes back to “runable”.
6
Examples
Non-blocking Stack. Listing 1.1 shows an implementation of a non-blocking
stack using our proposed new syntax for concurrent objects.
1
s u b t y p e Data i s Integer ;
2
3
4
5
6
7
8
9
t y p e List ;
t y p e List_P i s a c c e s s List ;
t y p e List i s
record
D : Data ;
Next : List_P ;
end r e c o r d ;
10
11
Empty : e x c e p t i o n ;
12
13
14
15
16
17
18
19
20
21
22
c o n c u r r e n t Lock_Free_Stack
is
e n t r y Push ( D : Data ) ;
e n t r y Pop ( D : o u t Data ) ;
private
Head : List_P w i t h R ea d M o d i f y W r i te ,
Memory Order Read = > R el a x ed ,
Memory Order Write Success => R elea s e ,
Memory Order Write Failure => Relaxed;
end L o c k _ F r e e _ S t a c k;
23
24
25
26
27
28
29
30
31
c o n c u r r e n t body L o c k _ F r e e _ S t a c k i s
e n t r y Push ( D : Data )
u n t i l Head = Head ’ OLD i s
New_Node : List_P := new List ;
begin
New_Node . a l l := ( D = > D , Next = > Head ) ;
Head := New_Node ;
end Push ;
32
33
34
35
36
37
38
39
40
41
42
43
44
e n t r y Pop ( D : o u t Data )
u n t i l Head = Head ’ OLD i s
Old_Head : List_P ;
begin
Old_Head := Head ;
i f Old_Head /= n u l l t h e n
Head := Old_Head . Next ;
D := Old_head . D ;
else
r a i s e Empty ;
end i f ;
end Pop ;
12
45
end L o c k _ F r e e _ S t a c k;
Listing 1.1: Non-blocking Stack Implementation Using Proposed New Syntax
Implementation of entry Push (lines 25–31) behaves as follows. In Line 29 a
new element is inserted at the head of the list. Pointer Next of this element is
set to the current head. The next statement (Line 30) assigns the new value to
the head of the list. Since variable Head has aspect Read Modify Write (line 18),
this is done with RMW semantics, i.e., if the value of Head has not been changed
(since the execution of Push has started) by a different thread executing Push
or Pop (i.e., Head = Head’OLD), then the RMW operation succeeds and execution proceeds at Line 31, i.e., Push returns. If the value of Head has been
changed (Head /= Head’OLD), then the RMW operation fails and entry Push is
re-executed starting from Line 29. Line 27 is not re-executed as it is not data
dependent on Head.
Several memory order attributes apply to the RMW operation (Line 30)
which are given in lines 19–21: In case of a successful execution of the RMW, the
value of Head is released such that other threads can read its value via memory
order acquire. In the failure case the new value of Head is assigned to the “local
copy” of Head (i.e., Head’OLD) via relaxed memory order. “Relaxed” is enough
because RMW semantics will detect if the value of Head has been changed by a
different thread anyway. The same applies to “Relaxed” in Line 27.
Implementation of entry Pop (lines 33–44) follows along the same lines.
Memory management needs special consideration: In our case it is enough to
use a synchronized counter that counts the number of threads inside Pop. If the
counter equals 1, memory can be freed. Ada’s storage pools are a perfect means
for doing this without spoiling the code.
This example also shows how easy it is to migrate from a (working) blocking
to a (working) non-blocking implementation of a program. Assume that a working implementation with a protected object exists, then one has to follow these
steps:
1. Replace keyword protected by keyword concurrent.
2. Replace protected operations by DRF concurrent operations, thereby adding
appropriate guards to the concurrent entries.
3. Test the non-blocking program which now has default memory order
sequentially consistent.
4. Carefully relax the memory ordering requirements: Add memory order aspects and/or attributes Acquire, Release, and/or Relaxed to improve performance but without violating memory consistency.
Generic Release-Acquire Object. Listing 1.2 shows how release-acquire semantics can be implemented for general data structures with help of one synchronized Boolean.
1
2
3
4
generic
t y p e Data i s p r i v a t e ;
package Generic_Release_Acquire i s
13
5
6
7
8
9
10
11
12
13
14
c o n c u r r e n t RA
is
p r o c e d u r e Write ( d : Data ) ;
e n t r y Get ( D : o u t Data ) ;
private
Ready : Boolean := false w i t h S y n c h r o n i z e d ,
Memory Order Read = > Acq u i r e ,
Memory Order Write = > R e l e a s e ;
Da : Data ;
end RA ;
15
16
end G e n e r i c _ R e l e a s e _ A c q u i r e;
17
18
p a c k a g e body G e n e r i c _ R e l e a s e _ A c q u i r e i s
19
20
c o n c u r r e n t body RA i s
21
22
23
24
25
26
p r o c e d u r e Write ( D : Data ) i s
begin
Da := D ;
Ready := true ;
end Write :
27
28
29
30
31
32
33
34
35
e n t r y Get ( D : o u t Data )
when Ready i s
-- spin - lock until released , i . e . , Ready = true ;
-- only sync . v a r i a b l e s and c o n s t a n t s allowed in guard e x p r e s s i o n
begin
D := Da ;
end Get ;
end RA ;
36
37
end G e n e r i c _ R e l e a s e _ A c q u i r e;
Listing 1.2: Generic Release-Acquire Object
7
API
As already pointed out, we feel that providing concurrent objects as first-class
citizens is the right way to enhance Ada with non-blocking synchronization on an
adequate memory model. On the other hand, if the programmer needs synchronization on a lower level than concurrent objects provide, an API-based approach
(generic function Read Modify Write in package Memory Model) would be a viable alternative. Listing 1.3 shows such a predefined package Memory Model. It
contains the specification of generic function Read Modify Write, which allows
to use the read-modify-write operation of the underlying computer hardware.
Exposing sync loops to the programmer makes it necessary to introduce a new
aspect sync loop to let the runtime perform the state change to “in sync loop”
(cf. Section 5). Because nobody can force the programmer to use this aspect
correctly, the information transferred to the runtime may be false or incomplete, giving rise to concurrency defects such as deadlocks, livelocks, and other
problems.
1
package Memory_Model i s
2
3
4
5
6
7
8
type Memory_Order_Type i s (
Sequentially Consistent ,
R el a x ed ,
Acq u i r e ,
R e l e a s e );
14
9
s u b t y p e M e m o r y _ O r d e r _ S u c c e s s _ T y p e i s M e m o r y _ O r d e r _ T y p e;
10
11
12
subtype Memory_Order_Failure_Type i s Memory_Order_Type
r a n g e S e q u e n t i a l l y C o n s i s t e n t .. A c q u i r e ;
13
14
15
16
17
18
19
20
21
22
23
generic
type Some_Synchronized_Type i s p r i v a t e ;
w i t h f u n c t i o n Update r e t u r n S o m e _ S y n c h r o n i z e d _ T y p e;
R e a d _ M o d i f y _ W r i t e _ V a r i a b l e: i n o u t S o m e _ S y n c h r o n i z e d _ T y p e
with Read Modify Write ;
M emo r y Or d er S u cces s : M e m o r y _ O r d e r _ S u c c e s s _ T y p e :=
Sequentially Consistent ;
M e m o r y O r d e r F a i l u r e : M e m o r y _ O r d e r _ F a i l u r e _ T y p e :=
Sequentially Consistent ;
f u n c t i o n R e a d M o d i f y W r i t e r e t u r n Boolean ;
24
25
end M e m o r y _ M o d e l;
Listing 1.3: Package Memory Model
8
Conclusion and Future Work
We have presented an approach for providing safe non-blocking synchronization
in Ada 202x. Our novel approach is based on introducing concurrent objects for
encapsulating non-blocking data structures on a high abstraction level. In addition, we have presented synchronized and read-modify-write types which support
the expression of memory ordering operations at a sufficient level of detail. Concurrent objects provide SC for programs without data races. This SC-for-DRF
memory model is well-aligned with Ada’s semantics for blocking synchronization
via protected objects, which requires legal programs to be without synchronization errors ([8, 9.10§11]).
Although Ada 2012 provides the highly abstract protected object monitorconstruct for blocking synchronization, there was previously no programming
primitive available to match this abstraction level for non-blocking synchronization. The proposed memory model in conjunction with our concurrent object
construct for non-blocking synchronization may bar users from having to invent ad-hoc synchronization solutions, which have been found error-prone with
blocking synchronization already [20].
Until now, all previous approaches are based on APIs. We have listed a
number of advantages that support our approach of making non-blocking data
structures first class language citizens. In contrast, our approach for Ada 202x
encapsulates non-blocking synchronization inside concurrent objects. This safe
approach makes the code easy to understand. Note that concurrent objects are
not orthogonal to objects in the sense of OOP (tagged types in Ada). However,
this can be achieved by employing the proposed API approach (cf. Section 7).
In addition, it is not difficult to migrate code from blocking to non-blocking
synchronization. Adding memory management via storage pools integrates well
with our modular approach and does not clutter the code.
A lot of work remains to be done. To name only a few issues: Non-blocking
barriers (in the sense of [8, D.10.1]) would be useful; details have to be elaborated.
Fully integrating concurrent objects into scheduling and dispatching models and
integrating with the features for parallel programming planned for Ada 202x
have to be done carefully.
15
9
Acknowledgments
This research was supported by the Austrian Science Fund (FWF) project
I 1035N23, and by the Next-Generation Information Computing Development
Program through the National Research Foundation of Korea (NRF), funded by
the Ministry of Science, ICT & Future Planning under grant NRF2015M3C4A7065522.
References
1. Working Draft, Standard for Programming Language C++. ISO/IEC N4296, 2014.
2. S. V. Adve and H.-J. Boehm. Memory models: A case for rethinking parallel
languages and hardware. Commun. ACM, 53(8):90–101, Aug. 2010.
3. S. V. Adve and K. Gharachorloo. Shared memory consistency models: A tutorial.
Computer, 29(12):66–76, Dec. 1996.
4. S. V. Adve and M. D. Hill. Weak ordering—a new definition. In Proceedings of the
17th Annual International Symposium on Computer Architecture, ISCA ’90, pages
2–14, New York, NY, USA, 1990. ACM.
5. J. Barnes. Ada 2012 Rationale: The Language – the Standard Libraries. Springer
LNCS, 2013.
6. H.-J. Boehm. Threads cannot be implemented as a library. SIGPLAN Not.,
40(6):261–268, June 2005.
7. H.-J. Boehm and S. V. Adve. Foundations of the C++ concurrency memory model.
In Proceedings of the 29th ACM SIGPLAN Conference on Programming Language
Design and Implementation, PLDI ’08, pages 68–78, New York, NY, USA, 2008.
ACM.
8. R. L. Brukardt, editor. Annotated Ada Reference Manual, ISO/IEC 8652:2012/Cor
1:2016. 2016.
9. K. Gharachorloo, D. Lenoski, J. Laudon, P. Gibbons, A. Gupta, and J. Hennessy.
Memory consistency and event ordering in scalable shared-memory multiprocessors. SIGARCH Comput. Archit. News, 18(2SI):15–26, May 1990.
10. L. Guerby. Ada 95 Rationale – The Language – The Standard Libraries. Springer,
1997.
11. M. Herlihy and N. Shavit. The Art of Multiprocessor Programming. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2012.
12. C. A. R. Hoare. Monitors: An operating system structuring concept. Commun.
ACM, 17(10):549–557, Oct. 1974.
13. L. Lamport. How to make a multiprocessor computer that correctly executes
multiprocess programs. IEEE Trans. Comput., 28(9):690–691, Sept. 1979.
14. J. Manson, W. Pugh, and S. V. Adve. The Java memory model. In Proceedings
of the 32nd ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, POPL ’05, pages 378–391, New York, NY, USA, 2005. ACM.
15. M. L. Scott. Shared-Memory Synchronization. Synthesis Lectures on Computer
Architecture. Morgan & Claypool Publishers, 2013.
16. H. Simpson. Four-slot fully asynchronous communication mechanism. Computers
and Digital Techniques, IEE Proceedings E, 137:17–30, 02 1990.
17. D. J. Sorin, M. D. Hill, and D. A. Wood. A Primer on Memory Consistency and
Cache Coherence. Number 16 in Synthesis Lectures on Computer Architecture.
Morgan & Claypool, 2011.
16
18. J. Ševčı́k and D. Aspinall. On validity of program transformations in the Java
memory model. In Proceedings of the 22nd European Conference on Object-Oriented
Programming, ECOOP ’08, pages 27–51, Berlin, Heidelberg, 2008. Springer-Verlag.
19. A. Williams. C++ Concurrency in Action. Manning Publ. Co., Shelter Island,
NY, 2012.
20. W. Xiong, S. Park, J. Zhang, Y. Zhou, and Z. Ma. Ad hoc synchronization considered harmful. In Proceedings of the 9th USENIX Conference on Operating Systems
Design and Implementation, OSDI’10, pages 163–176, Berkeley, CA, USA, 2010.
USENIX Association.
21. A. Zwinkau. A memory model for X10. In Proceedings of the 6th ACM SIGPLAN
Workshop on X10, X10 2016, pages 7–12, New York, NY, USA, 2016. ACM.
| 6cs.PL
|
arXiv:1802.02423v1 [stat.AP] 3 Feb 2018
On the Generalizability of Linear and Non-Linear
Region of Interest-Based Multivariate Regression
Models for fMRI Data
Ethan C. Jackson
James Alexander Hughes
Mark Daley
Computer Science Department
University of Western Ontario
London, Ontario, Canada N6A 3K7
[email protected]
Computer Science Department
University of Western Ontario
London, Ontario, Canada N6A 3K7
[email protected]
Computer Science Department
University of Western Ontario
London, Ontario, Canada N6A 3K7
[email protected]
Abstract—In contrast to conventional, univariate analysis,
various types of multivariate analysis have been applied to
functional magnetic resonance imaging (fMRI) data. In this paper, we compare two contemporary approaches for multivariate
regression on task-based fMRI data: linear regression with ridge
regularization and non-linear symbolic regression using genetic
programming. The data for this project is representative of
a contemporary fMRI experimental design for visual stimuli.
Linear and non-linear models were generated for 10 subjects,
with another 4 withheld for validation. Model quality is evaluated
by comparing R scores (Pearson product-moment correlation)
in various contexts, including single run self-fit, within-subject
generalization, and between-subject generalization. Propensity
for modelling strategies to overfit is estimated using a separate
resting state scan. Results suggest that neither method is objectively or inherently better than the other.
Index Terms—fMRI, Linear Regression, Non-Linear Regression, Symbolic Regression, Genetic Programming
I. I NTRODUCTION
fMRI is a non-invasive neuroimaging technique used in
research and clinical applications to record relative changes
in metabolic rate associated with neural activity via the blood
oxygenation level dependent (BOLD) signal [3]. This signal
measures the relative levels of oxygenated and deoxygenated
blood throughout the whole brain, and is a known metabolic
proxy for neural activity.
The neuroimaging field has been very successful in using
linear statistical methods for analysis. The general linear model
(GLM), in particular, has been fruitfully applied to identify
relationships between experimental conditions and localized
BOLD signal. To fit a GLM, neuroscientists create a predictive
time series model (predictor) for each experimental condition
of interest. Usually, these predictors are square-wave time
series corresponding to stimulus onset and duration convolved
with a model function for haemodynamic response (HR). For
each voxel (a volumetric pixel) or aggregate region of interest
(ROI), a multiple linear regression is performed to compute
a linear combination of predictors that optimally models the
recorded BOLD signal. A voxel or ROI’s relationship to
an experimental condition can be inferred by comparing the
relative weights (scaling factors) for each of the predictors.
A variety of statistical methods can be applied to perform
comparisons over all voxels or ROIs, though paired t-tests are
commonly used.
In contrast to conventional univariate analysis, multivariate
relational modelling has also been developed. For example, in
Hughes and Daley’s work on relational modelling between
ROI average time series ([6], [7]), average BOLD in one
ROI (dependent variable) is modelled as a non-linear function
of BOLD signal in other ROIs (independent variables). In
order to make a more direct comparison to conventional linear
regression and the GLM, the dependent regression variable can
be replaced with a hypothesized, stimulus-driven HR function.
In this configuration, regression models HR as a spatially
distributed combination of BOLD signal — an approach that
has applications in multivoxel pattern analysis [9].
Naturally one might ask whether conventional, linear regression or much more elaborate non-linear symbolic regression is
better suited for this type of modelling. This paper aims to answer that question by comparing the generalizability of models
generated using each method. We assess generalizability based
on within- and between-subject analytical experiments, and
we further estimate each method’s lower bound for propensity
to overfit using an additional resting state scan. The fMRI
data consists of scans from 15 subjects in total — a decidedly
small but generally acceptable sample size in the neuroimaging
community.
In the following sections, we first give details about the
data used in our experiments. Next, we give an overview of
the experiments themselves, followed by sections explaining
details about linear and non-linear models, respectively. Then
we present results, discussion, and finally avenues for future
work.
II. DATA AND P REPROCESSING
For all experimental runs, we used a 10-subject subset of
fMRI data collected for didactic purposes. Validation uses four
additional experimental scans and one additional resting state
scan.
All functional scans were collected using the same scanner and parameters: Siemens Magnetom Prisma 3T, 2.5-mm
isotropic resolution with matrix size of 84×84 over 52 slices,
TR = 1000ms, multi-band gradient-echo echoplanar pulse
sequence. Anatomical scans were collected using a matrix size
of 248×256 over 176 1mm slices. More information about the
scanner and facilities can be found at http://cfmm.robarts.ca/
tools/3t-mri/.
The experimental stimuli used were designed to localize
brain regions selective to three categories of visual stimuli:
faces, hands, and bodies. Scrambled versions of the images
formed a fourth stimulus category. Participants were instructed
to maintain their gaze on a central fixation point, which was
also presented on a blank screen during baseline periods.
Images were presented to subjects according to a conventional block design, with a block duration of 16s. Four
cycles of four blocks (faces, hands, bodies, scrambled images)
were presented in each run. Baseline blocks occurred at the
beginning and end of each run and between cycles. The total
run duration was 336 s (21 blocks x 16 s/block). Each stimulus
block consisted of 16 stimuli, each presented for 0.8 s with a
0.2 s intertrial interval. The order of blocks within cycles was
balanced such that for a run, each category was presented once
in each of the first, second, third and fourth part of the cycle.
Two runs with two different orders were collected. Participants
monitored the stream of visual stimuli for repetitions (a oneback task) to maintain attention.
Data were preprocessed in Brain Voyager 20.6 commercial
software ([5], [4]). Preprocessing consisted of functionalanatomical alignment, three-dimensional motion correction,
temporal high-pass filtering (X cycles/run), and transformation
into 2-mm isotropic resolution, warping into stereotaxic space
using the MNI-152 template. Data were loaded into Python
using a Brain Voyager-compatible fork of NiBabel [1].
For analysis, we considered two types partitioning. Though
various transformations could be applied to yield an aggregate
time series from each ROI, we chose to use the average
activation of the whole ROI. This is the same approach used
in [6] and [7]. Many methods for defining partition boundaries
exist. Conventional partitioning of cortex, for example, is
performed using static maps (atlases) on a reference brain.
In our experiments, we used two atlases: the Harvard-Oxford
Cortical Structural (H-O) atlas, and the Talairach Daemon
Labels (TalROI) atlas. Both atlases are provided with FSL [8],
and are visualized by Figures 4 and 1, respectively. The HO atlas defines 48 ROIs, while TalROI defines 986. To avoid
overfitting in linear regression and unreasonable computational
complexity for non-linear regression, an additional preprocessing step was applied to TalROI.
Since TalROI defines 986 ROIs and because we only used
this atlas in experiments on response to face stimuli, we used
a Neurosynth mask to eliminate ROIs not associated with
face selectivity in its body of literature. Neurosynth [10] is
a web tool that uses natural language processing to identify
associations between brain areas and cognitive function based
on a large corpus of neuroimaging research papers. A Neu-
rosynth search for ‘face’ or ‘FFA’ (fusiform face area) yields
a mask that can be overlaid onto a functional scan in MNI
space. The mask used in this work is visualized by Figure 2.
Application of this mask not only reduces the number of ROIs
in TalROI to 50, but focuses model generation on ROIs known
to demonstrate selectivity for faces.
Altogether, the fMRI data consists of 20 experimental
runs, 10 each for two orders of the block-designed stimulus
presentation scheme labelled ‘Loc1’ and ‘Loc2’ (for Localisers
1 and 2). Additionally, a resting state scan (stimulus- and
task-free, eyes closed, instructed not to fixate) was collected
from one independent subject. Each run consists of 340 time
points. Each of the two atlases was used to partition and
aggregate the each run into ROI averages. Thus, we have
two transformations (H-O and TalROI) for each of the 20
experimental runs and one resting state scan. All data are
stored as CSV files with the number of rows and columns
equal to the number of time points and ROIs, respectively.
These files, in addition to all Python code, masks, and atlases,
are available as part of the digital appendix [2].
III. M ODELS
Before describing the experiments themselves, we first
describe how linear and non-linear models were constructed
for analysis.
A. Linear Models
Linear models were all constructed using linear regression
with ridge or L2 regularization. This method computes the
linear combination of ROI time series that optimally models
a hypothesized HR using ordinary least squares estimation.
Thus for each run, linear models are composed of n scaling
factors plus one constant term, where n is equal to the number
of ROIs used for partitioning. Results based on linear methods
are labelled ‘L’.
B. Non-Linear Models
Non-linear models were constructed using Hughes and Daley’s GP-based system designed for symbolic regression. The
number of degrees of freedom is exponentially proportional
to the size of the GP language and to the number of ROIs.
We used two methods for assessing the quality of non-linear
models.
First, to obtain one model per run (fMRI scan, transformed)
in an unbiased manner, we computed 100 independent GP
runs for 10,000,000 generations each. Of the 100 candidate
models, we took the model of best fit as estimated by mean
squared error at the end of evolution. Thus for each run,
non-linear models are expressed as a non-linear function of
n variables, where n is again equal to the number of ROIs
used for partitioning. It is important to note that the resulting
non-linear models are temporally independent: the non-linear
functions must be applied independently at each time point in
order to reconstruct a time series. Results using this method
are labelled ‘N-L’. The GP language or set of basis functions,
settings, and parameters are summarized by Table I.
TABLE I: Parameter settings for GP. The last four parameters
are specific to the improvements made in Hughes and Daley’s
GP system. Improvements include the use of an acyclic graph
representation, fitness predictors, and island-model migration
strategies.
Elitism
Population
Subpopulations
Migrations
Generations
Crossover
Mutation
Fitness Metric
Language
Trainers
Predictors
Predictor Pop. Size
Max # Graph Nodes
1
101/subpopulation (707 total)
7
1,000
1,000 per migration (10,000,000 total)
80%
10% (x2 chances)
Mean Squared Error: n1 Pni=1 (Yˆi − Yi )2
+, − ,∗ , /, exp, abs, sin, cos, tan
8
20
20% of Dataset
140
The second method is the same as the first, except it
considers all 100 candidate models for a run produced by GP.
This allows us to plot and analyze a distribution of R scores
(rather than a single value) to better understand whether GP
can produce better models than linear regression. Using this
method, we can search a distribution of candidate models,
and select the single, best-scoring non-linear model. Though
heavily biased, this analysis is useful to evaluate potential
non-linear model quality. For example in between-subject
experiments, we could select the model for subject A that
generalizes best to subject B. Results using this method are
labelled ‘N-L (best)’.
IV. A NALYTICAL E XPERIMENTS
To compare the quality of models obtained from multivariate linear regression to symbolic non-linear regression, and to
estimate overfitting, we designed the following experiments.
Experiments are categorized as either all-stims or faces. For
the former, models fit data to a hypothesized HR to any
stimulus category, as visualized by Figure 5. For the latter,
models fit data to a hypothesized HR for the faces category
of stimuli only, as visualized by Figure 6. In practice, the set
of experiments for the latter is much more interesting since
it will gauge stimulus specific selectivity. The all-stims and
faces experiments exclusively use the H-O and TalROI atlases,
respectively.
For all experiments, we first estimate generalizability using
mean R scores and their standard deviations. There are three
analysis categories: L, N-L, and N-L (best). For validation,
we apply L and N-L models to data for four withheld
subjects. Validation is not possible for N-L (best) since those
models would depend on the withheld data, but an alternative
approach for non-linear model selection was tested. Rather
than selecting the top model during GP evolution, we selected
the model which best generalized within-subject. For withinsubject validation, these unbiased models are labelled NL(UnB).
A. Within-Subject Generalizability
Within-subject generalizability is estimated by first applying
models generated for one subject-run to data for the other,
independent run, and then computing the R score between the
model output and the relevant hypothesized HR.
B. Between-Subject Generalizability
Between-subject generalizability is estimated in two ways.
First, given one of the two runs, we apply the models generated
for one subject to all other subjects’ data. We then compute the
R score between model outputs and the hypothesized HR. We
call this approach pairwise between-subject generalizability.
In a second approach, we use model output averaging
to combine multiple models into one average model. This
approach simply applies multiple models to the same data and
computes the average output. With 10 subjects, we generate
10 different average models, each with one subject left out.
An average model is evaluated by applying it to the left-out
subject’s data and comparing the R score between its output
and the hypothesized HR.
Validation was possible for between-subject generalizability
due to there being four other subjects that were withheld
from modelling. This was because of the computational cost
for generating non-linear models. Pairwise between-subject
generalizability is evaluated by applying generated models to
withheld subjects’ data and computing the R score between the
output and the relevant hypothesized HR. A similar validation
strategy is taken for average models.
C. Resting State Validation
In an effort to compare how well linear and non-linear
regression are able to fit task-free data to an unrelated hypothesized HR, we fit four models to a separate resting state
scan. These consist of a linear and non-linear model for each
of the two hypothesized HRs used in other experiments. This
is especially important to do for non-linear regression since
the number of degrees of freedom is so large. The R score
between the fitted model’s output and the hypothesized HR
provides an estimate of a modelling strategy’s propensity to
overfit. It would be insightful to have many more resting state
scans available for analysis, but only one resting state scan
was collected as part of the funded project.
V. R ESULTS
A. Within-Subject Generalizability
An estimation of the relative within-subject generalizability
for linear and non-linear regression is summarized by Table
II. We did not observe any significant difference between
L and N-L models, however N-L (best) scores significantly
higher than both L and N-L. It is important to note that
all we can infer from this result is that GP produces, on
average, at least one model per run that generalizes better
than our linear models. A typical instance of this scenario
is visualized by Figure 7. Notice in this histogram that the
majority of non-linear models generalized worse than their
linear counterpart. However, in this case, and in fact in all
L
N-L
N-L (best)
all-stims
x̄ = 0.6035 s = 0.1516
x̄ = 0.5470 s = 0.2151
x̄ = 0.7132 s = 0.1302
faces
x̄ = 0.2265 s = 0.1229
x̄ = 0.1621 s = 0.1522
x̄ = 0.3240 s = 0.1152
TABLE II: Mean R scores (x̄) and standard deviation (s)
for within-subject generalizability. all-stims In independent ttests, L and N-L models are not significantly different, though
N-L (best) scores significantly higher than L and N-L with
p < 0.022 and p < 0.006, respectively. faces: In independent t-tests, scores between L and N-L are not significantly
different, though again N-L (best) scores significantly higher
than both L and N-L with p < 0.016 and p < 1 × 10−16 ,
respectively.
L
N-L
N-L (best)
all-stims
x̄ = 0.4703 s = 0.1959
x̄ = 0.3864 s = 0.2327
x̄ = 0.6399 s = 0.1739
faces
x̄ = 0.1013 s = 0.1182
x̄ = 0.0754 s = 0.1045
x̄ = 0.2814 s = 0.1302
TABLE III: Mean R scores (x̄) and standard deviation (s)
for pairwise between-subject generalizability. all-stims: In
independent t-tests, L scores significantly higher than N-L with
p = 0.2×10−3 , N-L (best) scores significantly higher than N-L
with p < 1 × 10−27 , and N-L (best) scores significantly higher
than L with p = 0.2 × 10−3 . faces: In independent t-tests,
L scores significantly higher than N-L with p < 0.029, N-L
(best) scores significantly higher than N-L with p < 1×10−47 ,
and N-L (best) scores significantly higher than L with p <
1 × 10−35 .
L
N-L
N-L (best)
all-stims
x̄ = 0.8183 s = 0.0774
x̄ = 0.7104 s = 0.1889
x̄ = 0.8737 s = 0.0620
faces
x̄ = 0.2389 s = 0.0992
x̄ = 0.1438 s = 0.1293
x̄ = 0.2659 s = 0.1410
TABLE IV: Mean R scores (x̄) and standard deviation (s)
for between-subject (average models) generalizability. allstims: In independent t-tests, L scores higher than N-L with
p < 0.027 and N-L (best) scores higher than both L and N-L
with p < 0.02 and < 0.001, respectively. faces: In independent
t-tests, none of the scores are significantly different from each
other.
L
N-L
N-L(UnB)
all-stims
x̄ = 0.5287 s = 0.1091
x̄ = 0.4736 s = 0.1637
x̄ = 0.5084 s = 0.1821
faces
x̄ = 0.0550 s = 0.1040
x̄ = 0.0431 s = 0.0832
x̄ = 0.0454 s = 0.0831
TABLE V: Mean R scores (x̄) and standard deviation (s)
for pairwise between-subject validation on data from four
withheld subjects. Within each stimulus category, there is no
significant difference between modelling strategies.
Average model validation results are summarized by Table
VI. We compared these results to those in Table IV and found
that, in validation, generalizability tended to be worse than
during model construction. Models for all-stims, however, tend
to be a good fit for the hypothesized HR, using any of the three
modelling strategies. This result is visualized by Figure 8.
C. Resting State Validation
other observed instances, there was at least one non-linear
model with a better R score than for the linear model.
Due to limitations on the number of models for which we
could generate non-linear models, a separate validation set was
not used for within-subject generalizability.
B. Between-Subject Generalizability
An estimation of the relative pairwise between-subject generalizability for linear and non-linear regression is summarized
by Table III. We observed in both categories that L models
scored significantly higher than N-L models. And once again,
we observed that N-L (best) generalizes better than both L and
N-L.
An estimation of average model generalizability is summarized by Table IV. We observed improvements over pairwise
between-subject generalizability in almost all cases. The only
exception was for N-L (best) for faces, which was not significantly different. Based on these results, we suggest that
model averaging could be a useful technique for improving
signal-to-noise ratios in between-subject modelling.
Pairwise between-subject generalizability validation results
are summarized by Table V. We compared these results to
those in Table III. According to independent t-tests, a significant difference between results for the modelled subjects and
validation subjects was only observed for N-L, all-stims, where
the mean R score was higher in validation with p = 0.034.
An estimation of each model’s propensity to overfit, as
estimated using independent resting state data, is summarized
by Table VII and is visualized by Figure 9. Though only
one resting state scan is used, we can conclude that for both
modelling strategies and for both stimulus categories, mean
R scores are significantly greater for stimulus-related models
than for at least one resting state scan. This provides at least
some confidence that models are not unreasonably overfit.
L
N-L
N-L(UnB)
all-stims
x̄ = 0.6026 s = 0.1288
x̄ = 0.6491 s = 0.1221
x̄ = 0.6591 s = 0.1101
faces
x̄ = 0.1199 s = 0.0477
x̄ = 0.0893 s = 0.0534
x̄ = 0.0670 s = 0.0555
TABLE VI: Mean R scores (x̄) and standard deviation (s) for
between-subject (average model) validation on data from
four withheld subjects. Within each stimulus category, there is
no significant difference between modelling strategies.
L
N-L
all-stims
x̄exp = 0.83 Rrest = 0.45
x̄exp = 0.86 Rrest = 0.35
faces
x̄exp = 0.59 Rrest = −0.08
x̄exp = 0.56 Rrest = 0.35
TABLE VII: Mean R scores for models fit using experimental data (x̄exp ) compared to R scores for models fit using
2
unrelated resting state data (Rrest
). In all cases, according
to independent one-sample t-tests, resting state R scores
are significantly lower than mean experimental scores with
p < 1 × 10−6 .
VI. D ISCUSSION
VIII. ACKNOWLEDGEMENTS
We found no evidence that non-linear regression is generally
more prone to overfitting than linear regression. In resting
state validation (see Table VII) linear regression produced a
better-fit model than non-linear regression between completely
independent time series. In practice, we can conclude only that
both methods can be similarly prone to overfitting.
For within-subject generalizability, we observed that in all
cases, at least one out of 100 GP runs produced a better model
than linear regression. This is reflected by the consistently
good results for N-L (best) models. The fact that such bestgeneralizing models exist does not help us select them in
an unbiased manner, though. We were, however, able to use
such results in an unbiased way by selecting the non-linear
model that best generalized within-subject for between-subject
analysis. Selecting such a model, however, did not yield
significant improvement over the other methods we tested.
Generalizability scores were poor for faces. Though all
models were significantly and positively correlated with the
hypothesized HR, further work must be done to evaluate their
specificity. The poor scores observed here could be due to
using ROIs that are too large to capture face-selectivity. Even
with the TalROI atlas, ROIs contained hundreds of voxels. A
finer partitioning of face-selective areas or other atlases may
yield better results.
This research was supported in part by the Natural Sciences
and Engineering Research Council of Canada (NSERC).
This research was enabled in part by support provided by
Compute Ontario (computeontario.ca), Westgrid (westgrid.ca),
and Compute Canada (computecanada.ca).
Data collection costs were funded by a grant from the
Canada First Research Excellence Fund (BrainsCAN: Brain
Health for Life) to the University of Western Ontario.
fMRI experimental design and stimuli were designed by
Jody Culham and Rainer Goebel; implemented by Kevin
Stubbs.
VII. C ONCLUSIONS AND F UTURE W ORK
The all-stims hypothesized HR was very successfully modelled using either linear or non-linear regression. In experiments for both within- and between-subject generalizability,
all models reproduced the hypothesized HR with mean R
scores reaching 0.65 in validation, and up to 0.81 in unbiased
experiments. Furthermore, we found no evidence to suggest
that non-linear regression over fit the data worse than linear
regression. These results suggest that our linear and non-linear
models generalized similarly well, and that neither was overfit
more than the other.
In validation, we observed that model averaging significantly improves model quality for all-stims, for all methods,
when compared to pairwise between-subject generalizability.
This suggests that all models feature a noise component that is
partially subdued by averaging across subjects. This result is
unsurprising, but provides further evidence that signal-to-noise
ratios can be significantly improved by considering ROI-based
averages across multiple subjects.
Overall, we found that while they are not any more likely
to over-fit, non-linear models did not generalize significantly
better than linear models when unbiased model selection methods are applied. Though non-linear regression consistently
produced some model that is better than its linear counterpart,
it is difficult to select and expensive to produce. As a result, we
conclude that for small sample sizes, it seems far more sensible
to use linear methods for multivariate regression modelling of
fMRI data, pending improved methods for non-linear model
selection.
R EFERENCES
[1] 2017. URL https://github.com/CNMaastricht/nibabel
BrainVoyager.
[2] February 2018. URL https://github.com/ethancjackson/
fMRI-CompRegression.
[3] Elizabeth A Disbrow, Daniel A Slutsky, Timothy PL
Roberts, and Leah A Krubitzer. Functional mri at 1.5
tesla: a comparison of the blood oxygenation leveldependent signal and electrophysiology. Proceedings of
the National Academy of Sciences, 97(17):9718–9723,
2000.
[4] Elia Formisano, Francesco Di Salle, and Rainer Goebel.
Fundamentals of data analysis methods in fmri. In Advanced Image processing in magnetic resonance imaging.
Marcel Dekker, 2006.
[5] Rainer Goebel, Fabrizio Esposito, and Elia Formisano.
Analysis of functional image analysis contest (fiac) data
with brainvoyager qx: From single-subject to cortically
aligned group general linear model analysis and selforganizing group independent component analysis. Human brain mapping, 27(5):392–401, 2006.
[6] James Alexander Hughes and Mark Daley. Finding
nonlinear relationships in fmri time series with symbolic
regression. In Proceedings of the 2016 on Genetic
and Evolutionary Computation Conference Companion,
pages 101–102. ACM, 2016.
[7] James Alexander Hughes and Mark Daley. Searching
for nonlinear relationships in fmri data with symbolic regression. In Proceedings of the Genetic and Evolutionary
Computation Conference, pages 1129–1136. ACM, 2017.
[8] Mark Jenkinson, Christian F Beckmann, Timothy EJ
Behrens, Mark W Woolrich, and Stephen M Smith. Fsl.
Neuroimage, 62(2):782–790, 2012.
[9] Kenneth A Norman, Sean M Polyn, Greg J Detre, and
James V Haxby. Beyond mind-reading: multi-voxel pattern analysis of fmri data. Trends in cognitive sciences,
10(9):424–430, 2006.
[10] Tal Yarkoni, Russell A Poldrack, Thomas E Nichols,
David C Van Essen, and Tor D Wager. Large-scale
automated synthesis of human functional neuroimaging
data. Nature methods, 8(8):665, 2011.
Fig. 1: Talairach standard atlas used to partition the whole
brain into 986 ROIs. Statistical transformations on smaller
ROIs are more likely to capture highly stimulus selective
metabolic responses.
Fig. 2: Two dimensional slice of Neurosynth mask for FFA
overlaid on TalROI atlas. Areas coloured in grey scale indicate voxels in MNI152 space that are associated with FFA
according to the Neurosynth corpus. Higher intensity (brightness) indicates stronger confidence. MNI152 coordinates =
(41.7, −51.6, −20.9)
Fig. 3: Hypothesized haemodynamic responses to stimuli
in arbitrary units over time in seconds for ‘Loc1’. Stimuli
(images) were presented to subjects for a constant duration
of 0.8 seconds followed by a 0.2 second gap. The order of
stimulus category presentation is reflected by the colours of
the response curves. The other experimental run, ’Loc2’ is a
variation of ’Loc1’ in which the order of stimulus category
presentation was changed.
Fig. 4: Harvard-Oxford atlas used to partition the cortical brain
into 48 ROIs. Large ROIs are less sensitive than smaller ones
to partitioning error due to anatomical differences between
subjects.
Fig. 5: Hypothesized haemodynamic response to all-stimuli in
arbitrary units over time in seconds for ‘Loc1’. This curve
is obtained by taking the sum of the four stimulus-specific
hypothesized HRs.
Fig. 6: Hypothesized haemodynamic response to face stimuli
in arbitrary units over time in seconds for ‘Loc1’.
Fig. 8: Average model validation for all-stims. Superimposition of all models applied to validation subjects (translucent,
many colours) against the hypothesized HR. Time series data
are shown in arbitrary units over time in seconds. In validation, there was no significant difference between modelling
strategies.
Fig. 7: Distribution of non-linear model R scores for withinsubject generalizability. x-axis: R scores. y-axis: Quantity of
non-linear models in score interval. Example of the distribution of R scores between non-linear models produced by GP
and the hypothesized HR for ’Loc2’ (all-stims). The vertical
black line denotes the score observed for the corresponding
linear model. In this and most other cases, the number of
non-linear models worse than linear models forms a majority,
but never a totality.
Fig. 9: Visualization of outputs from models fitted using experimental data (translucent, multiple colours), the output from
models fitted using unrelated resting state data (orange), and the stimulus-related hypothesized HR (blue). Associated R scores
are summarized by Table VII. Time series data are shown in arbitrary units over time in seconds.
| 1cs.CV
|
Provable learning of Noisy-or Networks
Sanjeev Arora∗
Rong Ge†
Tengyu Ma
‡
Andrej Risteski§
arXiv:1612.08795v1 [cs.LG] 28 Dec 2016
December 30, 2016
Abstract
Many machine learning applications use latent variable models to explain structure in
data, whereby visible variables (= coordinates of the given datapoint) are explained as a
probabilistic function of some hidden variables. Finding parameters with the maximum
likelihood is NP-hard even in very simple settings. In recent years, provably efficient
algorithms were nevertheless developed for models with linear structures: topic models,
mixture models, hidden markov models, etc. These algorithms use matrix or tensor
decomposition, and make some reasonable assumptions about the parameters of the
underlying model.
But matrix or tensor decomposition seems of little use when the latent variable model
has nonlinearities. The current paper shows how to make progress: tensor decomposition is applied for learning the single-layer noisy or network, which is a textbook example
of a Bayes net, and used for example in the classic QMR-DT software for diagnosing
which disease(s) a patient may have by observing the symptoms he/she exhibits.
The technical novelty here, which should be useful in other settings in future, is analysis of tensor decomposition in presence of systematic error (i.e., where the noise/error
is correlated with the signal, and doesn’t decrease as number of samples goes to infinity).
This requires rethinking all steps of tensor decomposition methods from the ground up.
For simplicity our analysis is stated assuming that the network parameters were
chosen from a probability distribution but the method seems more generally applicable.
∗
Princeton University, Computer Science Department. [email protected]
Duke University, Computer Science Department. [email protected]
‡
Princeton University, Computer Science Department. [email protected]
§
Princeton University, Computer Science Department. [email protected]
†
1
Introduction
Unsupervised learning is important and potentially very powerful because of the availability of
the huge amount of unlabeled data — often several orders of magnitudes more than the labeled
data in many domains. Latent variable models, a popular approach in unsupervised learning,
model the latent structures in data: the “structure”corresponds to some hidden variables, which
probabilistically determine the values of the visible coordinates in data. Bayes nets model the
dependency structure of latent and observable variables via a directed graph. Learning parameters
of a latent variable model given data samples is often seen as a canonical definition of unsupervised
learning. Unfortunately, finding parameters with the maximum likelihood is NP-hard even in very
simple settings. However, in practice many of these models can be learnt reasonably well using
algorithms without polynomial runtime guarantees, such as expectation-maximization algorithm,
Markov chain Monte Carlo, and variational inference. Bridging this gap between theory and practice
is an important research goal.
Recently it has become possible to use matrix and tensor decomposition methods to design polynomial-time algorithms to learn some simple latent variable models such as topic models [AGM12, AGH+ 13], sparse coding models [AGMM15, MSS16], mixtures of Gaussians [HK13,
GHK15], hidden Markov models [MR05], etc. These algorithms are guaranteed to work if the
model parameters satisfy some conditions, which are reasonably realistic. In fact, matrix and tensor decomposition are a natural tool to turn to since they appear to be a sweet spot for theory
whereby non-convex NP-hard problems can be solved provably under relatively clean and interpretable assumptions. But the above-mentioned recent results suggest that such methods apply
only to solving latent variable models that are linear: specifically, they need the marginal of the
observed variables conditioned on the hidden variables to depend linearly on the hidden variables.
But many settings seem to call for nonlinearity in the model. For example, Bayes nets in many
domains involve highly nonlinear operations on the latent variables, and could even have multiple
layers. The study of neural networks also runs into nonlinear models such as restricted Boltzmann
machines (RBM) [Smo86, HS06]. Can matrix factorization (or related tensor factorization) ideas
help for learning nonlinear models?
This paper takes a first step by developing methods to apply tensor factorization to learn possibly the simplest nonlinear model, a single-layer noisy-or network. This is a direct graphical model
with hidden variable d ∈ {0, 1}m , and observation node s ∈ {0, 1}n . The hidden variables d1 , . . . , dm
are independent and assumed to have Bernoulli distributions. The conditional distribution Pr[s|d]
is parameterized by a non-negative weight matrix W ∈ Rn×m . We use W i to denote the i-row of
W . Conditioned on d, the observations s1 , . . . , sn are assume to be independent with distribution
Pr [si = 0 | d] =
m
Y
exp(−Wij dj ) = exp(−hW i , di) .
(1.1)
j=1
We see that 1 − exp(−Wji dj ) can be thought of as the probability that dj activates symptom si ,
and si is activated if one of dj ’s activates it — which explains the name of the model, noisy-or. It
follows that the conditional distribution s | d is
Pr[s | d] =
n
Y
i=1
s
1−si
1 − exp(−hW i , di) i exp(−hW i , di)
.
One canonical use of this model is to model the relationship between diseases and symptoms,
as in the classical human-constructed tool for medical diagnosis called Quick Medical Reference
1
(QMR-DT) by (Miller et al.[MPJM82], Shwe et al. [SC91]) This textbook example ([JGJS99]) of
a Bayes net captures relationships between 570 binary disease variables (latent variables) and 4075
observed binary symptom variables, with 45, 470 directed edges, and the Wij ’s are small integers.1
The name “noisy-or ”derives from the fact that Q
the probability that the OR of m independent
binary variables y1 , y2 , . . . , ym is 1 is exactly 1 − j (Pr[yj = 0]). Noisy-or models are implicitly
using this expression; specifically, for the i-th symptom we are considering the OR of m events
where the jth event is “Disease j does not cause symptom i” and its probability is exp(−Wij dj ).
Treating these events as independent leads to expression (1.1).
The parameters of the QMR-DT network were hand-estimated by consulting human experts,
but it is an interesting research problem whether such networks can be created in an automated
way using only samples of patient data (i.e., the s vectors). Previously there were no approaches for
this that work even heuristically at the required problem size (n = 4000). (This learning problem
should not be confused with the simpler problem of infering the latent variables given the visible
ones, which is also hard but has seen more work, including reasonable heuristic methods [JGJS99]).
Halpern et al.[HS13, JHS13] have designed some algorithms for this problem. However, their first
paper [HS13] assumes the graph structure is given. The second paper [JHS13] requires the Bayes
network to be quartet-learnable, which is a strong assumption on the structure of the network.
Finally, the problem of finding a “best-fit” Bayesian network according to popular metrics2 has
been shown to be NP-complete by [Chi96] even when all of the hidden variables are also observed.
Our algorithm and analysis. Our algorithm uses Taylor expansion —on a certain correlation
measure called PMI, whose use in this context is new—to convert the problematic exponential into
an infinite sum, where we can ignore all but the first two terms. This brings the problem into the
realm of tensor decomposition but with several novel twists having to do with systematic error (see
the overview in Section 2). Our recovery algorithm makes several assumptions about W , the matrix
of connection weights, listed in Section 3. We verified some of these on the QMR-DT network, but
the other assumptions are asymptotic in nature. Thus the cleanest description of our algorithm
is in a clean average-case setting. First, we assume all latent variables are iid Bernoulli Ber(ρ)
for some ρ, which should be thought of as small (In the QMR-DT application, ρ is like O(1/m).)
Next we assume that the ground truth W ∈ Rn×m is created by nature by picking its entries in iid
fashion using the following random process:
(
0,
with probability 1 − p
Wij =
f
Wij , with probability p
fij ’s are upper bounded by νu for some constant νu and are identically distributed according
where W
to a distribution D which satisfies that for some constant νl > 0,
h
i
fij2 ) ≤ 1 − νl .
exp(−W
(1.2)
E
fij ∼D
W
fij is bounded away from 0. We will assume that
The condition (1.2) intuitively requires that W
p ≤ 1/3 and νu = O(1), νl = Ω(1). (Again, these are realistic for QMR-DT setting).
1
We thank Randolph Miller and Vanderbilt University for providing the current version of this network for our
research.
2
Researchers resort to these metrics as when the graph structure is unknown, multiple structures may have the
same likelihood, so maximum likelihood is not appropriate.
2
Theorem 1.1 (Informally stated). There exists a polynomial time algorithm (Algorithm 1) that,
given polynomially many samples from the noisy OR network described in the previous paragraph,
e √pm) relative error in ℓ2 -norm in each column.
recovers the weight matrix W with O(ρ
Recall that we mostly thought of the prior of the diseases ρ as being on the order O(1/m). This
√
means that even if p is on the order of 1, our relative error bound equals to O(1/ m) ≪ 1.
2
Preliminaries and overview
We denote by 0 the all-zeroes vector and 1 the all-ones vector. A+ will denote the Moore-Penrose
pseudo-inverse of a matrix A, and for symmetric matrices A, we use A−1/2 as a shorthand for
(A+ )1/2 . The least non-zero singular value of matrix A is denoted σmin (A).
For matrices A, B we define the Kronecker product ⊗ as (A ⊗ B)ijkl = Aij Bkl . A useful identity
is that (A⊗B)·(C ⊗D) = (AC)⊗(BD) whenever the matrix multiplications are defined. Moreover,
Ai will denote the i-th column of matrix A and Ai the i-th row of matrix A.
We write A . B if there exists a universal constant c such that A ≤ cB and we define &
similarly.
The pointwise mutual information of two binary-valued random variables x and y is
E[xy]
P M I2(x, y) , log E[x]
. Note that it is positive iff E[x, y] > E[x] E[y] and thus is used as a
E[y]
measure of correlation in many fields. This concept can be extended in more than one way to a
triple of boolean variables x, y, z and we use
P M I3(x, y, z) , log
E[xy] E[yz] E[zx]
.
[xyz]
E
E[x] E[y] E[z]
(2.1)
(We will sometimes shorten PMI3 and PMI2 to PMI when this causes no confusion.)
2.1
The Algorithm in a nutshell
Our algorithm is given polynomially many samples from the model (each sample describing which
symptoms are or are not present in a particular patient). It starts by computing the following
matrix n × n PMI and and n × n × n tensor PMIT, which tabulate the correlations among all pairs
and triples of symptoms (specifically, the indicator random variable for the symptom being absent):
PMIij , P M I2(1 − si , 1 − sj ).
PMITi,j,k , P M I3(1 − si , 1 − sj , 1 − sj )
(2.2)
(2.3)
The next proposition makes the key observation that the above matrix and tensor are close
to rank m, which we recall is much smaller than n. Here and elsewhere, for a matrix or vector
A we use exp(A) for the matrix or vector obtained by taking the entries-wise exponential. For
convenience, we define F, G ∈ Rn×m as
F , 1 − exp(−W )
G , 1 − exp(−2W ) .
3
(2.4)
(2.5)
Proposition 2.1 (Informally stated). Let Fk , Gk denote the kth columns of the above F, G. Then,
PMI ≈ ρ F F ⊤ + ρGG⊤
PMIT ≈ ρ
m
X
k=1
|
=
{z
m
X
Fk Fk⊤ + ρ2
|
}
m
X
k=1
m
X
Gk G⊤
k
(2.6)
k=1
k=1
Fk ⊗ Fk ⊗ Fk + ρ
:=S
ρ
Gk ⊗ Gk ⊗ Gk .
{z
:=E
(2.7)
(2.8)
}
The proposition is proved later (with precise statement) in Section A by computing the moments
by marginalization and using Taylor expansion to approximate the log of the moments, and ignoring
terms ρ3 and smaller. (Recall that ρ is the probability that a patient has a particular disease,
which should be small, of the order of O(1/n). The dependence of the final error upon ρ appears in
Section 3.) Since the tensor PMIT can be estimated to arbitrary accuracy given enough samples,
the natural idea to recover the model parameters W is to use Tensor Decomposition. This is what
our algorithm does as well, except the following difficulties have to be overcome.
Difficulty 1: Suppose in equation (2.8) we view the first summand S, which is rank m with components Fk ’s as the signal term. In all previous
P polynomial-time algorithms for tensor decomposition,
the tensor is required to have the form m
k=1 Fk ⊗ Fk ⊗ Fk + noise. To make our problem fit this
template we could consider the second summand E as the “noise”, especially since it is multiplied
by ρ ≪ 1 which tends to make E have smaller norm than S. But this is naive and incorrect, since
E is a very structured matrix: it is more appropriate viewed as systematic error. (In particular
this error doesn’t go down in norm as the number of samples goes to infinity.) In order to do tensor
decomposition in presence of such systematic error, we will need both a delicate error analysis and
a very robust tensor decomposition algorithm. These will be outlined in Section 2.3.
Difficulty 2: To get our problem into a form suitable for tensor decomposition requires a whitening
step, which uses the robust estimate of the whitening matrix from the second moment matrix. In
this case, the whitening matrix has to be extracted out of the PMI matrix, which itself suffers from
a systematic error. This also is not handled in previous works, and requires a delicate control of
the error. See Section 2.4 for more discussion.
Difficulty 3: There is another source of inexactness in equation (2.8), namely the approximation
is only true for those entries with distinct indices — for example, the diagonal entry PMIii has
completely different formula from that for PMIij when i 6= j. This will complicate the algorithm,
as described in Subsections 2.3 and 2.4.
The next few Subsections sketch how to overcome these difficulties, and the details appear in
the rest of the paper.
2.2
Recovering matrices in presence of systematic error
In this Section we recall the classical method of approximately recovering a matrix given noisy
estimates of its entries. We discuss how to adapt that method to our setting where the error in
the estimates is systematic and does not go down even with many more samples. The next section
sketches an extension of this method to tensor decomposition with systematic error.
In the classical setting, there is an unknown n × n matrix S of rank m and we are given
S + E where E is an error matrix. The method to recover S is to compute the best rank-m
approximation to S +E. The quality of this approximation was studied by Davis and Kahan [DK70]
and Wedin [Wed72], and many subsequent authors. The quality of the recovery depends upon the
4
ratio ||E||/σm (S), where σm (·) denotes m-th largest singular value and || · || denotes the spectral
norm. To make this familiar lemma fit our setting more exactly, we will phrase the problem as
trying to recover a matrix S given noisy estimate SS ⊤ + E. Now one can only recover S up to
rotation, and the following lemma describes the error in the Davis-Kahan recovery. It also plays a
key role in the error analysis of the usual algorithm for tensor decomposition.
b the subspace of the top m eigenvectors of SS ⊤ and
Lemma 2.2. In the above setting, let K, K
SS ⊤ + E. Let ε be such that kEk ≤ ε · σm (SS ⊤ ). Then IdK − IdKb . ε where Id is the identity
transformation on the subspace in question.
The Lemma thus treats ||E||/σm (SS ⊤ ) as the definition of noise/signal ratio. Before we generalize the definition and the algorithm to handle systematic error it is good to get some intuition,
from looking at (2.6): PMI ≈ ρ(F F T + ρGGT ). Thinking of the first term as signal and the
second as error, let’s check how bad is the noise/signal ratio defined in Davis-Kahan. The “signal”is σm (F F ⊤ ), which is smaller than n since the trace of F F ⊤ is of the order of mn in our
probabilistic model for the weight matrix. The “noise” is the norm of ρGG⊤ , which is large since
the Gk ’s are nonnegative vectors with P
entries of the order of 1, and therefore the quadratic form
h √1n 1, ρGG⊤ √1n 1i can be as large as ρ k hGk , √1n 1i2 ≈ ρmn. Thus the Davis-Kahan noise/signal
ratio is ρm, and so when ρm ≪ 1, it allows recovering the subspace of F with error O(ρm). Note
that this is a vacuous bound since ρ needs to be at least 1/m so that the hidden variable d contains
1 non-zero entry in average. We’ll argue that this error is too pessimistic and we can in fact drive
the estimation error down to close to ρ.
Definition 2.3 (spectral boundedness). Let n ≥ m. Let E ∈ Rn×n be a symmetric matrix and
S ∈ Rn×m . Then, we say E is τ -spectrally bounded by S if
E τ (SS ⊤ + σm (SS ⊤ ) · Idn )
(2.9)
The smallest such τ is the “error/signal ratio” for this recovery problem.
This definition differs from Davis-Kahan’s because of the τ SS ⊤ term on the right hand side
of (2.9). This allows, for any unit vector x, the quadratic form value xT Ex to be as large as
τ (xT SS ⊤ x + σm (SS ⊤ )). Thus for example the 1 vector no longer causes a large noise/signal ratio
since both quadtratic forms F F ⊤ and GG⊤ have large values on it.
This new error/signal ratio is no larger than the Davis-Kahan ratio, but can potentially be
much smaller. Now we show how to do a better analysis of the Davis-Kahan recovery in terms of
it. The proof of this theorem appears in Section 4.
Theorem 2.4 (matrix perturbation theorem for systematic error). Let n ≥ m. Let S ∈ Rn×m be
of full rank. Suppose positive semidefinite matrix E ∈ Rn×n is ε-spectrally bounded by S ∈ Rn×m
b the subspace of the top m eigenvectors of SS ⊤ and SS ⊤ + E. Then,
for ε ∈ (0, 1). Let K, K
IdK − IdKb . ε .
Finally, we should consider what this new definition of noise/signal ratio achieves. The next
proposition (whose proof appears in Section B) shows that that under the generative model for W
√
sketched earlier, τ = O(log n). Therefore, ρG is Õ(ρ)-bounded by F , and the recovery error of
the subspace of F from F F ⊤ + ρGG⊤ is Õ(ρ) (instead of O(ρm) using Davis-Kahan).
Proposition 2.5. Under the generative model for W , w.h.p, the matrix G = 1 − exp(−2W ) is
τ -spectrally bounded by F = 1 − exp(−W ), with τ = Õ(1).
5
Empirically, we can compute the τ value for the weight matrix W in the QMR-DT
dataset [SC91], which is a textbook application of noisy OR network. For the QMR-DT dataset, τ
is under 6. This implies that the recovery error of the subspace of F guaranteed by Theorem 2.4
is bounded by O(τ ρ) ≈ ρ, whereas the error bound by Davis-Kahan is O(ρm).
2.3
Tensor decomposition with systematic error
Now we extend the insight from the matrix case to tensor recovery under systematic error. In turns
out condition (2.9) is also a good measure of error/signal for the tensor recovery problem of (2.8).
Specifically, if G is τ -bounded by F , then we can recover the components Fk ’s from the PMIT with
√
column-wise error O(ρτ 3/2 m). This requires a non-trivial algorithm (instead of SVD), and the
additional gain is that we can recover Fk ’s individually, instead of only obtaining the subspace with
the PMI matrix.
First we recall the prior state of the art for the error analysis of tensor decomposition with
Davis-Kahan type bounds. The best error bounds involve measuring the magnitude of the noise
matrix Z in a new way. For any n1 × n2 × n3 tensor T , we define the k·k{1}{2,3} norm as
X
xi yjk Tijk .
(2.10)
kT k{1}{2,3} :=
sup
x∈Rn1 ,y∈Rn2 n3
kxk=1,kyk=1
i∈[n1 ]
(j,k)∈[n2 ]×[n3 ]
Note that this norm is in fact the spectral norm of the flattening of the tensor (into a n1 × n2 n3
dimensional matrix). This norm is larger than the injective norm3 , but recently [MSS16] shows that
ε-error in this norm implies O(ε)-error in the recovery guarantees of the components, whereas if one
uses injective norm, the guarantees often pick up an dimension-dependent factor [AGH+ 14]. We
define k·k{2}{1,3} norm similarly. As is customary in tensor decomposition, the theorem is stated
for tensors of a special form, where the components {ui }, {vi }, {wi } are orthonormal families of
vectors. This can be ensured without loss of generality using a procedure called whitening that
uses the 2nd moment matrix.
Theorem 2.6 (Extension of [MSS16, Theorem 10.2]). There is a polynomial-time algorithm (Algorithm 2 later) which has the following guarantee. Suppose tensor T is of the form
T =
r
X
i=1
ui ⊗ vi ⊗ wi + Z
where {ui }, {vi }, {wi } are three collections of orthonormal vectors in Rd , and kZk{1}{2,3} ≤ ε,
kZk{2}{1,3} ≤ ε. Then, it returns {(ũi , ṽi , w̃i )} in polynomial time that is O(ε)-close to {(ui , vi , wi )}
in ℓ2 norm up to permutation. 4
But in our setting the noise tensor has systematic error. An analog of Theorem 2.4 in this
setting is complicated because even the whitening step is nontrivial. Recall also the inexactness
in Proposition 2.1 due to the diagonal terms, which we earlier called Difficulty 3. We address this
difficulty in the algorithm by setting up the problem using a sub-tensor of the PMI tensor. Let
Sa , Sb , Sc be a uniformly random equipartition of the set of indices [n]. Let
ak = Fk,Sa ,
3
bk = Fk,Sb ,
ck = Fk,Sc ,
The injective norm of the tensor T is defined as kT k{1}{2,3} := supx∈Rn1 ,y∈Rn2 z∈Rn3
kxk=1,kyk=1,kzk=1
4
(2.11)
P
i∈[n1 ],j∈[n2 ],k∈[n3 ]
xi yj zk Tijk .
Precisely, here we meant that there exists a permutation π such that for every i, max{kũπ(i) − ui k, kṽπ(i) −
vi k, kw̃π(i) − wi k} ≤ O(ε)
6
where Fk,S denotes the restriction of vector Fk to subset S. Moreover, let
γk = Gk,Sa ,
δk = Gk,Sb ,
θk = Fk,Sc .
(2.12)
Then, since the sub-tensor PMITSa ,Sb ,Sc only contains entries with distinct indices, we can use
Taylor expansion (see Lemma A.1) to obtain that
X
X
γk ⊗ δk ⊗ θk + higher order terms .
PMITSa ,Sb ,Sc = ρ
ak ⊗ bk ⊗ ck + ρ2
k∈[m]
k∈[m]
Here the second summand on the RHS corresponds to the second order term in the Taylor expansion.
It turns out that the higher order terms are multiplied by ρ3 and thus have negligible Frobenius
norm, and therefore discussion below will focus on the first two summands.
For simplicity, let T = PMITSa ,Sb ,Sc . Our goal is to recover the components ak , bk , ck from the
approximate low-rank tensor T .
The first step is to whiten the components ak ’s, bk ’s and ck ’s. Recall that ak = Fk,Sa is a
non-negative vector. This implies the matrix A = [a1 , . . . , am ] must have a significant contribution
in the direction of the vector 1, and thus is far away from being well-conditioned. For the purpose
of this section, we assume for simplicity that we can access the covariance matrix defined by the
vector ak ’s,
X
(2.13)
Q̄a := AA⊤ =
ak a⊤
k .
k∈[m]
Similarly we assume the access of Q̄b and Q̄c which are defined analogously. In Section 2.4 we
discuss how to obtain approximately these three matrices.
Then,
we can compute the whitened tensor by applying transformation
1/2 , (Q̄+ )1/2 , (Q̄+ )1/2 along the three modes of the tensor T ,
)
(Q̄+
c
a
b
1/2
1/2
1/2
⊗ (Q̄+
·T =ρ
(Q̄+
⊗ (Q̄+
c )
a)
b )
k∈[m]
+ ρ2
|
X
1/2
1/2
1/2
bk ⊗ (Q̄+
ck
(Q̄+
ak ⊗ (Q̄+
c )
a)
b )
X
1/2
1/2
1/2
(Q̄+
γk ⊗ (Q̄+
δk ⊗ (Q̄+
θk +negligible terms
a)
c )
b )
k∈[m]
{z
}
:=Z
1/2 a ’s are orthonormal
Now the first summand is a low rank orthogonal tensor, since (Q̄+
k
a)
vectors. However, the term Z is a systematic error and we use the following Lemma to control its
k·k{1}{2,3} norm.
Lemma 2.7. Let n ≥ m and A, B, C ∈ Rn×m be full rank matrices and let Γ, ∆, Θ ∈ Rd×ℓ . Let
γi , δi , θi be the i-th column of Γ, ∆, Θ, respectively. Let Q̄a = AA⊤ , Q̄b = BB ⊤ , Q̄c = CC ⊤ . Suppose
ΓΓ⊤ (and ∆∆⊤ , ΘΘ⊤ ) is τ -spectrally bounded by A (and B, C respectively), then,
X
1/2
1/2
1/2
(Q̄+
γi ⊗ (Q̄+
δi ⊗ (Q̄+
θi
c )
a)
b )
i∈[ℓ]
≤ (2τ )3/2 .
{1}{2,3}
Lemma 2.7 shows that to give an upper bound on the k·k{1}{2,3} norm of the error tensor Z,
it suffices to show that the square of the components of the error, namely, ΓΓ⊤ , ∆∆⊤ , ΘΘ⊤ are
7
τ -spectrally bounded by the components of the signal A, B, C respectively. This will imply that
kZk{1}{2,3} ≤ (2τ )3/2 ρ2 .
Recall that A and Γ are two sub-matrices of F and G. We have shown that GG⊤ is τ -spectrally
bounded by F in Proposition 2.5. It follows straightforwardly that the random sub-matrices also
have the same property.
Proposition 2.8. In the setting of this section, under the generative model for W , w.h.p, we have
that ΓΓ⊤ is τ -spectrally bounded by A with τ = O(log n). The same is true for the other two
modes.
Using Proposition 2.8 and Lemma 2.7, we have that
kZk{1}{2,3} . ρ2 log3/2 (n) .
1/2 ⊗ (Q̄+ )1/2 ⊗ (Q̄+ )1/2 · T , we can recover the comThen using Theorem 2.6 on the tensor (Q̄+
c
a)
b
+ 1/2
+
1/2
+
1/2
ponents (Q̄a ) ak ’s, (Q̄b ) bk ’s, and (Q̄c ) ck ’s. This will lead us to recover ak ,bk and ck , and
finally to recover the weight matrix W .
2.4
Robust whitening
In the previous subsection, we assumed the access to Q̄a , Q̄b , Q̄c (defined in (2.13)) which turns out
to be highly non-trivial. A priori, using equation (2.6), noting that A = [F1,Sa , . . . , Fm,Sa ], we have
PMISa ,Sa /ρ ≈ Q̄a + error .
However, this approximation can be arbitrarily bad for the diagonal entries of PMI since equation (2.6) only works for entries with distinct indices. (Recall that this is why we divided the
indices set into Sa , Sb , Sc and studied the asymmetric tensor in the previous subsection). Moreover,
the diagonal of the matrix Q̄a contributes to its spectrum significantly and therefore we cannot get
meaningful bounds (in spectral norm) by ignoring the diagonal entries.
This issue turns out to arise in most of the previous tensor papers and the solution was to
compute AA⊤ by using the asymmetric moments AB ⊤ , BC ⊤ , CA⊤ ,
AA⊤ = (AB ⊤ )(CB ⊤ )+ (CA⊤ ) .
Typically AB ⊤ , BC ⊤ , CA⊤ can be estimated with arbitrarily small error (as number of samples
go to infinity) and therefore the equation above leads to accurate estimate to AA⊤ . However,
in our case the errors in the estimate PMISa ,Sb ≈ AB ⊤ , PMISb ,Sc ≈ BC ⊤ , PMISc ,Sa ≈ CA⊤ are
systematic. Therefore, we need to use a more delicate analysis to control how the error accumulates
in the estimate,
Q̄a ≈ PMISa ,Sb · PMI−1
Sb ,Sc · PMISc ,Sa .
Here again, to get an accurate bound, we need to understand how the error in PMISa ,Sb −
AB ⊤ behaves relatively compared with AB ⊤ in a direction-by-direction basis. We generalized
Definition 2.3 to capture the asymmetric spectral boundedness of the error by the signal.
Definition 2.9 (Asymmetric spectral boundedness). Let n ≥ m and B, C ∈ Rn×m . We say a
matrix E ∈ Rn×n is ε-spectrally bounded by (B, C) if E can be written as:
⊤
E = B∆1 C ⊤ + B∆⊤
2 + ∆3 C + ∆4 .
(2.14)
Here ∆1 ∈ Rm×m , ∆2 , ∆3 ∈ Rn×m and ∆4 ∈ Rn×n are matrices whose spectral norms are bounded
by: k∆1 k ≤ ε, k∆2 k ≤ εσmin (C), k∆3 k ≤ εσmin (B) and k∆4 k ≤ εσmin (B)σmin (C).
8
Let K be the column subspace of B and H be the column subspace of C. Then we have
∆1 = B + E(C ⊤ )+ , ∆2 = B + EIdH ⊥ , ∆3 = IdK ⊥ E(C ⊤ )+ , ∆4 = IdK ⊥ EIdH ⊥ . Intuitively, they
measure the relative relationship between E and B, C in different subspaces. For example, ∆1 is
the relative perturbation in the column subspace of K and row subspace of H. When B = C,
this is equivalent to the definition in the symmetric setting (this will be clearer in the proof of
Theorem 2.4).
Theorem 2.10 (Robust whitening theorem). Let n ≥ m and A, B, C ∈ Rn×m .
Σab , Σbc , Σca ∈ Rn×n are of the form,
Suppose
Σab = AB ⊤ + Eab , Σbc = BC ⊤ + Ebc , and Σca = CA⊤ + Eca .
where Eab , Ebc , Eca are ε-spectrally bounded by (A, B), (B, C), (C, A) respectively. Then, the matrix
matrix
+
Qa = Σab [Σ⊤
bc ]m Σca
+
⊤
is a good approximation of AA⊤ in the sense that Qa = Σab [Σ⊤
bc ]m Σca − AA is O(ε)-spectrally
bounded by A. Here [Σ]m denotes the best rank-m approximation of Σ.
The theorem is non-trivial even if the we have an absolute error assumption, that is, even
if kEbc k ≤ τ σmin (B)σmin (C), which is stronger condition than Ebc is τ -spectrally bounded by
⊤ +
⊤
(B, C). Suppose we establish bounds on kΣab − AB ⊤ k, kΣ+⊤
bc − (BC ) k and kΣab − AB k
+
⊤
individually, and then putting them together in the obvious way to control the error Σab [Σbc ]m Σca −
AB ⊤ (BC ⊤ )+ CA⊤ . Then the error will be too large for us. This is because standard matrix
⊤ −1
⊤ −1 2 .
perturbation theory gives that kΣ−⊤
bc − (BC ) k can be bounded by O kEbc kk(BC ) k
ε/[σmin (B)σmin (C)], which is tight. Then we multiply the error with the norm of the rest of the
(B)σmax (C)
. That is, we will loss a condition number of
two terms, the error will be roughly ε · σσmax
min (B)σmin (C)
B, C, which can be dimension dependent for our case.
+
The fix to this problem is to avoid bounding each term in Σab [Σ⊤
bc ]m Σca individually. To do
this, we will take the cancellation of these terms into account. Technically, we re-decompose the
+
+
+
⊤ +
product Σab [Σ⊤
bc ]m Σca into a new product of three matrices (Σab B )(B[Σbc ]m C)(C Σca ), and then
bound the error in each of these terms instead. See Section C for details.
1/2 a ’s are indeed approximately
As a corollary, we conclude that the whitened vectors (Q+
i
a)
orthonormal.
1/2 A contains approximately
Corollary 2.11. In the setting of Theorem 2.10, we have that (Q+
a)
orthonormal vectors as columns, in the sense that
1/2
1/2
k(Q+
AA⊤ (Q+
− Idk . ε .
a)
a)
Therefore we have found an approximate whitening matrix for A even though we do not have
access to the diagonal entries.
3
Main Algorithms and Results
As sketched in Section 2, our main algorithm (Algorithm 1) uses tensor decomposition on the PMI
tensor. In this section, we describe the different steps and how the fit together. Subsequently, all
steps will be analyzed in separate sections.
9
Algorithm 1 Learning Noisy-Or Networks via Decomposing PMI Tensor
Inputs: N samples generated from a noisy-or network, disease prior ρ
c.
Outputs: Estimate of weight matrix W
d PMIT
\ using equation (E.1).
1. Compute the empirical PMI matrix and tensor PMI,
2. Choose a random equipartition Sa , Sb , Sc of [n].
\ via Algorithm 4 for the partitioning
3. Obtain approximate whitening matrices for PMIT
Sa , Sb , Sc
4. Run robust tensor-decomposition Algorithm 3 to obtain vectors âi , b̂i , ĉi , i ∈ [m]
1/3 â , 1 − ( 1−ρ )1/3 b̂ , 1 − ( 1−ρ )1/3 ĉ .
5. Let Yi be the concatenation of the three vectors 1 − ( 1−ρ
i
i
i
ρ )
ρ
ρ
(Recall that âi , b̂i , ĉi are of dimension n/3 each.)
c , where
6. Return W
ci,j
W
(
− log((Yi )j ), if (Yi )j > exp(−νu ) :
=
exp(−νu ),
otherwise
Theorem 3.1 (Main theorem, random weight matrix). Suppose the true W is generated from
the random model in Section 1 with ρpm ≤ c for some sufficiently small constant c. Then given
c in polynomial
N = poly(n, 1/p/, 1/ρ) number of examples, Algorithm 1 returns a weight matrix W
time that satisfies
ci − Wi k2 ≤ O(η
e √pn) ,
∀i ∈ [m], kW
√
where η = Õ mpρ .
√
Note that the column ℓ2 norm of Wi is on the order of pn, and thus η can be thought of as the
relative error in ℓ2 norm. Note also that Pr[si = 0] = 1 − Pr[si = 1] ≈ 1 − pmρ, so ρpm = o(1) is
necessary purely for sample complexity reasons. Finally, we can also state a result with a slightly
weaker guarantee, but with only deterministic assumptions on the weight matrix W . Recall that
F = 1 − exp(−W ) and G = 1 − exp(−2W ). We will also define third and fourth-order terms
H = 1 − exp(−3W ), L = 1 − exp(−4W ).
We also define the incoherence of a matrix F . Roughly speaking, it says that the left singular
vectors of F don’t correlate with any of the natural basis vector much more than the average.
Definition 3.2 (Incoherence:). Let Fp∈ Rn×m have singular value decomposition F = U ΣV ⊤ . We
say F is µ-incoherent if maxi kUi k ≤ µm/n. where Ui is the i-th row of U .
We assume the weight matrix W satisfies the following deterministic assumptions,
1. GG⊤ , HH ⊤ , LL⊤ is τ -spectrally
p bounded by F for τ ≥ 1.
e n/m).
2. F is µ-incoherent with µ ≤ O(
3. If maxi kFi k0 ≤ pn, with high probability over the choice of a subset Sa , |Sa | = n/3, σmin (FSa ) &
√
np and ρpm ≤ c for some sufficiently small constant c.
Theorem 3.3 (Main theorem, deterministic weight matrix). Suppose the matrix W satisfies the
c in polynomial
conditions 1-3 above. Given polynomial number of samples, Algorithm 1 returns W
time, s.t.
ci − Wi k2 ≤ O(η
e √np) .
∀i ∈ [m], kW
10
for η =
√
mpρτ 3/2
√
√
Since the ℓ2 norm of Wi is on the order of np, the relative error in ℓ2 -norm is as most mρτ 3/2 ,
which mirrors the randomized case above.
The proofs of Theorems uses the overall strategy of Section 2, and is deferred to Section D. We
give a high level outline that demonstrates how the proofs depends on the machinery built in the
subsequent sections.
Both Theorem 3.1 and Theorem 3.3 are similarly proved – the only technical difference being
how the third and higher order terms are bounded. (Because of generative model assumption, for
Theorem 3.1 we can get a more precise control on them.) Hence, we will not distinguish between
them in the coming overview.
Overall, we will follow the approach outlined in Section 2. Let us step through Algorithm 1 line
by line:
1. The overall goal will be to recover the leading terms of the PMI tensor. Of course, we get
samples only, so can merely get an empirical version of it. In Section E, we show that the
simple plug-in estimator does the job – and does so with polynomially many samples.
2. Recall Difficulty 3 from Section 2 : the PMI tensor and matrix expression is only accurate
on the off-diagonal entries. In order to address this, in Section 2.3 we passed to a sub-tensor
of the original tensor by partitioning the symptoms into three disjoint sets, and considering
the induced tensor by this partition.
3. In order to apply the robust tensor decomposition algorithm from Section 5, we need to first
calculate whitening matrices. This is necessarily complicated by the fact that the diagonals
of the PMI matrix are not accurate, as discussed in Section 2.4. Section C gives guarantees
on the procedure for calculating the whitening matrices.
4. This is main component of the algorithm: the robust tensor decomposition machinery. In
Section 5, the conditions and guarantees for the success of the algorithm are formalized.
There, we deal with the difficulties layed out in Section 2.2 : namely that we have a substantial
systematic error that we need to handle. (Both due to higher-order terms, and due to the
missing diagonal entries)
5. This step, along with Step 6, is a post-processing step – which allows us to recover the weight
matrix W after we have recovered the leading terms of the PMI tensor.
We also give a short quantitative sense of the guarantee of the algorithm. (The reader can find
the full proof in Section D.)
To get quantitative bounds, we will first need a handle on spectral properties of the random
model: these are located in Section B. As we mentioned above, the main driver of the algorithm is
step 4, which uses our robust tensor decomposition machinery in Section 5. To apply the machinery,
we first need to show that the second (and higher) order terms of the PMI tensor are spectrally
bounded. This is done by applying Proposition B.4, which roughly shows the higher-order terms are
O(ρ log n)-spectrally bounded by ρF F ⊤ . The whitening matrices are calculated using machinery
in Section C. We can apply these tools since the random model gives rise to a O(1)-incoherent F
matrix as shown in Lemma B.1.
To get a final sense of what the guarantee is, the l2 error which step 4 gives, via Theorem 5.4
√
roughly behaves like σmax τ 3/2 , where σmax is the spectral norm of the whitening matrices and
τ is the spectral boundedness parameter. But, by Lemma D.1 σmax is approximately the spectral
11
norm of ρF F ⊤ – which on the other hand by Lemma B.1 is on the order of mnp2 ρ. Plugging in
these values, we get the theorem statement.
4
Finding the Subspace under Heavy Perturbations
In this section, we show even if we perturb a matrix SS ⊤ with an error whose spectral norm might
be much larger than σmin (SS ⊤ ), as long as E is spectrally bounded the top singular subspace of
S is still preserved. We defer the proof of the asymmetric case (Theorem 2.10) to Section C. We
note that such type of perturbation bounds, often called relatively perturbation bounds, have been
studied in [Ips98, Li98a, Li98b, Li97]. The results in these papers either require the that signal
matrix is full rank, or the perturbation matrix has strong structure. We believe our results are new
and the way that we phrase the bound makes the application to our problem convenient. We recall
Theorem 2.4, which was originally stated in Section 2.
Theorem 2.4 (matrix perturbation theorem for systematic error). Let n ≥ m. Let S ∈ Rn×m be
of full rank. Suppose positive semidefinite matrix E ∈ Rn×n is ε-spectrally bounded by S ∈ Rn×m
b the subspace of the top m eigenvectors of SS ⊤ and SS ⊤ + E. Then,
for ε ∈ (0, 1). Let K, K
IdK − IdKb . ε .
Proof. We can assume ε ≤ 1/10 since otherwise the statement is true (with a hidden constant 10).
Since E is a positive semidefinite matrix, we write E = RR⊤ where R = E 1/2 . Since A has full
column rank, we can write R = AS + B where S ∈ Rm×n and the columns of B are in the subspace
K ⊥ . (Specifically, we can choose S = A+ R and B = R − AA+ R = IdK ⊥ B.) By the definition of
spectral boundedness, we have
⊤
⊤
⊤
⊤
BB = IdK ⊥ RR IdK ⊥ IdK ⊥ ε AA + σm (AA )Idn IdK ⊥ .
= εσm (AA⊤ )IdK ⊥ .
Therefore, we have that kBk2 ≤ εσmin (AA⊤ ). Moreover, we also have
IdK RR⊤ IdK εAA⊤ + εσmin IdK ,
It follows that
ASS ⊤ A⊤ ≤ 2εAA⊤ .
which implies
Let P = Idm + SS ⊤
1/2
kSS ⊤ k ≤ ε.
. Then we write AA⊤ + E as,
AA⊤ + E = AA⊤ + RR⊤ = AA⊤ + (AS + B)(AS + B)⊤
= A(Id + SS ⊤ )A⊤ + ASB ⊤ + BS ⊤ A⊤ + BB ⊤
= (AP + BS ⊤ P −1 )(AP + BS ⊤ P −1 )⊤ + BB ⊤ − BS ⊤ P −2 SB ⊤
(4.1)
b = (AP + BS ⊤ P −1 ). Let K ′ be the column span of A.
b We first prove that K
b is close to K ′ .
Let A
Note that
BB ⊤ − BS ⊤ P −2 SB ⊤ . kBk2 + kBk2 S ⊤ P −2 S . kBk2
. εσmin (AA⊤ ) .
12
(since P = Id + SS ⊤ SS ⊤ )
bA
b⊤ ) = σmin (A)
b 2 = σmin (AP ) − kBS ⊤ P −1 k 2 ≥ (1 − O(ε))σmin (A)2 .
Moreover, we have σmin (A
Therefore, using Wedin’s Theorem (Lemma F.2) on equation (4.1), we have that
kIdKb − IdK ′ k . ε .
Next we show K ′ and K are also close. We have
p
b − AP k ≤ kBS ⊤ P −1 k ≤ ε σmin (A)2
kA
(4.2)
(since kSk .
√
ε, kBk .
√
ε)
b is close to the
Therefore, by Wedin’s Theorem, K ′ , as the span of top m left singular vectors of A,
span of the top left singular vector of AP , namely, K
kIdK − IdK ′ k . ε .
(4.3)
Therefore using equation (4.2) and (4.3) and triangle inequality, we complete the proof.
5
Robust Tensor Decomposition with Systematic Error
In this section we discuss how to robustly find the tensor decomposition even in presence of systematic error. We first illustrate the main techniques in an easier setting of orthogonal tensor
decomposition (Section 5.1), then we describe how it can be generalized to the general setting that
we require for our algorithm (Section 5.2).
5.1
Warm-up: Approximate Orthogonal Tensor Decomposition
We start with decomposing an orthogonal tensor with systematic error. The algorithm we use here
is a slightly more general version of an algorithm in [MSS16].
Algorithm 2 Robust orthogonal tensor decomposition
Inputs: Tensor T ∈ Rd×d×d , number δ, ε ∈ (0, 1).
Outputs: Set S = {(ãi , b̃i , c̃i )}
1. S = ∅
2. For s = 1 to O(d1+δ log d)
3.
4.
5.
Draw g ∼ N (0, Idn ), and compute M = Idn ⊗ Idn ⊗ g⊤ · T .5
Compute the top left and right singular vectors u, v ∈ Rd of M . Let z = (u⊤ ⊗v ⊤ ⊗Idn )·T .
If (u⊤ ⊗ v ⊤ ⊗ z ⊤ ) · T ≥ 1 − ζ, where ζ = O(ε), and u is 1/2-far away from any of ui ’s
with (ui , vi , wi ) ∈ S, then add (u, v, w) to S.
6. Return S
Theorem 5.1 (Stronger version of Theorem 2.6). Suppose {ui }, {vi }, {wi } are three collection
ε-approximate orthonormal vectors. Suppose tensor T is of the form
T =
r
X
i=1
5
ui ⊗ vi ⊗ wi + Z
Recall that product of two tensor (A ⊗ B ⊗ C) · (E ⊗ D ⊗ F ) = AE ⊗ BD ⊗ CF
13
with kZk{2}{1,3} ≤ τ and kZk{1}{2,3} ≤ τ . Then, with probability at least 0.9, Algorithm 2 returns
S = {(ũi , ṽi , w̃i )} which is guaranteed to be O((τ + ε)/δ)-close to {(ui , vi , wi )} in ℓ2 -norm up to
permutation.
Proof Sketch of Theorem 5.1. The Theorem is a direct extension of [MSS16, Theorem 10.2] to
asymmetric and approximate orthogonal case. We only provide a proof sketch here. We start by
writing
m
X
⊤
hg, wi iui vi⊤ + (Idn ⊗ Idn ⊗ g⊤ ) · Z
·T =
M = Id ⊗ Id ⊗ g
|
{z
}
i=1
:=Mg
{z
}
|
(5.1)
:=Ms
Since kZk{2}{1,3} ≤ τ and kZk{1}{2,3} ≤ τ , [MSS16, Theorem 6.5] implies that with probability at
least 1 − d2 over the choice of g,
p
(Idn ⊗ Idn ⊗ g⊤ ) · Z ≤ 2 log d · τ
√
Let t = 2 log d. We have that with probability 1/(d1+δ logO(1) d), hg, w1 i ≥ (1 + δ/3)t and
hg, wj i ≤ t for every j 6= 1. We condition on these events. Let ūi be a set of orthonormal vectors
such that Eu = [u1 , . . . , um ] − [ū1 , . . . , ūm ] satisfies kEu k ≤ ε (we can take ūi ’s to be the whitening
of uP
i ’s). Similarly define v̄i ’s. Then we have that the
Pterm (defined in equation (5.1)) can be written
′
′
as i hg, w1 iūi v̄i + E where kEk . ε. Let M̄S = i hg, w1 iūi v̄i . Then M̄S has top singular value
hg, w1 i ≥ (1 + δ/3)t, and second singular value at most t. Moreover, the term Mg + E ′ has spectral
norm bounded by O(τ + ε). Thus by Wedin’s Theorem (Lemma F.2), the top left and right singular
vectors u, v of MS + Mg = M̄S + Mg + E ′ are O((τ + ε)/δ)-close to ū1 and v̄1 respectively. They
are also O((τ + ε)/δ)-close to u1 , v1 since u1 is close to ū1 . Moreover, we have (u⊤ ⊗ v ⊤ ⊗ Id) · T is
O(τ /δ)-close to w1 .
Therefore, with probability 1/(d1+δ logO(1) d), each round of the for loop in Algorithm 2 will find
u1 , v1 , w1 . Line 5 is used to verify if the resulting vectors are indeed good using the injective norm
as a test. It can be shown that if the test is passed then (u, v, z) is close to one of the component.
Therefore, after d1+δ logO(1) d iterations, with high probability, we can find all of the components.
5.2
General tensor decomposition
In many previous works, general tensor decomposition is reduced to orthogonal tensor decomposition via a whitening procedure. However, here in our setting we cannot estimate the exact whitening
matrix because of the systematic error. Therefore we need a more robust version of approximate
whitening matrix, which we define below:
Definition 5.2. Let r ≤ d. A collection of r vectors {a1 , . . . , ar } is ε-approximately orthonormal if
the matrix A with ai as columns satisfies
A⊤ A − Id ≤ ε
(5.2)
Definition 5.3. Let d ≥ r and A = [a1 , . . . , ar ] ∈ Rd×r . A PSD matrix Q ∈ Rd×d is an εapproximate whitening matrix for A if (Q+ )1/2 A is ε-approximately orthonormal .
With this in mind, we can state the guarantee on the tensor decomposition algorithm (Algorithm 3).
14
Algorithm 3 Tensor decomposition with systematic error
Inputs: Tensor T ∈ Rn1 ×n2 ×n3 and ε-approximate whitening matrices Qa , Qb , Qc ∈ Rd×d .
Outputs: {âi , bˆi , ĉi }i∈[r]
1/2 ⊗ (Q+ )1/2 ⊗ (Q+ )1/2 · T
1. Compute T̃ = (Q+
c
a)
b
2. Run orthogonal tensor decomposition (Algorithm 2) with input T̃ , and obtain {ăi , b̆i , c̆i }
1/2
1/2
1/2
3. Return: {Qa ăi , Qb b̆i , Qc c̆i }
Theorem 5.4. Let d ≥ r, and A, B, C ∈ Rd×r be full rank matrices. Let Γ, ∆, Θ ∈ Rd×ℓ . Let
ai , bi , ci , γi , δi , θi be the columns of A, B, C, Γ, ∆, Θ respectively. Suppose tensor T is of the form
T =
r
X
i=1
ai ⊗ bi ⊗ ci +
ℓ
X
i=1
γi ⊗ δi ⊗ θi + E
(5.3)
Suppose matrices Qa ∈ Rd×d , Qb ∈ Rd×d , Qc ∈ Rd×d are ε-approximate whitening matrices for
1/2
1/2
1/2
A, B, C, and suppose Γ, ∆, Θ are τ -spectrally bounded by Qa , Qb , Qc , respectively. Then,
Algorithm 3 returns âi , b̂i , ĉi that are O(η)-close to ai , bi , ci in Õ(d4+δ ) time with
η . max(kQa k , kQb k , kQc k)1/2 · τ 3/2 + σ −3/2 kEk{1,2}{3} + ε · 1/δ
where σ = min(σmin (Qa ), σmin (Qb ), σmin (Qc )).
Note that in our model, the matrix E has very small spectral norm as it is the third order term
in ρ (and ρ = O(1/n)). The spectral boundedness of Γ, ∆, Θ are discussed in Section B. Therefore
we can expect the RHS to be small.
In order to prove this theorem, we show after we apply whitening operation using the approximate whitening matrices, the tensor is still close to an orthogonal tensor. To do that, we need the
following lemma which is a useful technical consequence of the condition (2.9).
Lemma 5.5. Suppose F is τ -spectrally bounded by g. Then,
kG⊤ (F F ⊤ )+ Gk ≤ 2τ .
(5.4)
Proof. Let K be the column span of F . Let Q = F F ⊤ . Multiplying (Q+ )1/2 on both sides of
equation (2.9), we obtain that
(Q+ )1/2 GG⊤ (Q+ )1/2 τ (IdK + σm (Q)Q+ )
τ (IdK + σm (Q)kQ+ kIdK )
It follows that k(Q+ )1/2 Gk
kG⊤ (Q+ )1/2 (Q+ )1/2 Gk ≤ 2τ .
≤
√
2τ IdK
2τ , which in turns implies that kG⊤ (F F ⊤ )+ Gk
=
We also need to bound the {1, 2}{3} norm of the following systematic error tensor. This is
important because we want to bound the spectral norm of the perturbation after the whitening
operation.
15
Lemma 5.6 (Variant of [MSS16, Theorem 6.1]). Let Γ, ∆, Θ ∈ Rd×ℓ . Let γi , δi , θi be the i-th
column of Γ, ∆, Θ, respectively. Then,
X
i∈[ℓ]
γi ⊗ δi ⊗ θi
≤ kΓk · kΘk · k∆k1→2 ≤ kΓk · kΘk · k∆k
(5.5)
{1,2}{3}
Proof of Lemma 5.6. Using the definition of k·k{1,2}{3} we have that
X
i∈[]
γi ⊗ δi ⊗ θi
=
X
(γi ⊗ δi )θi⊤
(5.6)
i∈[ℓ]
{1,2}{3}
≤
=
X
1/2
(γi ⊗ δi )(γi ⊗ δi )
X
1/2
θi θi⊤
i∈[ℓ]
i∈[ℓ]
(by Cauchy-Schwarz inequality)
1/2
X
(γi γi⊤ )
i∈[ℓ]
⊗
(δi δi⊤ )
kΘk
Next observe that we have that for any i, δi δi⊤ (max kδi k2 )Id and therefore,
(γi γi⊤ ) ⊗ (δi δi⊤ ) γi γi⊤ ⊗ (max kδi k2 )Id .
(5.7)
It follows that
X
i∈[r]
γi ⊗ δi ⊗ θi
≤
{1,2}{3}
X
i∈[r]
1/2
2
γi γi⊤ ⊗ (max kδi k )Id
kΘk
= kΓk · kΘk · k∆k1→2 .
With this in mind, we prove the main theorem:
1/2 A, B̃ = (Q+ )1/2 B, C̃ = (Q+ )1/2 C. Moreover, let Γ̃ =
Proof of Theorem 5.4. Let à = (Q+
a)
c
b
1/2 Γ and define ∆,
˜
(Q+
)
Θ̃
similarly.
Let
ã
,
b̃
,
c̃
,
γ̃
,
δ̃
,
θ̃
be
their
columns.
Then we have that T̃
1
i
i
i
i
i
a
as defined in Algorithm 3 satisfies
T̃ =
r
X
i=1
ãi ⊗ b̃i ⊗ c̃i +
ℓ
X
i=1
γ̃i ⊗ δ̃i ⊗ θ̃i + Ẽ
(5.8)
1/2 ⊗ (Q+ )1/2 ⊗ (Q+ )1/2 · E. We will show that T̃ meets the condition of Thewhere Ẽ = (Q+
c
a)
b
1/2 A is
orem 2.6. Since Qa is an ε-approximate whitening matrix of A, by Definition, Ã = (Q+
a)
ε-approximately orthonormal . Similarly, B̃, C̃ are ε-approximately orthonormal . √
Γ is τ -spectrally
bounded by Qa , hence by Lemma 5.5, we have that kΓ̃k ≤ 2τ . Similarly,
√
kΘk, k∆k ≤ 2τ . Applying Lemma 5.6, we have,
ℓ
X
i=1
γ̃i ⊗ δ̃i ⊗ θ̃i
{1,2}{3}
16
≤ (2τ )3/2
(5.9)
−3/2 kEk
1/2 k · k(Q+ )1/2 k · k(Q+ )1/2 kkEk
Moreover, we have kẼk{1,2}{3} ≤ k(Q+
{1,2}{3} ≤ σ
{1,2}{3} ,
c
c
a)
where σ = min{σmin (Qa ), σmin (Qb ), σmin (Qc )}.PTherefore, using Theorem 2.6 (with ai , bi , ci there
replaced by ãi , b̃i , c̃i , and Z there replaced by ℓi=1 γ̃i ⊗ δ̃i ⊗ θ̃i + Ẽ), we have that a set of vectors
{ăi , b̆i , c̆i } that are ε-close to {ãi , b̃i , c̃i } with ε = (2τ )3/2 + σ −3/2 kẼk{1,2}{3} . Therefore, we obtain
that kai − Q1/2 ăi k ≤ kQa k1/2 ε. Similarly we can control the error for bi and ci and complete the
proof.
6
Conclusions
We have presented theoretical progress on the longstanding open problem of presenting a
polynomial-time algorithm for learning noisy-or networks given sample outputs from the network.
In particular it is enouraging that linear algebraic methods like tensor decomposition can play a
role. Earlier there were no good approaches for this problem; even heuristics fail for realistic sizes
like n = 1000.
Can sample complexity be reduced, say to subcubic? (Cubic implies more than one billion
examples for networks with 1000 outputs.) Possibly this requires exploiting some hierarchichal
structure –e.g. groupings of diseases and symptoms— in practical noisy-OR networks but exploring
such possibilities using the current version of QMR-DT is difficult because it has been scrubbed of
labels for diseases and symptoms.)
Various more practical versions of our algorithm are also easy to conceive and will be tested
in the near future. This could be somewhat analogous to topic modeling, for which discovery of
provable polynomial-time algorithms soon led to very efficient algorithms.
Acknowledgments:
We thank Randolph Miller and Vanderbilt University for providing the current version of this
network for our research.
References
[AGH+ 13] Sanjeev Arora, Rong Ge, Yonatan Halpern, David M Mimno, Ankur Moitra, David
Sontag, Yichen Wu, and Michael Zhu. A practical algorithm for topic modeling with
provable guarantees. In ICML (2), pages 280–288, 2013.
[AGH+ 14] Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models. Journal of Machine
Learning Research, 15(1):2773–2832, 2014.
[AGM12]
Sanjeev Arora, Rong Ge, and Ankur Moitra. Learning topic models–going beyond svd.
In Foundations of Computer Science (FOCS), 2012 IEEE 53rd Annual Symposium on,
pages 1–10. IEEE, 2012.
[AGMM15] Sanjeev Arora, Rong Ge, Tengyu Ma, and Ankur Moitra. Simple, efficient, and neural
algorithms for sparse coding. In Proceedings of The 28th Conference on Learning
Theory, pages 113–149, 2015.
[Chi96]
David Maxwell Chickering. Learning bayesian networks is np-complete. In Learning
from data, pages 121–130. Springer, 1996.
17
[DK70]
Chandler Davis and William Morton Kahan. The rotation of eigenvectors by a perturbation. iii. SIAM Journal on Numerical Analysis, 7(1):1–46, 1970.
[GHK15]
Rong Ge, Qingqing Huang, and Sham M Kakade. Learning mixtures of gaussians in
high dimensions. In Proceedings of the Forty-Seventh Annual ACM on Symposium on
Theory of Computing, pages 761–770. ACM, 2015.
[HK13]
Daniel Hsu and Sham M Kakade. Learning mixtures of spherical gaussians: moment
methods and spectral decompositions. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, pages 11–20. ACM, 2013.
[HS06]
Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data
with neural networks. Science, 313(5786):504–507, 2006.
[HS13]
Yoni Halpern and David Sontag. Unsupervised learning of noisy-or bayesian networks.
In Uncertainty in Artificial Intelligence, page 272. Citeseer, 2013.
[Ips98]
Ilse CF Ipsen. Relative perturbation results for matrix eigenvalues and singular values.
Acta numerica, 7:151–201, 1998.
[JGJS99]
Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An
introduction to variational methods for graphical models. Machine learning, 37(2):183–
233, 1999.
[JHS13]
Yacine Jernite, Yonatan Halpern, and David Sontag. Discovering hidden variables in
noisy-or networks using quartet tests. In Advances in Neural Information Processing
Systems, pages 2355–2363, 2013.
[Li97]
Ren-Cang Li. Relative perturbation theory. iii. more bounds on eigenvalue variation.
Linear algebra and its applications, 266:337–345, 1997.
[Li98a]
Ren-Cang Li. Relative perturbation theory: I. eigenvalue and singular value variations.
SIAM Journal on Matrix Analysis and Applications, 19(4):956–982, 1998.
[Li98b]
Ren-Cang Li. Relative perturbation theory: Ii. eigenspace and singular subspace variations. SIAM Journal on Matrix Analysis and Applications, 20(2):471–492, 1998.
[MPJM82] Randolph A Miller, Harry E Pople Jr, and Jack D Myers. Internist-i, an experimental computer-based diagnostic consultant for general internal medicine. New England
Journal of Medicine, 307(8):468–476, 1982.
[MR05]
Elchanan Mossel and Sébastien Roch. Learning nonsingular phylogenies and hidden
markov models. In Proceedings of the thirty-seventh annual ACM symposium on Theory
of computing, pages 366–375. ACM, 2005.
[MSS16]
T. Ma, J. Shi, and D. Steurer. Polynomial-time Tensor Decompositions with Sum-ofSquares. ArXiv e-prints, October 2016.
[SC91]
Michael Shwe and Gregory Cooper. An empirical analysis of likelihood-weighting simulation on a large, multiply connected medical belief network. Computers and Biomedical
Research, 24(5):453–475, 1991.
[Smo86]
Paul Smolensky. Information processing in dynamical systems: Foundations of harmony theory. Technical report, DTIC Document, 1986.
18
[Ste77]
GW Stewart. On the perturbation of pseudo-inverses, projections and linear least
squares problems. SIAM review, 19(4):634–662, 1977.
[Ste90]
Gilbert W Stewart. Matrix perturbation theory. 1990.
[Wed72]
Per-Åke Wedin. Perturbation bounds in connection with singular value decomposition.
BIT Numerical Mathematics, 12(1):99–111, 1972.
A
Formal expression for the PMI tensor
In this section we formally derive the expressions for the PMI tensors and matrices, which we only
informally did in Section 2.
As a notational convenience for l ∈ N, we will denote by P̃l the matrix which has as columns the
vectors 1 − exp(−lWk ), k ∈ [m]. Furthermore, for a subset Sa ⊆ [n], we will introduce the notation
⊤
X
X
=
(1 − exp(−lWk )Sa ) (1 − exp(−lWk )Sa )⊤
(P̃l )k,Sa (P̃l )k,Sa
Pl,Sa =
k∈[m]
k∈[m]
These matrices will appear naturally in the expressions for the higher-order terms in the Taylor
expansion for the PMI matrix and tensor.
We first compute the formally the moments of the noisy-or model.
Lemma A.1. We have
log Pr[si = 0] =
X
k∈[m]
∀i 6= j log Pr [si = 0 ∧ sj = 0] =
∀ distinct i, j, k ∈ [n], log Pr [si = 0 ∧ sj = 0 ∧ sℓ = 0] =
X
k∈[m]
X
k∈[m]
log (1 − ρ(1 − exp(Wik )))
log (1 − ρ(1 − exp(Wik + Wjk )))
log (1 − ρ(1 − exp(Wik + Wjk + Wℓk )))
Proof of Lemma A.1. We only give the proof for the second equation. The rest can be shown
analogously.
h
i
log Pr [si = 0 ∧ sj = 0] = E [Pr[si = 0|d] · Pr[sj = 0|d]] = E exp(−(Wi + Wj )⊤ d)
Y
=
E [exp(−(Wik + Wjk )dk ))]
k∈[m]
=
Y
k∈[m]
(1 − ρ(1 − exp(−(Wik + Wjk )))) .
With this in mind, we give the expression for the PMI tensor along with all the higher-order
terms.
Proposition A.2. For any equipartition Sa , Sb , Sc of [n], the restriction of the PMI tensor
PMITSa ,Sb ,Sc satisfies, for any L ≥ 2,
PMITSa ,Sb ,Sc
L
X
X
ρ
Fk,Sa ⊗Fk,Sb ⊗Fk,Sc + (−1)l+1
=
1−ρ
k∈[m]
l=2
19
1
l
ρ
1−ρ
l ! X
k∈[m]
(P̃l )k,Sa ⊗(P̃l )k,Sb ⊗(P̃l )Sc +EL
(A.1)
where
kEL k{1,2},{3} ≤
(mn)3
L
1−
ρ
1−ρ
L
ρ
1−ρ
L
Proof. The proof will proceed by Taylor expanding the log terms. Towards that, using Lemma A.1,
we have :
PMITijl =
X
(1 − ρ(1 − exp(−Wik − Wjk ))) (1 − ρ(1 − exp(−Wik − Wlk ))) (1 − ρ(1 − exp(−Wjk − Wlk )))
log
(1 − ρ(1 − exp(−Wik − Wjk − Wlk ))) (1 − ρ(1 − exp(−Wik ))) (1 − ρ(1 − exp(−Wjk ))) (1 − ρ(1 − exp(−Wlk )))
k∈[m]
By the Taylor expansion of log(1 − x), we get that
PMIijl =
∞
X
1 X t
−
ρ (((1 − exp(−Wik − Wjk )))t + ((1 − exp(−Wik − Wlk )))t + ((1 − exp(−Wjk − Wlk )))t −
t
t=1
k∈[m]
(1 − exp(−Wik ))t − (1 − exp(−Wjk ))t − (1 − exp(−Wlk ))t − (1 − exp(−Wik − Wjk − Wlk ))t )
Furthermore, note that
((1 − exp(−Wik − Wjk )))t + ((1 − exp(−Wik − Wlk )))t + ((1 − exp(−Wjk − Wlk )))t −
(1 − exp(−Wik ))t − (1 − exp(−Wjk ))t − (1 − exp(−Wlk ))t − (1 − exp(−Wik − Wjk − Wlk ))t =
t
X
t
(−1)l (1 − exp(−lWik )) (1 − exp(−lWjk )) (1 − exp(−lWlk ))
l
l=1
by simple regrouping of the terms. By exchanging l and t, we get
PMITSa ,Sb ,Sc
X
t1 t
(1 − exp(−lWk ))Sa ⊗ (1 − exp(−lWk ))Sb ⊗ (1 − exp(−lWk ))Sc
ρ
=
(−1)
t l
l=1 t≥l
k∈[m]
l ! X
∞
X
1
ρ
(1 − exp(−lWk ))Sa ⊗ (1 − exp(−lWk ))Sb ⊗ (1 − exp(−lWk ))Sc
=
(−1)l+1
l 1−ρ
∞ X
X
l+1
l=1
k∈[m]
(A.2)
where the last equality holds by noting that
X 1 t 1 ρ l
ρt
=
t l
l 1−ρ
t≥l
The term corresponding to t = 1 is easily seen to be
ρ X
Fk,Sa ⊗ Fk,Sb ⊗ Fk,Sc
1−ρ
k∈[m]
therefore we to show the statement of the lemma, we only need bound the contribution of the terms
with l ≥ L.
20
Toward that, note that ∀l, kk1 − exp(−lWk )k ≤ n. Hence, we have by Lemma 5.6,
m
X
k=1
(1 − exp(−lWk ))Sa ⊗ (1 − exp(−lWk ))Sb ⊗ (1 − exp(−lWk ))Sc
{1,2},{3}
≤ (mn)3
Therefore, subadditivity of the {12}, {3} norm gives
l
ρ
∞
X
1−ρ X
(1 − exp(−lWk ))Sa ⊗ (1 − exp(−lWk ))Sb ⊗ (1 − exp(−lWk ))Sc
(−1)l+1
l
l=L
k∈[m]
∞
X
≤ (mn)
3
l=L
ρ
1−ρ
l
l
l
∞
(mn)3
ρ
(mn)3 X
=
≤
L
1−ρ
L
1−
l=L
ρ
1−ρ
L
ρ
1−ρ
{1,2},{3}
L
which gives us what we need.
A completely analogous proof gives a similar expression for the PMI matrix:
Proposition A.3. For any subsets Sa , Sb of [n], s.t. Sa ∩ Sb = ∅, the restriction of the PMI matrix
PMISa ,Sb satisfies, for any L ≥ 2,
PMISa ,Sb
L
X
X
ρ
⊤
+
(−1)l+1
Fk,Sa Fk,S
=
b
1−ρ
k∈[m]
l=2
1
l
ρ
1−ρ
where
kEL k{1,2},{3} ≤
B
(mn)2
L
1−
l ! X
(P̃l )k,Sa ((P̃l )k,Sb )⊤ + EL
(A.3)
k∈[m]
ρ
1−ρ
L
ρ
1−ρ
L
Spectral properties of the random model
The goal of this section is to prove that the random model specified in Section 1 satisfies the
incoherence property 3.2 on the weight matrix and the spectral boundedness property of the PMI
tensor. (Recall, the former is required for the whitening algorithm, and the later for the tensor
decomposition algorithm.)
Before delving into the proofs, we will need a few simple bounds on the singular values of Pl .
Lemma B.1. Let Sa ⊆ [n], s.t. |Sa | = Ω(n). With probability 1 − exp(− log2 n) over the choice of
W , and for all l = O(poly(n)),
σmin (Pl,Sa ) & np
and
σmax (Pl,Sa ) . mnp2
21
Proof. Let us proceed to the lower bound first.
If we denote by L the matrix which has as columns (P̃l )k,Sa , k ∈ [m], it’s clear that Pl,Sa = LL⊤ .
Since
σmin (LL⊤ ) = σmin (L⊤ L)
we will proceed to bound the smallest eigenvalue of L⊤ L.
Note that
X
L⊤ L =
(1 − exp(−lW k ))(1 − exp(−lW k ))⊤
k∈Sa
Since the matrices (1 − exp(−lW k ))(1 − exp(−lW k ))⊤ are independent, the bound will follow
from a matrix Bernstein bound. Denoting
i
h
Q = E (1 − exp(−lW k ))(1 − exp(−lW k ))⊤
by a simple calculation we have
h
i2
i
h
Q = p2 E 1 − exp(−lW̃ ) 11⊤ + pE (1 − exp(−lW̃ ))2 − p2 E [1 − exp(−lw̃)]2 Idm
(B.1)
where 1 is the all-ones vector of size m, and Idm is the identity of the same size. Furthermore, W̃
is a random variable following the distribution D of all the W̃i,j .
Note that (1.2) together with the assumption ν = Ω(1) gives
= Ω(p)
Pσmin (Q)
i
i ⊤ = |S | · Id , and
i
−1/2
Let Z = Q
(1 − exp(−lWi )). Then we have that E
a
m
i∈Sa Z (Z )
with high probability, kZ i k2 ≤ 1/σmin (Q) · kF i k2 . m and it’s a sub-exponential random variable.
Moreover,
#
"
2
X
X X
Z i (Z i )⊤ ≤ mn .
.m E
Z i (Z i )⊤
r2 =
E
i
i
Therefore, by Bernstein inequality we have that w.h.p,
X
i∈Sa
Z i (Z i )⊤ − nIdm .
It follows that
X
i∈Sa
p
r 2 log n + max kZ i k2 log n =
p
mn log n + m log n .
p
mn log n Idm nIdm .
Z i (Z i )⊤ n − O
which in turn implies that Pl,Sa nQ. But this immediately implies σmin (Pl,Sa ) & np with high
probability. Union bounding over all l, we get the first part of the lemma.
The upper bound will be proven by a Chernoff bound. Note that the matrices
(1 − exp(−lWk ))Sa (1 − exp(−lWk ))⊤
Sa , k ∈ [m]
2
are independent. Furthermore, k(1−exp(−lWk ))Sa (1−exp(−lWk ))⊤
Sa k ≤ pn with high probability,
2
⊤
and the variable k(1 − exp(−lWk ))Sa (1 − exp(−lWk ))Sa k is sub-exponential. Finally,
#
"m
X
2
((1 − exp(−lWk ))Sa (1 − exp(−lWk ))⊤
r2 = E
Sa )
k=1
≤
m
X
k=1
i
h
2
E ((1 − exp(−lWk )Sa )(1 − exp(−lWk ))⊤
)
Sa
h
i
2 2
≤ m k1 − exp(−lWk )Sa k2 E ((1 − exp(−lWk ))Sa (1 − exp(−lWk ))⊤
Sa ) ≤ mn p
22
Similarly as in the lower bound,
h
i2
i
h
E[Pl,Sa ] = p2 E 1 − exp(−lW̃ ) 11⊤ + pE (1 − exp(−lW̃ ))2 − p2 E [1 − exp(−lw̃)]2 Id|Sa |
where 1 is the all-ones vector of size |Sa |, and Id|Sa | is the identity of the same size. Again, W̃ is a
random variable following the distribution D of all the W̃i,j . This immediately gives
p
Pl E[Pl ] + r log nId|Sa | mnp2 + mn2 p2 log nId|Sa | O(mnp2 )Id|Sa |
A union bound over all values of l gives the statement of the Lemma.
B.1
Incoherence of matrix F
First, we proceed to show the incoherence property 3.2 on the weight matrix.
Lemma B.2. Suppose n is a multiple of 3. Let F = U ΣV be the singular value decomposition of
F . Let Sa , Sb , Sc be a uniformly random equipartition of the rows of [n]. Suppose F is µ-incoherent
with µ ≤ n/(m log n). Then, with high probability over the choice of and Sa , Sb , Sc , we have for
every i ∈ {a, b, c},
r
1
µm
Si ⊤ Si
U − Id .
log n .
U
3
n
Proof. Let S = Sa . Then, since U ⊤ U = Idm , we have
#
"
X
1
(U i )(U i )⊤ = · Idm .
E
3
i∈S
By the assumption on the row norms of U , kU i (U i )⊤ k2 = kU i k2 ≤ µ m
n . By the incoherence
assumption, we have that maxi kU i k2 ≤ µm/n.
We also note that U i ’s are negatively associated random variables. Therefore by the matrix
Chernoff inequality for negatively associated random variables, we have with high probability,
r
i
h
µm log n
S ⊤ S
S ⊤ S
.
(U ) U − E (U ) U
.
n
But, an analogous argument holds for Sb , Sc as well – so by a union bound over k, we complete
the proof.
Lemma B.3. Suppose n & m log n. Under the generative assumption in Section 1 for W , we have
that we have that F = 1 − exp(−W ) is O(1)-incoherent.
Proof. We have that F F ⊤ = U Σ2 U T and therefore, kF i k2 = Σ2i,i kU i k2 . This in turn implies that
kU i k2 ≤
1
1
Fi = 2
kF i k2 .
min Σ2ii
σmin (F )
√
2
Since F i ≤ pm+ pm ≤ 2pm with high probability, we only need to bound σmin (F ) from below.
2 (F ) = σ
⊤
⊤
Note that σmin
min (F F ). Therefore it suffices to control σmin (F F ). But by Lemma B.1
2
we have σmin (F ) & np.
Therefore, we have that
m
1
kF i k2 = O( )
kU i k2 ≤ 2
n
σmin (F )
23
B.2
Spectral boundedness
The main goal of the section is to show that the bias terms in the PMI tensor are spectrally
bounded by the PMI matrix (which we can estimate from samples). Furthermore, we show that
we can calculate an approximate whitening matrix for the leading terms of the PMI tensor using
the machinery in Section .
The main proposition we will show is the following:
Proposition
B.4.
Let W be sampled according to the random model in Section 1 with ρ =
log n
1
. Let Sa ⊆ [n], |Sa | = Ω(n). If RSa is the matrix that has as columns
o log n , p = ω √
mn
l 1/3
ρ
the vectors 1l 1−ρ
(P̃l )j,Sa , l ∈ [2, L], j ∈ [m] for L = O(poly(n)), and A is the matrix that
1/3
ρ
(P̃1 )j,Sa for j ∈ [m], with high probability it holds that RSa RS⊤a
has as columns the vectors 1−ρ
is O(ρ2/3 log n)-spectrally bounded by A.
The main element for the proposition is the following Lemma:
Lemma B.5. For any set Sa ⊆ [n], |Sa | = Ω(n), with high probability over the choice of W , for
⊤ is O(log n)-spectrally bounded by P ′
every ℓ, ℓ′ = O(poly(n)) Pl,Sa Pl,S
l ,Sa .
a
Before proving the Lemma, let us see how the proposition follows from it:
Proof of B.4. We have
RSa RS⊤a
=
L
X
l=2
1
l
ρ
1−ρ
l !2/3
Pl,Sa
By Lemma B.5, we have that ∀l > 1, P̃l,Sa is τ -spectrally bounded by P̃1,Sa , for some τ =
O(log n). Hence,
RSa RS⊤a
∞
X
l=2
1
l
ρ
1−ρ
l !2/3
τ (P1,Sa + σmin (P1,Sa ))
- ρ4/3 τ (P1,Sa + σmin (P1,Sa ))
ρ 2/3
) P1,Sa , the claim of the Proposition follows.
Since AA⊤ = ( 1−ρ
It is clear analogous statements hold for Sb and Sc .
Finally, we proceed to show Lemma B.5.
For notational convenience, we will denote by Jm×n the all ones matrix with dimension m × n.
(We will omit the dimensions when clear from the context.)
This statement will immediately follow from the following two lemmas:
Lemma B.6. For any set Sa ⊆ [n], |Sa | = Ω(n), with probability 1 − exp(− log2 n) over the choice
of W , for all ℓ ≤ O(poly(n)),
5
Pl,Sa 10np log nId + mp2 J
2
24
Lemma B.7. For any set Sa ⊆ [n], |Sa | = Ω(n), with probability 1 − exp(− log2 n) over the choice
of W , ∀ℓ = poly(n),
Pl,Sa + 6np log nId - mp2 J
Before showing these lemmas, let us see how Lemma B.5 is implied by them:
Proof of Lemma B.5. Let κ be the constant in B.7, s.t. Pl,Sa + 6np log nId - mp2 J. Putting the
bounds from Lemmas B.6 and B.7 together along with a union bound, we have that with high
probability, ∀l, l′ = O(poly(n))
15
5
Pl,Sa − κPl′ ,Sa 10np log n + κnp log n Id O(mp log n)Id
2
2
But, note that σmin (Pl′ ) = Ω(np), by Lemma B.1. Hence, Pl − 52 κPl′ ,Sa r log nσmin (Pl′ ), for
some sufficiently large constant r. This implies
5
Pl − r log nPl′ ,Sa Pl − κPl′ r log nσmin (Pl′ ,Sa )
2
from which the statement of the lemma follows.
We proceed to the first lemma:
Proof of Lemma B.6. To make the notation less cluttered, we will drop l and Sa and use P = Pl,Sa .
Furthermore, we will drop Sa when referring to columns of P̃ so we will denote P̃k = P̃k,Sa .
Let’s denote by e = √ 1 1 . Let’s furthermore denote Id1 = ee⊤ , and Id−1 = Id − ee⊤ . Note
|Sa |
first that trivially, since Id1 + Id−1 = Id,
P = (Id1 + Id−1 )P (Id1 + Id−1 )
(B.2)
Furthermore, it also holds that
1
1
0 (2Id−1 − Id1 )P (2Id−1 − Id1 )
2
2
1
= Id1 P Id1 + 4Id−1 P Id−1 − Id−1 P Id1 − Id1 P Id−1
4
where the first inequality holds since
1
1
(2Id−1 − Id1 )P (2Id−1 − Id1 ) =
2
2
1
(2Id−1 − Id1 )P̃
2
1
(2Id−1 − Id1 )P̃
2
⊤
From this we get that
Id−1 P Id1 + Id1 P Id−1
1
Id1 P Id1 + 4Id−1 P Id−1
4
(B.3)
We proceed to upper bound both the terms on the RHS above. More precisely, we will show
that
Id1 P Id1 ≤ 2mp2 nJ
(B.4)
Id−1 P Id−1 2np log nId
25
(B.5)
Let us proceed to showing (B.4). The LHS can be rewritten as
Id1 P Id1 = ee⊤ e⊤ P̃ P̃ ⊤ e
Note that
m
X
1
e⊤ P̃ P̃ ⊤ e =
n
k=1
h1, P̃k i2
!
All the terms h1, P̃k i2 are independent and satisfy and E[h1, P̃k i2 ] ≤ (E[h1, P̃k i])2 ≤ p2 n2 . By
Chernoff, we have that
m
X
p
h1, P̃k i2 ≤ mp2 n2 + mp2 n2 log n ≤ 2mp2 n2
k=1
log n
).
with probability at least 1 − exp(− log2 n), where the second inequality holds because p = ω( √
mn
Hence, e⊤ P̃ P̃ ⊤ e ≤ 2mp2 n with high probability.
We proceed to (B.5), which will be shown by a Bernstein bound. Towards that, note that
E[(Id−1 P̃k (Id−1 P̃k )⊤ ] = Id−1 E[P̃k (P̃k )⊤ ]Id−1
= Id−1 p2 E[1 − exp(W̃ )]2 11⊤ + (p − p2 )E[(1 − exp(W̃ ))2 ]Id Id−1
pId
where W̃ is a random variable following the distribution D of all the W̃i,j . The second line can be
seen to follow from the independence of the coordinates of P̃k according to our model.
Furthermore, with high probability kId−1 P̃k k22 ≤ kP̃k k22 ≤ np and the random variable
kId−1 P̃k k2 is sub-exponential Finally,
#
"m
#
"m
2
X
X
2
(Id−1 P̃k )(Id−1 Γk )P̃
≤ P̃k
r2 = E
(I−1 P̃k )(Id−1 P̃k )⊤
E
2
k=1
k=1
≤ npmkE[(Id−1 P̃k )(Id−1 P̃k )⊤ ]k ≤ np2 m
Therefore, applying a matrix Bernstein bound, we get
Id−1 P Id−1 mpId + np log nId + r log nId 2np log nId
with high probability.
Combining this with (B.2) and (B.3), we get
5
P Id1 P Id1 + 5Id−1 P Id−1
4
5
mp2 J + 10np log nId
2
Let us proceed to the second inequality, which essentially follows the same strategy:
26
Proof of Lemma B.7. Similarly as in the proof of Lemma B.6, for notational convenience, let’s
denote by P̃ the matrix which has column k the vector (1 − exp(lWk )).
Reusing the notation from Lemma B.7, we have that
P = (Id1 + Id−1 )P (Id1 + Id−1 )
(B.6)
and
1
1
0 ( Id1 + 2Id−1 )P ( Id1 + 2Id−1 )
2
2
1
′
= Id−1 P Id1 + Id1 P Id−1 + Id1 P Id1 + 4Id−1 P Id−1
4
for similar reasons as before. From this we get that
1
Id−1 P Id1 + Id1 P Id−1 − Id1 P Id1 − 4Id−1 P Id−1
4
(B.7)
Putting (B.6) and (B.7) together, we get that
3
P + 3Id−1 P Id−1 Id1 P Id1
4
(B.8)
We will proceed to show an upper bound Id−1 P Id−1 2np log nId on second term of the LHS.
We will do this by a Bernstein bound as before. Namely, analogously as in Lemma B.7,
Id−1 P Id−1 =
m
X
k=1
(Id−1 P̃k )(Id−1 P̃k )⊤
and E[(Id−1 P̃k )(Id−1 P̃k )⊤ ] pId and and kId−1 P̃k k22 ≤ kP̃k k22 ≤ np are satisfied so
"m
#
2
X
(Id−1 P̃k )(Id−1 P̃k )⊤
≤ np2 m
r2 = E
k=1
Therefore, applying a matrix Bernstein bound, we get
Id−1 P Id−1 mpId + np log nId + r 2np log nId
with high probability.
Plugging this in in (B.8), we get
3
3
3
Id1 P Id1 = ee⊤ P̃ P̃ ⊤ ee⊤ = ee⊤ e⊤ P̃ P̃ ⊤ e
4
4
4
P
m
2
with the goal of applying Chernoff, we will lower
Since we have e⊤ P̃ P̃ ⊤ e = n1
k=1 hP̃k , 1i
P + 6np log nId
bound E[h1, Ak i2 ]. More precisely, we will show E[h1, P̃k i]2 = Ω(n2 p2 ). In order to do this, we have
X
X
E[(P̃k )2j ] +
E[h1, P̃k i2 ] =
E[(P̃k )j ]E[(P̃k )j ′ ]
j6=j ′ ;j,j ′∈Sa
j∈Sa
=
X
j∈Sa
≥
X
j∈Sa
˜ )2 ] +
pE[(1 − exp(−lW
pE[(1 − exp(−W̃ )2 ] +
2 2
= Ω(n p )
27
X
j6=j ′ ;j,j ′∈Sa
X
j6=j ′ ;j,j ′∈Sa
˜ )]2
p2 E[1 − exp(−lW
p2 E[1 − exp(−W̃ )]2
where W̃ is a random variable following the distribution D of all the W̃i,j . and the last inequality
holds because of (1.2).
p
So by Chernoff, we get that e⊤ P̃ P̃ ⊤ e = n1 (Ω(mn2 p2 ) − n2 p2 m) = Ω(mnp2 ) with high probability. Altogether, this means
P + 6np log nId % mp2 J
C
Robust whitening
Algorithm 4 Obtaining whitening matrices
d
Inputs: Random partitioning Sa , Sb , Sc of [n]. Empirical PMI matrix PMI.
d×d
Outputs: Whitening matrices Qa , Qb , Qc ∈ R
1. Output
+
d Sc ,Sa ,
d Sa ,S (PMI
d S ,S )⊤ PMI
Qa = ρ−1/3 PMI
b
b c
d S ,Sc (PMI
d + )⊤ PMI
d Sa ,S ,
Qb = ρ−1/3 PMI
Sc ,Sa
b
b
⊤ d
d Sc ,Sa (PMI
d+
Qc = ρ−1/3 PMI
Sa ,Sb ) PMISb ,Sc
⊤ d
d Sa ,S (PMI
d+
In this section, we show the formula Qa = ρ−1/3 PMI
Sb ,Sc ) PMISc ,Sa computes an
b
⊤
approximation of the true whitening matrix AA , so that the error is ε-spectrally bounded by A.
We recall Theorem 2.10.
Theorem 2.10. Let n ≥ m and A, B, C ∈ Rn×m . Suppose Σab , Σbc , Σca ∈ Rn×n are of the form,
Σab = AB ⊤ + Eab , Σbc = BC ⊤ + Ebc , and Σca = CA⊤ + Eca .
where Eab , Ebc , Eca are ε-spectrally bounded by (A, B), (B, C), (C, A) respectively. Then, the matrix
matrix
+
Qa = Σab [Σ⊤
bc ]m Σca
⊤
+
is a good approximation of AA⊤ in the sense that Qa = Σab [Σ⊤
bc ]m Σca − AA is O(ε)-spectrally
bounded by A. Here [Σ]m denotes the best rank-m approximation of Σ.
Towards proving Theorem 2.10, an intermediate step is to understand the how the space of
singular vectors of BC ⊤ are aligned with the noisy version Σbc . The following explicitly represent
BC ⊤ + E as the form B ′ R(C ′ )⊤ + ∆′ . Here the crucial benefit to do so is that the resulting ∆′ is
small in every direction. In other words, we started with a relative error guarantees on E and the
Lemma below converts to it an absolute error guarantees on ∆′ (though the signal term changes
slightly).
Lemma C.1. Suppose B, C are n × m matrices with n ≥ m. Suppose a matrix E is ε-spectrally
bounded by (B, C), then BC ⊤ + E can be written as
BC ⊤ + E = (B + ∆B )RBC (C + ∆C )⊤ + ∆′BC ,
28
where ∆B , ∆C , ∆′BC are small and RBC is close to identity in the sense that,
k∆B k ≤ O(εσmin (B))
k∆C k ≤ O(εσmin (C))
k∆′BC k ≤ O(εσmin (B)σmin (C))
kRBC − Idk ≤ O(ε)
Proof. The key intuition is if the perturbation is happening in the span of columns of B and C,
they cannot change the subspace. By Definition 2.9, we can write E as
⊤
E = B∆1 C ⊤ + B∆⊤
2 + ∆3 C + ∆4 .
Now since k∆1 k ≤ ε < 1, we know (Id − ∆1 ) is invertible, so we can write
(BC ⊤ + E) = (B + ∆3 (Id + ∆1 )−1 )(Id + ∆1 )(C + ∆2 (Id − ∆1 )−⊤ )⊤ − ∆3 (Id − ∆1 )−1 ∆⊤
2 + ∆4 .
This is already in the desired form as we can let ∆B = ∆3 (Id + ∆1 )−1 , RBC = (Id + ∆1 ),
∆C = ∆2 (Id − ∆1 )−⊤ , and ∆′BC = −∆3 (Id − ∆1 )−1 ∆⊤
2 + ∆4 . By Weyl’s Theorem we know
−1
ε
σmin (B). Other terms can be
σmin (Id + ∆1 ) ≥ 1 − ε, therefore k∆B k ≤ k∆3 kσmin
(Id + ∆) ≤ 1−ε
bounded similarly.
Now we prove that the top m approximation of BC ⊤ + E has similar column/row spaces as
BC ⊤ . Let UB be the column span of B, UB′ be the column span of (B + ∆B ), and UB′′ be the top m
left singular subspace of (BC ⊤ + E). Similarly we can define UC , UC′ , UC′′ to be the column spans
of C, C + ∆C and the top m right singular subspace of (BC ⊤ + E).
For B + ∆B , we can apply Weyl’s Theorem and Wedin’s Theorem. By Weyl’s Theorem we
know σmin (B + ∆B ) ≥ σmin (B) − k∆B k ≥ (1 − O(ε))σmin (B). By Wedin’s Theorem we know UB′
is O(ε)-close to UB . Similar results apply to C + ∆C .
Now we know σmin ((B + ∆B )RBC (C + ∆C )⊤ )) ≥ σmin (B + ∆B )σmin (RBC )σmin (C + ∆C ) ≥
Ω(σmin (B)σmin (C). Therefore we can again apply Wedin’s Theorem, considering (B+∆B )RBC (C+
∆C )⊤ ) as the original matrix and ∆′BC as the perturbation. As a result, we know UB′′ is O(ε) close
to UB′ , UC′′ is O(ε) close to UB′ . The distance between UB , UB′′ (and UC , UC′′ ) then follows from
triangle inequality.
As a direct corollary of Lemma C.1, we obtain that the BC ⊤ and BC ⊤ + E have similar
subspaces of singular vectors.
Corollary C.2. In the setting of Lemma C.1, let [BC ⊤ + E]m be the best rank-m approximation
of BC ⊤ + E. Then, the span of columns of [BC ⊤ + E]m is O(ε)-close to the span of columns of
B, span of rows of [BC ⊤ + E]m is O(ε)-close to the span of columns of C.
Furthermore, we can write [BC ⊤ + E]m = (B + ∆B )RBC (C + ∆C )⊤ + ∆BC . Here ∆B , ∆C and
RBC as defined in Lemma C.1 and ∆BC satisfies k∆BC k ≤ O(εσmin (B)σmin (C)).
Proof. Since [BC ⊤ + E]m is the best rank-m approximation, because (B + ∆B )RBC (C + ∆C )⊤ ) is
a rank m matrix, in particular we have
kBC ⊤ + E − [BC ⊤ + E]m k ≤ kBC ⊤ − (B + ∆B )RBC (C + ∆C )⊤ )kk∆′BC k.
29
Therefore
k∆BC k = k[BC ⊤ + E]m − (B + ∆B )RBC (C + ∆C )⊤ )k
≤ kBC ⊤ + E − [BC ⊤ + E]m k + kBC ⊤ + E − (B + ∆B )RBC (C + ∆C )⊤ )k
≤ 2k∆′BC k.
+
In order to fix this problem, we notice that the matrix [Σ⊤
bc ]m is multiplied by Σab on the left
⊤
⊤
+
and Σca on the right. Assuming Σab = AB , Σca = CA , we should expect [Σ⊤
bc ]m to “cancel” with
the B ⊤ factor on the left and the C factor on the right, giving us AA⊤ . Therefore, we should really
⊤
+
measure the error of the middle term [Σ⊤
bc ]m after left multiplying with B and right multiplying
with C. We formalize this in the following lemma:
+
⊤ +
Lemma C.3. Suppose Σbc is as defined in Theorem 2.10, let ∆ = [Σ⊤
bc ]m − [CB ] , then we have
kB ⊤ ∆Ck = O(ε),
k∆Ck ≤ O(
ε
),
σmin (B)
ε
),
σmin (C)
ε
k∆k ≤ O(
).
σmin (B)σmin (C)
kB ⊤ ∆k ≤ O(
We will first prove Theorem 2.10 assuming Lemma C.3.
Proof of Theorem 2.10. By Lemma C.1, we know Σab can be written as
(A + ∆1A )RAB (B + ∆1B )⊤ + ∆AB .
Similarly Σca can be written as
(C + ∆3C )RCA (A + ∆3A )⊤ + ∆CA .
Here the ∆ terms and R terms are bounded as in Lemma C.1.
+
Now let us write the matrix Σab [Σ⊤
bc ]m Σca as
(A + ∆1A )RAB (B + ∆1B )⊤ + ∆AB ([CB ⊤ ]+ + ∆BC ) (C + ∆3C )RCA (A + ∆3A )⊤ + ∆CA
+
We can now view Σab [Σ⊤
bc ]m Σca as the product of three terms, each term is the sum of two
matrices. Therefore we can expand the product into 8 terms. In each of the three pairs, we will
call the first matrix the main matrix, and the second matrix the perturbation.
In the remaining proof, we will do calculations to show the product of the main terms is close
to AA⊤ , and all the other 7 terms are small.
Before doing that, we first prove several Claims about PSD matrices
Claim C.4. If k∆k ≤ ε, then A∆A⊤ εAA⊤ . If kΓk ≤ εσmin (A), then
2 (A)Id .
εAA⊤ + εσmin
1
⊤
2 (AΓ
+ ΓA⊤ )
Proof. Both inequalities can be proved by consider the quadratic form. We know for any x,
x⊤ A∆A⊤ x ≤ k∆kkA⊤ xk2 ≤ εx⊤ AA⊤ x, so the first part is true.
For the second part, for any x we can apply Cauchy-Schwartz inequality
√
1
2
x⊤ (AΓ⊤ + ΓA⊤ )x = h εA⊤ x, ε−1/2 Γ⊤ xi ≤ εkA⊤ xk2 + ε−1 kΓ⊤ xk2 = x⊤ (εAA⊤ + εσmin
(A)Id)x.
2
30
Now, we will first prove the product of three main matrices is close to AA⊤ :
Claim C.5. We have (A + ∆1A )RAB (B + ∆1B )⊤ (CB ⊤ )+ (C + ∆3C )RCA (A + ∆3A )⊤ = AA⊤ +
EA , where EA is O(ε)-spectrally bounded by AA⊤ .
Proof. We will first prove the middle part of the matrix (B +∆1B )⊤ (CB ⊤ )+ (C +∆3C ) is O(ε) close to
identity matrix Id. Here we observe that both B, C have full column rank so (CB ⊤ )+ = (B ⊤ )+ C + .
Therefore we can rewrite the product as (Id + B + ∆1B )⊤ (Id + C + ∆3C ). Since k∆1B k ≤ O(εσmin (B))
by Lemma C.1 (and similarly for C), we know kB + ∆1B k ≤ O(ε). Therefore the middle part is O(ε)
bAB = RAB (B + ∆1 )⊤ (CB ⊤ )+ (C + ∆3 )RCA is O(ε)-close
close to Id. Now since ε ≪ 1 we know R
B
C
to Id.
bAB (A + ∆3 )⊤ , for this matrix we know
Now we are left with (A + ∆1A )R
A
bAB (A + ∆3A )⊤ − AA⊤ = A(R
bAB − Id)A⊤ + ∆1A R
bAB A⊤ + AR
bAB (∆3A )⊤ + ∆1A R
bAB (∆3A )⊤ .
(A + ∆1A )R
3 ⊤
bAB − Id)A⊤ O(ε)AA⊤ (Claim C.4); the fourth term ∆1 R
b
The first term A(R
A AB (∆A )
2 (A))Id (by the norm bounds of ∆1 and ∆3 . For the cross terms, we can bound them
O(εσmin
A
A
using the second part of Claim C.4.
Next we will try to prove the remaining 7 terms are small. We partition them into three types
depending on how many ∆ factors they have. We proceed to bound them in each of these cases.
For the terms with only one ∆, we claim:
Claim C.6. The three terms ∆AB (CB ⊤ )+ (C + ∆3C )RCA (A + ∆3A )⊤ , (A + ∆1A )RAB (B +
∆1B )⊤ ∆AB (C + ∆3C )RCA (A + ∆3A )⊤ , (A + ∆1A )RAB (B + ∆1B )⊤ (CB ⊤ )+ ∆CA are all O(ε) spectrally bounded by AA⊤ .
Proof. For the first term, note that both B, C have full column rank, and hence (CB ⊤ )+ =
(B ⊤ )+ C + . Therefore the first term can be rewritten as
[∆AB (B ⊤ )+ ][(Id + C + ∆3C )RCA ](A + ∆3A )⊤ .
By Lemma C.1, we have spectral norm bounds for ∆AB , ∆3C , ∆3A , RCA . Therefore we know
k∆AB (B ⊤ )+ k ≤ O(εσmin (A)) and [(Id + C + ∆3C )RCA ] is O(ε) close to Id.
Therefore
2 (A)) is trivially O(ε) spectrally bounded, and
k[∆AB (B ⊤ )+ ][(Id + C + ∆3C )RCA ](∆3A )⊤ k ≤ O(εσmin
[∆AB (B ⊤ )+ ][(Id + C + ∆3C )RCA ]A⊤ is O(ε) spectrally bounded by Claim C.4. The third term is
exactly symmetric.
b BC = (B+∆1 )⊤ ∆BC (C +
For the second part, we will first prove the middle part of the matrix ∆
B
∆3C ) has spectral norm O(ε). This can be done y expanding it to the sum of 4 terms, and use
appropriate spectral norm bounds on ∆BC and its products with B ⊤ and C from Lemma C.3.
b BC RCA (A + ∆3 )⊤ is O(ε) spectrally bounded by the first part
Now we can show (A + ∆1A )RAB ∆
A
of Claim C.4.
Next we try to bound the terms with two ∆ factors.
Claim C.7. The three terms
∆AB ∆BC (C + ∆3C )RCA (A + ∆3A )⊤ , ∆AB (CB ⊤ )+ ∆CA ,
(A + ∆1A )RAB (B + ∆1B )⊤ ∆BC ∆CA are all O(ε2 ) spectrally bounded by AA⊤ .
Proof. For the first term, notice that k∆BC (C + ∆3C )k is bounded by O(ε/σmin (B)) by
Lemma C.3, and k∆AB k = O(εσmin (A)σmin (B)). Therefore we know k∆AB ∆BC (C + ∆3C )RCA k ≤
O(ε2 σmin (A)), so by Claim C.4 we know this term is O(ε2 ) spectrally bounded by AA⊤ . Third
term is symmetric.
31
2 (A)),
For the second term, by Lemma C.1 we can directly bound its spectral norm by O(ε2 σmin
so it is trivially O(ε2 ) spectrally bounded by AA⊤ .
Finally, for the product ∆AB ∆BC ∆CA , we can get the spectral norm for the three factors by
2 (A)) which is trivially O(ε3 )
Lemma C.1 and Lemma C.3. As a result k∆AB ∆BC ∆CA k ≤ O(ε3 σmin
⊤
spectrally bounded by AA .
Combining the bound for all of the terms we get the theorem.
With that, we now try to prove Lemma C.3
We first prove a simpler version where the perturbation is simply bounded in spectral norm
Lemma C.8. Suppose B, C are n × m matrices and n ≥ m. Let R be an n × n matrix such that
kR − Idk ≤ ε, and E is a perturbation matrix with kEk ≤ εσmin (B)σmin (C) and (CRB ⊤ + E) is
also of rank m.
Now let ∆ = (CRB ⊤ + E)+ − (CRB ⊤ )+ , then when ε ≪ 1 we have
kB ⊤ ∆Ck = O(ε),
k∆Ck ≤ O(
ε
),
σmin (B)
ε
),
σmin (C)
ε
k∆k ≤ O(
).
σmin (B)σmin (C)
kB ⊤ ∆k ≤ O(
Proof. We first give the proof for kB ⊤ ∆Ck. Other terms are similar.
Let UB be the column span of B, and UB′ be the row span of (CRB ⊤ + E). Similarly let UC
be the column span of C and UC′ be the column span of (CRB ⊤ + E). By Wedin’s theorem, we
know UB′ is O(ε) close to UB and UC′ is O(ε) close to UC . As a result, suppose the SVD of B is
UB DB VB⊤ , we know
σmin (B ⊤ UB′ ) = σmin (VB DB UB⊤ UB′ ) ≥ (1 − O(ε))σmin (B).
The same is true for C: σmin (C ⊤ UC′ ) ≥ (1 − O(ε))σmin (C).
By the property of pseudoinverse, the column span of (CRB ⊤ + E)+ is UB′ , and the row span
of (CRB ⊤ + E)+ is UC′ , further, (CRB ⊤ + E)+ = UB′ [(UC′ )⊤ (CRB ⊤ + E)UB′ ]−1 UC′ , therefore we
can write
B ⊤ (CRB ⊤ + E)+ C = B ⊤ UB′ [(UC′ )⊤ (CRB ⊤ + E)UB′ ]−1 (UC′ )⊤ C.
Note that now the three matrices are all n×n and invertible! We can write B ⊤ UB′ = ((B ⊤ UB′ )−1 )−1
(and do the same thing for (UC′ )⊤ C. Using the fact that P −1 Q−1 = (QP )−1 , we have
B ⊤ (CRB ⊤ + E)+ C = (((UC′ )⊤ C)−1 (UC′ )⊤ (CRB ⊤ + E)UB′ (B ⊤ UB′ )−1 )−1
= (R + ((UC′ )⊤ C)−1 (UC′ )⊤ EUB′ (B ⊤ UB′ )−1 )−1 =: (R + X)−1 .
Here we defined X = ((UC′ )⊤ C)−1 (UC′ )⊤ EUB′ (B ⊤ UB′ )−1 . The spectral norm of X can be
bounded by
kXk ≤ k((UC′ )⊤ C)−1 kkEkk(B ⊤ UB′ )−1 k
−1
−1
′
= kEkσmin
(B ⊤ UB′ )σmin
(C ⊤ CB
)
≤ O(ε).
We can write B ⊤ ∆C = B ⊤ (CRB ⊤ + E)+ C − Id = (Id + (R − Id + X))−1 − Id, and we now know
k(R − Id + X)k ≤ O(ε), as a result kB ⊤ ∆Ck ≤ O(ε) as desired.
32
For the term kB ⊤ ∆k, by the same argument we have
B ⊤ (CRB ⊤ + E)+ = ((UC′ )⊤ (CRB ⊤ + E)UB′ (B ⊤ UB′ )−1 )−1 (UC′ )⊤
= ((UC′ )⊤ CR + (UC′ )⊤ EUB′ (B ⊤ UB′ )−1 )−1 (UC′ )⊤
= ((UC′ )⊤ C(R + X))−1 (UC′ )⊤
= (R + X)−1 ((UC′ )⊤ C)−1 (UC′ )⊤ .
On the other hand, we know B ⊤ (CRB ⊤ )+ = R−1 C + = R−1 ((UC )⊤ C)−1 UC⊤ . We can match
the three factors:
kR−1 − (R + X)−1 k ≤ O(ε),
k((UC )⊤ C)−1 − ((UC′ )⊤ C)−1 k ≤ O(ε/σmin (C)),
kUC − UC′ k ≤ O(ε),
kR−1 k ≤ 1 + O(ε)
k((UC )⊤ C)−1 k = O(1/σmin (C))
kUC k = 1.
Here, first and third bound are proven before. The second bound comes if we consider the SVD of
C = UC DC VC⊤ and notice that k(UC′ )⊤ UC − Idk ≤ O(ε). We can write ∆1 = R−1 − (R + X)−1 ,
∆2 = ((UC )⊤ C)−1 − ((UC′ )⊤ C)−1 , ∆3 = UC − UC′ , then we have
B ⊤ ∆ = B ⊤ (CRB ⊤ + E)+ − B ⊤ (CRB ⊤ )+
= (R−1 − ∆1 )(((UC )⊤ C)−1 − ∆2 )(UC − ∆3 )⊤ − R−1 ((UC )⊤ C)−1 UC⊤ .
Expanding the last equation, we get 7 terms and all of them can be bounded by O(ε/σmin (C)).
The bounds on k∆Ck and k∆k can be proved using similar techniques.
Finally we are ready to prove the main Lemma C.3:
⊤ , we can write the matrix before pseudoinverse
Proof of Lemma C.3. Using Lemma C.1, let E = Ebc
as
[CB ⊤ + E]m = (C + ∆C )RBC (B + ∆B )⊤ + ∆BC .
We can then apply Lemma C.8 on (C + ∆C )RBC (B + ∆B )⊤ + ∆BC . As a result, we know if we
⊤ +
let ∆′ = [CB ⊤ + E]+
m − ((C + ∆C )RBC (B + ∆B ) ) , we have the desired bound if we left multiply
⊤
with (B + ∆B ) or right multiply with (C + ∆C ).
We will now show how to prove the first bound, all the other bounds can be proved using the
same strategy:
First, we can write
′
⊤ ′
B ⊤ ∆′ C = −(B + ∆B )⊤ ∆′ (C + ∆C ) + (B + ∆B )⊤ ∆′ C + ∆⊤
B ∆ (C + ∆C ) − ∆B ∆ ∆C .
All the four terms on the RHS can be bounded by Lemma C.8 so we know kB ⊤ ∆′ Ck ≤ O(ε).
On the other hand, let ∆′′ = ((C + ∆C )RBC (B + ∆B )⊤ )+ − (C ⊤ B)+ = ∆ − ∆′ . We will prove
kB ⊤ ∆′′ Ck ≤ O(ε) and then the bound on kB ⊤ ∆Ck follows from triangle inequality.
For B ⊤ ∆′′ C, we know it is equal to
−1
(C + ∆C )+ C − Id
B ⊤ [(B + ∆B )⊤ ]+ RAB
−1
(C + ∆C )+ C − Idk ≤ O(ε)
Claim C.9. kB ⊤ [(B + ∆B )⊤ ]+ RAB
33
−1
this follows
Proof. We will show all three factors in the first term are O(ε) close to Id. For RAB
+
immediately from Lemma C.1. For (C + ∆C ) C, we know
(C + ∆C )+ C − Id = −(C + ∆C )+ ∆C .
−1
Therefore its spectral norm bound is bounded by k∆C kσmin
(C + ∆C ) = O(ε) (where the bound
on k∆C k comes from Lemma C.1).
With the claim we have now proven kB ⊤ ∆′′ Ck ≤ O(ε), therefore
kB ⊤ ∆Ck ≤ kB ⊤ (∆′ + ∆′′ )Ck ≤ kB ⊤ ∆′ Ck + kB ⊤ ∆′′ Ck ≤ O(ε).
C.1
Spectrally Boundedness and Incoherence
Here we will show under mild incoherence conditions (defined below), if an error matrix E is
ε-spectrally bounded by F F ⊤ , then the partial matrices satisfy the requirement of Theorem 2.10.
q
Theorem C.10. If F is µ-incoherent for µ ≤ n/m log2 n, then when n ≥ Ω(m log2 m), with high
probability over the random partition of F into A, B, C, we know σmin (A) ≥ σmin (F )/3 (same is
true for B, C).
As a corollary, if E is ε-spectrally bounded by F . Let a, b, c be the subsets corresponding to
A, B, C, and let Ea,b be the submatrix of E whose rows are in set a and columns are in set b. Then
Ea,b (also Eb,c , Ec,a ) is O(ε)-spectrally bounded by the corresponding asymmetric matrices AB ⊤
(BC ⊤ , CA⊤ ).
Proof. Consider the singular value decomposition of F : F = U DV ⊤ . Here U is a n × m matrix
whose columns are orthonormal, V is an m × m orthonormal matrix and D is a diagonal matrix
whose smallest diagonal entry is σmin (F ).
Consider the following way of partitioning the matrix: for each row of F , we put it into A, B
or C with probability 1/3 independently.
Now, let Xi = 1 if row i is in the matrix A, and 0 otherwise. Then Xi ’s are Bernoulli random
variables with probability 1/2. Suppose S is the set of rows in A, let UA be U restricted to rows in
A, then we have A = UA DV ⊤ . We will show with high probability
σmin (A) ≥ 1/3.
P
The key observation here is the expectation of UA⊤ UA = ni=1 Xi Ui Ui⊤ , where Ui is the i-th row
of U (represented as a column vector). Since Xi ’s are Bernoulli random variables, we know
n
n
X
1
1X
Ui Ui⊤ = Id.
Xi Ui Ui⊤ ] =
E[UA⊤ UA ] = E[
3
3
i=1
i=1
Therefore we can hope to use matrix concentration to prove that UA⊤ UA is close to its expectation.
Let Mi = Xi Ui Ui⊤ − 1/3Ui Ui⊤ . Clearly E[Mi ] = 0. By the Incoherence assumption, we know
kUi k ≤ 1/ log n. Therefore we know kMi k ≤ O(1/ log n). Also, we can bound the variance
n
n
n
X
X
X
⊤
⊤
2
⊤
Ui Ui⊤ k ≤ O(1/ log 2 n).
Xi Ui Ui Ui Ui ]k ≤ max kUi k k
Mi Mi ]k ≤ kE[
kE[
i=1
i=1
i=1
Here the last inequality is because
Pn
⊤
⊤
i=1 Xi Ui Ui Ui Ui
34
kUi k2 Ui Ui⊤ .
Therefore by Matrix Bernstein’s inequality we know with high probability k
When this happens we know
kUA⊤ UA k
≥
σmin (E[UA⊤ UA ])
−k
n
X
i=1
Pn
i=1 Mi k
≤ 1/6.
Mi k ≥ 1/6.
p
Hence we have σmin (UA ) ≥ 1/6 > 1/3, and σmin (A) ≥ σmin (UA )σmin (D) ≥ σmin (F )/3. Note
that matrices B, C have exactly the same distribution as A so the bounds for B,C follows from
union bound.
For the corollary, if a matrix E is ε spectrally bounded, we can write it as F ∆1 F ⊤ + F ∆⊤
2 +
2 (F ). This can be done by
∆2 F ⊤ + ∆4 , where k∆1 k ≤ ε, k∆2 k ≤ εσmin (F ) and k∆4 k ≤ εσmin
considering different projections of E: let U be the span of columns of F , then F ∆1 F ⊤ term
⊤
corresponds to ProjU E ProjU ; F ∆⊤
2 term corresponds to ProjU E ProjU ⊥ ; ∆2 F term corresponds
to ProjU ⊥ E ProjU ; ∆4 term corresponds to ProjU ⊥ E ProjU ⊥ . The spectral bounds are necessary
for E to be spectrally bounded.
⊤
Now for Ea,b , we can write it as A∆1 B ⊤ + A(∆2 )⊤
b + (∆2 )a B + (∆4 )a,b , where we also take
the corresponding submatrices of ∆’s. Since the spectral norm of a submatrix can only be smaller,
we know k∆1 k ≤ ε, k(∆2 )b k ≤ εσmin (F ) ≤ 3εσmin (B), k(∆2 )a k ≤ εσmin (F ) ≤ 3εσmin (A) and
2 (F ) ≤ 9εσ
k(∆2 )a,b k ≤ εσmin
min (A)σmin (B). Therefore by Definition 2.9 we know Ea,b is 9ε
⊤
spectrally bounded by AB .
D
Proof of Theorem 3.1 and Theorem 3.3
In this section, we provide the full proof of Theorem 3.1. We start with a simple technical Lemma.
Lemma D.1. If Q is an ε-approximate whitening matrix for A, then kQk ≤
1
⊤
1−ε kAA k
1
⊤
1−ε kAA k,
σmin (Q) ≥
Proof. By the definition of approximate-whitening, we have
1 − ε ≤ σmin ((Q+ )1/2 A⊤ A(Q+ )1/2 ), σmax ((Q+ )1/2 A⊤ A(Q+ )1/2 ) ≤ 1 + ε
which implies that
1 − ε ≤ σmin ((Q+ )1/2 AA⊤ (Q+ )1/2 ), σmax ((Q+ )1/2 AA⊤ (Q+ )1/2 ) ≤ 1 + ε
by virtue of the fact that (Q+ )1/2 AA⊤ (Q+ )1/2 =
semidefinite-order notation, we get that
(Q+ )1/2 A⊤ A(Q+ )1/2
⊤
.
Rewriting in
(1 − ε)Id (Q+ )1/2 AA⊤ (Q+ )1/2 (1 + ε)Id
Multiplying on the left and right by Q1/2 , we get
(1 − ε)Q AA⊤ (1 + ε)Q
This directly implies
1
⊤
1+ε AA
Q
1
⊤
1−ε AA
which is equivalent to the statement of the lemma.
Towards proving Theorem 3.1, we will first prove the following proposition, which shows that
we recover the exp(−W ) matrix correctly:
35
Proposition D.2 (Recovery of exp(−W )). Under the random generative model defined in Section 1, if the number of samples satisfies
N = poly(n, 1/p/, 1/ρ)
√
the vectors W̃i , i ∈ [m] in Algorithm 1 are O(η np)-close to exp(−Wi ) where
√
η = Õ ( mpρ)
Proof. The proof will consist of checking the conditions for Algorithms 4 and 3 to work, so that we
can apply Theorems 5.4 and 2.10.
To get a handle on the PMI tensor, by Proposition A.2, for any equipartition Sa , Sb , Sc of [n],
we can it as
PMITSa ,Sb ,Sc =
L
X
ρ X
(−1)l+1
Fk,Sa ⊗ Fk,Sb ⊗ Fk,Sc +
1−ρ
k∈[m]
l=2
1
l
ρ
1−ρ
l ! X
(P̃l )k,Sa ⊗ (P̃l )k,Sb ⊗ (P̃l )k,Sc + EL
k∈[m]
1 1
We can choose L = poly(log(n, , )) to ensure
ρ p
√
kEL k = o p5/2 ρ7/3 mn2
(D.1)
Having an explicit form for the tensor, we proceed to check the spectral boundedness condition
for Algorithm 4.
Let Sa , Sb , Sc be a random equipartition. Let RSa be the matrix that has as columns the vectors
l !1/3
ρ
1
(P̃l )j,Sa , for all l ∈ [2, L], j ∈ [m] and let A be the matrix that has as columns the
l 1−ρ
1/3
ρ
Fj,Sa , for all j ∈ [m]. Since L is polynomially bounded in n, by Proposition B.4
vectors 1−ρ
we have that with high probability RSa is τ -spectrally bounded by A, for a τ = O(ρ2/3 log n).
Analogous statements hold for Sb , Sc .
Next, we verify the conditions for calculating approximate whitening matrices (Algorithm 4).
Towards applying Theorem C.10, note that if R[n] is the matrix that has as columns the vectors
l !1/3
ρ
1
(P̃l )j , for all l ∈ [2, L], j ∈ [m], and D is the matrix that has as columns the
l 1−ρ
1/3
ρ
Fj , for all j ∈ [m], R[n] then is τ -spectrally bounded by D for τ = O(ρ2/3 log n).
vectors 1−ρ
Furthermore, the matrix F is O(1)-incoherent with high probability by Lemma B.1. Hence, we can
apply Theorem C.10, the output of Algorithm 4 are matrices Qa , Qb , Qc which are τ -approximate
whitening matrices for A, B, C respectively.
Next, we will need bounds on
min(σmin (Qa ), σmin (Qb ), σmin (Qc )), max(σmax (Qa ), σmax (Qb ), σmax (Qc ))
to plug in the guarantee of Algorithm 4.
36
By Lemma D.1, we have
σmax (Qa ) ≤
1
1
kAA⊤ k . (1 + τ )kAA⊤ k, σmin (Qa ) ≥
σmin (AA⊤ ) & (1 − τ )σmin (AA⊤ )
1−τ
1+τ
However, for the random model, applying Lemma B.1,
⊤
σmin (AA ) ≥
ρ
1−ρ
2/3
np & ρ
2/3
⊤
np, σmax (AA ) ≤
ρ
1−ρ
2/3
mnp2 . ρ2/3 mnp2
(D.2)
Analogous statements hold for B and C.
Finally, we bound the error due to empirical estimates. Since ρpm = o(1),
Pr[si = 0 ∧ sj = 0 ∧ sk = 0] ≥ 1 − Pr[si = 1] − Pr[sj = 1] − Pr[sk = 1] ≥ 1 − 3pmρ = Ω(1)
Hence, by Corollary E.3, with a number of samples as stated in the theorem,
√
ˆ Sa ,S ,Sc − PMITSa ,S ,Sc k{1,2},{3} . p5/2 ρ5/2 mn2
kPMIT
b
b
(D.3)
as well.
With that, invoking Theorem 5.4 (with kEk{1,2},{3} taking into account both the EL term above,
and the above error due to sampling), the output of Algorithm 3 will produce vectors vi , i ∈ [m],
1/3
ρ
(1 − exp(−Wi )), for
s.t. vi is O(η ′ )-close to 1−ρ
ˆ Sa ,S ,Sc − PMITSa ,S ,Sc k{1,2},{3} )
η ′ . max(kQa k , kQb k , kQc k)1/2 · τ 3/2 + σ −3/2 (kEL k{1,2},{3} + kPMIT
b
b
where σ = min(σmin (Qa ), σmin (Qb ), σmin (Qc )).
Plugging in the estimates from (D.1), (D.2), (D.3) as well as τ = O(ρ2/3 log n), we get:
q
√
1/2 3/2
max(kQa k , kQb k , kQc k) τ
. mnp2 ρ2/3 (ρ2/3 log n)3/2 = ρ4/3 mnp log3/2 n
√
1 3/2
) kEL k{1,2},{3} . ρ4/3 mnp log3/2 n
σ −3/2 kEL k{1,2},{3} . (
ρnp
√
−3/2
ˆ Sa ,S ,Sc − PMITSa ,S ,Sc k{1,2},{3} . ρ4/3 mnp log3/2 n
σ
kPMIT
b
b
1/3
ρ
(1 − exp(Wi )), for all i ∈ [m].
which implies the vectors (âi , b̂i , ĉi ) are O(η ′ )-close to 1−ρ
1/3
However, this directly implies that 1−ρ
(âi , b̂i , ĉi ) are O(η ′ /ρ1/3 )-close to 1 − exp(Wi ),
ρ
i ∈ [m], which in turn implies (ãi , b̃i , c̃i ) are O(η ′ /ρ1/3 ) close to exp(Wi ).
This implies the statement of the Lemma.
Given that, we prove the main theorem. The main issue will be to ensure that taking log of the
values
√
Proof of Theorem 3.1. By Proposition D.2, the vectors Yi , i ∈ [m] in Algorithm 1 are O(η np)
√
close to exp(−Wi ) with η = Õ mpρ . Let (Yi′ )j = (Yi )j if (Yi )j > exp(−νu ) and otherwise
(Yi )′j = exp(−νu ).
Then we have that kYi′ − Wi k ≤ kYi − Wi k. By the Lipschitzness of log(·) in the region [νi , ∞]
37
we have that
It follows that
ci )j − (Wi )j | = | log(Y ′ )j − log(Wi )j | . |(Yi )′ − (Wi )j |
|(W
i
j
ci − Wi k = klog Yi′ − log Wi k . kYi′ − Wi k
kW
√
Therefore recalling kYi′ − Wi k ≤ kYi − Wi k ≤ O(η np) we complete the proof.
Proof of Theorem 3.3. The proof will follow the same outline as the proof of Theorem 3.1. The
difference is that since we only have a guarantee on the spectral boundedness of the second and
third-order term, we will need to bound the higher-order terms in a different manner. Given that
we have no information on them in this scenario, we will simply bound them in the obvious manner.
We proceed to formalize this.
The sample complexity is polynomial for the same reasons as in the proof of Theorem 3.1, so
we will not worry about it here.
We only need to check the conditions for Algorithms 4 and 3 to work, so that we can apply
Theorems 5.4 and 2.10.
Towards that, first we claim that we can write the PMI tensor for any equipartition Sa , Sb , Sc
of [n] as
PMITSa ,Sb ,Sc =
2 ! X
1
ρ
ρ X
Fk,Sa ⊗ Fk,Sb ⊗ Fk,Sc −
Gk,Sa ⊗ Gk,Sb ⊗ Gk,Sc
1−ρ
2 1−ρ
k∈[m]
k∈[m]
3 ! X
ρ
1
Hk,Sa ⊗ Hk,Sb ⊗ Hk,Sc + E
+
3 1−ρ
(D.4)
k∈[m]
where kEk{1,2},{3} ≤ ρ4 m(np)3/2 . Towards achieving this, first we claim that Proposition 5.6
implies that for any subsets Sa , Sb , Sc ,
m
X
k=1
(1 − exp(−lWk ))Sa ⊗ (1 − exp(−lWk ))Sb ⊗ (1 − exp(−lWk ))Sc
{1,2},{3}
≤ m(np)3/2
Indeed, if we put γk = (1 − exp(−lWk ))Sa , δk = (1 − exp(−lWk ))Sb , θk = (1 − exp(−lWk ))Sc ,
P
√
then we have k k γk γk⊤ k ≤ mnp, and similarly for δk . Since maxk kθk k ≤ (np)1/2 , the claim
immediately follows. Hence, (D.4) follows.
l 1/3
ρ
1
Next, let RSa be the matrix that has as columns the vectors l 1−ρ
(P̃l )j,Sa , l ∈
1/3
ρ
(P̃1 )j,Sa for j ∈ [m] for
[2, L], j ∈ [m] and A is the matrix that has as columns the vectors 1−ρ
some L = O(poly(n)), similarly as in the proof of Theorem 3.1.
We claim that RSa RS⊤a is τ spectrally bounded by ρF F ⊤ .
l 1/3
√
ρ
1
Indeed, for any l > 2, we have k l 1−ρ
P̃l k . ρl/3 kP̃l k . ρl/3 mnp Hence,
RSa RS⊤a ρ2/3 GG⊤ + ρ4/3 HH ⊤ + ρ2 LL⊤ +
3ρ
2/3
τ (F F
⊤
⊤
+ σmin (F F )) + ρ
X
8/3
38
ρ2l/3 mnp2
l≥4
mnp2 - ρ2/3 τ (F F ⊤ + σmin (F F ⊤ ))
(D.5)
where the first inequality holds since HH ⊤ , GG⊤ , LL⊤ are τ -spectrally bounded bounded by F and
the second since σmin (F F ⊤ ) & np and τ ≥ 1. Let τ ′ = ρ2/3 τ . Since we are assuming the matrix F
is O(1)-incoherent, we can apply Theorem C.10, and claim the output of Algorithm 4 are matrices
Qa , Qb , Qc which are τ -approximate whitening matrices for A, B, C respectively.
By Lemma D.1, we have again
1
1
. (1 + τ ′ )kAA⊤ k, σmin (Qa ) ≥
& (1 − τ ′ )σmin (AA⊤ )
σmax (Qa ) ≤
1 − τ′
1 + τ′
Then, applying Theorem 5.4, we get that we recover vectors (âi , b̂i , ĉi ) are O(η ′ )-close to
1/3
ρ
(1 − exp(Wi )), for all i ∈ [m]. for
1−ρ
3/2
η ′ . max(kQa k , kQb k , kQc k)1/2 · τ ′ + σ −3/2 kEk{1,2},{3}
√
Recall that τ ′ = ρ2/3 τ , and kQa k ≤ ρ2/3 σmax (F ) . ρ2/3 mnp and kEk{1,2},{3} ≤ ρ4 m(np)3/2 and
σ & ρ2/3 np, we obtain that
!
4 m(np)3/2
√
√
ρ
. ρ4/3 mnpτ 3/2
η ′ . ρ1/3 mnp (τ ρ2/3 )3/2 + 2/3
3/2
(ρ np)
where the last inequality holds since ρ3 m = o(1) = o(τ ).
1/3 (â , b̂ , ĉ ) are O(η ′ /ρ1/3 ) = O(η)-close to 1 −
However, this directly implies that ( 1−ρ
i i i
ρ )
exp(−Wi ), i ∈ [m], which in turn implies (ãi , b̃i , c̃i ) are O(η) close to exp(−Wi ).
Argument for recovering Wi from exp(−Wi ) is then exactly the same as the one in Theorem
3.1.
E
Sample complexity and bias of the PMI estimator
Finally, we consider the issue of sample complexity. The estimator we will use for the PMI matrix
will simply be the plug-in estimator, namely:
ˆ i,j = log P̂r[si = 0 ∧ sj = 0]
PMI
ˆ j = 0]
P̂r[si = 0]Pr[s
(E.1)
Notice that this estimator is biased, but as the number of samples grows, the bias tends to zero.
Formally, we can show:
Lemma E.1. If the number of samples N satisfies
1
1
log m
N≥
mini6=j {Pr[si = 0 ∧ sj = 0} δ2
ˆ i,j − PMIi,j | ≤ δ, ∀i 6= j.
with high probability |PMI
Proof. Denoting ∆i,j = P̂r[si = 0 ∧ sj = 0] − Pr[si = 0 ∧ sj = 0] and ∆i = P̂r[si = 0] − Pr[si = 0],
we get that
Pr[si = 0 ∧ sj = 0] + ∆i,j
ˆ i,j = log P̂r[si = 0 ∧ sj = 0] = log
PMI
ˆ
(Pr[s
P̂r[si = 0]Pr[sj = 0]
i = 0] + ∆i )(Pr[sj = 0] + ∆j )
∆i,j
∆i
∆j
= PMIi,j + log 1 +
− log 1 +
− log 1 +
Pr[si = 0 ∧ sj = 0]
Pr[si = 0]
Pr[sj = 0]
39
Furthermore, we have that
2
3x
2x
2+x
≤ log(1 + x) ≤
√x ,
x+1
for x ≥ 0, which implies that when x ≤ 1,
≤ log(1 + x) ≤ x. From this it follows that if
∆i
∆i,j
, max
≤δ
max max
i Pr[si = 0]
i,j Pr[si = 0 ∧ sj = 0]
we have
δ
ˆ i,j ≤ PMIi,j + δ
≤ PMI
3
Note that it suffices to show that if N > 1−4pmax1 mρmax δ12 log m, we have
∆i
∆i
> (1 + δ) ∨
< (1 − δ) ≤ exp(− log2 m)
Pr
Pr[si = 0]
Pr[si = 0]
PMIi,j −
and
Pr
∆i,j
∆i,j
> (1 + δ) ∨
< (1 − δ) ≤ exp(− log2 m)
Pr[si = 0 ∧ sj = 0]
Pr[si = 0 ∧ sj = 0]
since this implies
(E.2)
(E.3)
∆i,j
∆i
max max
, max
≤δ
i,j Pr[si = 0 ∧ sj = 0]
i Pr[si = 0]
with high probability by a simple union bound.
Both (E.2) and (E.3) will follow by a Chernoff bound.
Indeed, consider (E.2) first. We have by Chernoff
s
!
#
"
log N
Pr[si = 0] ≤ exp(− log2 N )
Pr ∆i > 1 +
N Pr[si = 0]
Hence, if N >
2
1
1
Pr[si =0] δ2
log m, we get that 1 − δ ≤
∆i
Pr[si =0]
≤ 1 + δ with probability at least
1 − exp(log m).
The proof of (E.3) is analogous – the only difference being that the requirement is that N >
1
1
Pr[si =0] δ2 log m which gives the statement of the lemma.
Virtually the same proof as above shows that:
Lemma E.2. If the number of samples N satisfies
N≥
1
1
log m
mini6=j6=k {Pr[si = 0 ∧ sj = 0 ∧ sk = 0} δ2
ˆ i,j,k − PMITi,j,k | ≤ δ, ∀i 6= j 6= k.
with high probability |PMIT
As an immediate corollary, we get:
Corollary E.3. If the number of samples N satisfies
N≥
1
1
log m
mini6=j6=k {Pr[si = 0 ∧ sj = 0 ∧ sk = 0} δ2
1
1
log m
mini6=j6=k {Pr[si = 0 ∧ sj = 0 ∧ sk = 0} δ2
with high probability for any equipartition Sa , Sb , Sc
N≥
ˆ Sa ,S ,Sc − PMITSa ,S ,Sc k{1,2},{3} . n3 δ
kPMIT
b
b
40
F
Matrix Perturbation Toolbox
In this section we discuss standard matrix perturbation inequalities. Many results in this section
b = A + E, the perturbation in individual singular
can be found in Stewart and Sun [Ste77]. Given A
values can be bounded by Weyl’s theorem:
b = A+E, we know σk (A)−kEk ≤ σk (A)
b ≤ σk (A)+kEk.
Theorem F.1 (Weyl’s theorem). Given A
For singular vectors, the perturbation is bounded by Wedin’s Theorem:
Lemma F.2 (Wedin’s theorem; Theorem 4.1, p.260 in [Ste90]). Given matrices A, E ∈ Rm×n with
m ≥ n. Let A have the singular value decomposition
Σ1 0
A = [U1 , U2 , U3 ] 0 Σ2 [V1 , V2 ]⊤ .
0
0
b = A + E, with analogous singular value decomposition. Let Φ be the matrix of canonical
Let A
b1 , and Θ be the matrix of canonical angles
angles between the column span of U1 and that of U
between the column span of V1 and that of Vb1 . Suppose that there exists a δ such that
min |[Σ1 ]i,i − [Σ2 ]j,j | > δ,
i,j
and
min |[Σ1 ]i,i | > δ,
i,i
then
k sin(Φ)k2 + k sin(Θ)k2 ≤ 2
kEk2
.
δ2
Perturbation bound for pseudo-inverse When we have a lowerbound on σmin (A), it is easy
to get bounds for the perturbation of pseudoinverse.
Theorem F.3 (Theorem 3.4 in [Ste77]). Consider the perturbation of a matrix A ∈ Rm×n : B =
A + E. Assume that rank(A) = rank(B) = n, then
√
kB † − A† k ≤ 2kA† kkB † kkEk.
Note that this theorem is not strong enough when the perturbation is only known to be τ spectrally bounded in our definition.
41
| 8cs.DS
|
HILBERT SERIES OF RESIDUAL INTERSECTIONS
arXiv:1304.0044v1 [math.AC] 29 Mar 2013
Marc Chardin, David Eisenbud♦ , and Bernd Ulrich♦
♥
Abstract We find explicit formulas for the Hilbert series of residual intersections of a scheme in
terms of the Hilbert series of its conormal modules. In a previous paper we proved that such
formulas should exist. We give applications to the dimension of secant varieties of surfaces and
three-folds.
Introduction
Let M = ⊕i∈Z Mi be a finitely generated graded module over the homogeneous coordinate ring
R of a projective variety over a field k. The Hilbert series (sometimes called the Hilbert-Poincaré
series) of M , which we will denote [[M ]], is the Laurent series
[[M ]] =
X
(dim Mi )ti .
If Z ⊂ Pn := Pnk is a scheme, then the Hilbert series of Z is by definition the Hilbert series of the
homogeneous coordinate ring of Z. Of course this Hilbert series contains the data of the Hilbert
polynomial of Z as well.
Sometimes interesting geometric data (such as the dimension of a secant variety) can be described in terms of residual intersections in the sense of Artin and Nagata [2], and the purpose of
this paper is to compute the Hilbert series of such schemes. Here is the definition: let X ⊂ Y ⊂ Pn
be closed subschemes of Pn , let R be the homogeneous coordinate ring of Y , and let IX ⊂ R be
the ideal of X in Y . A scheme Z ⊂ Y is called an s-residual intersection of X in Y if Z is defined
by an ideal of the form IZ = (f1 , . . . , fs ) :R IX , with f1 , . . . , fs homogeneous elements in IX , and
Z is of codimension at least s in Y .
We wish to derive formulas for the Hilbert series of Z in terms of information about X and
the degrees of the polynomials fi . In our previous paper [6] we showed that this is sometimes
possible in principle: under certain hypotheses the Hilbert series of Z does not vary if we change
the polynomials fi , keeping their degrees fixed. In this paper we make this more precise by giving
formulas—under somewhat stronger hypotheses—for the Hilbert series of Z in terms of the degrees
m
of the fi and the Hilbert series of finitely many modules of the form ωR /IX
ωR , where ωR denotes
the canonical module of R.
♦
partially supported by the National Science Foundation.
We are all grateful to MSRI for supporting the commutative algebra years in 2002-03 and
2012-13, where we worked on this paper.
AMS 2010 Subject Classification: Primary, 13D40, 13C40; Secondary, 13H15, 13M06, 14C17,
14N15
♥
1
For example, suppose that Y = Pn and X is locally a complete intersection (for instance,
smooth). If f1 . . . , fs are homogeneous elements of degree d of I = IX such that R := (f1 , . . . , fs ) : I
has codimension > s then the Hilbert series of the homogeneous coordinate ring of the scheme Z
defined by R differs from that of a complete intersection defined by s forms of degree d by
s
X
n+j s
(−1)
tjd [[ωR /I j−g+1 ωR ]](t−1 ) + a polynomial.
j
j=g
The polynomial remainder term is present because we have made assumptions only on the scheme,
and not on the homogeneous coordinate ring. Here the expression [[R/I j−g+1 ]](t−1 ) denotes the
Laurent series obtained by writing [[R/I j−g+1 ]] as a rational function in t, substituting t−1 for t,
and rewriting the result as a Laurent series. Up to a small shift in notation, this is the formula
given in Remark 1.5b (where the case of forms fi of different degrees is also treated.)
In applications, one sometimes needs to know only whether the s-residual intersection Z actually has codimension exactly s in Y ; for example, we will use such information in Section 3 to say
when the secant varieties of certain (possibly singular) surfaces and smooth 3-folds have dimension
less than the expected dimension. For this purpose it is enough to know just one coefficient of
the Hilbert polynomial of Z, that corresponding to the degree of the codimension s component of
Z. More generally, we show how to use partial information about X to compute just the first k
coefficients of the Hilbert polynomial of Z.
Consider the case where Y is Gorenstein and X is Cohen-Macaulay. Suppose, further, that
locally in codimension i < s, the subscheme X ⊂ Y can be defined by i equations, and that, for
j 6 s − g,
j
j+1
depth IX
/IX
> dim X − j.
If Z is any s-residual subscheme of X in Y , then the Hilbert polynomial of Z may be written in
j
terms of the Hilbert polynomials of IX
for j 6 s − g + 1 (the explicit formula is given in Theorem
1.9). Moreover, if X and Y satisfy some of our hypotheses only up to some codimension r, then
the formula gives the first r coefficients of the Hilbert polynomial of Z.
Our formulas are derived in Section 1, which is the technical heart of the paper. To prove them
we need to adapt the arguments of Ulrich [19]. The delicate point is the use in that paper of the
isomorphism R ≃ ωR that holds for a Gorenstein ring. Since our rings are Gorenstein only up to
a certain codimension r, we know only that ωR is a line bundle locally in codimension r, and this
does not suffice to determine its Hilbert series. Thus the module ωR must be brought into play.
For other work along these lines, see Cumming [7].
j
j+1
In case where Y is Gorenstein, the sheaves IX
/IX
themselves play the crucial role in our
2
formulas. If X were locally a complete intersection scheme, then IX /IX
would be a vector bundle
j
j+1
and IX /IX would be its j-th symmetric power, so it is reasonable to hope that for “nice” ideals
I the Hilbert series of the first few conormal modules should determine the rest. We prove a
general theorem of this kind in Section 2, and carry out the reduction in some particular cases. For
instance, if s = codim(X) then the degree of an s-residual intersection scheme Z in Y = Pn may be
Q
calculated immediately from Bézout’s Theorem: deg Z = ( i deg fi ) − deg X. This was extended
to a formula in the case s = codimY X + 1 by Stückrad [18] and to the case s = codimY X + 2 by
2
Huneke and Martin [16]. Our formula gives an answer in general, and we work this out explicitly
for the case s = codimY X + 3.
In Section 3 we apply our results to the study of secant loci. Our general theorems imply
conditions in terms of Chern classes of a smooth embedded three-fold for the degeneracy of the
secant locus in terms of Chern classes and in terms of certain Hilbert coefficients. We also recover
the analogous criteria for surfaces with mild singularities, a case already by Dale [7a] and others.
1. Formulas for the Hilbert series of residual intersections
To deal with rings R that are not equidimensional, we define the true codimension of a prime
ideal p ⊂ R to be dim R − dim R/p. We say that an ideal I satisfies the condition ∗Gs if, for every
prime p in V (I) of true codimension < s, the minimal number of generators of Ip is at most dim Rp
(the usual condition Gs is the same but for codimension instead of true codimension ).
Definition 1.1. Let R be a graded ring and M, N graded R-modules. We say that M and
N are equivalent up to true codimension r, and write M ∼
= N , if there exist graded R-modules
r
W1 , . . . , Wn with W1 = M , Wn = N and homogeneous maps Wi → Wi+1 or Wi+1 → Wi , for
1 6 i 6 n − 1, which are isomorphisms locally in true codimension r. A homogeneous map which
∼
is an isomorphism up to true codimension r will be denoted by −→.
r
Saying that M ∼
= N is of course much stronger than saying that M and N are isomorphic locally
r
at each prime of true codimension 6 r. For example, any two modules M and N that represent
line bundles on a projective variety of dimension r satisfy the latter condition, but M ∼
= N implies
r
that they represent isomorphic line bundles! The need to provide explicit maps that are locally
isomorphisms in some true codimension between modules that are not in fact isomorphic is what
makes the work in this section delicate.
If R is a Noetherian standard graded algebra over a field and M a finitely generated graded
R-module, we will denote the Hilbert series of M by [[M ]]. If M ∼
= N then [[M ]] − [[N ]], written as
r
a rational function, has a pole of order less than dim R − r at 1; we will write this as [[M ]] ≡ [[N ]]
r
and say that these series are r-equivalent. Thus if r = dim R − 1 =: d then [[M ]] ≡ [[N ]] means that
r
the Hilbert polynomials of M and N agree, and in general if r < dim R then [[M ]] ≡ [[N ]] means
r
that the Hilbert polynomials of M and N , written in the form
d+t
d−1+t
ad
+ ad−1
+ ···,
d
t
have the same coefficients of as for s > dim R − 1 − r.
We will also extend the notation ≡ to arbitrary series that are rational functions with no pole
r
outside 1, by the same requirement, as soon as dim R is clear from the context.
The substitution t 7→ t−1 is a well-defined automorphism of the ring Z[t, t−1 , (1 − t)−1 ], since
(1 − t)−1 = −t(1 − t)−1 ∈ Z[t, t−1 , (1 − t)−1 ].
Lemma 1.2. Let R be a positively graded Noetherian algebra over a field k. If for each prime
p of dimension > dim R − r, the ring Rp is Cohen-Macaulay of dimension dim R − dim R/p, then
[[ωR ]](t) ≡ (−1)dim R [[R]](t−1 ).
r
3
Proof. Dualize a free resolution of R over a polynomial ring S, and note that all the homology
codimS (R)
modules other than ExtS
(R, ωS ) are supported in codimension > r.
We next adapt some results of [6] and [20] to our context.
Suppose that R is a local Cohen-Macaulay ring and I is an ideal of height g. In the rest of this
section we will often use the condition that depth R/I j > dim R/I − j + 1 for 1 6 j 6 s − g. These
conditions are satisfied if I satisfies Gs and if moreover, I has the sliding depth property or, more
restrictively, is strongly Cohen-Macaulay (which means that for every i, the i-th Koszul homology
Hi of a generating set h1 , . . . , hn of I satisfies depth Hi > dim R − n + i or is Cohen-Macaulay,
respectively) ([13, 3.3] and [15, 3.1]). The latter condition always holds if I is a Cohen-Macaulay
almost complete intersection or a Cohen-Macaulay deviation 2 ideal of a Gorenstein ring [3]. It
is also satisfied for any ideal in the linkage class of a complete intersection [14, 1.11]. Standard
examples include perfect ideals of grade 2 ([1] and [9]) and perfect Gorenstein ideals of grade 3 [21].
Lemma 1.3. Let R be a finitely generated positively graded algebra over a factor ring of a local
Gorenstein ring and write ω = ωR . Let I be a homogeneous ideal of height g, let f1 , . . . , fs be forms
contained in I of degrees d1 , . . . , ds , write Ai := (f1 , . . . , fi ), A := As , and Ri = Ai : I. Assume
that ht Ri > i for 1 6 i 6 s and ht I + Ri > i + 1 for 1 6 i 6 s − 1. Further suppose that, locally
off V (A), the elements f1 , . . . , fs form a weak regular sequence on R and on ω, and that locally in
true codimension r in R along V (A), the ring R is Gorenstein and depth R/I j > dim R/I − j + 1
for 1 6 j 6 s − g. Then :
∼
(a) (R/Ri−1 )(−di ) −→ Ai /Ai−1 via multiplication by fi for 1 6 i 6 s.
r
·fi
j
(b) 0 → (ωI /ωAi−1 I j−1 )(−di ) −→ ωI j+1 /ωAi−1 I j −→ ωI j+1 /ωAi I j → 0 is a complex that
is exact locally in true codimension r for 1 6 i 6 s and min{1, i − g} 6 j 6 s − g.
(c) ExtiR (R/Ri , ω) ∼
= (ωI i−g+1 /ωAi I i−g )(d1 + · · · + di ) for 0 6 i 6 s, if locally in true
r
codimension r in R along V (A), depth R/I j > dim R/I − j + 1 for 1 6 j 6 s − g + 1.
Proof. Adjoining a variable to R and to I and localizing, we may suppose that grade I > 0.
Furthermore, our assumptions imply that I is Gs and satisflies the Artin-Nagata condition ANs−1
locally in true codimension r; see [20, 2.9(a)]. Likewise, in the setting of (c), ANs holds locally in
true codimension r.
Part (a) holds along V (A) by [20, 1.7(g)], and off V (A) because f1 , . . . , fi form a weak regular
sequence. Moreover, the sequence of (b) is obviously a complex and it is exact in true codimension
r by [20, 2.7(a)].
For the proof of (c), we induct on i. For i = 0, our assertion is clear since grade I > 0 and
therefore R/R0 = R. Assuming that the assertion holds for Ri = R/Ri for some i, 0 6 i 6 s − 1,
we are going to prove our claim for Ri+1 = R/Ri+1 . To this end we may suppose r > i + 1.
We first wish to prove that
(1)
i
i
∼
Exti+1
R (Ri /(fi+1 Ri : IRi ), ω) = [I ExtR (Ri , ω)/fi+1 ExtR (Ri , ω)](di+1 ).
r
Using the exact sequence
.fi+1
0 → Ri /(0 :Ri fi+1 Ri )(−di+1 ) −→ Ri −→ Ri /fi+1 Ri → 0
4
we obtain a long exact sequence
.fi+1
i+1
· · · ExtiR (Ri , ω) −→ ExtiR (Ri /(0 :Ri fi+1 Ri ), ω)(di+1 ) → Exti+1
R (Ri /fi+1 Ri , ω) → ExtR (Ri , ω) · · ·
Since the support in R of (0 :Ri fi+1 ) has true codimension > r + 1 > i by part (a), we have
∼
∼
ExtiR (Ri /(0 :Ri fi+1 Ri ), ω) −→ ExtiR (Ri , ω) via the natural map. Furthermore Exti+1
R (Ri , ω) = 0;
r
r
as locally up to true codimension r on V (A), Ri is Cohen-Macaulay of true codimension i by
[20, 1.7(a)], whereas locally up to true codimension r off V (A), Ri is defined by the weak regular
sequence f1 , . . . , fi and hence has projective dimension at most i. Therefore
ExtiR (Ri , ω)/fi+1 ExtiR (Ri , ω) (di+1 ) ∼
= E := Exti+1
R (Ri /fi+1 Ri , ω).
r
Hence, to prove (1), it suffices to show that
i+1
∼
I Exti+1
R (Ri /fi+1 Ri , ω) = ExtR (Ri /(fi+1 Ri : IRi ), ω).
r
The map Ri /fi+1 Ri → Ri /(fi+1 Ri :Ri IRi ) induces a map
i+1
φ : Exti+1
R (Ri /(fi+1 Ri : IRi ), ω) → ExtR (Ri /fi+1 Ri , ω).
We prove that locally in true codimension r in R, the map φ is injective, and its image coincides
with I E, which gives
∼
∼
im φ −→ im φ + I E ←− I E.
r
r
This is trivial locally off V (A) because on this locus I = R and fi+1 Ri : IRi = fi+1 Ri . Therefore,
we may localize to assume that R is a local ring of dimension at most r and A 6= R. Of course we
may suppose Ri 6= 0. In this case R is Gorenstein, Ri is Cohen-Macaulay of codimension i, and
fi+1 is a non zerodivisor on Ri by [20, 1.7(f)]. Let S = Ri /fi+1 Ri .
The natural equivalence of functors Exti+1
R (—, ω) ≃ HomS (—, ωS ) together with the exact
sequence
0 → 0 :S IS −→ S −→ S/0 :S IS −→ 0
yield a commutative diagram with an exact row
Exti+1
R (Ri /(fi+1 Ri : IRi ), ω)
φ
≃
0
/ HomS (S/0 :S IS, ωS )
/E
≃
/ ω S
ψ
/ HomS (0 :S IS, ωS )
/0
The last map is surjective because S/(0 :S IS) = Ri+1 by [20,1.7(f)] and Ri+1 is a maximal CohenMacaulay S-module. Now φ is injective and the desired equality im φ = I E follows once we have
shown that im ψ = IωS . For this it suffices to prove
(2)
coker ψ ≃ ωS /IωS ;
5
for then IωS ⊂ im ψ and we have the natural epimorphism of isomorphic modules ωS /IωS →
coker ψ, which is necessarily an isomorphism. We first argue that ωS /IωS is a maximal CohenMacaulay S-module. Indeed, [20, 2.7(c)] gives Ri ∩ I i−g+2 = Ai I i−g+1 , which implies
Ai I i−g ∩ I i−g+2 = Ai I i−g+1 .
(3)
Hence by our induction hypothesis,
IωS ≃ I i−g+2 /(Ai I i−g ∩ I i−g+2 + fi+1 I i−g+1 ) = I i−g+2 /Ji+1 I i−g+1 .
But the latter is indeed a maximal Cohen-Macaulay S-module according [20, 2.7(b)]. Thus
ωS /IωS ≃ HomS (HomS (ωS /IωS , ωS ), ωS ) ≃ HomS (HomS (S/IS, S), ωS ) ≃ HomS (0 :S IS, ωS ).
This completes the proof of (2), and hence of (1).
Now
IExtiR (Ri , ω)/fi+1 ExtiR (Ri , ω) ∼
= (ωI i−g+2 /(ωAi I i−g ∩ ωI i−g+2 + ωfi+1 I i−g+1 ))(d1 + · · · + di )
r
by our induction hypothesis, and using (3) one sees that
IExtiR (Ri , ω)/fi+1 ExtiR (Ri , ω) ∼
= (ωI i−g+2 /ωAi+1 I i−g+1 )(d1 + · · · + di ).
(4)
r
∼
On the other hand, Ri+1 −→ Ri /(fi+1 Ri : IRi ) according to [20, 1.7(f)], and hence
r
i+1
∼
Exti+1
R (Ri+1 , ω) = ExtR (Ri /(fi+1 Ri : IRi ), ω).
(5)
r
Now combining (5), (1), (4) concludes the proof of part (c).
We write σm (t1 , . . . , ts ) for the m-th elementary symmetric function.
Theorem 1.4. Let R be a standard graded Noetherian algebra over a field. Write n = dim R
and ω = ωR , and let I be a homogeneous ideal of height g satisfying ∗Gs . Let f1 , . . . , fs be forms
Qs
contained in I of degrees d1 , . . . , ds , write ∆s := i=1 (1 − tdi ), A = (f1 , . . . , fs ), R = A : I, and
assume that ht R > s. For each prime of true codimension 6 r suppose:
• If p ∈
/ V (A), then the elements f1 , . . . , fs form a weak regular sequence on Rp and on ωp .
• If p ∈ V (A), then the ring Rp is Gorenstein of dimension equal to the true codimension of p
and depth Rp /Ipj > dim Rp /Ip − j + 1 for 1 6 j 6 s − g.
(a)
n−g
[[R/A]](t) ≡ ∆s [[R]](t) − (−1)
r
s−g
X
(−1)j σg+j (td1 , . . . , tds )[[ω/I j ω]](t−1 ).
j=1
6
(b) If furthermore, locally in true codimension r in R along V (A), depth R/I s−g+1 > dim R/I −
s + g, then
n−g
[[R/R]](t) ≡ ∆s [[R]](t) − (−1)
r
s−g+1
X
(−1)j−1 σg+j−1 (td1 , . . . , tds )[[ω/I j ω]](t−1 ).
j=1
Proof. For 0 6 i 6 s, write Ai = (f1 , . . . , fi ), Ri = Ai : I. Applying repeatedly a general
position argument (see [6, 2.5]), we may assume that ht Ri > i and ht I +Ri > i+1 for 0 6 i 6 s−1.
We first notice that for 0 6 i 6 s and i − g 6 j 6 s − g,
[[ωAi I j ]] ≡
r
i
X
(−1)ℓ+1 σℓ (td1 , . . . , tdi )[[ωI j−ℓ+1 ]],
ℓ=1
which can be easily deduced from Lemma 1.3(b) using induction on i.
Now let 0 6 i 6 s − 1 for (a), or 0 6 i 6 s for (b), respectively. Then by Lemma 1.3(c),
[[ExtiR (R/Ri , ω)]](t) ≡ t−(d1 +···+di ) ([[ωI i−g+1 ]] − [[ωAi I i−g ]])(t)
r
(1)
≡t
−(d1 +···+di )
r
i
X
(−1)ℓ σℓ (td1 , . . . , tdi )[[ωI i−g−ℓ+1 ]](t).
ℓ=0
Locally in true codimension r in R, R/Ri is either zero or Cohen-Macaulay of codimension i
by [20, 2.9 and 1.7(a)]. Therefore
[[ExtiR (R/Ri , ω)]](t) ≡ (−1)n−i [[R/Ri ]](t−1 ),
r
as can be easily seen by dualizing a homogeneous finite free resolution of R/Ri over a polynomial
ring and using Lemma 1.2. Now by Lemma 1.3(c) and (1),
[[R/Ri ]](t) ≡ (−1)n−i [[ExtiR (R/Ri , ω)]](t−1 )
r
≡ (−1)n−i td1 +···+di
r
i
X
(−1)ℓ σℓ (t−d1 , . . . , t−di )[[ωI i−g−ℓ+1 ]](t−1 )
ℓ=0
= (−1)n−i
i
X
(−1)ℓ σi−ℓ (td1 , . . . , tdi )[[ωI i−g−ℓ+1 ]](t−1 ).
ℓ=0
Therefore
(2)
n−g
[[R/Ri ]](t) ≡ (−1)
r
i−g+1
X
(−1)j+1 σg+j−1 (td1 , . . . , tdi )[[ωI j ]](t−1 ).
j=−g+1
Now part (b) follows since
∆i [[R]](t) ≡ ∆i (−1)n [[ω]](t−1 ) = (−1)n−i
r
i
X
(−1)ℓ σi−ℓ (td1 , . . . , tdi )[[ω]](t−1 )
ℓ=0
n−g
= (−1)
i−g
X
j=−g
7
(−1)j σg+j (td1 , . . . , tdi )[[ω]](t−1 ).
To see (a) notice that by Lemma 1.3(a),
[[R/Ai ]](t) = [[R/Ai−1 ]](t) − tdi [[R/Ri−1 ]](t)
for 1 6 i 6 s. Now by induction on i using (2), one shows that
i−g
X
n−g
[[R/Ai ]](t) ≡ (−1)
r
(−1)j σg+j (td1 , . . . , tdi )[[ωI j ]](t−1 ),
j=−g
from which (a) can be easily deduced.
Remark 1.5. If in Theorem 1.4, R is a polynomial ring in n variables, then the formulas of
that theorem take the following form:
(a)
−n
[[R/A]](t) ≡ ∆s [[R]](t) − (−t)
r
s−g
X
(−1)g+j σg+j (t)[[R/I j ]](t−1 );
j=1
s−g+1
(b)
X
−n
[[R/R]](t) ≡ ∆s [[R]](t) − (−t)
r
(−1)g+j−1 σg+j−1 (t)[[R/I j ]](t−1 ).
j=1
Lemma 1.6. Write ∆s (t) = (1 − t)s
P
k>0 ck (d1 , . . . , ds )(1
X
k
ck (d1 , . . . , ds ) = (−1)
i1 >1,...,is >1
i1 +···+is =k+s
− t)k . Then
s
Y
dj
j=1
ij
.
Q
Pdj −1 ℓ
t and notice that ∆s (t) = (1 − t)s sj=1 Pj (t).
Proof. Write Pj (t) = ℓ=0
Pdj −1
(m)
ℓ
dj
m! m
= m! m+1
. Hence
Now Pj (1) = ℓ=0
s
Y
j=1
Pj
(k)
X
(1) =
m1 >0,...,ms >0
m1 +···+ms =k
= k!
X
s
Y
k!
(m )
P j (1)
m1 ! · · · ms ! j=1 j
m1 >0,...,ms >0
m1 +···+ms =k
This yields our formula since ck (d1 , . . . , ds ) =
(−1)k
k!
s
Y
j=1
Q
s
dj
.
mj + 1
j=1 Pj
(k)
(1).
Pm
.
Lemma 1.7. Let P be a numerical polynomial written in the form P (t) = i=0 (−1)i ei t+m−i
m−i
Pm
t+m−i
i
For an integer d define the polynomial Q(t) = P (−t + d), and write Q(t) = i=0 (−1) hi m−i .
Then
i
X
k d+m+1−k
m
(−1)
hi = (−1)
ek .
i−k
k=0
8
Proof. We first notice that for integers r and n > 0, one has the following identities of
numerical polynomials :
(3)
−t + n
n
(4)
t+r+n
n
n
= (−1)
=
t−1
,
n
n
X
r−1+ℓ t+n−ℓ
ℓ
ℓ=0
n−ℓ
,
where the first is obvious and the second can be easily proved by induction on n.
Now
m
X
−t + d + m − k
m−k
k=0
m
X
t−d−1
m
ek
by (3)
= (−1)
m−k
k=0
m
X
t + (−d − 1 − m + k) + (m − k)
m
ek
= (−1)
m−k
k=0
m
m−k
X
X −d − 1 − m + k − 1 + ℓt + m − k − ℓ
ek
= (−1)m
ℓ
m−k−ℓ
k=0
ℓ=0
!
i
m
X
X
−d − m − 2 + i
t+m−i
m
= (−1)
ek
,
i−k
m−i
i=0
Q(t) =
(−1)k ek
by (4)
k=0
where
−d−m−2+i
i−k
= (−1)i+k
d+m+1−k
i−k
by (3).
Recall that the Hilbert series [[M ]] of a finitely generated graded module M over a homogeneous
ring over a field is element of the ring Z[t, t−1 , (1−t)−1 ] ⊂ Z[[t]][t−1 ]. In general, any S ∈ Z[t, t−1 , (1−
t)−1 ] can be written uniquely in the form
S(t) =
D−1
X
(−1)i ei
i=0
1
+F
(1 − t)D−i
where ei ∈ Z and F ∈ Z[t, t−1 ]. The coefficients ei can be computed as ei (M ) =
P (t) = S(t)(1 − t)D . We call
Q(t) =
D−1
X
i=0
i
(−1) ei
∂iP
i!
(1), where
t+D−1−i
∈ Q[t]
D−1−i
the polynomial associated to S. Its significance is that if we write S =
for i ≫ 0.
9
P
i∈Z ci t
i
, then ci = Q(i)
Remark 1.8. In the case where S(t) = [[M ]](t), we can take D to be any integer > dim M ,
D
and we define eD
i (M ) := ei . If D = dim M , we simply set ei (M ) := ei (M ). Notice that e0 (M ) is
the multiplicity (or degree) of M . The polynomial associated to [[M ]](t) is the Hilbert polynomial
of M , which we denote by [M ](t).
Theorem 1.9. Write eℓ (d1 , . . . , ds ) =
P
i1 >1,...,is >1
i1 +···+is =ℓ+s
(a) With the assumptions of Theorem 1.4(a),
en−s
(I/A) =
i
i
X
Qs
j=1
dj
ij
.
ei−k (d1 , . . . , ds )ek (R) − (−1)s−g es−g+i (R/I)
k=0
s−g
− (−1)
s−g s−g+i
X
X
(−1)
j=1 k=0
X
j+k
16i1 <···<ig+j 6s
di1 + · · · + dig+j + n − g − k
i+s−g−k
ek (ω/I j ω)
ek (ω/I j ω)
for 0 6 i 6 r − s.
(b) With the assumptions of Theorem 1.4(b),
en−s
(R/R) =
i
i
X
ei−k (d1 , . . . , ds )ek (R)
k=0
s−g+1 s−g+i
X
s−g
+ (−1)
j=1
X
X
j+k
(−1)
16i1 <···<ig+j−1 6s
k=0
di1 + · · · + dig+j−1 + n − g − k
i+s−g−k
for 0 6 i 6 r − s.
Proof. We only prove part (a). First write
n−g
(−1)
s−g
X
j=1
n−g−1
j
(−1)
X
j
[ω/I ω](−t + di1 + · · · + dig+j ) =
X
(−1) hℓ
(−1)ℓ h∗ℓ
ℓ
ℓ=0
16i1 <···<ig+j 6s
=
n−s−1
X
ℓ=0
t+n−g−1−ℓ
n−g−1−ℓ
t+n−s−1−ℓ
n−s−1−ℓ
and notice that h∗i = (−1)s−g hi+s−g . Lemma 1.7 gives
hi+s−g
s−g
X
(−1)j
=−
j=1
X
i+s−g
X
16i1 <···<ig+j 6s k=0
di1 + · · · + dig+j + n − g − k
(−1)k ek (ω/I j ω).
i+s−g−k
Now our assertion follows from Theorem 1.4(a) together with Lemma 1.6.
The proof of part (b) is similar, using Theorem 1.4(b) in place of Theorem 1.4(a).
Remark 1.10. If in Theorem 1.9, R is a polynomial ring in n variables, then the formula in
that theorem takes the following form:
10
en−s
(I/A) = ei (d1 , . . . , ds ) − (−1)s−g es−g+i (R/I)
i
s−g
− (−1)
s−g s−g+i
X
X
j+k
(−1)
j=1 k=0
X
16i1 <···<ig+j 6s
di1 + · · · + dig+j − g − k
i+s−g−k
ek (R/I j )
for 0 6 i 6 r − s;
en−s
(R/R) = ei (d1 , . . . , ds )
i
s−g+1 s−g+i
s−g
+ (−1)
X
j=1
X
j+k
(−1)
X
16i1 <···<ig+j−1 6s
k=0
di1 + · · · + dig+j−1 − g − k
i+s−g−k
ek (R/I j )
for 0 6 i 6 r − s.
Proof. Notice that ω/I j ω ≃ R/I j (−n) and proceed as in the proof of Theorem 1.9.
Corollary 1.11. Let R be a homogeneous ring over a field, write n = dim R, ω = ωR , and
assume that R is Gorenstein locally in true codimension r > s. Let I be a homogeneous ideal
of height g satisfying Gs , let f1 , . . . , fs be forms contained in I of degrees d1 , . . . , ds , write A =
(f1 , . . . , fs ), R = A : I, and assume that locally in true codimension r, depth R/I j > dim R/I −j +1
for 1 6 j 6 s − g.
Then ht R > r + 1 if and only if ht R > s and
(−1)s−g e0 (R)
s
Y
dj = es−g (R/I)
j=1
+
s−g s−g
X
X
j+k
(−1)
j=1 k=0
X
16i1 <···<ig+j 6s
di1 + · · · + dig+j + n − g − k
ek (ω/I j ω).
s−g−k
Proof. One uses Theorem 1.9(a) and [20, 1.7(a)].
2. Hilbert series of powers of ideals and degrees of residual intersections.
2.1 Computing Hilbert series of powers.
Motivated by Remark 1.5, showing the usefulness of the Hilbert series of the powers of an
ideal, we will focus here on the following question: to what extent does the Hilbert series of the
first powers of an ideal determine the Hilbert series of the next powers?
The following Lemma tells us a useful property of a general set of elements of an ideal.
Lemma 2.1. Let R be a standard graded Cohen-Macaulay ring over an infinite field k, let
I a homogeneous ideal, generated by forms of degrees at most d, and let r be an integer with
0 6 r 6 dim R . Further assume that I satisfies Gr+1 and has sliding depth locally in codimension
r.
Given di > d for 1 6 i 6 r + 1], there exists a Zariski dense open subset Ω of the affine k-space
Pr+1
Id1 × · · · × Idr+1 ≃ AN
k (where N =
i=1 dimk Idi ) such that if (f1 , . . . , fr+1 ) ∈ Ω :
11
(a) The ideal (f1 , . . . , fr+1 ) coincides with I locally up to codimension r.
(b) f1 , . . . , fr+1 is a d-sequence locally up to codimension r.
Proof. For part (a) we refer to [2, 1.6 (a)] and [10, 3.9], while (b) follows from [6, 3.6(b)].
We now choose some degrees (e.g. di = d for all i) and polynomials fi as above. Part two of
the lemma implies that the approximation complexes corresponding to f1 , . . . , fr+1 , and therefore
their different graded components,
Mp :
0 → Hr+1 ⊗ Sp−r−1 → Hr ⊗ Sp−r → · · · → H0 ⊗ Sp → 0
have homology in positive degrees supported in codimension at least r + 1. Moreover, H0 (Mp )
coincides with I p /I p+1 locally up to codimension r. See [12] for all of these facts.
Recall that in this sequence, Hq stands for the q-th homology module of the Koszul complex
K(f1 , . . . , fr+1 ; R) and Sq is the free R-module generated by monomials of degree q in r+1 variables.
The maps are homogeneous of degree 0 in the graded case, with the usual weights on the Koszul
complex and similarely the weight of si1 · · · siq ∈ Sq is deg(fi1 ) + · · · + deg(fiq ).
Therfore we have the following equalities :
(1)p
p
[[I /I
p+1
]](t) ≡
r
p
X
(−1)i sp−i (td1 , . . . , tdr+1 )[[Hi ]](t)
i=0
for every p > 0, where sj stands for the sum of all the monomials of degree j in r + 1 variables
(complete symmetric functions). These express the Hilbert series of the modules H0 , . . . , Hp in
terms of the ones of R/I, . . . , I p /I p+1 , and vice versa.
If I has height g, then Hq = 0 for q > r + 1 − g. We therefore immediately see that the
Hilbert series of all the modules I p /I p+1 are determined, up to r-equivalence, by the knowledge of
the Hilbert series of I p /I p+1 for 0 6 p 6 r + 1 − g, up to r-equivalence.
We now assume that, in addition, R is Gorenstein with a-invariant a := a(R) and is strongly
Cohen-Macaulay locally in codimension r. In this case we can use the self-duality of the homology
of the Koszul complex to see that only half of the information about the Koszul homology is needed.
Indeed, the structure of graded alternating algebra on the homology of the Koszul complex gives a
graded map of degree 0,
Hp −→ HomR/I (Hr+1−g−p , Hr+1−g ),
which is an isomorphism up to codimension r by a theorem of Herzog (see [11, 2.4.1]).
Now we have a collection of graded maps of degree 0 that connect the following modules,
g
HomR/I (Hr+1−g−p , Hr+1−g ) ∼
= HomR/I (Hr+1−g−p , ExtR (R/I, R)[−(d1 + · · · + dr+1 )])
r
∼
= HomR/I (Hr+1−g−p , ωR/I [−a − (d1 + · · · + dr+1 )])
r
∼
= ωHr+1−g−p [a + (d1 + · · · + dr+1 )].
r
From the Cohen-Macaulayness of the modules Hp , locally in codimension at most r, we therefore
have:
(2)p
[[Hp ]](t) ≡ ta+(d1 +···+dr+1 ) (−1)dim R/I [[Hr+1−g−p ]](t−1 ).
r
12
Moreover, the Euler characteristic of the homology of the Koszul complex depends only upon
the degrees d1 , . . . , dr+1 , namely
r+1−g
X
(3)
(−1)p [[Hp ]] = ∆r+1 [[R]] ≡ 0.
r
p=0
We now have put together all the formulas needed to effectively compute what we state in the
next theorem.
Remark that the result of the computation does not depend on the choice of the di ’s. Thus
we may choose di = 0 for all i, the intermediate steps have no meaning (e.g. [[Hp ]] may not have
positive coefficients), but the information that we extract from the computation is the same. Out
of this remark, one may use the following :
(1)p
(2)p
(3)
p
[[I /I
p+1
]](t) ≡
r
p
X
i
(−1)
i=0
r+p−i
[[Hi ]](t),
r
[[Hp ]](t) ≡ ta (−1)dim R/I [[Hr+1−g−p ]](t−1 ),
r
r+1−g
X
(−1)p [[Hp ]] ≡ 0.
r
p=0
Theorem 2.2.
Let R be a homogeneous algebra, I an homogeneous R-ideal, and let us
suppose that, locally in true codimension r, I is licci purely of true codimension g and satisfies
, up to r-equivalence, the Hilbert series
Gr+1 . Given the Hilbert series of I p /I p+1 for 0 6 p 6 r−g
2
of I p /I p+1 can be computed for all p, up to r-equivalence, by the formulas above.
Proof. Let us choose some sufficiently big integers d1 , . . . , dr+1 (one can treat them as unknowns, or take all of them equal to some fixed number or unknown d). Let us put q = r + 1 − g.
First, using (1)p for 0 6 p 6 q−1
, we get the Hilbert series of Hp , for p in the same range,
2
up to true codimension > r terms.
Then, from (2)p for the same p’s, we get the series of Hq , . . . , Hq−[ q−1 ] . Therefore, we get the
2
series of all the Hp ’s (up to r-equivalence) if q is odd ; and all of them but one, namely Hq/2 , if q
is even. In case q is even, we get the series of Hq/2 using (3).
Therefore we know the series of all the modules Hp in every case, and can use (1)p to get the
ones of I p /I p+1 for any p.
Notice that the di ’s appear in every formula, but should disappear at the end !
Example 2.3. If X ⊆ Pn is an equidimensionnal locally complete intersection scheme, and IX
p
p+1
the corresponding ideal sheaf, the Hilbert polynomials of the sheaves IX
/IX
are all determined
dim X
by the ones for 0 6 p 6
.
2
Proof. It is the case where r = n and g = n − dim X.
13
Example 2.4. If X ⊆ Pn is an equidimensionnal locally complete intersection threefold,
p
p+1
and IX the corresponding ideal sheaf, the Hilbert polynomial of the sheaves IX
/IX
are all
2
determined by the Hilbert polynomials of X and the one of the conormal bundle IX /IX . Moreover
the coefficients of terms in degrees 3 and 2 in the Hilbert polynomial of the conormal bundle are
determined by those of the Hilbert polynomial for X.
Informally, we have the following picture for the determination of the highest r−g+1 coefficients
of the Hilbert polynomial of the powers of an ideal I of codimension g that is locally complete
intersection up to codimension r :
HP coef.
R/I
I/I 2
I 2 /I 3
I 3 /I 4
I 4 /I 5
..
.
···
···
···
···
···
·
·
g
·
·
·
·
r
·
·
0
0
0
0
0
..
.
0
0
0
0
0
..
.
..
.
..
.
..
.
..
.
..
.
..
.
?
?
?
?
?
..
.
?
?
?
?
?
..
.
···
···
···
···
···
: needed as input.
: may be computed from the others.
? : not concerned.
2.2 Hilbert polynomials of powers of an ideal
We will treat the example of an equidimensional locally complete intersection threefold in
projective n-space.
Let us abreviate ei (p) = ei (I p /I p+1 ). Theorem 2.2 asserts that all the coefficients ei (p) are
determined by six of them.
With the help of a computer algebra system, one gets the following formulas.
Formulas 2.5
g+p−1
e0 (p) =
e0 (0),
p
g+p−1
(g + 2p) g + p
e1 (p) = g
e0 (0) +
e1 (0),
p−1
(g + p)
p
g+p−2
g(g + 1) g + p − 1
e0 (0) + (g + 1)
e1 (0)
e2 (p) =
2
p−2
p−2
g+p
(p − 1)g g + p + 1
e2 (0) +
−
e2 (1),
(g + p)
p
p−1
g(g + 1)(g + 2) g + p − 1
(g + 1)(g + 2) g + p − 1
e3 (p) =
e0 (0) −
e1 (0)
6
2
p−3
p−2
g+p
p(p − 1)(g + 2) g + p + 1
−
e2 (0) + (g + 2)
e2 (1)
(g + p)
p−2
p−2
(g + 2p) g + p
(p − 1)(g + 2p) g + p + 1
e3 (0) +
e3 (1).
+
(g + p)
(g + 2) p − 1
p
14
Notice that these formulas remains valid for the case of any scheme X ⊆ Pn that is locally a
complete intersection in dimension > dim X − 3.
As a first guess, one may hope that, at least with some strong hypotheses on X, the Hilbert
polynomial of the powers are determined by the Hilbert polynomial of X. This is even not true for
complete intersections, due to the following computation.
Suppose that X is a global complete intersection of codimension g and denote by σ1 , . . . , σg
the symmetric functions on the degrees of the defining equations of X. Setting
α1 = σ1 − g,
α2 = σ12 − 2σ2 − g,
α3 = σ13 − 3σ1 σ2 + 3σ3 − g,
one gets,
e0 (0) = σg ,
σg
e1 (0) =
α1 ,
2
σg
(3α21 − 6α1 + α2 ),
e2 (0) =
24
σg 3
(α − 6α21 + 8α1 + α1 α2 − 2α2 ).
e3 (0) =
48 1
Notice that these formulas imply that e3 (0), the fourth coefficient of the Hilbert polynomial, is a
rational function of the first three :
Remark 2.6 If ei denotes the i-th coefficient of the Hilbert polynomial of a global complete
intersection of dimension at least 3 in a projective space, then
e3 = e2 −
e1 e2
e1
e2
e3
+
− 1 + 12 .
e0
6
2e0
3e0
Now, using the expansion
td1 + · · · + tdg = g + (g + α1 )(t − 1) + (α2 − α1 )
(t − 1)2
(t − 1)3
+ (α3 − 3α2 + 2α1 )
+ ···,
2
6
one can compute the coefficients ei (1) for 1 6 i 6 3. The only place where α3 appears is in
σ
e3 (1) = 6g α3 + · · ·.
If one chooses two collections of degrees such that the first, second and 4-th symmetric functions are equal but the third one differs, one gets an example of two complete intersections of
dimesion three in P7 having the same Hilbert polynomials (but distinct Hilbert functions !) such
that the constant term of the Hilbert polynomials of their conormal bundle are distinct. Such examples were given to us by Benjamin de Weger, the two “smallest” ones are (1,6,7,22)-(2,2,11,21)
and (2,6,7,15)-(3,3,10,14). He also gave an infinite collection of them, and Noam Elkies gave a
rational parametrization of all the solutions (after a linear change of coordinates the solutions are
parametrized by a quadric in P5 ).
15
2.3 The degree of the residual
Let us suppose that the projective scheme X of dimension D is locally a complete intersection
in codimension at most s, and use our formulas and the above computations to derive the degree
of the codimension s part of an s-residual intersection.
We will treat the cases where δ = s − g is less or equal to three. As before σi stands for the i-th
Pp
symmetric function on d1 , . . . , ds . We will also set, to simplify some formulas, e′i (p) = j=0 ei (j),
p
which is the i-th coefficient of the Hilbert polynomial R/IX
.
• If δ = 0, eD
0 (R/R) = σs − e0 (0) (Bézout).
• If δ = 1, e0D−1 (R/R) = σs − (σ1 − g)e0 (0) + 2e1 (0), as proved by Stückrad in [18].
• If δ = 2, using the Taylor expansion of σ1 (td1 , . . . , tds ),
(t − 1)2
2
(t − 1)3
+ ···,
+ (σ13 − 3σ12 + 2σ1 − 3σ2 (σ1 − 2) + 3σ3 )
6
σ1 (td1 , . . . , tds ) = s + σ1 (t − 1) + (σ12 − σ1 − 2σ2 )
one recovers the formula given by Huneke and Martin in [16],
e0D−2 (R/R)
g+1
= σs − σ3 − gσ2 +
e0 (0) + (2σ1 − (g + 1))e1 (0) + (g + 1)e2 (0) − e′2 (1).
1
• If δ = 3, using the following Taylor expansion,
s(s − 1)
(t − 1)2
+ (s − 1)σ1 (t − 1) + ((s − 1)(σ12 − σ1 − 2σ2 ) + 2σ2 )
2
2
(t − 1)3
+ ···
+ ((s − 1)(σ13 − 3σ12 + 2σ1 ) − 3(s − 2)σ2 (σ1 − 2) + 3(s − 4)σ3 )
6
σ2 (td1 , . . . , tds ) =
one gets from Theorem 1.4,
e0D−3 (R/R)
g+1
g+2
= σs − σ3 − gσ2 +
σ1 −
e0 (0) − (2σ2 − (g + 1)σ1 )e1 (0)
2
3
+ ((g + 1)σ1 − (g + 2)(g + 3))e2 (0) − (σ1 − (g + 2))e′2 (1) + 2(g + 1)e3 (0) + 2e′3 (1).
3. Applications to secant varieties
Theorem 3.1. Let k be a perfect field, X ⊂ PN
k an eqidimensional subscheme of dimension
two with at most isolated licci Gorenstein singularities, A its homogeneous coordinate ring, ω := ωA
the canonical module, and Ω := ΩA/k the module of differentials.
(a) One has
e0 (A)2 + 14e0 (A) − 16e1 (A) + 4e2 (A) > e2 (ω ⊗A ω) + e2 (Ω).
16
(b) In case the singularities of X have embedding codimension at most two, then equality holds
in (a) if and only if the secant variety of X is deficient, i.e.
dim Sec(X) < 5.
Proof. We may assume that k is infinite. We define the ring R and the R-ideal I via the exact
sequence
mult
0 −→ I −→ R := A ⊗k A −→ A −→ 0.
Recall that Ω ≃ I/I 2 . The ring R is a standard graded k-algebra of dimension 6 with ωR = ω ⊗k ω
and codim N G(R) > 2. The ideal I has height 3, and is generated by linear forms. Moreover, I
satisfies G5 and, in the setting of (b), even G6 . Indeed, for any p ∈ V (I), one has µ(Ip ) = µ(Ωp ) 6
ecodim(Ap ) + dim A 6 dim Rp if dim Rp 6 4 or, in the setting of (b), dim Rp 6 5. In addition, for
every p ∈ V (I) with dim Rp 6 5, we have depth (I/I 2 )p = depth Ωp > dim Ap − 1. To see this, we
write Ap ≃ S/B with S a regular local ring and B a licci Gorenstein ideal. This is possible because
Ap is licci and Gorenstein. By [4, 6.2.11 and 6.2.12] the module B/B2 is Cohen-Macaulay. Thus
the natural complex
0 → B/B2 → ΩS/k ⊗S Ap ≃ ⊕Ap → ΩAp /k ≃ Ωp → 0
is exact and shows that depth Ωp > dim Ap − 1.
Now let A be an R-ideal generated by 5 general linear forms in I. Notice that the five general
linear forms in I that generate A are a weak R-regular sequence off V (I), hence off V (A). Since
codim N G(R) > 0 and codim N G(R) ∩ V (I) > 5 it follows that codim N G(R) ∩ V (A) > 5. By [2,
1.4] one has ht (A : I) > 5 as I satisfies G5 , and ht (I + (A : I)) > 6 in (b) as I is G6 . Thus in the
setting of (b), [19] shows that ht (A : I) > 6 if and only if the analytic spread ℓ(I) is at most 5.
On the other hand dim Sec(X) = ℓ(I) − 1 according to [17]. Hence dim Sec(X) < 5 if and only if
e10 (I/A) = 0.
We now apply Theorem 1.9(a) with r = s = 5 and g = 3. The theorem yields
(1)
e10 (I/A) = e0 (R) − e2 (A) −
P2
j=1
P2
j+k 5
k=0 (−1)
j+3
6+j−k
2−k
ek (ωR /I j ωR ).
Thus the present theorem follows once we have shown that the right hand side of (1) equals
(2)
e0 (A)2 + 14e0 (A) − 16e1 (A) + 4e2 (A) − e2 (ω ⊗A ω) − e2 (Ω).
From [5, IX 2.1] one obtains the isomorphisms of R-modules
ωR /IωR ∼
= ωR ⊗R R/I ∼
= (ω ⊗k ω) ⊗A⊗k A A ∼
= ω ⊗A (ω ⊗A A) ∼
= ω ⊗A ω
and, since codim N G(R) ∩ V (I) > 5,
IωR /I 2 ωR ∼
= (ω ⊗k ω) ⊗A⊗k A Ω ∼
= ωR ⊗R I/I 2 ∼
= ω ⊗A (ω ⊗A Ω) ∼
= (ω ⊗A ω) ⊗A Ω.
5
Therefore the right hand side of (1) becomes
(3)
e0 (A)2 − 7e0 (A) − e2 (A) − 23e1 (ω ⊗2 ) + 4e2 (ω ⊗2 ) + 7e1 (ω ⊗2 ⊗ Ω) − e2 (ω ⊗2 ⊗ Ω).
17
Here and in what follows tensor products are taken over the ring A.
We are now going to express the Hilbert coefficients e1 (ω ⊗2 ), e1 (ω ⊗2 ⊗ Ω) and e2 (ω ⊗2 ⊗ Ω) in
terms of the Hilbert coefficients of A, e2 (ω ⊗2 ) and e2 (Ω). First notice that for any finitely generated
graded A-module M ,
(4)
ei (M (−1)) = ei (M ) + ei−1 (M ).
Hence, by Lemma 1.7 and Remark 1.8,
(5)
e1 (ω) = 3e0 (A) − e1 (A) and
e2 (ω) = 3e0 (A) − 2e1 (A) + e2 (A).
Since ω is free of rank 1 locally in codimension 1, there is a complex of graded A-modules
(6)
0 −→ Z −→ A(−a)2 −→ ω −→ 0
for some a ≫ 0, that is exact in codimension 1. It induces complexes
(7)
0 −→ Z ⊗ A(−ja + a)j −→ A(−ja)j+1 −→ Symj (ω) ∼
= ω ⊗j −→ 0
1
that are likewise exact locally in codimension 1. Now (6) yields e1 (Z) = 2e1 (A(−a)) − e1 (ω) and
then (4), (5) and (7) show that for every j > 0,
(8)
e1 (ω ⊗j ) = 3je0 (A) − (2j − 1)e1 (A).
Now, we treat the first Hilbert coefficient of ω ⊗2 ⊗ Ω. Since ω is free of rank 1 locally in
∼
codimension 2, we also have ω ∗ −→ HomA (ω ⊗2 , ω), which by Lemma 1.7 and Remark 1.8 gives
2
(9)
e1 (ω ∗ ) = 3e0 (ω ⊗2 ) − e1 (ω ⊗2 ) and e2 (ω ∗ ) = 3e0 (ω ⊗2 ) − 2e1 (ω ⊗2 ) + e2 (ω ⊗2 ).
Furthermore,
(10)
∼
∼
2
2
ω ⊗2 ⊗ ω ∗ −→ HomA (ω, ω ⊗2 ) ←− ω.
As Ω is free of rank 3 locally in codimension 1, there is an exact sequence of graded A-modules
0 −→ A(−1)2 −→ Ω −→ C −→ 0
where C is free of rank 1 locally in codimension 1. Thus
C⊗
which gives a complex
2
^
∼
2
(A(−1) ) −→
1
3
^
∼
Ω −→ ω,
1
0 −→ A(−1)2 −→ Ω −→ ω(2) −→ 0
that is exact in codimension 1. Tensoring with ω ⊗2 we obtain
0 −→ ω ⊗2 (−1)2 −→ ω ⊗2 ⊗ Ω −→ ω ⊗3 (2) −→ 0.
18
Since this complex is exact in codimension 1, (4) and (8) imply that
(13)
e1 (ω ⊗2 ⊗ Ω) = 21e0 (A) − 11e1 (A).
Next, we turn to the second Hilbert coefficient of ω ⊗2 ⊗ Ω. Write —∗ := HomA (—, A) and
e := N − 2. Increasing N if needed, we may assume e > 2. We define a graded A-module E via
the exact sequence
(14)
0 −→ E −→ A(−1)e+3 −→ Ω −→ 0.
Notice that E has rank e, and is free locally in codimension 1 and Cohen-Macaulay locally in
codimension 2. Futhermore (14) gives
(15)
V3
Ve
∼
∼
( E)∗∗ −→ ( Ω)∗ (−e − 3) ←− ω ∗ (−e − 3).
2
2
As E ∗ is free locally in codimension 1 and rk E ∗ − 1 > 1, there exists a homogeneous element
f ∈ E ∗ of degree c ≫ 0 whose order ideal (E ∗ )∗ (f ) has height at least 2 (see [8]). However,
the ideals E ∗∗ (f ) and J := f (E) coincide locally in codimension 1 since E is reflexive locally in
codimension 1. Hence ht J = ht E ∗∗ (f ) > 2. The map f induces an exact sequence of graded
A-modules
0 −→ Ee−1 −→ Ee := E −→ Je (ce ) := J(c) −→ 0.
Repeating this procedure, if needed, we obtain a filtration
(16)
E1 ⊂ E2 ⊂ · · · ⊂ Ee
with
Ei /Ei−1 ∼
= Ji (ci ),
where Ji are homogeneous A-ideals of height at least 2. Thus Ei has rank i, is free in codimension
1 and Cohen-Macaulay in codimension 2, and
i−1
^
(
∼
i−1
^
Ei−1 )∗∗ (ci ) ←− (
2
i
^
∼
Ei−1 )∗∗ ⊗ (Ji (ci ))∗∗ −→ ( Ei )∗∗ .
2
Since E1 is reflexive locally in codimension 2, it follows that
e
e
X
^
∼
E1 −→ E1∗∗ ∼
ci ),
= ( E)∗∗ (−
2
2
i=2
which together with (15) implies
(17)
Pe
E1 ∼
= ω ∗ (−e − 3 − i=2 ci ).
2
The exact sequence
0 −→ Ji (ci ) −→ A(ci ) −→ (A/Ji )(ci ) −→ 0
yields a complex
0 −→ ω ⊗2 ⊗ Ji (ci ) −→ ω ⊗2 (ci ) −→ (ω ⊗2 /ω ⊗2 Ji )(ci ) −→ 0
19
that is exact in codimension 2. Since ht Ji > 2 and ω ⊗2 is free of rank 1 locally in codimension 2,
it follows that e10 ((A/Ji )(ci )) = e10 ((ω ⊗2 /ω ⊗2 Ji )(ci )). We conclude
(18)
e2 (Ji (ci )) − e2 (ω ⊗2 ⊗ Ji (ci )) = e2 (A(ci )) − e2 (ω ⊗2 (ci )).
Tensoring (14) and (16) with ω ⊗2 and using (18) and (17) we obtain
e2 (ω ⊗2 ⊗ Ω) − e2 (Ω) = e2 (ω ⊗2 (−1)e+3 ) − e2 (A(−1)e+3 ) +
e
X
(e2 (A(ci )) − e2 (ω ⊗2 (ci )))
i=2
(19)
+ e2 (ω ∗ (−e − 3 −
e
X
ci )) − e2 (ω 2 ⊗ ω ∗ (−e − 3 −
i=2
e
X
ci )).
i=2
Combining (19) with (4), (8), (9), (10), (5), we deduce
(20)
e2 (ω ⊗2 ⊗ Ω) = −12e0 (A) + 8e1 (A) − 5e2 (A) + 5e2 (ω ⊗2 ) + e2 (Ω).
Substituting (8), (13), (20) into (3), we conclude that (3) and (2) coincide.
Corollary 3.2. Let k be a perfect field, X ⊂ P4k an equidimensional subscheme of dimension
two with at most isolated Gorenstein singularities and A its homogeneous coordinate ring. One
has
e0 (A)2 + 14e0 (A) − 16e1 (A) + 4e2 (A) = e2 (ωA ⊗A ωA ) + e2 (ΩA/k ).
Remark. The inequality in Theorem 3.1 can be replaced by
∗
e0 (A)2 + 5e0 (A) − 10e1 (A) + 4e2 (A) > e2 (ωA
) + e2 (ΩA/k ).
Proof. Use equalities (9) and (8) in the proof of Theorem 3.1.
Corollary 3.3. Let k be a field, X ⊂ PN
k an equidimensional smooth subscheme of dimension
two, H the class of the hyperplane section, K the canonical divisor, and c2 the second Chern class
of the cotangent bundle of X. One has
(H 2 )2 > 10H 2 + 5HK + K 2 − c2 ,
and equality holds if and only if dim Sec(X) < 5.
Proof. The Riemannn-Roch theorem in dimension two gives
χ(X, E) =
1
[c1 (E)2 − 2c2 (E) − c1 (E)KX ] + rk Eχ(X, OX ).
2
If D is a divisor this equality specializes to
1 2 2
1
1
H n + (DH − KH)n + (D 2 − KD) + χ(X, OX )
2
2
2
n+1
n
+
2
1
2
2
=H
− (3H + KH − 2DH)
1
2
2
1 2
+ (H + KH − 2DH − KD + D 2 ) + χ(X, OX ).
2
χ(D + nH) =
20
For a rank two vector bundle E the formula reads
1
χ(E + nH) = H 2 n2 + (c1 (E)H − KH)n + (c1 (E)2 − Kc1 (E)) − c2 (E) + 2χ(X, OX )
2
n
+
2
n+1
2
2
= 2H
− (3H + KH − c1 (E)H)
2
1
1
1
+ H 2 + KH − c1 (E)H − Kc1 (E) + c1 (E)2 − c2 (E) + 2χ(X, OX ).
2
2
Taking D = 0 we obtain
e0 (OX ) = H 2
3
e1 (OX ) = H 2 +
2
1
e2 (OX ) = H 2 +
2
and for D = 2K
⊗2
e0 (ωX
) = H2
3
⊗2
e1 (ωX
) = H2 −
2
1
⊗2
e2 (ωX
) = H2 −
2
1
KH
2
1
KH + χ(X, OX ),
2
3
KH
2
3
KH + K 2 + χ(X, OX ).
2
Finally, taking E = ΩX , the cotangent sheaf of X, and using the fact that c1 (ΩX ) = K we deduce
e0 (ΩX ) = 2H 2
e1 (ΩX ) = 3H 2
e2 (ΩX ) = H 2 − c2 (ΩX ) + 2χ(X, OX ).
Now the assertion of the remark follows from the Theorem since ei (A) = ei (OX ), ei (ω) =
ei (ωX ) and ei (Ω) = ei (ΩX ) + ei (OX ).
Theorem 3.4. Let k be a field, X ⊂ PN
k an equidimensional smooth subscheme of dimension
three, H the class of the hyperplane section, K the canonical divisor, and c2 and c3 the second and
third Chern class of the cotangent bundle of X. One has
(H 3 )2 6 35H 3 − 11H 2 K − 9K 2 H + c2 H − K 3 −
1
1
Kc2 + c3 ,
12
2
and equality holds if and only if dim Sec(X) < 7.
Proof. We use the notation of Theorem 3.1, taking A to be generated by 7 general linear forms
in the ideal I of the diagonal. Recall that dim Sec(X) < 7 if and only if ht(A : I) > 8.
We will apply Theorem 1.9 (a) with r = s = 7 and g = 4. Because X is smooth we have
IωR /I 2 ωR ∼
= ω2 ⊗ Ω
7
I ωR /I ωR ∼
= ω 2 ⊗ S2 Ω.
2
3
7
21
As X is smooth, we can apply the Riemann-Roch formula as in the proof of Corollary 3.3,
to derive the following formulas, which express the Hilbert coefficients used in Theorem 1.9(a) in
terms of the numbers that appear in our desired formula:
e0 (A) = H 3
e1 (A) = 2H 3 − 23 KH 2
1
e2 (A) = 12
(14H 3 + 9KH 2 + K 2 H + c2 H)
1
(4H 3 + 6KH 2 + 2K 2 H + 2c2 H + Kc2 )
e3 (A) = 24
e0 (Ω) = 4H 3
e1 (Ω) = 8H 3 + KH 2
e2 (Ω) = 61 (28H 3 + 9KH 2 + 2K 2 H − 4c2 H)
1
(8H 3 + 6KH 2 + 4K 2 H − 8c2 H + Kc2 − 6c3 )
e3 (Ω) = 12
e0 (ω ⊗2 ) = H 3
e1 (ω ⊗2 ) = 2H 3 − 32 KH 2
1
(14H 3 − 27KH 2 + 13K 2 H + c2 H)
e2 (ω ⊗2 ) = 12
1
e3 (ω ⊗2 ) = 24
(4H 3 − 18KH 2 + 26K 2 H + 2c2 H − 12K 3 − 3Kc2 )
e0 (ω ⊗2 ⊗ Ω) = 4H 3
e1 (ω ⊗2 ⊗ Ω) = 8H 3 − 7KH 2
e2 (ω ⊗2 ⊗ Ω) = 61 (28H 3 − 63KH 2 + 38K 2 H − 4c2 H)
1
e3 (ω ⊗2 ⊗ Ω) = 12
(8H 3 − 42KH 2 + 76K 2 H − 8c2 H − 48K 3 + 17Kc2 − 6c3 )
e0 (ω ⊗2 ⊗ S2 Ω) = 10H 3
e1 (ω ⊗2 ⊗ S2 Ω) = 20H 3 − 19KH 2
e2 (ω ⊗2 ⊗ S2 Ω) = 61 (70H 3 − 171KH 2 + 119K 2 H − 25c2 H)
1
(20H 3 − 114KH 2 + 238K 2 H − 50c2 H − 186K 3 + 125Kc2 − 42c3 )
e3 (ω ⊗2 ⊗ S2 Ω) = 12
Corollary 3.5. Let k be a perfect field, X ⊂ PN
k an equidimensional smooth subscheme of
dimension three, A its homogeneous coordinate ring, ω := ωA the canonical module, and Ω := ΩA/k
the module of differentials. One has
e0 (A)2 + 391e0 (A) − 246e1 (A) + 66e2 (A) + 50e3 (A) > 18e2 (ω ⊗A ω) − 2e3 (Ω) − 2e3 (ω ⊗A ω)
and equality holds if and only if dim Sec(X) < 7.
Proof. Use the formulas given in Theorem 3.4 to express all the necessary Hilbert coefficients
in terms of e0 , . . . , e3 .
Remark 3.6. If H is the class of the hyperplane section, K the canonical divisor, D = H 3
the degree of X, the above inequality is equivalent to:
D 2 > 7(5D + 3KH 2 + K 2 H − c2 (ΩX )H) − 2c2 (ΩX )K + K 3 + c3 (ΩX ).
22
In other words, if C and S are respectively a curve and a surface obtained by taking general linear
sections of X, the formula reads
3
D 2 > 7(5D + 3χC + 12χ(OS ) − 2χS ) − 48χ(OX ) + KX
+ χX .
References
[1] R. Apéry, Sur les courbes de première espèce de l’espace à trois dimensions, C. R. Acad.
Sci. Paris, t. 220, Sér. I, p. 271–272, 1945.
[2] M. Artin and M. Nagata, Residual intersections in Cohen-Macaulay rings, J. Math. Kyoto
Univ. 12 (1972), 307–323.
[3] L. Avramov and J. Herzog, The Koszul algebra of a codimension 2 embedding, Math. Z.
175 (1980), 249–280.
4] R. Buchweitz, Contributions à la théorie des singularités, These d’Etat, Université Paris 7,
1981. Available at https://tspace.library.utoronto.ca/handle/1807/16684.
[5] H. Cartan and S. Eilenberg, Homological Algebra, Princeton University Press, Princeton,
NJ.
[6] M. Chardin, D. Eisenbud, and B. Ulrich Hilbert functions and residually S2 ideals, Compositio 125 (2001), 193–219.
[7] C. Cumming, Residual intersections in Cohen-Macaulay rings. J. Algebra 308 (2007), no.
1, 91–106.
[7a] M. Dale, Severi’s theorem on the Veronese-surface. J. London Math. Soc. 32 (1985)
419–425.
[8] D. Eisenbud and E. G. Evans, Generating modules efficiently: theorems from algebraic
K-theory. J. Algebra (1973), 278–305.
[9] F. Gaeta, Quelques progrès récents dans la classification des variétés algébriques d’un espace
projectif, Deuxième Colloque de Géometrie Algébrique, Liège, 1952.
[10] S. H. Hassanzadeh, Cohen-Macaulay residual intersections and their Castelnuovo-Mumford
regularity. Trans. Amer. Math. Soc. 364 (2012) 6371–6394.
[11] J. Herzog, Komplexe, Auflösungen und Dualität in der lokalen Algebra. Habilitationsschrift (1974).
[12] J. Herzog, A. Simis, and W.V. Vasconcelos, Koszul homology and blowing-up rings, in
Commutative Algebra, eds. S. Greco and G. Valla, Lecture Notes in Pure and Appl. Math. 84,
Marcel Dekker, New York, 1983, 79–169.
[13] J. Herzog, W.V. Vasconcelos, and R. Villarreal, Ideals with sliding depth, Nagoya Math.
J. 99 (1985), 159–172.
23
[14] C. Huneke, Linkage and Koszul homology of ideals, Amer. J. Math. 104 (1982), 1043–
1062.
[15] C. Huneke, Strongly Cohen-Macaulay schemes and residual intersections, Trans. Amer.
Math. Soc. 277 (1983), 739–763.
[16] C. Huneke and H. Martin, Residual Intersection and the Number of Equations Defining
Projective Varieties, Comm. Algebra 23 (1995), no. 6, 2345–2376.
[17] A. Simis and B. Ulrich, On the ideal of an embedded join. J. Algebra 226 (2000), no. 1,
114.
[18] J. Stückrad, On quasi-complete intersections, Arch. Math. 58 (1992), 529–538.
[19] B. Ulrich, Remarks on residual intersections, in Free Resolutions in Commutative Algebra
and Algebraic Geometry, Sundance 1990, eds. D. Eisenbud and C. Huneke, Res. Notes in Math.
2, Jones and Bartlett Publishers, Boston-London, 1992, 133–138.
[20] B. Ulrich, Artin-Nagata properties and reductions of ideals, Contemp. Math. 159 (1994),
373–400.
[21] J. Watanabe, A note on Gorenstein rings of embedding codimension three, Nagoya Math.
J. 50 (1973), 227–232.
Marc Chardin, Institut de Mathématiques, CNRS & Université Pierre et Marie Curie
[email protected]
David Eisenbud, Department of Mathematics, University of California, Berkeley
[email protected]
Bernd Ulrich, Department of Mathematics, Purdue University,
[email protected]
24
| 0math.AC
|
Under review as a conference paper at ICLR 2018
A C LASSIFICATION –BASED P ERSPECTIVE ON
GAN D ISTRIBUTIONS
arXiv:1711.00970v3 [cs.LG] 19 Nov 2017
Shibani Santurkar, Ludwig Schmidt, Aleksander Madry
˛
Massachusetts Institute of Technology
{shibani,ludwigs,madry}@mit.edu
A BSTRACT
A fundamental, and still largely unanswered, question in the context of Generative Adversarial Networks (GANs) is whether GANs are actually able to capture
the key characteristics of the datasets they are trained on. The current approaches
to examining this issue require significant human supervision, such as visual inspection of sampled images, and often offer only fairly limited scalability. In this
paper, we propose new techniques that employ a classification–based perspective
to evaluate synthetic GAN distributions and their capability to accurately reflect
the essential properties of the training data. These techniques require only minimal human supervision and can easily be scaled and adapted to evaluate a variety
of state-of-the-art GANs on large, popular datasets. Our analysis indicates that
GANs have significant problems in reproducing the more distributional properties
of the training dataset. In particular, when seen through the lens of classification,
the diversity of GAN data is orders of magnitude less than that of the original data.
1
I NTRODUCTION
Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) have garnered a significant
amount of attention due to their ability to learn generative models of multiple natural image datasets
(Radford et al., 2015; Denton et al., 2015; Zhang et al., 2016). Since their conception, a fundamental
question regarding GANs is to what extent they truly learn the underlying data distribution. This
is a key issue for multiple reasons. From a scientific perspective, understanding the capabilities of
common GANs can shed light on what precisely the adversarial training setup allows the GAN to
learn. From an engineering standpoint, it is important to grasp the power and limitations of the
GAN framework when applying it in concrete applications. Due to the broad potential applicability
of GANs, researchers have investigated this question in a variety of ways.
When we evaluate the quality of a GAN, an obvious first check is to establish that the generated
samples lie in the support of the true distribution. In the case of images, this corresponds to checking
if the generated samples look realistic. Indeed, visual inspection of generated images is currently
the most common way of assessing the quality of a given GAN. Individual humans can performs
this task quickly and reliably, and various GANs have achieved impressive results for generating
realistic-looking images of faces and indoor scenes (Salimans et al., 2016; Denton et al., 2015).
Once we have established that GANs produce realistic-looking images, the next concern is that the
GAN might simply be memorizing the training dataset. While this hypothesis cannot be ruled out
entirely, there is evidence that GANs perform at least some non-trivial modeling of the unknown
distribution. Previous studies show that interpolations in the latent space of the generator produce
novel and meaningful image variations (Radford et al., 2015), and that there is a clear disparity
between generated samples and their nearest neighbors in the true dataset (Arora & Zhang, 2017).
Taken together, these results provide evidence that GANs could constitute successful distribution
learning algorithms, which motivates studying their distributions in more detail. The direct approach
is to compare the probability density assigned by the generator with estimates of the true distribution
(Wu et al., 2016). However, in the context of GANs for high-dimensional image distributions, this
is complicated by two factors. First, GANs do not naturally provide probability estimates for their
samples. Second, estimating the probability density of the true distribution is a challenging problem
1
Under review as a conference paper at ICLR 2018
itself (the adversarial training framework specifically avoids this issue). Hence prior work has only
investigated the probability density of GANs on simple datasets such as MNIST (Wu et al., 2016).
Since reliably computing probability densities in high dimensions is challenging, we can instead
study the behavior of GANs in low-dimensional problems such as two-dimensional Gaussian mixtures. Here, a common failure of GANs is mode collapse, wherein the generator assigns a disproportionately large mass to a subset of modes from the true distribution (Goodfellow, 2016). This raises
concerns about a lack of diversity in the synthetic GAN distributions, and recent work shows that
the learned distributions of two common GANs indeed have (moderately) low support size for the
CelebA dataset (Arora & Zhang, 2017). However, the approach of Arora & Zhang (2017) heavily
relies on a human annotator in order to identify duplicates. Hence it does not easily scale to comparing many variants of GANs or asking more fine-grained questions than collision statistics. Overall,
our understanding of synthetic GAN distributions remains blurry, largely due to the lack of versatile
tools for a quantitative evaluation of GANs in realistic settings. The focus of this work is precisly to
address this question:
Can we develop principled and quantitative approaches to study synthetic GAN distributions?
To this end, we propose two new evaluation techniques for GAN distributions. Our methods are
inspired by the idea of comparing moments of distributions, which is at the heart of many methods in classical statistics. Although simple moments of high-dimensional distributions are often not
semantically meaningful, we can extend this idea to distributions of realistic images by leveraging
image statistics identified using convolutional neural networks. In particular, we train image classifiers in order to construct test functions corresponding to semantically meaningful properties of the
distributions. An important feature of our approach is that it requires only light human supervision
and can easily be scaled to evaluating many GANs and large synthetic datasets.
Using our new evaluation techniques, we study five state-of-the-art GANs on the CelebA and LSUN
datasets, arguably the two most common testbeds for advanced GANs. We find that most of the
GANs significantly distort the relative frequency of even basic image attributes, such as the hair
style of a person or the type of room in an indoor scene. This clearly indicates a mismatch between
the true and synthetic distributions. Moreover, we conduct experiments to explore the diversity of
GAN distributions. We use synthetic GAN data to train image classifiers and find that these have
significantly lower accuracy than classifiers trained on the true data set. This points towards a lack
of diversity in the GAN data, and again towards a discrepancy between the true and synthetic distributions. In fact, our additional examinations show that the diversity in GANs is only comparable to
a subset of the true data that is 100× smaller.
2
U NDERSTANDING GAN S THROUGH THE L ENS OF C LASSIFICATION
When comparing two distributions, a common first test is to compute low-order moments such as
the mean and the variance. If the distributions are simple enough, these quantities provide a good
understanding for how similar they are. Moreover, low-order moments have a precise definition and
are usually quick to compute. On the other hand, low-order moments can also be misleading for
more complicated, high-dimensional distributions. As a concrete example, consider a generative
model of digits (such as MNIST). If a generator produces digits that are shifted by a significant
amount yet otherwise perfect, we will probably still consider this as a good approximation of the
true distribution. However, the expectation (mean moment) of the generator distribution can be very
different from the expectation of the true data distribution. This raises the question of what other
properties of high-dimensional image distributions are easy to test yet semantically meaningful.
In the next two subsections, we describe two concrete approaches to evaluate synthetic GAN data
that are easy to compute yet capture relevant information about the distribution. The common theme
is that we employ convolutional neural networks in order to capture properties of the distributions
that are hard to describe in a mathematically precise way, but usually well-defined for a human (e.g.,
what fraction of the images shows a smiling person?). Automating the process of annotating images
with such high-level information will allow us to study various aspects of synthetic GAN data.
2.1
Q UANTIFYING M ODE C OLLAPSE
Mode collapse refers to the tendency of the generator to concentrate a large probability mass on a
few modes of the true distribution. While there is ample evidence for the presence of mode-collapse
2
Under review as a conference paper at ICLR 2018
in GANs (Goodfellow, 2016; Arora & Zhang, 2017; Metz et al., 2016), elegant visualizations of this
phenomena are somewhat restricted to toy problems on low-dimensional distributions (Goodfellow,
2016; Metz et al., 2016). For image datasets, it is common to rely on human annotators and therein
derived heuristics (see Section 2.3). While these methods have their merits, they are restrictive both
in the scale and granularity of testing. Here we propose a classification-based tool to assess how
good GANs are at assigning the right mass across broad concepts/modes. To do so, we use a trained
classifier as an expert “annotator” that labels important features in synthetic data, and then analyze
the resulting distribution. Specifically, our goal is to investigate if a GAN trained on a well-balanced
dataset (i.e., contains equal number of samples from each class) can learn to reproduce this balanced
N
structure. Let D = (X, Y ) = {(xi , yi )}i=1 represent a dataset of size N with C classes, where
(xi , yi ) denote an image-label pair drawn from true data. If the dataset D is balanced, it contains
N/C images per class. The procedure for computing class distribution in synthetic data is:
1. Train an annotator (a multi-class classifier) using the dataset D.
2. Train an unconditional GAN on the images X from dataset D, without using class labels.
3. Create a synthetic dataset by sampling N images from a GAN and labeling them using the
annotator from Step 1.
The annotated data generated via the above procedure can provide insight into the GAN’s class
distribution at the scale of the entire dataset. Moreover, we can vary the granularity of mode analysis
by choosing richer classification tasks, i.e., more challenging classes or a larger number of them. In
Section 3.3, we use this technique to visualize mode collapse in several state-of-the-art GANs on the
CelebA and LSUN datasets. All the studied GANs show significant mode collapse and the effect
becomes more pronounced when the granularity of the annotator is increased (larger number of
classes). We also investigate the temporal aspect of the GAN setup and find that the dominant mode
varies widely over the course of training. Our approach also enables us to benchmark and compare
GANs on different datasets based on the extent of mode collapse in the learned distributions.
2.2
M EASURING DIVERSITY
Our above method for inspecting distribution of modes in synthetic data provides a coarse look at the
statistics of the underlying distribution. While the resulting quantities are semantically meaningful,
they capture only simple notions of diversity. To get a more holistic view on the sample diversity
in the synthetic distribution, we now describe a second classification-based approach for evaluating
GAN distributions. The main question that motivates it is: Can GANs recover the key aspects of
real data to enable training a good classifier? We believe that this is an interesting measure of sample diversity for two reasons. First, classification of high-dimensional image data is a challenging
problem, so a good training dataset will require a sufficiently diverse sample from the distribution.
Second, augmenting data for classification problems is one of the proposed use cases of GANs (e.g.,
see the recent work of Shrivastava et al. (2017)).
If GANs are truly able to capture the quality and diversity of the underlying data distribution, we
expect almost no gap between classifiers trained on true data and synthetic data from a GAN. A
generic method to produce data from GANs for classification is to train separate GANs for each
class in the dataset D.1 Samples from these class-wise GANs can then be pooled together to get a
labeled synthetic dataset. Note that the labels are trivially determined based on the class modeled
by the particular GAN from which a sample is drawn. We perform the following steps to assess the
classification performance of synthetic data vs. true data:
1. Train a classifier on the true data D (from Section 2.1) as a benchmark for comparison.
2. Train C separate unconditional GANs, one per class in dataset D.
3. Generate a balanced synthetic labeled dataset of size N by consolidating an equal number of
samples drawn from each of these C GANs. The labels obtained by aggregating samples from
per-class GANs are designated as “default" labels for the synthetic dataset. Note that by design,
both true and synthetic datasets have N samples, with N/C examples per class.
4. Use synthetic labeled data from Step 3 to train classifier with the same architecture as Step 1.
Comparing the classifiers from Steps 1 and 4 can then shed light on the disparity between the two
distributions. Radford et al. (2015) conducted an experiment similar to Step 2 on the MNIST dataset
1
We tried the alternate approach of using class-conditional GANs to get labeled datasets. This method did
not yield good samples since most common GANs have not been designed with a conditional structure in place.
3
Under review as a conference paper at ICLR 2018
using a conditional GAN. They found that samples from their DCGAN performed comparably to
true data on nearest neighbor classification. We obtained similar good results on MNIST, which
could be due to the efficacy of GANs in learning the MNIST distribution or due to the ease of
getting good accuracy on MNIST even with a small training set (Rolnick et al., 2017). To clarify
this question, we restrict our analysis to more complex datasets, specifically CelebA and LSUN.
We evaluate the two following properties in our classification task:
(i) How well can the GANs recover nuances of the decision boundary, which is reflected by how
easily the classifier can fit the training data?
(ii) How does the diversity of synthetic data compare to that of true data when measured by classification accuracy on a hold-out set of true data?
We observe that all the studied GANs have very low diversity in this metric. In particular, the
accuracy achieved by a classifier trained on GAN data is comparable only to the accuracy of a
classifier trained on a 100× (or more) subsampled version of the true dataset. Even if we draw more
samples from the GANs to produce a training set several times larger than the true dataset, there is
no improvement in performance. Looking at the classification accuracy gives us a way to compare
different models on a potential downstream application of GANs. Interestingly, we find that visual
quality of samples does not necessarily correlate with good classification performance.
2.3
R ELATED W ORK
In GAN literature, it is common to investigate performance using metrics that involve human supervision. Arora & Zhang (2017) proposed a measure based on manually counting duplicates in GAN
samples as a heuristic for the support or diversity of the learned distribution. In Wu et al. (2016),
manual classification of a small sample (100 images) of GAN generated MNIST images is used as
a test for the GAN is missing certain modes. Such annotator-based metrics have clear advantages
in identifying relevant failure-modes of synthetic samples, which explains why visual inspection
(eyeballing) is still the most popular approach to assess GAN samples.
There have also been various attempts to build good metrics for GANs that are not based on manual
heuristics. Parzen window estimation can be used to approximate the log-likelihood of the distribution, though it is known to work poorly for high-dimensional data (Theis et al., 2016). Wu et al.
(2016) develop a method to get a better estimate for log-likelihood using annealed importance sampling. Salimans et al. (2016) propose a metric known as Inception Score, where the entropy in the
labels predicted by a pre-trained Inception network is used to assess the diversity in GAN samples.
3
E XPERIMENTS
In the following sub-sections we describe the setup and results for our classification-based GAN
benchmarks. Additional details can be found in Section 5 in the Appendix.
3.1
DATASETS
GANs have shown promise in generating realistic samples, resulting in efforts to apply them to a
broad spectrum of datasets. However, the Large-scale CelebFaces Attributes (CelebA) (Liu et al.,
2015) and Large-Scale Scene Understanding (LSUN) (Yu et al., 2015) datasets remain the most
popular and canonical ones in developing and evaluating GAN variants. Conveniently, these datasets
also have rich annotations, making them particularly suited for our classification–based evaluations.
Details on the setup for classification tasks for these datasets are given in the Appendix.
3.2
M ODELS
Using our framework, we perform a comparative study of several popular variants of GANs:
1. Deep Convolutional GAN (DCGAN): Convolutional GAN trained using a Jensen–Shannon
divergence–based objective (Goodfellow et al., 2014; Radford et al., 2015).
2. Wasserstein GAN (WGAN): GAN that uses a Wasserstein distance–based objective (Arjovsky
et al., 2017).
3. Adversarially Learned Inference (ALI): GAN that uses joint adversarial training of generative
and inference networks (Dumoulin et al., 2017).
4
Under review as a conference paper at ICLR 2018
Female
+
Mouth Closed
True Data
DCGAN
WGAN
ALI
BEGAN
Improved GAN
Female
+
Mouth Open
Male
+
Mouth Closed
Male
+
Mouth Open
4-class dataset from CelebA for attributes Male, Mouth open.
Figure 1: Illustration of a dataset used in the proposed classification benchmarks. Shown alongside
are images sampled from various unconditional GANs trained on this dataset. Labels for the GAN
samples are obtained using a pre-trained classifier as an annotator.
4. Boundary Equilibrium GAN (BEGAN): Auto-encoder style GAN trained using Wasserstein distance objective (Berthelot et al., 2017).
5. Improved GAN (ImGAN): GAN that uses semi-supervised learning (labels are part of GAN
training), with various other architectural and procedural improvements (Salimans et al., 2016).
All the aforementioned GANs are unconditional, however, ImGAN has access to class labels as a
part of the semi-supervised training process. We use standard implementations for each of these
models, details of which are provided in the Appendix (Section 5). We also used the prescribed
hyper-parameter settings for each GAN, including number of iterations we train them for. Our
analysis is based on 64 × 64 samples, which is a size at which GAN generated samples tend to be of
high quality. We also use visual inspection to ascertain that the perceptual quality of GAN samples
in our experiments is comparable to those reported in previous studies. We demonstrate sample
images in Figures 1, 4 and 5. BEGAN did not converge in our experiments on the LSUN dataset
and hence is excluded from the corresponding analysis.
In our study, we use two types of classification models:
1. ResNet: 32-Layer Residual network He et al. (2016). This model is chosen as it is a standard
classifier in vision and yields high accuracy on various datasets, making it a reliable baseline.
2. Linear Model: This is a network with one-fully connected layer between the input and output (no
hidden layers) with a softmax non-linearity. If the dimensions of input x and output ŷ, are D and
C (number of classes) respectively, then linear models implement the function ŷ = σ(W T x+b),
where W is a D × C matrix, b is a C × 1 vector and σ(·) is the softmax function. Due to it’s
simplicity, this model will serve as a useful baseline in some of our experiments.
We always train the classifiers to convergence, with decaying learning rate and no data augmentation.
3.3
E XAMINATION OF M ODE C OLLAPSE
Experimental results for quantifying mode collapse through classification tasks, described in Section 2.1, are presented below. Table 1 gives details on datasets (subsets of CelebA and LSUN)
used in our analysis, such as size (N ), number of classes (C), and accuracy of the annotator, i.e.,
a classifier pre-trained on true data, which is then used to label the synthetic, GAN-generated data.
Figure 2 presents class distribution in the synthetic data, as determined by using these annotators.
The left panel compares the relative distribution of modes in true data (uniform) with that in various
GAN-generated datasets. Each of these datasets is created by drawing N samples from the GAN
after it was trained on the corresponding true dataset. The right panel illustrates the evolution of
class distributions in various GANs over the course of training2 .
Results: These visualization lead to the following findings:
2
Temporal evaluation of ALI class distribution is absent in the analysis due to absence of periodic checkpointing provisions in the code.
5
Under review as a conference paper at ICLR 2018
Dataset
CelebA: Makeup, Smiling
CelebA: Male, Mouth Open
CelebA: Bangs, Smiling
LSUN: Bedroom, Kitchen, Classroom
LSUN: Bedroom, Conference Room, Dining Room,
Kitchen, Living Room
N
102,436
115,660
45,196
150,000
250,000
C
4
4
4
3
5
Annotator’s Accuracy (%)
90.9, 92.4
97.9, 93.5
93.9, 92.4
98.7
93.7
Table 1: Details of CelebA and LSUN subsets used for the studies on mode collapse in Section 3.3.
Here, we use a classifier trained on true data as an annotator that let’s us infer label distribution for
the synthetic, GAN-generated data. N is the training set size and C is the number of classes for both
the true and synthetic datasets. Annotator’s accuracy refers to the accuracy of the classifier on a test
set of true data. For CelebA, we use a combination of attribute-wise binary classifiers as annotators
due their higher accuracy compared to a single classifier trained jointly on all the four classes.
CelebA: Makeup, Smiling
CelebA: Male, Mouth Open
CelebA: Bangs, Smiling
LSUN: Bedroom, Kitchen, Classroom
LSUN: Bedroom, Conference Room, Dining Room, Kitchen, Living Room
Figure 2: Visualizations of mode collapse in the synthetic, GAN-generated data produced after training on our chosen subsets of CelebA and LSUN datasets. Left panel shows the relative distribution
of classes in samples drawn from synthetic datasets extracted at the end of the training process, and
compares is to the true data distribution (leftmost plots). On the right, shown is the evolution of analogous class distribution for different GANs over the course of training. BEGAN did not converge
on the LSUN tasks and hence is excluded from the corresponding analysis.
6
Under review as a conference paper at ICLR 2018
• All GANs seem to suffer from significant mode-collapse. This becomes more apparent when the
annotator granularity is increased, by considering a larger set of classes. For instance, one should
compare the relatively balanced class distributions in the 3-class LSUN task to the near-absence
of some modes in the 5-class task.
• Mode collapse is prevalent in GANs throughout the training process, and does not seem to recede
over time. Instead the dominant mode(s) often fluctuate wildly over the course of the training.
• For each task, often there is a common set of modes onto which distinct GANs exhibit collapse.
In addition to viewing our method as an approach to analyze the mode collapse, we can also use it
as a benchmark for GAN comparison. From this perspective, we can observe that ALI consistently
shows lowest mode collapse amongst all the studied GANs. The behavior of the other GANs appears
to vary based on the dataset – on CelebA, DCGAN learns a somewhat balanced distributions, while
WGAN, BEGAN and ImGAN show prominent mode collapse. This is in contrast to the results
obtained LSUN, where, for example, WGAN exhibit relatively small mode collapse, while DCGAN
and ImGAN show significant mode collapse. This points to a general challenge in real world applications of GANs: they often perform well on the datasets they were designed for (e.g. WGAN on
LSUN), but extension to new datasets is not straightforward. Temporal analysis of mode-collapse
shows that there is wide variation in the dominant mode for WGAN and Improved GAN, whereas
for BEGAN, the same mode(s) often dominates the entire training process.
3.4
D IVERSITY E XPERIMENTS
Using the procedure outlined in Section 2.2, we perform a quantitative assessment of sample diversity in GANs on the CelebA and LSUN datasets. We restrict our experiments to binary classification
as we find they have sufficient complexity to highlight the disparity between true and synthetic data.
Results for classification-based evaluation of GANs are presented in Table 2 and Figure 3.
As a preliminary check, we inspect the quality of our labeled GAN datasets. For this, we use
high-accuracy annotators from Section 2.1 to predict labels for GAN generated data and measure
consistency between the predicted and default labels (label correctness). We also inspect confidence
scores, defined as the softmax probabilities for predicted class, of the annotator. The motivation
behind these metrics is that if the classifier can correctly and with high-confidence predict labels
for labeled GAN samples, then it is likely that they are convincing examples of that class, and
hence of good “quality". Empirical results for label agreement and annotator confidence of GAN
generated datasets are shown in Table 2, and Figure 6 in the Appendix. In Table 2, we also report
an equivalent Inception Score (Salimans et al., 2016), similar to that described in Section 2.3. Using
the Inception network to get the label distribution may not be meaningful for face or scene images.
Instead, we compute the Inception Score using the label distribution predicted from the annotator
networks. Score is computed as exp(Ex [KL(p(y|x))||p(y)]), where y refers to label predictions
from the annotators 3 .
Next, we train classifiers using the true and labeled GAN-generated datasets and study their performance in terms of accuracy on a hold-out set of true data. ResNets (and other deep variants) yield
good classification performance on true data, but suffer from severe overfitting on the synthetic data,
leading to poor test accuracy (see Table 2). This already indicates a possible problem with GANs
and the diversity of the data they generate. But to highlight this problem better and avoid the issues
that stem from overfitting, we also look for a classifier which does not always overfit on the synthetic
data. We, however, observed that even training simple networks, such as one fully connected layer
with few hidden units, led to overfitting on synthetic data. Hence, we resorted to a very basic linear
model described in Section 3.2. Table 2 shows results from binary classification experiments using
linear models, with the training and test accuracies of the classifier on various datasets.
Finally, to get a better understanding of the underlying "diversity" of synthetic datasets, we train
linear models using down-sampled versions of true data (no augmentation), and compare this to
the performance of synthetic data, as shown in Table 2. Down-sampling the data by a factor of
M , denoted as ↓M implies selecting a random N/M subset of the data D. Visualizations of how
GAN classification performance compares with (down-sampled) true data are in Figure 3. Here,
the bold curve shows test accuracy of a classifier trained on true data as a function of true dataset
3
Code from https://github.com/openai/improved-gan
7
Under review as a conference paper at ICLR 2018
Task
CelebA
Male (Y/N)
# Images: 136522
CelebA
Smiling (Y/N)
# Images: 156160
CelebA
Black Hair (Y/N)
# Images: 77812
LSUN
Bedroom/Kitchen
# Images: 200000
Data Source
True
True ↓64
True ↓256
True ↓512
True ↓1024
DCGAN
WGAN
ALI
BEGAN
Improved GAN
True
True ↓64
True ↓256
True ↓512
True ↓1024
DCGAN
WGAN
ALI
BEGAN
Improved GAN
True
True ↓64
True ↓256
True ↓512
True ↓1024
DCGAN
WGAN
ALI
BEGAN
Improved GAN
True
True↓512
True ↓1024
True ↓2048
True ↓4096
DCGAN
WGAN
ALI
Improved GAN
Classification Performance
Label
Accuracy (%)
Inception Score
Correctness
Linear model
(%)
(µ ± σ)
↑1
↑10
Train Test Train Test
88.1 88.8
89.6 88.7
97.9
1.98 ± 0.0033 91.6 86.9
96.3 83.8
100.0 83.1
98.2
1.97 ± 0.0013 100.0 79.2 100.0 79.6
98.3
1.97 ± 0.0013 96.7 84.0 96.7 83.9
99.2
1.99 ± 0.0008 95.8 86.7 95.8 86.7
99.3
1.99 ± 0.0006 97.9 78.0 98.0 78.2
99.8
1.99 ± 0.0004 100.0 75.6 100.0 71.0
85.7 85.6
87.6 85.0
92.4
1.69 ± 0.0074 91.5 82.4
93.7 80.0
95.0 76.2
96.1
1.67 ± 0.0028 100.0 77.1 100.0 77.1
98.2
1.68 ± 0.0031 96.8 83.4 96.8 83.5
93.3
1.71 ± 0.0027 94.5 80.1 95.0 82.4
93.5
1.74 ± 0.0028 98.5 69.5 98.5 69.6
98.4
1.88 ± 0.0021 100.0 70.2 100.0 70.1
76.4 76.5
79.7 75.4
84.5
1.68 ± 0.0112 86.3 72.6
89 68.7
100.0 65.4
86.7
1.68 ± 0.0040 100.0 70.9 100.0 70.5
76.0
1.60 ± 0.0055 94.4 73.7 94.3 73.4
79.4
1.63 ± 0.0028 94.9 71.0 94.9 70.2
87.6
1.74 ± 0.0028 94.1 67.6 94.1 67.7
86.7
1.64 ± 0.0045 100.0 70.3 100.0 69.1
64.7 64.1
64.7 64.0
98.2
1.94 ± 0.0217 65.2 64.0
98.7 56.2
100.0 55.1
92.7
1.85 ± 0.0036 90.8 56.5 91.2 56.3
87.8
1.70 ± 0.0023 86.2 58.2 96.3 54.1
94.1
1.86 ± 0.0021 82.5 60.0 82.0 59.7
84.2
1.68 ± 0.0030 91.6 55.9 90.8 56.5
ResNet
↑1
Test
97.9
92.9
89.8
82.6
81.4
56.4
50.0
58.9
55.4
71.7
92.4
87.8
82.1
77.8
71.2
63.3
65.3
55.8
64.1
61.6
84.5
80.0
75.8
73.9
72.7
53.4
58.5
55.7
67.2
70.2
99.1
76.4
66.9
56.5
55.1
51.2
55.7
56.2
51.2
Table 2: Results from a comparative study on classification performance of true data vs. GAN generated synthetic data for the CelebA and LSUN datasets. Label correctness measures the agreement
between default labels for the synthetic datasets, and those predicted by the annotator, a classifier
trained on the true data. Shown alongside are the equivalent inception scores computed using labels
predicted by the annotator (instead of the Inception Network). Training and test accuracies for a
linear model classifier on the various true and synthetic datasets are reported. Also presented are the
corresponding accuracies for a linear model trained on down-sampled true data (↓M ) and oversampled synthetic data (↑L ). Test accuracy for ResNets trained on these datasets is also shown (training
accuracy was always 100%), though it is noticeable that deep networks suffer from issues when
trained on synthetic datasets.
8
Under review as a conference paper at ICLR 2018
CelebA: Male/Female
CelebA: Smiling/Not Smiling
LSUN: Bedroom/Kitchen
CelebA: Black Hair/Not Black Hair
Figure 3: Illustration of the classification performance of true data vs. GAN-generated synthetic
data. Classification is performed using a basic linear model, and performance is reported in terms of
accuracy on the test set of true data. The bold curve shows variation of the classification performance
of models trained on true data with the size of the training set (maximum size is N ). Here, training
sets of various sizes are obtained by down-sampling the true dataset at random. Dashed lines show
the performance of classifiers trained on various GAN-generated datasets of size N .
size (maximum size being N ). The dashed curves show test accuracy of classifiers trained on GAN
datasets, obtained by drawing N samples from GANs at the culmination of the training process.
A natural argument in the defense of GANs is that we can oversample them, i.e. generate datasets
much larger than the size of training data. Results for linear models trained using a 10-fold oversampling of GANs (drawing 10N samples), denoted by ↑10 , are show in Table 2.
Results: The major findings from these experiments are:
• Based on Table 2, and Figure 6, we see strong agreement between annotator labels and true
labels for synthetic data, on par with the scores for the test set of true data. It is thus apparent that
the GAN images are of high-quality, as expected based on the visual inspection. These scores
are lower for LSUN than CelebA, potentially due to lower quality of generated LSUN images.
From these results, we can get a broad understanding of how good GANs are at producing
convincing/representative samples from different classes across datasets. This also shows that
simple classification-based benchmarks can highlight relevant properties of synthetic datasets.
• The equivalent inception score is not very informative and is similar for the true (test set) and
synthetic datasets. This is not surprising given the simple nature of our binary classification task
and the fact that the true and synthetic datasets have almost a uniform distribution over labels.
• It is evident from Table 2 that there is a large performance gap between true and synthetic data
on classification tasks. Inspection of training accuracies shows that linear models are able to
nearly fit the synthetic datasets, but are grossly underfitting on true data. Given the high scores
of synthetic data on the previous experiments to assess dataset ‘quality’ (Table 2 and Figure 6),
it is likely that the poor classification performance is more indicative of lack of ‘diversity’.
• Comparing GAN performance to that of down-sampled true data reveals that the learned distribution, which was trained on datasets that have around hundred thousand data points exhibits
diversity that is on par with what only mere couple of hundreds of true data samples constitute! This shows that, at least from the point of view of classification, the diversity of the GAN
generated data is severely lacking.
9
Under review as a conference paper at ICLR 2018
• Oversampling GANs by 10-fold to produce larger datasets does not improve classification performance. The disparity between true and synthetic data remains nearly unchanged even after
this significant oversampling, further highlighting the lack of diversity in GANs.
In terms of the conclusions of relative performance of various GANs, we observe that WGAN and
ALI perform consistently better than the other GANs on both the CelebA and LSUN datasets. While
BEGAN samples have good perceptual quality (see Figures 1 and 4), it performs badly on our
classification tasks. On the other hand, WGAN samples have relatively poor visual quality but seem
to outperform most other GANs in classification tasks. This is a strong indicator of the need to
consider other metrics, such as the ones proposed in this paper, in addition to visual inspection to
study GANs. For LSUN, the gap between true and synthetic data is much larger, with the classifiers
getting near random performance on all the synthetic datasets. Note that these classifiers get poor
test accuracy on LSUN but are not overfitting on the training data. In this case, we speculate the
lower performance could be due to both lower quality and diversity of LSUN samples.
In summary, our key experimental finding is that even simple classification–based tests can hold
tremendous potential to shed insight on the learned distribution in GANs. This not only helps us
to get a deeper understanding of many of the underlying issues, but also provides with a more
quantitative and rigorous platform on which to compare different GANs. Our techniques could,
in principle, be also applied to assess other generative models such as Variational Auto-Encoders
(VAEs) Kingma & Welling (2014). However, VAEs have significant problems in generating realistic
samples on the datasets used in our analysis in the first place – see Arora & Zhang (2017).
4
C ONCLUSIONS
In this paper, we put forth techniques for examining the ability of GANs to capture key characteristics of the training data, through the lens of classification. Our tools are scalable, quantitative and
automatic (no need for visual inspection of images). They thus are capable of studying state-ofthe-art GANs on realistic, large-scale image datasets. Further, they serve as a mean to perform a
nuanced comparison of GANs and to identify their relative merits, including properties that cannot
be discerned from mere visual inspection.
We used the developed techniques to perform empirical studies on popular GANs on the CelebA and
LSUN datasets. Our examination showed that mode collapse is indeed a prevalent issue for GANs.
Also, we observed that synthetic GAN-generated datasets have significantly reduced diversity, at
least when examined from a classification perspective. In fact, the diversity of such synthetic data is
often few orders of magnitude smaller than that of the true data. Furthermore, this gap in diversity
does not seem to be bridged by simply producing much larger datasets by oversampling GANs.
Also, our methods can be viewed as benchmarks for GANs. From this perspective, we observed
that ALI consistently outperforms all the other studied GANs, both in terms of the extent of modecollapse, as well as, the diversity of the learned distribution when assessed from the classification
point-of-view. Finally, we noticed that good perceptual quality of samples does not necessarily correlate – and might sometimes even anti-correlate – with distribution diversity. These findings suggest
that we need to go beyond the visual inspection–based evaluations and look for more quantitative
tools for assessing quality of GANs, such as the ones presented in this paper.
10
Under review as a conference paper at ICLR 2018
R EFERENCES
Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks.
In International Conference on Machine Learning (ICML), pp. 214–223, 2017.
Sanjeev Arora and Yi Zhang. Do GANs actually learn the distribution? An empirical study. arXiv
preprint arXiv:1706.08224, 2017.
David Berthelot, Tom Schumm, and Luke Metz. BEGAN: Boundary equilibrium generative adversarial networks. arXiv preprint arXiv:1703.10717, 2017.
Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a Laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems
(NIPS), pp. 1486–1494, 2015.
Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Alex Lamb, Martin Arjovsky, Olivier Mastropietro, and Aaron Courville. Adversarially learned inference. In International Conference on Learning Representations (ICLR), 2017.
Ian Goodfellow.
Nips 2016 tutorial:
arXiv:1701.00160, 2016.
Generative adversarial networks.
arXiv preprint
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NIPS), pp. 2672–2680, 2014.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference
on Learning Representations (ICLR), 2014.
Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.
In International Conference on Computer Vision (ICCV), 2015.
Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial
networks. arXiv preprint arXiv:1611.02163, 2016.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
David Rolnick, Andreas Veit, Serge Belongie, and Nir Shavit. Deep learning is robust to massive
label noise. arXiv preprint arXiv:1705.10694, 2017.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training GANs. In Advances in Neural Information Processing Systems
(NIPS), pp. 2234–2242, 2016.
Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb.
Learning from simulated and unsupervised images through adversarial training. In Computer
Vision and Pattern Recognition (CVPR), July 2017.
L. Theis, A. van den Oord, and M. Bethge. A note on the evaluation of generative models. In
International Conference on Learning Representations (ICLR), 2016.
Yuhuai Wu, Yuri Burda, Ruslan Salakhutdinov, and Roger Grosse. On the quantitative analysis of
decoder-based generative models. arXiv preprint arXiv:1611.04273, 2016.
Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. LSUN: Construction
of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint
arXiv:1506.03365, 2015.
Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaolei Huang, Xiaogang Wang, and Dimitris
Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial
networks. arXiv preprint arXiv:1612.03242, 2016.
11
Under review as a conference paper at ICLR 2018
5
A PPENDIX
5.1
E XPERIMENTAL S ETUP
5.1.1
DATASETS FOR C LASSIFICATION TASKS
To assess GAN performance from the perspective of classification, we construct a set of classification tasks on the CelebA and LSUN datasets. In the case of the LSUN dataset, images are annotated
with scene category labels, which makes it straightforward to use this data for binary and multiclass classification. On the other hand, each image in the CelebA dataset is labeled with 40 binary
attributes. As a result, a single image has multiple associated attribute labels. Here, we can construct classification tasks by considering binary combinations of an attribute(s) (examples are shown
in Figures 1 and 4). Attributes used in our experiments were chosen such that the resulting dataset
was large, and classifiers trained on true data got high-accuracy so as to be good annotators for the
synthetic data.
True Data
DCGAN
WGAN
ALI
BEGAN
Improved GAN
No Makeup
+
Not Smiling
No Makeup
+
Smiling
Makeup
+
Not Smiling
Makeup
+
Smiling
(a) 4-class dataset from CelebA for attributes Makeup, Smiling.
True Data
DCGAN
WGAN
ALI
BEGAN
Improved GAN
No Bangs
+
Not Smiling
No Bangs
+
Smiling
Bangs
+
Not Smiling
Bangs
+
Smiling
(b) 4-class dataset from CelebA for attributes Bangs, Smiling.
Figure 4: Illustration of datasets from CelebA used in proposed classification-based benchmarks to
evaluate GANs. Shown alongside are images sampled from various unconditional GANs trained on
this dataset. Labels for the GAN samples are obtained using a pre-trained classifier as an annotator.
5.1.2
M ODELS
Benchmarks were performed on standard implementations •
•
•
•
•
DCGAN: https://github.com/carpedm20/DCGAN-tensorflow
WGAN: https://github.com/martinarjovsky/WassersteinGAN
ALI: https://github.com/IshmaelBelghazi/ALI
BEGAN :https://github.com/carpedm20/BEGAN-tensorflow
Improved GAN: https://github.com/openai/improved-gan
12
Under review as a conference paper at ICLR 2018
True Data
DCGAN
WGAN
ALI
Improved GAN
Bedroom
Classroom
Kitchen
(a) 3-class dataset from LSUN for Bedroom, Classroom, Kitchen.
True Data
DCGAN
WGAN
ALI
Improved GAN
Bedroom
Conference
Room
Dining
Room
Kitchen
Living
Room
(b) 5-class dataset from LSUN for Bedroom, Conference Room, Dining Room, Kitchen, Living Room.
Figure 5: Illustration of datasets from LSUN used in proposed classification-based benchmarks to
evaluate GANs. Shown alongside are images sampled from various unconditional GANs trained on
this dataset. Labels for the GAN samples are obtained using a pre-trained classifier as an annotator.
• ResNet Classifier: Variation of the standard TensorFlow ResNet https://github.com/
tensorflow/models/blob/master/research/resnet/resnet_model.py
5.2
5.2.1
A DDITIONAL E XPERIMENTAL R ESULTS
S AMPLE Q UALITY
For each of our benchmark experiments, we ascertain that the visual quality of samples produced
by the GANs is comparable to that reported in prior work. Examples of random samples drawn for
multi-class datasets from both true and synthetic data are shown in Figures 1 and 4 for the CelebA
dataset, and in Figure!5 for the LSUN dataset.
5.2.2
D IVERSITY E XPERIMENTS
Sections 2.2 and 3.4 describe techniques to study classification performance of labeled GANgenerated datasets against that of true data. In order to assess the quality of our GAN-generated
synthetic datasets, we evaluate label agreement between the default labels of the synthetic datasets,
13
Under review as a conference paper at ICLR 2018
Figure 6: Histograms of annotator confidence (softmax probability) during label prediction on true
data (test set) and synthetic data for tasks on the CelebA and LSUN datasets (see Section 3.4).
and that obtained using the annotators described in Section 2.1 as shown in Table 2. In Figure 6,
we inspect the confidence scores, defined as the softmax probabilities for predicted class, of the annotator while making these predictions. As can be seen from these results, the annotator confidence
for the synthetic data is comparable to that on the test set of true data. Thus, it seems likely that the
GAN generated samples are of good quality and are truly representative examples of their respective
classes, as expected based on visual inspection.
14
| 1cs.CV
|
Application of Deep Learning in Neuroradiology: Automated
Detection of Basal Ganglia Hemorrhage using 2D-Convolutional
Neural Networks.
Vishal Desai,1 Adam E. Flanders,1 Paras Lakhani1*
1.
Department of Radiology, Thomas Jefferson University Hospital, Sidney Kimmel Jefferson Medical College, Philadelphia, PA, U.S.A.
*Corresponding Author, email: [email protected]
Abstract— Purpose: Deep learning techniques have achieved
high accuracy in image classification tasks, and there is interest
in applicability to neuroimaging critical findings. This study
evaluates the efficacy of 2D deep convolutional neural networks
(DCNNs) for detecting basal ganglia (BG) hemorrhage on
noncontrast head CT. Methods: 170 unique de-identified
HIPAA-compliant noncontrast head CTs were obtained, those
with and without BG hemorrhage. 110 cases were held-out for
test, and 60 were split into training (45) and validation (15),
consisting of 20 right, 20 left, and 20 no BG hemorrhage. Data
augmentation was performed to increase size and variation of
the training dataset by 48-fold. Two DCNNs were used to
classify the images—AlexNet and GoogLeNet—using untrained
networks and those pre-trained on ImageNet. Area under the
curves (AUCs) for the receiver-operator characteristic (ROC)
curves were calculated, using the DeLong method for statistical
comparison of ROCs. Results: The best performing model was
the pre-trained augmented GoogLeNet, which had an AUC of
1.00 in classification of hemorrhage. Preprocessing
augmentation increased accuracy for all networks (p<0.001),
and pretrained networks outperformed untrained ones
(p<0.001) for the unaugmented models. The best performing
GoogLeNet model (AUC 1.00) outperformed the best
performing AlexNet model (AUC 0.95)(p=0.01). Conclusion:
For this dataset, the best performing DCNN identified BG
hemorrhage on noncontrast head CT with an AUC of 1.00.
Pretrained networks and data augmentation increased
classifier accuracy. Future prospective research would be
important to determine if the accuracy can be maintained on a
larger cohort of patients and for very small hemorrhages.
Index Terms—deep learning, intracranial hemorrhage, basal
ganglia, convolutional neural networks
I. INTRODUCTION
Basal ganglia hemorrhage is a type of intracerebral
hemorrhage (ICH), and is associated with long-standing
hypertension.1–3
Along with other forms of ICH, it is
considered a neurologic emergency. The imaging modality
of choice to detect such hemorrhages is non-contrast head
CT given its wide availability, speed at which it can be
performed, and the very high sensitivity and specificity for
detecting acute hemorrhage.4
The American Heart Association and American Stroke
Association note that timely diagnosis and aggressive early
management is very important in ICH, as affected patients
commonly deteriorate within the first few hours after onset. 5
As such, an automated solution to identify such hemorrhages
may be helpful to decrease time to diagnosis, and more
readily triage appropriate care. Prior work with computer
aided detection (CAD) has shown success in automated
segmentation of ICH.6 In addition, one study using CAD
achieved 95% sensitivity and 89% specificity for detection
of ICH using pre-processing techniques and a knowledgebased classification system.7 Another study using machine
learning techniques was able to achieve 96% accuracy for
detecting ICH using pre-processing techniques, feature
selection, and a fuzzy classifier .8
Based on the recent success in the ImageNet Large Scale
Visual Recognition Competition, deep convolutional neural
networks (DCNNs) are considered state-of-the-art for image
classification.9 While convolutional neural networks have
been around for many years, only recently have we seen
their increase used due to availability of high-performing
graphics processing units (GPU) and availability of many
open-source frameworks. Since 2012, deep convolutional
neural networks (DCNN) have been utilized by all winning
entries in ILSVRC, resulting in a drop in top-5 classification
error rate from 25% in 2011 to 3% more recently with some
of the latest architectures.10
Application of deep learning in radiology has found
promising results in multiple modalities. Some recent
examples include brain segmentation on MRI,11 pancreatic
segmentation,12 knee cartilage evaluation,13 detection of
pleural effusion and cardiomegaly on chest radiographs,14
detection of tuberculosis on chest radiographs,15 and
intracranial critical findings.16
In this study, we focus on identifying a common site of
hemorrhagic stroke, the basal ganglia. We evaluate the
efficacy of automated detection of basal ganglia hemorrhage
on noncontrast head CTs using two DCNNs – AlexNet17 and
GooLeNet,18 the winners of the 2012 and 2014 ImageNet
Page 1 of 7
Datasets:
This was a retrospective study using de-identified and
Health Insurance Portability and Accountability Act
(HIPAA) compliant datasets from Thomas Jefferson
University Hospital, PA, U.S.A. To create the datasets, the
institutional Radiology Information System (RIS) database
was searched over a 12-month period between May 2016 May 2017 for patients with basal ganglia hemorrhage and
for patients without any acute intracranial pathology on noncontrast head CT. All CT studies were verified by two
board-certified radiologists, the author of the final report,
and by an independent board-certified radiologist (P.L.). The
terms “basal ganglia, hemorrhage, hematoma, and bleed”
were entered into the database using multiple permutations.
Patients with brain tumors and prior cranial surgery were
excluded.
170 unique patients were identified. 110 of these patients
were held out and placed in the test dataset, consisting of 55
with BG hemorrhage, and 55 without hemorrhage. The
remaining 60 patients were used for training and validation,
divided in a 75/25 ratio, with 45 cases for training and 15
cases for validation. Of these 60 training/validation cases,
20 consisted of right basal ganglia hemorrhage (15 training +
5 validation), 20 left basal ganglia hemorrhage (15 training +
5 validation), and 20 no intracranial hemorrhage (15 training
+ 5 validation). Seven cases with basal ganglia calcifications
(mimic of BG hemorrhage) were included in the datasets,
four for training and 3 in the test-dataset. The training set
was used to train the algorithm, validation set for model
selection, and test set for assessment of the final chosen
model. 110 test cases were chose to provide a 95%
confidence interval of ±7.5% based on estimated accuracy of
the best-performing models on the accuracy on the
validation datasets.19
Study Design:
For each head CT in the training dataset, several axial key
images were obtained near or at the level of the basal
ganglia, which were determined by two radiologists (V.D.,
P.L.). This sometimes included images one slice above
and/or below the basal ganglia, and approximately 3-5
images at the level of the basal ganglia. Approximately 4-6
images were saved per study depending on the scan angle,
slice thickness (range 2.5-5 mm), and, if present, size of the
hemorrhage. A total of 308 unique DICOM (Digital Imaging
and Communications in Medicine) images at the levels of
interest were obtained from 60 non-contrast head CTs (20
with right basal ganglia hemorrhage, 20 with left basal
GoogLeNet
II. METHODS
ganglia hemorrhage, and 20 without hemorrhage). The
images were resized to 256 x 256 pixels (from 512 x 512
originally) and converted into Portable Network Graphics
(PNG) format. The images were loaded onto a workstation
running the DIGITS deep learning GPU training system
(DIGITS 4.0, Nvidia Corporation, Santa Clara, CA) running
Ubuntu 14.04, Caffe deep learning framework (Nvidia fork),
CUDA 8.0, and cuDNN dependencies (Nvidia Corporation,
Santa Clara, CA) for graphics processing unit (GPU)
acceleration. The computer contained an Intel i5 3570k
3.4gHz processor, 4TB hard disk space, 32gb RAM, and a
CUDA-enabled NVIDIA Maxwell Titan X 12Gb GPU
(Nvidia Corporation, Santa Clara, CA).
AlexNet
Large Scale Visual Recognition Competition respectively.
Untrained
Pre-trained
Untrained
Pre-trained
Unaugmented
Dataset
Classifer 1
Augmented
Dataset
Classifier 2
Unaugmented
Dataset
Classifier 3
Augmented
Dataset
Classifier 4
Unaugmented
Dataset
Classifier 5
Augmented
Dataset
Classifier 6
Unaugmented
Dataset
Classifier 7
Augmented
Dataset
Classifier 8
Figure 1: Eight different models were evaluated, using two
different architectures (AlexNet and GoogLeNet), including
pretrained and untrained networks, and with and without extra
augmentation.
Multiple sequential pre-processing augmentation techniques
were performed, including three different CT window-width
(WW) and window-levels (WL) ( “brain” 80/35 WW/WL,
“acute blood” 220/95 WW/WL, and “stroke” 27/33
WW/WL), 10% magnification, rotation of 15 & 30 degrees,
blur and edge enhancement. This pre-processing was
performed using ImageJ v1.50i (National Institutes of
Health, U.S.A.) and XnConvert v1.73 (XnSoft, France).
Images were also augmented in real-time during the training
process using prebuilt options within the Caffe framework, 16
including random cropping of 227 x 227 pixels for AlexNet
and 224 x 224 pixels for GoogLeNet, and mean wholeimage subtraction. After image augmentation using preprocessing detailed above, the new training dataset totaled
11,088 images, which was a 48-fold increase in number of
images from the originals.
Page 2 of 7
Two deep convolutional neural networks architectures were
used—AlexNet and GoogLeNet, using unaugmented and
augmented datasets, and those untrained and pre-trained on
ImageNet. Taking the permutation of the two datasets
(unaugmented and augmented datasets), two neural networks
(AlexNet and GoogLeNet), and two neural network types
(pre-trained and untrained), a total of 8 trained models were
created (Figure 1). Pre-trained networks were obtained from
the Caffe Model Zoo (http://caffe.berkeleyvision.org), Berkeley,
CA), an open-access repository of pre-trained models for use
with Caffe,20 and were previously trained on over 1 million
every-day color images from ImageNet.9
The following solver parameters were used for training
AlexNet: 90 epochs, base learning rate of 0.01 for untrained
models and 0.001 for pre-trained models, stochastic gradient
descent, step-down 33%, and gamma of 0.5. For
GoogLeNet, the parameters were: 40 epochs, base learning
rate of 0.01 for untrained models and 0.001 for pre-trained
models, stochastic gradient descent, step-down 33%, and
gamma of 0.5. Categorical cross-entropy was used for the
loss function. The solver parameters including number of
epochs were determined after reviewing the training and
validation loss curves after multiple runs of the data, with
the goal of achieving the lowest validation loss, and stopping
training after plateau in the loss. For pretrained networks, we
randomized the weights of the final fully-connected layer of
the networks to learn from the CT images, and employed a
fine-tuning strategy where all layers were open to learn at a
reduced base learning rate.
dataset had significantly better AUC than those trained with
the smaller, unaugmented dataset in all scenarios–pre-trained
GoogLeNet and AlexNet and untrained GoogLeNet and
AlexNet (Figure 2 & 3, Table 2). The pre-trained classifier
was significantly better than the untrained classifier for
GoogLeNet when using the unaugmented dataset (Table 3).
There was no significant difference in AUC between the
untrained and pretrained GoogLeNet and AlexNet classifiers
when using the augmented dataset (Table 4).
The best performing classifiers were the augmented
untrained and pre-trained GoogLeNet DCNN, which had an
AUC of 0.99 and 1.0 respectively in identifying BG
hemorrhage (Table 1). The augmented GoogLeNet
classifiers correctly identified 55/55 cases of BG
hemorrhage in the test dataset, including the laterality of the
bleed (sensitivity 100%). The pre-trained classifier correctly
labeled 55/55 cases without hemorrhage (specificity 100%)
while the untrained classifier had one false positive in a case
with asymmetric basal ganglia calcifications (specificity
96.4%).
The augmented untrained AlexNet classifier retained a high
AUC (0.96), with sensitivity for hemorrhage at 100% but
with a lower specificity at 80%, particularly mislabeling
basal ganglia calcifications as hemorrhages (Figure 4).
Overall, the augmented pre-trained GoogLeNet classifier
had greater accuracy compared to the AlexNet classifiers
(Table 5, Figure 3). The best performing GoogLeNet model
(AUC 1.00) outperformed the best performing AlexNet
model (AUC 0.95)(p=0.01).
Statistical Analysis:
Receiver operating characteristic (ROC) curves and area
under the curves (AUC) were determined using the pROC
package (ver. 1.7.3) for R (ver. 3.3.1), utilizing the DeLong
method for statistical comparison of ROCs.21,22 P-values <
0.05 were considered statistically significant.
IRB approval:
This study was approved by the Institutional Review Board
at Thomas Jefferson University Hospital, Philadelphia, PA,
U.S.A.
III. RESULTS
The average size of the hemorrhages for the test dataset was
2.8±1.9cm in the transverse dimension, and 3.6±2.7 cm in
the anteroposterior dimension.
The smallest two
hemorrhages were 0.5 x 1.1cm and 0.7 x 0.7cm, and the
largest was 7.1 x 11.2cm.
Table 1 contains area under the curve (AUC) calculations
for all classifiers. The classifiers trained with the augmented
Table 1: AUC for All Classifiers
Classifier
Area Under the Curve
Unaugmented AlexNet-Untrained
0.57 (0.46-0.68)
Unaugmented AlexNet-Pretrained
Unaugmented GoogLeNetUntrained
Unaugmented GoogLeNetPretrained
Augmented AlexNet-Untrained
0.81 (0.74-0.89)
Augmented AlexNet-Pretrained
0.92 (0.87-0.97)
0.60 (0.49-0.70)
0.89 (0.83-0.95)
0.95 (0.92-0.99)
Augmented GoogLeNet-Untrained
0.99 (0.96-1.00)
Augmented GoogLeNet-Pretrained
1.00 (1.00-1.00)
Parentheses reflects 95% confidence interval
IV. DISCUSSION:
Recent advances in technology have enabled the use of
machine learning, specifically deep learning, to be applied to
radiologic images. Machine learning allows computers to
analyze data and perform tasks without being explicitly
Page 3 of 7
programmed to do so.23 Deep artificial neural networks, on
the other hand, is a relatively newer branch of machine
learning, which consists of multiple hidden layers, and
excels with high dimension datasets such as images. For this
study, we used supervised (pre-labeled) images to train our
added laterality to our classifier, which is a simple method to
verify appropriate training. So, if the classifier is erroneously
labeling right basal ganglia hemorrhages as left, we can
assume that the network is inappropriately trained.
Table 2: AUC Comparison of Classifiers with Unaugmented and
Augmented Datasets
Classifier
Unaugmented
AUC
Augmented
AUC
p-value
AlexNetUntrained
AlexNetPretrained
GoogLeNetUntrained
GoogLeNetPretrained
0.57
0.95
p<0.001
0.82
0.92
p=0.001
0.60
0.99
p<0.001
0.89
1.00
p<0.001
basal ganglia hemorrhage classifiers. DCNNs consist of a
varying number of interconnected layers, and each layer
contains numerous independent “nodes” or “neurons” which
analyze a certain feature. Since all layers are interconnected,
this creates a network where each data point is constantly
referenced against all other layers to provide the best
possible prediction.
Deep neural networks have been described as “black boxes,”
where the reasoning behind a network’s prediction may not
be readily apparent.24 For medical imaging, ensuring the
network has been properly trained and performs accurately
in unknown cases is crucial. The black box scenario is
compounded by the fact that DCNNs are extremely large
and complex.
Table 3: AUC Comparison of Untrained and Pre-trained
Classifiers with Unaugmented Dataset
Classifier
AlexNet,
Unaugmented
GoogLeNet,
Unaugmented
Untrained
AUC
0.57
Pre-trained
AUC
0.81
p<0.001
0.60
0.89
p<0.001
Figure 2. Receiver operating characteristic (ROC) curve
comparing the Untrained AlexNet with and without extra
augmentation, with an AUC of 0.95 and 0.57 respectively,
p<0.0001.
Activations within a layer can also offer confidence that the
network is being activated by the area of interest.26 Figure 6
demonstrates a strong activation in 2nd convolutional layer of
the pretrained GoogLeNet classifier followed by a correct
prediction of left basal ganglia hemorrhage with 100%
confidence.26
p-value
However, several strategies do exist to get insight into what
is contributing to the network’s prediction.24-26 One
technique is to obscure the area of interest and then assess
how this affects the network’s prediction.25 For instance,
obscuring a basal ganglia hematoma by overlaying a gray
box should result in a prediction of “no basal ganglia
hemorrhage” (if the network has learned what was intended)
(Figure 5). Rather than manually processing images, we
One helpful step in building an accurate image classifier is
augmentation.27 The results of our study demonstrate the
statistically significant difference in AUCs among the
unaugmented and augmented datasets (Table 2). It has been
demonstrated that the more cases and variations supplied to
the neural network during training, the better the
performance and generalization of the DCNN.27
For the basal ganglia classifier, we ensured variety in the
initial, unaugmented dataset by including cases with basal
ganglia calcifications, different sized hemorrhages, multicompartmental hemorrhage, and hydrocephalus. Then, with
image augmentation we take that variety of cases and expose
it to the neural network in different ways by introducing
varying degrees of blur, edge enhancement, rotation, zoom,
Page 4 of 7
windowing and other image processing techniques. This
simulates variations in the CT scan that might be
encountered on a day-to-day basis, improving generalization
of our classifier without needing to acquire extra cases. A
degree of manual input is certainly required in the case
layers for AlexNet) and employs an “inception” module with
a “bottleneck” design.18,28 The inception module uses
smaller 1x1 convolutions to reduce the number of variables
that flow into the layer, before larger convolutions, which
improves computational efficiency and may improve
predictions.28
Table 5: AUC Comparison of GoogLeNet and AlexNet with
Augmented Dataset
Figure 3: ROC curves of the pretrained AlexNet and GoogLeNet
models with extra augmentation. The best-performing pretrained
GoogLeNet model (AUC 1.00) performed better than the
corresponding AlexNet model (AUC 0.92), p=0.006.
Classifier
AlexNet
GoogLeNet
p-value
Untrained,
Augmented
AUC
Pretrained,
Augmented
AUC
0.95
0.99
p=0.12
0.92
1.00
p=0.006
We also found that with small (unaugmented) datasets, the
networks pre-trained with everyday images on ImageNet
were superior to untrained networks. However, with larger
(augmented) datasets, there was no significant difference
between pre-trained and untrained networks (Table 4).
Since all images, including medical images, share basic lowlevel features such as edges, lines and blobs, a network pretrained on non-medical images can help prime the initial
layers of the network through transfer learning.29 Then, the
interconnected layers could re-learn from the CT images
provided. This appeared less significant with large datasets,
perhaps because there was enough input data to robustly
train the initial layers without pretraining.
selection process. For instance, we chose to specifically
include cases with basal ganglia calcifications to decrease
our false positive rate. The augmented GoogLeNet
classifiers successfully labeled basal ganglia calcifications as
“no hemorrhage” while AlexNet had difficulty with dense
asymmetric calcifications, possibly related to the extra depth
of the classifier.
Table 4: AUC Comparison of Untrained and Pre-trained
Classifiers with Augmented Dataset
Classifier
AlexNet,
Augmented
GoogLeNet,
Augmented
Untrained
AUC
0.95
Pre-trained
AUC
0.92
p-value
0.99
1.00
p=0.22
p=0.27
GoogLeNet was superior to AlexNet using the augmented
dataset (Table 5), although only statistically significant for
the pre-trained models. This is likely due to the fact that
GoogLeNet is a much deeper network (22 layers versus 8
Figure 4: Axial noncontrast CT through the basal ganglia region
(ww/wl: 80/35) demonstrating bilateral basal ganglia calcification
(white arrows). This image was incorrectly predicted as
hemorrhage by all classifiers except for the best-performing
pretrained, augmented GoogLeNet model.
Page 5 of 7
One of the common problems with deep learning is
overfitting, or more simply, lack of generalizability. 30 In this
scenario, the classifier maintains high prediction accuracy on
the training dataset but does poorer with unknown cases,
which can happen when the training dataset is small. The
models employed several strategies to help overcome this,
including increase in the size and variety of the dataset by
pre-processing augmentation, and use of dropout—model
regularization technique—which has been shown to help
with overfitting.27,30
Figure 5: Axial non-contrast CT through the basal ganglia region
(ww/wl: 220/95). The right image shows an acute hemorrhage in
the right basal ganglia (white arrow), which is correctly predicted
by the classifier (100% confidence). In the left image, the
hemorrhage is obscured by a gray box, which changes the
prediction to “no hemorrhage.” This technique can help
demonstrate that the model is assessing the appropriate region.
generalizable to new unforeseen cases.
There are limitations to this study. While the image classifier
we generated serves as a great proof-of-concept that deep
learning can accurately detect basal ganglia hemorrhage,
more research is needed to demonstrate its accuracy with
larger cohorts. It is unknown how well the model would do
in scenarios where there is significant artifact, prior
cerebrovascular conditions such as brain tumors, or altered
anatomy from prior surgery. In addition, it is unclear how
sensitive the algorithm would be on more subtle
subcentimeter hemorrhages. Most of the hemorrhages were
greater than 1.5cm, although the classifier did detect some
smaller hemorrhages measuring approximately 1cm (10 of
the 55 positive test cases), including one that was 0.5 x
1.1cm and another that was 0.7 x .7cm. Also, this study was
performed retrospectively and the performance of this
classifier prospectively would need to be determined.
Finally, this was a model that used 2D convolutions, and
makes an assessment on a slice-by-slice basis. More research
is needed to determine efficacy compared to 3D CNNs,
which would have the benefit of analyzing the whole head
but at greater computational costs and GPU memory needs.
Before utilization in a clinical scenario, the classifier will
need be trained to learn the normal anatomy of all parts of
the cranium including the vertex and skull bases, rather than
just the basal-ganglia region, so as not to flag normal
structures as hemorrhage.
Despite the limitations, this study establishes a starting point
for building an accurate CT-based image classifier for
hemorrhage. The next steps will be to build a moreencompassing intracranial hemorrhage algorithm, and
address the questions or limitations noted above.
Future work may also include a system where a deep
learning algorithm automatically flags studies as positive or
negative on a reading worklist, where one could then
evaluate its effect on work prioritization and result turnaround-time.
V. CONCLUSIONS:
Figure 6: On the left is an axial CT image at the level of the basal
ganglia (WW/WL: 80/35) containing a left-sided BG hemorrhage.
On the right is an accompanying activation from the 2 nd layer of the
best-performing GoogLeNet model. The small orange-white focus
(white arrow) represents an area activated by the network, and
corresponds to the location of the hemorrhage.
For this dataset, the best performing DCNN identified BG
hemorrhage on noncontrast head CT with an AUC of 1.00.
Pretrained networks and data augmentation increased
classifier accuracy. Future prospective research would be
important to determine if high accuracy can be maintained
on a larger cohort of patients and for very small
hemorrhages.
VI. REFERNCES:
Finally, we assess the trained models with test cases,
previously unseen by the classifiers, and analyze the
predictions (Table 1). This shows that the models can be
1.
Hanley DF, Awad IA, Vespa PM, Martin NA,
Zuccarello M. Hemorrhagic Stroke: Introduction.
Stroke. 2013;44(6 suppl 1).
Page 6 of 7
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
Sutherland GR, Auer RN. Primary intracerebral
hemorrhage. J Clin Neurosci. 2006;13(5):511-517.
doi:10.1016/j.jocn.2004.12.012.
Sahni R, Weinberger J. Management of intracerebral
hemorrhage. Vasc Health Risk Manag.
2007;3(5):701-709.
Rincon F, Mayer SA. Clinical review: Critical care
management of spontaneous intracerebral
hemorrhage. Crit Care. 2008;12(6):237.
doi:10.1186/cc7092.
Hemphill JC, Greenberg SM, Anderson CS, Becker
K, Bendok BR, Cushman M, Fung GL, Goldstein
JN, Macdonald RL, Mitchell PH, Scott PA.
Guidelines for the management of spontaneous
intracerebral hemorrhage. Stroke. 2015 Jul
1;46(7):2032-60.
Scherer M, Cordes J, Younsi A, Sahin YA, Götz M,
Möhlenbruch M, Stock C, Bösel J, Unterberg A,
Maier-Hein K, Orakcioglu B. Development and
Validation of an Automatic Segmentation Algorithm
for Quantification of Intracerebral Hemorrhage.
Stroke. 2016 Nov 1;47(11):2776-82.
Chan T. Computer aided detection of small acute
intracranial hemorrhage on computer
tomography of brain. Computerized Medical
Imaging and Graphics. 2007 Jul 31;31(4):285-98.
Maduskara P, Acharyyaa M. Automatic
identification of intracranial hemorrhage in noncontrast CT with large slice thickness for trauma
cases. InProc. of SPIE Vol 2009 Feb 26 (Vol. 7260,
pp. 726011-1
Russakovsky O, Deng J, Su H, et al. ImageNet
Large Scale Visual Recognition Challenge.
September 2014.
Szegedy, C., Ioffe, S., Vanhoucke, V. and Alemi, A.,
2016. Inception-v4, inception-resnet and the impact
of residual connections on learning. arXiv preprint
arXiv:1602.07261.
Li R, Zhang W, Suk H-I, et al. Deep learning based
imaging data completion for improved brain disease
diagnosis. Med Image Comput Comput Assist Interv.
2014;17(Pt 3):305-312.
Roth HR, Farag A, Lu L, Turkbey EB, Summers
RM. Deep convolutional networks for pancreas
segmentation in CT imaging. April 2015.
doi:10.1117/12.2081420.
Prasoon A, Petersen K, Igel C, Lauze F, Dam E,
Nielsen M. Deep Feature Learning for Knee
Cartilage Segmentation Using a Triplanar
Convolutional Neural Network.
Bar Y, Diamant I, Wolf L, Greenspan H. Deep
learning with non-medical training used for chest
pathology identification. In: Hadjiiski LM, Tourassi
GD, eds. ; 2015:94140V. doi:10.1117/12.2083124.
Lakhani P, Sundaram B. Deep Learning at Chest
Radiography: Automated Classification of
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
Pulmonary Tuberculosis by Using Convolutional
Neural Networks. Radiology. April 2017:162326.
doi:10.1148/radiol.2017162326.
Prevedello LM, Erdal BS, Ryu JL, Little KJ,
Demirer M, Qian S, White RD. Automated Critical
Test Findings Identification and Online Notification
System Using Artificial Intelligence in Imaging.
Radiology. 2017 Jul 3:162664.
Krizhevsky A, Sutskever I, Hinton GE. ImageNet
Classification with Deep Convolutional Neural
Networks. 2012:1097-1105.
Szegedy C, Liu W, Jia Y, et al. Going Deeper with
Convolutions.
Hastie T, Tibshirani R, Friedman JH, Jerome H. The
Elements of Statistical Learning : Data Mining,
Inference, and Prediction.
Jia Y, Shelhamer E, Donahue J, et al. Caffe:
Convolutional Architecture for Fast Feature
Embedding. June 2014.
Obuchowski NA. Receiver Operating Characteristic
Curves and Their Use in Radiology. Radiology.
2003;229(1):3-8. doi:10.1148/radiol.2291010898.
DeLong ER, DeLong DM, Clarke-Pearson DL.
Comparing the areas under two or more correlated
receiver operating characteristic curves: a
nonparametric approach. Biometrics.
1988;44(3):837-845.
Wang S, Summers RM. Machine learning and
radiology. Med Image Anal. 2012;16(5):933-951.
doi:10.1016/j.media.2012.02.005.
Yosinski J, Clune J, Nguyen A, Fuchs T, Lipson H.
Understanding Neural Networks Through Deep
Visualization. June 2015.
Zeiler MD, Fergus R. Visualizing and understanding
convolutional networks. arXiv preprint
arXiv:1311.2901. 2013 Nov 12.
Yeager L, Heinrich G, Mancewicz J, Houston M.
Effective Visualizations for Training and
Evaluating Deep Models. Proceedings of the 33rd
International Conference on Machine Learning, New
York, NY, USA, 2016. JMLR
https://icmlviz.github.io/assets/papers/16.pdf
Wu R, Yan S, Shan Y, Dang Q, Sun G, Research B.
Deep Image: Scaling up Image Recognition.
Yu D, Seltzer ML. Improved Bottleneck Features
Using Pretrained Deep Neural Networks.
Yosinski J, Clune J, Bengio Y, Lipson H. How
transferable are features in deep neural networks?
InAdvances in neural information processing
systems 2014 (pp. 3320-3328).
Srivastava N, Hinton G, Krizhevsky A, Sutskever I,
Salakhutdinov R. Dropout: A Simple Way to
Prevent Neural Networks from Overfitting. J Mach
Learn Res. 2014;15:1929-1958.
Page 7 of 7
| 1cs.CV
|
arXiv:1708.09313v1 [cs.IT] 30 Aug 2017
Rotation Symmetric Bent Boolean Functions for
n = 2p
T. W. Cusick
University at Buffalo, 244 Math. Bldg. Buffalo, NY 14260
Corresponding author- e-mail: [email protected]
E. M. Sanger
University at Buffalo, 244 Math. Bldg., Buffalo, NY 14260
e-mail: [email protected]
Abstract
It has been conjectured that there are no homogeneous rotation
symmetric bent Boolean functions of degree greater than two. In this
paper we begin by proving that sums of short-cycle rotation symmetric
bent Boolean functions must contain a specific degree two monomial
rotation symmetric Boolean function. We then prove most cases of the
conjecture in n = 2p, p > 2 prime, variables and extend this work to
the nonhomogeneous case.
1
Introduction
A Boolean function in n variables can be defined as a map from Vn , the
n-dimensional vector space over the two element field F2 , to F2 . If f is a
Boolean function in n variables, the truth table of f is defined to be the 2n tuple given by (f (v0 ), f (v1 ), . . . , f (v2n −1 )) where v0 = (0, . . . , 0, 0), v1 =
(0, . . . , 0, 1), . . . , v2n −1 = (1, . . . , 1, 1) are the 2n elements of Vn listed in
lexicographical order. The weight or Hamming weight of f (notation wt(f ))
is the number of 1’s that appear in the truth table of f .
As described in [1, pp. 5-6], every Boolean function on Vn can be expressed as a polynomial over F2 in n binary variables by:
X
an−1
f (x0 , . . . , xn−1 ) =
ca xa00 · · · xn−1
a∈Vn
where ca ∈ F2 and a = (a0 , . . . , an−1 ) with each ai equal to 0 or 1. The
above representation is the algebraic normal form (ANF) of f . Let di be the
1
number of variables in the i-th monomial of f , so di is the algebraic degree
(or just the degree) of the monomial. If we let D be the set of the distinct
degrees of the monomials in f which have non-zero coefficients, then the
degree (notation deg(f )) of f is given by max(D). If D contains only one element, then each monomial in f has the same degree and f is homogeneous.
If deg(f ) = 1, then f is affine, and if f is affine and homogeneous (i.e. the
constant term is 0), then f is linear.
A Boolean function f is rotation symmetric (RotS) if its ANF is invariant
under any power of the cyclic permutation ρ(x0 , . . . , xn−1 ) = (x1 , . . . , xn−1 , x0 ).
We will use the notation u ∼ v to indicate that there exists some 1 ≤ k ≤ n
such that ρk (u) = v. Clearly ∼ defines an equivalence relation on Vn .
Let On (x) denote the orbit of x under the action of ρk , 1 ≤ k ≤ n, and
let Gn be the set of representatives of all the orbits in Vn . Then a rotation
symmetric Boolean function f can be written as
a0 ⊕ a1 x 0 ⊕
n−1
X
a1j x0 xj ⊕ . . . ⊕ a12...n x0 x1 · · · xn−1 ,
j=1
where the coefficients a0 , a1 , a1j , . . . , a12...n ∈ F2 , and the existence of a
representative term x0 xi2 · · · xil implies the existence of all the terms from
On (x0 , xi2 , . . . , xil ) in the algebraic normal form. We call this representation
the short algebraic normal form (SANF) of f .
Unless otherwise specified, all subscripts in any monomial will be taken
mod (n) with entries in {0, 1, . . . , n − 1}. We may omit the modulus if it is
clear from the context.
Suppose f has SANF x0 xi1 . . . xil , then we say f is a monomial rotation
symmetric (MRS) function and call it full-cycle if |On (x0 , xi1 , . . . , xil )| =
n. Thus if f is full-cycle then it contains n monomials in its ANF. If
|On (x0 , xi1 , . . . , xil )| < n we say f is short-cycle. In this case f will contain n/w monomials in its ANF for some divisor w > 1 of n.
The Hamming distance between two Boolean functions f and g, denoted
d(f, g), is defined as d(f, g) = wt(f ⊕ g). Each Boolean function f has an
associated sign function, fˆ : R∗ → C∗ , defined by fˆ(x) = (−1)f (x) .
Definition 1.1. The Walsh transform of a function f : Vn → F2 is the map
W (f ) : Vn → R given by,
X
f (x)(−1)w·x
W (f )(w) =
x∈Vn
The Walsh spectrum of f is the list of the 2n Walsh coefficients given by
W (f )(w) as w varies.
2
Definition 1.2. A matrix H is called a Hadamard matrix of order n if it is
an n × n matrix of ±1s such that
HH t = nIn ,
where H t is the transpose of H and In is the n × n identity matrix.
Definition 1.3. A Boolean function f in n variables is called bent if and
only if the Walsh transform coefficients of fˆ are all ±2n/2 .
Definition 1.4. The nonlinearity of a function f, denoted by Nf , is defined
by
Nf = min d(f, g),
g∈An
where An is the set of all affine functions on Vn .
From [1, pp. 76-77] we have the following equivalent definition of the
bent property:
Theorem 1.1. Let f : Vn → F2 be a Boolean function. Then the following
are equivalent:
(i) f is bent.
(ii) Mfˆ = (fˆ(u ⊕ v)) is a Hadamard matrix.
(iii) The nonlinearity of f is Nf = 2n−1 − 2n/2−1 .
It is easy to observe that bent functions exist only for even dimensions.
Let n be even and v = (a0 , a1 , . . . , an−1 ); then we will use the notation
v′ = (a0 , . . . , a n2 −1 ) and v′′ = (a n2 , . . . , an−1 ).
Lemma 1.2. For n = 2, the degree of a bent function on V2 is 2. For
n > 2, the degree of a bent function is at most n/2.
Proof. The proof of this can be found in [1, pp. 80].
2
Rotation symmetric bent functions when n=2m
In this section we will consider rotation symmetric functions of any even
degree n. We will first prove that any degree 2 rotation symmetric bent
n/2−1
L
function must contain f0 =
xi x n2 +i and then extend this to the case
i=0
where f is composed of short-cycle MRS functions. We begin by proving
the well known fact that f0 is bent.
3
Lemma 2.1. Let n = 2m and f0 =
m−1
L
xi xm+i . Then wt(f0 ) = 22m−1 −
i=0
2m−1 and f0 is bent.
Proof. The fact that f is bent follows from [1, Cor. 5.23, p. 82], after a
permutation of the variables. Then a computation gives the weight.
To prove that any degree 2 rotation symmetric bent function, f , must
contain f0 , we will look at the matrix M = (fˆ(vi ⊕ vj ))(fˆ(vi ⊕ vj ))T to
determine when (fˆ(vi ⊕ vj )) is a Hadamard matrix. We will first simplify
the matrix M to a useful form.
Lemma 2.2. Let M = (fˆ(vi ⊕ vj ))(fˆ(vi ⊕ vj ))T where f is a Boolean
function in n variables, then
Mi,j =
n −1
2X
fˆ(vi ⊕ vk )fˆ(vk ⊕ vj ).
(1)
k=0
Proof. Let A = (fˆ(vi ⊕ vj )). Notice AT = A, so M = AA. Then
Mi,j = (AA)i,j =
n −1
2X
Aik Akj =
n −1
2X
fˆ(vi ⊕ vk )fˆ(vk ⊕ vj ).
k=0
k=0
Lemma 2.3. We have vi ∼ vj if and only if there exists k, 1 ≤ k ≤ n such
that 2k i ≡ j mod (2n − 1).
Proof. Let vi = (a0 , a1 , ..., an−2 , an−1 ), so i =
n
X
an−u 2u−1 . ⇒: Assume
u=1
vi ∼ vj , then there exists ρl , 1 ≤ l ≤ n, such that
ρl (vi ) = (ρl (a0 ), . . . , ρl (an−2 ), ρl (an−1 )) = (al , . . . , al+n−2 , al+n−1 ) = vj .
4
Thus, j =
n
X
al+n−u 2u−1 where 1 ≤ l ≤ n and
u=1
2l i =
n
X
an−u 2l+u−1
u=1
= an−1 2l + . . . + an−(n−l) 2l+(n−l)−1 + an−(n−l+1) 2l+(n−l+1)−1 +
an−(n−l+2) 2l+(n−l+2)−1 + . . . + a0 2l+n−1
= an−1 2l + . . . + al 2n−1 + al−1 2n + al−2 2n+1 . . . + a0 2n−1+l
≡ al−1 + al−2 2 + . . . + al 2n−1 mod (2n − 1)
≡ al+n−1 + al+n−2 2 + . . . + al+1 2n−2 + al 2n−1 mod (2n − 1)
≡ j mod (2n − 1),
since indices are taken mod n.
Thus there exists k, 1 ≤ k ≤ n such that 2k i ≡j mod (2n − 1).
⇐: Assume there exists k, 1 ≤ k ≤ n such that 2k i ≡ j mod (2n − 1).
Then
2k i =
n
X
an−u 2k+u−1
u=1
= an−1 2k + . . . + an−(n−k) 2k+(n−k)−1 + an−(n−k+1) 2k+(n−k+1)−1
+an−(n−k+2) 2k+(n−k+2)−1 + . . . + a0 2n−1+k
= an−1 2k + . . . + ak 2n−1 + ak−1 2n + ak−2 2n+1 + . . . + a0 2n−1+k
≡ ak−1 + ak−2 2 + . . . + ak 2n−1
mod (2n − 1)
= j.
Thus vj = (ak−1 , ak−2 , . . . , ak ) so ρk−1 (vi ) = vj and vi ∼ vj .
Lemma 2.4. We have va ∼ vb if and only if v2n −1−a ∼ v2n −1−b .
Proof. By Lemma 2.3,
va ∼ vb
⇐⇒
there exists k, 1 ≤ k ≤ n such that 2k a ≡ b mod (2n − 1)
⇐⇒
2n − 1 − 2k a ≡ 2n − 1 − b
⇐⇒
2k (2n − 1 − a) ≡ 2n − 1 − b
⇐⇒
v2n −1−a ∼ v2n −1−b .
5
Lemma 2.5.
M0,2n −1 =
2n−1
X−1
2fˆ(vk )fˆ(v2n −1−k ).
(2)
k=0
Proof. From (1) we have,
M0,2n −1 =
n −1
2X
fˆ(v0 ⊕ vk )fˆ(vk ⊕ v2n −1 ) =
n −1
2X
fˆ(vk )fˆ(v2n −1−k )
k=0
k=0
= fˆ(v0 )fˆ(v2n −1 ) + fˆ(v1 )fˆ(v2n −2 ) + . . . + fˆ(v2n−1 −1 )fˆ(v2n−1 )
+fˆ(v2n−1 )fˆ(v2n−1 −1 ) + . . . + fˆ(v2n −2 )fˆ(v1 ) + fˆ(v2n −1 )fˆ(v0 )
=
2n−1
X−1
2fˆ(vk )fˆ(v2n −1−k ).
k=0
Lemma 2.6. The function f0 =
m−1
L
xi xm+i is the only MRS short-cycle
i=0
function of degree 2 in n = 2m variables.
Proof. Let f be an MRS function with SANF x0 xk . Suppose f is short-cycle,
then f has 1 ≤ M < 2m monomials and
f = x0 xk ⊕ x1 xk+1 ⊕ . . . ⊕ xM −1 xk+M −1 .
Thus {0, k} = {M, k + M }. Since 1 ≤ M < 2m = n then we have the
following cases:
(i) 0 = M and k ≡ M + k mod (n). This is a contradiction since 1 ≤ M .
(ii) k = M and 0 ≡ M + k mod (n). Thus 0 ≡ M + k mod (n) ≡
2M mod (2m). So 2M = 2am for some a ∈ Z. Since 1 ≤ M < 2m,
then 2 ≤ 2am < 4m and so a = 1. Thus M = m.
Thus f is a short-cycle if and only if k = m which gives f = f0 .
We now have enough to prove that f0 is the only homogeneous MRS
bent function of degree 2:
Theorem 2.7. Let f be a rotation symmetric boolean function in n = 2m
variables which has the SANF x0 xk . Then f is bent if and only if k = m.
6
Proof. Assume k 6= m = n2 . Then by Lemma 2.6, f is full cycle and we write
n−1
L
f=
xi xk+i . Then each variable xi , 0 ≤ i ≤ n − 1, appears in two distinct
i=0
monomials in f and there are n (even) monomials in f . We will first show
f (vi ) = f (v2n −1−i ), 0 ≤ i ≤ n − 1, and hence fˆ(vi ) = fˆ(v2n −1−i ). Notice
f (vi ) = 0 when there are an even number of monomials in f which have a
value of 1 and f (vi ) = 1 when there are an odd number of monomials which
have a value of 1.
Fix i = I and let wt(vI ) = l. Suppose f (vI ) has r, s, and t monomials
of the form xj xh with {xj , xh } = {1, 1}, {1, 0}, and {0, 0} respectively, so
f (vI ) ≡ r mod 2. Since each variable xi appears in two distinct monomials
of f , then by counting the number of times xi = 1 in f we see 2r+s = 2l and
s must be even. Also, we know there are n monomials in f so r + s + t = n.
Clearly f (vI ) = 0 ⇐⇒ r is even ⇐⇒ t is even, since s, n are both even
and r + s + t = n. It is easy to see that f (v2n −1−I ) = f (v¯I ) ≡ t mod 2.
Thus f (v2n −1−I ) = 0 ⇐⇒ t is even.
Hence for any i, 0 ≤ i ≤ 2n − 1, we have f (vi ) = f (v2n −1−i ) and
therefore fˆ(vi ) = fˆ(v2n −1−i ). So (2) becomes:
M0,2n −1 =
2n−1
X−1
2(fˆ(vi ))2 = 2n .
i=0
Thus if k 6= m then M0,2n −1 6= 0 and we have that (fˆ(vi ⊕ vj )) is not
Hadamard. Thus by Theorem 1.1 f is not bent. The remainder of the
theorem follows from Lemma 2.1.
Corollary 2.8. If n = 2m, then any degree 2 rotation symmetric bent
m−1
L
function must contain f0 =
xi xm+i .
i=0
Proof. Assume f is a rotation symmetric function of degree 2. By [1, Th.
5.29, p. 83] we can assume f has only quadratic monomials, so f has SANF
n−1
L
ai x0 xi where the coefficients ai ∈ F2 . Let gk be the function with SANF
i=1
ak x0 xk where 1 ≤ k ≤ n − 1. Suppose f does not contain f0 , then a n2 = 0.
n−1
L
We have gk = ak
xi xk+i when 1 ≤ k ≤ n − 1 and from the proof of
i=0
Theorem 2.7, since an/2 = 0 we know gˆk (vi ) = gˆk (v2n −1−i ) for 1 ≤ i ≤ n−1.
\
n−1
n−1
n−1
n−1
Q
Q
Q
L
gˆk (v2n −1−i ) =
gˆk (vi ) =
gˆk . Thus fˆ(vi ) =
gk =
Also, fˆ =
k=1
k=1
k=1
7
k=1
fˆ(v2n −1−i ). So (2) gives:
M0,2n −1 =
2n−1
X−1
2(fˆ(vk ))2 = 2n .
k=0
Thus M0,2n −1 6= 0, so (fˆ(vi ⊕ vj )) is not Hadamard and therefore by Theorem 1.1 f is not bent.
Necessary and sufficient conditions for a bent quadratic function are
given in [6, Lemma 1, p. 4909]. The Lemma gives an alternate proof of
Corollary 2.8 as discussed in [6, Rmk. 1, p. 4910]. A characterization of
any bent RS function of degree 2 follows from [4, Th. 3.7, p. 6]. Examples
show that [4, Th. 3.1, p. 3] is false.
L
We will now look at rotation symmetric functions f =
fi of any degree
i
where each of the fi is a short-cycle MRS Boolean function. We will prove
that f is bent only if fi = f0 for some i. To do this we need the following
result from [5, pp. 218].
Theorem 2.9. Let f be a function on Vn and J ⊂ {0, 1, 2, . . . , n − 1} such
that f does not contain any term xj1 · · · xjt where t > 1 and j1 , . . . , jt ∈ J.
Then the nonlinearity of f satisfies Nf ≤ 2n−1 − 2s−1 , where s = |J|.
Theorem 2.10. Let n = 2m, then f0 =
m−1
L
xi xm+i is the only bent short-
i=0
cycle MRS Boolean function in n variables.
Proof. Let f have SANF x0 xa1 · · · xad−1 and suppose f is a short-cycle function. Then the number of monomials in f is k where 1 ≤ k < 2m. Thus,
f = x0 xa1 · · · xad−1 ⊕ x1 xa1 +1 · · · xad−1 +1 ⊕ · · · ⊕ xk−1 xa1 +k−1 · · · xad−1 +k−1
It is easy to see then that {0, a1 , . . . , ad−1 } = {k, a1 + k, . . . , ad−1 + k} =
. . . = {lk, a1 + lk, . . . , ad−1 + lk} where lk < 2m. By Lemma 3.1 we know
k must divide n = 2m, thus 1 ≤ k ≤ m. Thus by rotation symmetry
the tuple {ai , ai + k, . . . , ai + lk} appears in each monomial of f , for every
i, 0 ≤ i ≤ d − 1. If k = m then l = 1 and f has SANF
f = x0 xa1 xa2 . . . xag−1 xm xm+a1 . . . xm+ag−1 ,
(3)
where 0 < a1 < a2 < . . . < ag−1 < m. First consider g > 1. Suppose
ak < m and ak + 1 = m; then m + ak < 2m and m + ak + 1 = 2m. So
8
if xak +1 = xm then xm+ak +1 = x0 . Thus for each monomial in f, g of the
variables xi have i < m and g have i ≥ m. Let J = {m − 1, m, . . . , n − 1}.
Thus by Theorem 2.9 we have Nf ≤ 2n−1 − 2m < 2n−1 − 2m−1 , so f is not
bent by Theorem 1.1. If g = 1 then f = f0 which we know is bent. Thus if
k = m then f is bent only if f = f0 . If k < m then by regrouping the terms
we have that,
f=
k−1
M
xi xb1 +i · · · xbg +i xk+i xb1 +k+i · · · xbg +k+i · · · xlk+i xb1 +lk+i · · · xbg +lk+i
i=0
From the first term in the monomial, xi , we see that each monomial contains
at least one element of the set {x0 , x1 , . . . , xk−1 }. Since k < m then k − 1 <
m − 1. So if we let J = {m − 1, m, m + 1, . . . , n − 1}, then by Theorem 2.9
we have Nf ≤ 2n−1 − 2m < 2n−1 − 2m−1 . Thus f is not bent by Theorem
1.1.
Theorem 2.11. Let n =
L2m and let f in n variables be a sum of MRS shortcycle functions, so f =
fi , where each fi has fewer than n monomials. If
i
f is bent then one of the fi is f0 =
m−1
L
xi xm+i .
i=0
Proof. Assume f is bent and no fi is f0 . Then any short-cycle of the form
in equation (3) has g > 1 and from the proof of Theorem 2.10 we see that
each monomial in f contains at least one variable xi where i < m − 1.
Thus if we let J = {m − 1, m, . . . , n − 1} then f does not contain any
term xj1 · · · xjt where t > 1 and ji ∈ J; hence from Theorem 2.9 we have
Nf ≤ 2n−1 − 2m < 2n−1 − 2m−1 . Thus f cannot be bent, contradiction.
Therefore one of the fi is f0 .
3
Homogeneous rotation symmetric bent functions
when n=2p
Using ideas similar to those in section 2, we can prove that the only homogeneous MRS bent function in n = 2p variables, where p > 2 is prime, is
p−1
L
f0 =
xi xp+i . By a different method, this result was proven for any n by
i=0
Meng et al. [3, Th. 11, pp. 1114]
Notice we have already proven this for some cases in section 2. For the
remainder of the cases we must first further simplify (2). To do this we will
need a few facts about the rotation symmetric equivalence classes of Vn .
9
Lemma 3.1. |On (vi )| divides n.
Proof. On (vi ) is the orbit generated by vi ∈ Vn under the action of G
where G is the group of left cyclic shifts. Thus |On (vi )| = |G : Gvi |, where
Gvi = {ρ ∈ G|ρ(vi ) = vi }. Since |G| = n then |On (vi )| divides n.
Lemma 3.2. |On (vi )| = 2 ⇐⇒ vi ∼ v 2n −1 .
3
Proof. ⇒: |On (vi )| = 2 ⇐⇒ On (vi ) = {(1, 0, 1, 0, . . . , 1, 0), (0, 1, 0, 1, . . . , 0, 1)}.
Let vi = (0, 1, 0, 1, . . . , 0, 1) ⇒ i = 1 + 22 + 24 + . . . + 2n−2 ⇒
3i = (1 + 2)i = (1 + 2)(1 + 22 + 24 + . . . + 2n−2 )
= (1 + 22 + 24 + . . . + 2n−2 ) + (2 + 23 + 25 + . . . + 2n−1 )
= 1 + 2 + 22 + 23 + 24 + . . . + 2n−1
n
⇒ v3i = v2n −1 since v2n −1 = (1, 1, 1, . . . , 1). So we have i = 2 3−1 . Since
(0, 1, 0, 1, . . . , 0, 1) ∼ (1, 0, 1, 0, . . . , 1, 0) we have (1, 0, 1, 0, . . . , 1, 0) ∼ v 2n −1 .
3
⇐: Reverse the argument used above.
M0,2n −1
n
2
is an odd prime, then
X
fˆ(vk )fˆ(v2n −1−k )
= 2 + 2fˆ(v0 )fˆ(v2n −1 ) + 2n
Theorem 3.3. If n is even and
vk ∈Gn
|On (vk )|=n
vk ≁v2n −1−k
+n
X
fˆ(vk )fˆ(v2n −1−k ) + n
vk ∈Gn
|On (vk )|= n
2
X
(fˆ(vk ))2 . (4)
vk ∈Gn
|On (vk )|=n
vk ∼v2n −1−k
Proof. Let n = 2p where p is an odd prime. Then by Lemma 3.1, |On (vi )| =
1, 2, p, or n.
Let vi ∈ Vn such that |On (vi )| = 1, then vi = v0 or v2n −1 which appears
as the first term of (2) with k = 0.
n
n
If |On (vi )| = 2 then vi = v 2n −1 and 2n − 1 − 2 3−1 = 2 2 3−1 , so vi ∼
3
v2n −1−i and we have one 2(fˆ(v 2n −1 )2 term. Since fˆ(vi ) = ±1 for any
3
n
vi ∈ Vn then (fˆ(vi ))2 = 1 for any vi . So when k = 2 3−1 in (2), then
2fˆ(vk )fˆ(v2n −1−k ) = 2.
Now suppose vi = (a0 , a1 , . . . , an−1 ) ∈ Gn and |On (vi )| = p, then
vi′ = vi′′ . So to consider the cyclic shifts of vi we need only consider the cyclic
shifts of (a0 , a1 , . . . , ap−1 ). Notice v2n −1−i = (1 + a0 , 1 + a1 , . . . , 1 + an−1 )
and we have (1 + a0 , 1 + a1 , . . . , 1 + ap−1 ) = (1 + ap , 1 + ap+1 , . . . , 1 + an−1 )
10
so |On (v2n −1−i )| = p as well. Since p is odd (1 + a0 , 1 + a1 , . . . , 1 + ap−1 ) will
have a different number of 1’s than (a0 , a1 , . . . , ap−1 ). Hence vi ≁ v2n −1−i
since (a0 , a1 , . . . , ap−1 ) ≁ (1+a0 , 1+a1 , . . . , 1+ap−1 ). Thus from Lemma 2.4,
if |On (vi )| = p, then grouping the corresponding terms in (2) by the equivalence class representatives gives 2pfˆ(vi )fˆ(v2n −1−i ) = nfˆ(vi )fˆ(v2n −1−i ).
If |On (vi )| = n and vi ≁ v2n −1−i then grouping the corresponding terms
in (2) by the equivalence class representative gives the new coefficient 2n. If
vi ∼ v2n −1−i then this coefficient becomes 2 n2 = n.
Thus (2) reduces to the above equation.
Lemma 3.4. Let n = 2p, where p > 2 is prime. If f is a rotation symmetric
bent Boolean function in n variables, then fˆ(v0 ) = −fˆ(v2n −1 ).
Proof. From Theorem 1.1 f is bent if and only if M = 2n I2n . Thus if f is
bent then M0,2n −1 = 0. From (4) we see that in order for M0,2n −1 to be 0
we need the 2 which appears from the case vk = v 2n −1 to cancel. Let
3
X
a = fˆ(v0 )fˆ(v2n −1 ), b =
fˆ(vk )fˆ(v2n −1−k ),
vk ∈Gn
|On (vk )|=n
vk ≁v2n −1−k
c=
X
fˆ(vk )fˆ(v2n −1−k ), and d =
vk ∈Gn
|On (vk )|= n
2
X
(fˆ(vk ))2 .
vk ∈Gn
|On (vk )|=n
vk ∼v2n −1−k
Assume f is bent, then (4) becomes:
2 + 2a + 4pb + 2pc + 2pd = 0 ⇒ p(2b + c + d) = −1 − a.
If a = 1, then p(2b + c + d) = −2. Since p > 2 and 2b + c + d ∈ Z
this is a contradiction. Thus a = −1 and fˆ(v0 )fˆ(v2n −1 ) = −1. Therefore
fˆ(v0 ) = −fˆ(v2n −1 ) since fˆ(vi ) = ±1 for all i.
The following corollary now proves that full-cycle homogeneous MRS
Boolean functions in 2p variables cannot be bent.
Corollary 3.5. Let n = 2p where p is an odd prime. Let f be an MRS
boolean function with SANF x0 xa1 xa2 . . . xad−1 . If the number of monomials
in f is n, then f is not bent.
Proof. Using Theorem 1.1 we need only show M0,2n −1 6= 0. If the number
of monomials in f is n, then we have fˆ(v0 ) = fˆ(v2n −1 ) since n = 2p is even.
Thus M0,2n −1 6= 0 so f is not bent.
11
We can now prove that f0 is the only homogeneous MRS bent function.
Theorem 3.6. Let n = 2p, where p > 2 is prime. The only bent homogep−1
L
neous MRS Boolean function in n variables is f0 =
xi xp+i .
i=0
Proof. Any monomial homogeneous rotation symmetric function, f , has
SANF x0 xa1 . . . xad−1 and is either short-cycle or full-cycle. If f is fullcycle then from Corollary 3.5 f is not bent. If f is short-cycle then from
Theorem 2.10, f is bent if and only if f = f0 .
Theorem 3.7. Let n = 2p, where p > 2 is prime, and let f be a rotation
symmetric Boolean function in n variables with SANF x0 xa1 · · · xad−1 . If f
has p monomials, then d is even and f has SANF given by equation (3).
Proof. If the number of monomials in f is p then
f = x0 xa1 · · · xad−1 ⊕ x1 xa1 +1 · · · xad−1 +1 ⊕ · · · ⊕ xp−1 xa1 +p−1 · · · xad−1 +p−1
Thus {0, a1 , . . . , aad−1 } = {p, a1 + p, . . . , ad−1 + p} so each pair {xai , xai +p }
appears in every monomial and f is of the form in (3). The result that d is
even follows immediately since each monomial contains the pairs {xai , xai +p }.
Theorem 3.8. Let n = 2p, where p > 2 is prime. Let f =
s
L
fi where
i=1
each fi has SANF x0 xai1 · · · xaid−1 and deg(f ) = d. Let r be the number of
fi that are short-cycle and l be the number of fi that are long-cycle. If f is
bent then d is even. Furthermore, if d = 2 then f contains f0 and if d 6= 2
then r ≥ 1 is odd and l ≥ 1.
Proof. Let f = f1 ⊕ f2 ⊕ · · · ⊕ fs where each fi has SANF x0 xai1 . . . xaid−1 .
Thus deg(fi ) = d ∀i and so deg(f ) = d. Since n = 2p then GCD(n, d) =
1, 2, p, or n and from Lemma 3.1 we know that the number of monomials in
each fi is either 1, 2, p, or n.
(i) Suppose the number of monomials in fi for some i is 1, then since fi
is rotation symmetric then fi = x0 x1 . . . xn−1 and d = n.
(ii) Suppose the number of monomials in fi for some i is 2, then f =
x0 xai1 . . . xaid−1 ⊕ x1 xai1 +1 . . . xaid−1 +1 . It is easy to see by induction
that the first monomial contains all of the even indices and the second
monomial contains all of the odd indices. Thus d = p.
12
(iii) Suppose the number of monomials in fi for some i is p, then from
Theorem 3.7 fi is of the form in equation (3) and d is even.
Thus fi is a short-cycle function only if d = p, n, or is even. Let GCD(n, d) =
1, then the number of monomials in fi is n for all i. Thus, fˆi (v0 ) =
fˆi (v2n −1 ) ∀i and so fˆ(v0 ) = fˆ(v2n −1 ) and by Lemma 3.4 f is not bent.
If GCD(n, d) = p then fi either has 2 monomials or n monomials for each
i. Thus fˆ(v0 ) = fˆ(v2n −1 ) and so f is not bent. If GCD(n, d) = n then
d = n and f is not bent by Lemma 1.2. Now suppose GCD(n, d) = 2, thus
d is even. If d = 2 then by Corollary 2.8 f must contain f0 which is the
only degree 2 short-cycle. If d 6= 2 then by (iii), the short-cycle functions of
degree d are of the form in equation (3). By Theorem 2.11 we know that
any combination of these short-cycle functions is not bent, thus f must also
contain a long-cycle and so l ≥ 1. If r is even then,
fˆ(v0 ) =
s
Y
fˆi (v0 ) = (−1)r
s
Y
fˆi (v2n −1 ) = fˆ(v2n −1 )
i=1
i=1
since fˆi (v0 ) = −fˆi (v2n −1 ) when fi is short-cycle and fˆi (v0 ) = fˆi (v2n −1 )
when fi is long-cycle. Thus by Lemma 3.4 f is not bent.
Remark. The previous Theorem says that any homogeneous RotS bent
function must have even degree and contain an odd number of short-cycle
MRS functions. If the degree is greater than 2 then it must also contain at
least one long-cycle MRS function.
Lemma 3.9. Let n = 2p, p an odd prime, and f be a homogeneous RotS
Boolean function in n variables with deg(f ) = d = 2k where k > 1. Let
s
l
L
L
f =
fi ⊕
fi where for all 1 ≤ i ≤ s, fi is an MRS short-cycle
i=1
i=s+1
function and for all s + 1 ≤ i ≤ l, fi is an MRS full-cycle function. Suppose
fi has SANF x0 xai1 . . . xaid−1 . If 2 ≤ r ≤ d − 2 for all s + 1 ≤ i ≤ l, where
r is the number of odd indices in {ai1 , . . . , aid−1 }, then f is not bent.
Proof. Since the degree of f is even, we know from the proof of Theorem 3.8
that any short-cycle is of the form in (3). Thus no short-cycle contains a term
xj1 . . . xjt where t > 1 and j1 , . . . , jt are in the set J = {0, 2, 4, . . . , n−2, n−1}
since deg(f ) ≥ 4. If the number of odd indices, r, in the first monomial of
any full-cycle in f is 2 ≤ r ≤ d − 2, then each monomial in the full-cycles
has either r odd indices and d − r even indices or d − r odd indices and
r even indices. Since 2 ≤ r ≤ d − 2 then 2 ≤ d − r ≤ d − 2, thus each
13
monomial contains at least 2 odd indices. Thus, because J contains only
one odd number, n − 1, then no monomial, xj1 . . . xjt , in a long-cycle has
xj1 , . . . , xjt ∈ J where t > 1. Thus by Theorem 2.9, Nf ≤ 2n−1 − 2p <
2n−1 − 2p−1 and so f is not bent by Theorem 1.1.
L
Remark. Lemma 3.9 says that if f =
fi is a homogeneous RotS Boolean
i
function of even degree d, d > 2, in n = 2p variables then one of the fi must
have a monomial with exactly one odd index.
L
Let f be a homogeneous RotS function of degree d with SANF
βi ,
1≤i≤s
where βi = xk(i) xk(i) · · · xk(i) and
0
(i)
2
d−1
(i)
sj , j = 1, 2, . . . , d, by sj
(i)
(i)
k0
(i)
= 0 for all i. Define a sequence
(i)
(i)
= kj − kj−1 for 1 ≤ j ≤ d − 1, and sd =
(i)
k0 + n − kd−1 . Let sf be the largest distance between two consecutive
(i)
indices in all of the monomials in f , thus sf = max sj . Then from [3, Th.
i,j
13, pp. 1116] we have the following:
Theorem 3.10. Let f be a homogeneous RotS function with degree d ≥ 3
in n variables. If sf ≤ n2 , then f is not bent.
We can use Theorem 3.10 to get a useful bound on the degree of any
possible homogeneous RotS bent function in 2p variables:
Theorem 3.11. Let f be a homogeneous RotS bent function of degree d ≥ 3
in n = 2p variables, p an odd prime. Then d ≤ (p + 3)/2.
Proof. Suppose d > (p + 3)/2. We choose J = {0, 1, 2, 4, . . . , 2p − 2} (so
|J| = p + 1 and J has p even elements) in Theorem 2.9. The theorem
applies since now no monomial x0 xa1 . . . xad−1 can have all of its variables
xj in J and also have a gap of length ≥ (n/2) + 1 in its index set (necessary
for f to be bent by Theorem 3.10). Suppose the monomial x0 xa1 . . . xad−1
has all of its variables xj in J and d > p+3
2 . If there is a gap in the indices
greater than or equal to p + 1, then at least p−1
2 of the indices in J do not
p+3
appear in the monomial. Thus there are at most p + 1 − p−1
2 = 2 possible
indices in J which appear in the monomial. Thus d ≤ p+3
2 , contradiction.
Note, for example, that if the indices are 0, 1, 2, p + 3 p + 5, . . . , 2p − 2 then
| {z }
gap
we have a gap of length p + 1 but only (p + 3)/2 indices, contradicting our
assumption about d. Now Theorem 2.9 gives
14
Nf ≤ 2n−1 − 2p ≤ 2n−1 − 2n/2 < 2n−1 − 2n/2−1 ,
so f is not bent by Theorem 1.1, contradicting our hypothesis.
4
Nonhomogeneous rotation symmetric bent functions when n=2p
We have already shown in Theorem 2.11 that any bent function composed
only of short-cycle MRS functions must contain f0 .
We can now extend the ideas used in section 3 to show that in most
cases any bent rotation symmetric Boolean function in n = 2p, p > 2 prime,
variables must contain f0 .
Theorem 4.1. Let n = 2p where p is an odd prime. Let f = f1 ⊕ f2 ⊕
. . . ⊕ fs be a rotation symmetric Boolean function where each fi has SANF
x0 aa1 . . . xadi −1 . If di = 2 or the number of monomials in each fi is either
p−1
L
2 or n, then f is bent only if f contains f0 =
xj xp+j . If f contains an
j=0
even number of fi , where the number of monomials in fi is p and all other
fi contain 2 or n monomials, then f is not bent.
Proof. Let f = f1 ⊕ f2 ⊕ . . . ⊕ fs where for each fi , di = 2 or the number
of monomials in fi is 2 or n. Suppose f is bent and does not contain f0 .
Then since f0 is the only short-cycle RotS function with degree 2 then each
fi contains either 2 or n monomials. Thus from the proof of Corollary 3.5,
fˆi (v0 ) = fˆi (v2n −1 ) for all i. Thus
Y
Y
fˆ(v0 ) =
fˆi (v0 ) =
fˆi (v2n −1 ) = fˆ(v2n −1 )
and from Lemma 3.4 we see that f cannot be bent which is a contradiction.
Thus f must contain f0 .
Now suppose the number of monomials in f1 , f2 , . . . , f2r is p and
f2r+1 , f2r+2 , . . . , fs contain either 2 or n monomials. Then fˆi (v0 ) = −fˆi (v2n −1 )
for 1 ≤ i ≤ 2r and fˆi (v0 ) = fˆ(v2n −1 ) for 2r + 1 ≤ i ≤ s. Thus,
fˆ(v0 ) =
s
Y
fˆi (v0 ) =
i=1
2r
Y
−fˆi (v2n −1 )
s
Y
i=2r+1
i=1
= fˆ(v2n −1 ).
Thus from Lemma 3.4, f cannot be bent.
15
fˆi (v2n −1 ) = (−1)2r fˆ(v2n −1 )
Remark. If fi has 1 monomial, for some i, then deg(fi ) = n. Thus deg(f ) =
n and so f is not bent from Lemma 1.2.
Remark. The previous Theorem says that any nonhomogeneous RotS bent
function must contain an odd number of short-cycle MRS functions of even
degree and at least one function which is a long-cycle MRS function.
References
[1] T. W. Cusick and P. Stănică, Cryptographic Boolean Functions and
Applications (San Diego: Academic Press, 2009).
[2] H. Kim, S-M. Park and S. G. Hahn, On the weight and nonlinearity of homogeneous rotation symmetric Boolean functions of degree 2,
Discrete Applied Mathematics 157, 428-432 (2009)
[3] Q. Meng, L. Chen, F.-W. Fu, On homogeneous rotation symmetric bent
functions, Discrete Appl. Math. 158 (2010), 1111–1117.
[4] X. Zhang, G. Gao, On the conjecture about the nonexistence of rotation
symmetric bent functions, arXiv.org, arXiv:1303.2282v1, 7 pp. (2013).
[5] Y. Zheng, X-M. Zhang, H. Imai, Restriction, terms and nonlinearity of
Boolean functions, Theoretical Computer Science 226, 207-223 (1999)
[6] Guangpu Gao, Xiyong Zhang, Wenfen Liu, and Claude Carlet, Constructions of Quadratic and Cubic Rotation Symmetric Bent Functions,
IEEE Transactions on Information Theory 58, 4908-4913 (2012)
16
| 7cs.IT
|
Maximizing the information learned from finite
data selects a simple model
Henry H. Mattinglya , Mark K. Transtrumb , Michael C. Abbottc,1 , and Benjamin B. Machtad,1
a
Lewis-Sigler Institute and Department of Chemical and Biological Engineering, Princeton University, Princeton, NJ 08544, USA; b Department of Physics and Astronomy,
Brigham Young University, Provo, Utah 84602, USA; c Marian Smoluchowski Institute of Physics, Jagiellonian University, Ulica Łojasiewicza 11, 30-348 Kraków, Poland;
d
Lewis-Sigler Institute and Department of Physics, Princeton University, Princeton, NJ 08544, USA
arXiv:1705.01166v3 [physics.data-an] 14 Feb 2018
PNAS (7 February 2018) + SI ≈ arXiv:1705.01166v3
Earlier versions had the title “Rational Ignorance: Simpler Models Learn More Information from Finite Data”
We use the language of uninformative Bayesian prior choice to study
the selection of appropriately simple effective models. We advocate
for the prior which maximizes the mutual information between parameters and predictions, learning as much as possible from limited data.
When many parameters are poorly constrained by the available data,
we find that this prior puts weight only on boundaries of the parameter manifold. Thus it selects a lower-dimensional effective theory in a
principled way, ignoring irrelevant parameter directions. In the limit
where there is sufficient data to tightly constrain any number of parameters, this reduces to Jeffreys prior. But we argue that this limit is
pathological when applied to the hyper-ribbon parameter manifolds
generic in science, because it leads to dramatic dependence on effects invisible to experiment.
The theory and the experiment are together described by a
probability distribution p(x|θ), for each value of the theory’s
parameters θ ∈ Θ. This function encodes both the quality and
quantity of data to be collected.
The mutual information between the parameters and their
expected data is defined as MI = I(X; Θ) = S(Θ) − S(Θ|X),
where S is the Shannon entropy (23). The MI thus quantifies
the information which can be learned about the parameters
by measuring the data, or equivalently, the information about
the data which can be encoded in the parameters (24, 25).
Defining p? (θ) by maximizing this, we see:
i. The prior p? (θ) is almost always discrete (26–30), with
weight only on a finite number
PK K of points, or atoms
(Figures 1 and 2): p? (θ) = a=1 λa δ(θ − θa ).
Effective Theory | Model Selection | Renormalization Group | Bayesian
Prior Choice | Information Theory
ii. When data is abundant, p? (θ) approaches Jeffreys prior
pJ (θ) (31–33). As this continuum limit is approached,
the proper spacing of the atoms shrinks as a power law
(Figure 3).
P
hysicists prefer simple models not because nature is simple, but because most of its complication is usually irrelevant. Our most rigorous understanding of this idea comes from
the Wilsonian renormalization group (1–3), which describes
mathematically the process of zooming out and losing sight
of microscopic details. These details only influence the effective theory which describes macroscopic observables through
a few relevant parameter combinations, such as the critical
temperature, or the proton mass. The remaining irrelevant
parameters can be ignored, as they are neither constrained
by past data nor useful for predictions. Such models can now
be understood as part of a large class called sloppy models
(4–14), whose usefulness relies on a similar compression of a
large microscopic parameter space down to just a few relevant
directions.
This justification for model simplicity is different from the
one more often discussed in statistics, motivated by the desire
to avoid overfitting (15–21). Since irrelevant parameters have
an almost invisible effect on predicted data, they cannot be
excluded on these grounds. Here we motivate their exclusion
differently: we show that simplifying a model can often allow it
to extract more information from a limited data set, and that
this offers a guide for choosing appropriate effective theories.
We phrase the question of model selection as part of the
choice of a Bayesian prior on some high-dimensional parameter space. In a set of nested models, we can always move
to a simpler model by using a prior which is nonzero only on
some subspace. Recent work has suggested that interpretable
effective models are typically obtained by taking some parameters to their limiting values, often 0 or ∞, thus restricting to
lower-dimensional boundaries of the parameter manifold (22).
Our setup is that we wish to learn about a theory by
performing some experiment which produces data x ∈ X.
iii. When data is scarce, most atoms lie on boundaries of
parameter space, corresponding to effective models with
fewer parameters (Figure 4). The resulting distribution of
weight along relevant directions is much more even than
that given by Jeffreys prior (Figure 5).
Significance Statement
Most physical theories are effective theories, descriptions at
the scale visible to our experiments which ignore microscopic
details. Seeking general ways to motivate such theories, we
find an information theory perspective: if we select the model
which can learn as much information as possible from the
data, then we are naturally led to a simpler model, by a path
independent of concerns about overfitting. This is encoded as
a Bayesian prior which is nonzero only on a subspace of the
original parameter space. We differ from earlier prior selection
work by not considering infinitely much data. Having finite data
is always a limit on the resolution of an experiment, and in our
framework this selects how complicated a theory is appropriate.
H.H.M. and M.C.A. performed the numerical experiments. H.H.M., M.K.T., M.C.A. and B.B.M. designed the research, interpreted results, and contributed to writing. M.C.A. and B.B.M. led the
writing of the paper.
Present affiliation for H.H.M.: Department of Molecular Cellular and Developmental Biology, Yale
University, New Haven, CT, 06520
Present affiliations for B.B.M: Department of Physics, Yale University, New Haven, CT, 06520
Systems Biology Institute, Yale University, West Haven, CT, 06516
1
To whom correspondence should be addressed. E-mail: [email protected] and
[email protected]
A
B
C
Fig. 1. Optimal priors for the Bernoulli model (1). Red lines indicate the positions
of delta functions in p? (θ), which are at the maxima of fKL (θ), Eq. (3). As
m → ∞ these coalesce into Jeffreys prior pJ (θ).
After some preliminaries, we demonstrate these properties in
three simple examples, each a stylized version of a realistic
experiment. To see the origin of discreteness, we study the bias
of an unfair coin and the value of a single variable corrupted
with Gaussian noise. To see how models of lower dimension
arise, we then study the problem of inferring decay rates in a
sum of exponentials.
In the Appendix we discuss the alorithms used for finding
p? (θ) (Figures 2 and S3), and we apply some more traditional
model selection tools to the sum of exponentials example
(Figure S1).
Priors and Geometry
Bayes’ theorem tells us how to update our knowledge of θ
upon observing data x, from prior
p(θ) to posterior p(θ|x) =
R
p(x|θ)p(θ)/p(x), where p(x) = dθ p(θ) p(x|θ). In the absence
of better knowledge we must pick an uninformative prior
which codifies our ignorance. The naive choice of a flat prior
p(θ) = const. has undesirable features, in particular making
p(x) depend on the choice of parameterization, through the
measure dθ.
Jeffreys prior pJ (θ) is invariant under changes of parameterizaion, because it is constructedpfrom some properties of
the experiment (34). This pJ (θ) ∝ det gµν is, up to normalization, the volume form arising from the Fisher information
metric (FIM, often -matrix):
Bernardo defined a prior p? (θ) by maximizing the mutual
information between parameters Θ and the expected data X m
from m repetitions, and then a reference prior by taking the
limit m → ∞ (29, 31). Under certain benign assumptions,
this reference prior is exactly Jeffreys prior (31–33), providing
an alternative justification for pJ (θ).
We differ in taking seriously that the amount of data collected is always finite.∗ Besides being physically unrealistic,
the limit m → ∞ is pathological both for model selection
and prior choice. In this limit any number of parameters
can be perfectly inferred, justifying an arbitrarily complicated
model. In addition, in this limit the posterior p(θ|x) becomes
independent of any smooth prior.†
Geometrically, the defining feature of sloppy models is that
they have a parameter manifold with hyper-ribbon structure
(6–9): there are some long directions (corresponding to d relevant, or stiff, parameters) and many shorter directions (D − d
irrelevant, or sloppy, parameter combinations). These lengths
are often estimated using the eigenvalues of gµν , and have
logarithms that are roughly evenly-spaced over many orders
of magnitude (4, 5). The effect of coarse-graining is to shrink
irrelevant directions (here using the technical meaning of irrelevant: a parameter which shrinks under renormalization group
flow) while leaving relevant directions extended, producing a
sloppy manifold (8, 14). By contrast the limit m → ∞ has the
effect of expanding all directions, thus erasing the distinction
between directions longer and shorter than the critical length
scale of (approximately) one standard deviation.
On such a hyper-ribbon, Jeffreys prior has an undesirable
feature: since it is constructed from the D-dimensional notion
of volume, its weight along the relevant directions always
depends on the volume of the D − d irrelevant directions. This
gives it extreme dependence on which irrelevant parameters
are included in the model.‡ The optimal prior p? (θ) avoids
this dependence because it is almost always discrete, at finite
m.§ It puts weight on a set of nearly distinguishable points,
closely spaced along the relevant directions, but ignoring the
irrelevant ones. Yet being the solution to a reparameterizationinvariant optimization problem, the prior p? (θ) retains this
good feature of pJ (θ).
Maximizing the mutual information was originally done to
calculate the capacity of a communication channel, and we
can borrow techniques from rate-distortion theory here: the
algorithms we use were developed there (37, 38), and the discreteness we exploit was discovered several times in engineering
(26–28, 39). In statistics, this problem is more often discussed
as an equivalent minimax problem (40). Discreteness was
∗
~ =
gµν (θ)
Z
~
~
~ ∂ log p(x|θ) ∂ log p(x|θ) .
dx p(x|θ)
µ
ν
∂θ
∂θ
This Riemannian metric defines a P
reparameterization-invariant
D
distance between points, ds2 =
g dθµ dθν . It meaµ,ν=1 µν
sures the distinguishability of the data which θ and θ + dθ are
expected to produce, in units of standard deviations. Repeating an (identical and independently distributed)
experiment
Qm
m times means considering pm (~
x|θ) = j=1 p(xj |θ), which
m
leads to metric gµν
(θ) = m gµν (θ). However the factor mD/2
in the volume is lost by normalizing pJ (θ). Thus Jeffreys prior
depends on the type of experiment, but not the quantity of
data.
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
2
Interned for five years, John Kerrich only flipped his coin 104 times (35). With
computers we can do better, but even the LHC only generated about 1018 bits of
data (36).
†
For simplicity we consider only regular models, i.e. we assume all parameters are
structurally identifiable.
‡
See Figure 5 for a demonstration of this point. For another example, consider a
parameter manifold Θ which is a cone, with Fisher metric ds2 = (50 dϑ)2 +
ϑ2 dΩ2n /4: there is one relevant direction ϑ ∈ [0, 1] of length L = 50, and n
irrelevant directions forming a sphere of diameter ϑ. Then the prior on ϑ alone
~ is p(ϑ) = (n + 1)ϑn , putting most of the weight near to ϑ = 1,
implied by pJ (θ)
dramatically so if n = D − d is large. But since only the relevant direction is visible
to our experiment, the region ϑ ≈ 0 ought to be treated similarly to ϑ ≈ 1. The prior
~ has this property.
p? (θ)
§
We offer both numerical and analytic arguments for discreteness below. The exception
to discreteness is that if there is an exact continuous symmetry, p? (θ) will be constant
along it. For example if our Gaussian model Eq. (2) is placed on a circle (identifying
both θ ∼ θ + 1 and x ∼ x + 1) then the optimum prior is a constant.
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
also observed in other minimax problems (41–43), and later
in directly maximising mutual information (29, 30, 33, 44).
However it does not seem to have been seen as useful, and
none of these papers explicitly find discrete priors in dimension
D > 1, which is where we see attractive properties. Discreteness has been useful, although for different reasons, in the idea
of rational inattention in economics (45, 46). There, market
actors have a finite bandwidth for news, and this drives them
to make discrete choices despite all the dynamics being continuous. Rate-distortion theory has also been useful in several
areas of biology (47–49), and discreteness emerges in a recent
theoretical model of the immune system (50).
We view this procedure of constructing the optimal prior
as a form of model selection, picking out the subspace of Θ
on which p? (θ) has support. This depends on the likelihood
function p(x|θ) and the data space X, but not on the observed
data x. In this regard it is closer to Jeffreys’ perspective
on prior selection than to tools like the information criteria
and Bayes factors, which are employed at the stage of fitting
to data. We discuss this difference at more length in the
Appendix.
log p𝜏(𝜃)
𝜎 = 1/10
2
2
1
e−(x−θ) /2σ .
p(x|θ) = √
2πσ
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
𝜃, with steps Δ𝜃 = 1/300
�
�
log p⭑(𝜃)
dθ p(θ) fKL (θ),
[3]
Z
fKL (θ) = DKL p(x|θ) p(x) =
dx p(x|θ) log
p(x|θ)
p(x)
¶
where DKL is the Kullback–Leibler divergence.
MaximizR
ing MI over all functions p(θ) with dθ p(θ) = 1 gives
fKL (θ) = const. But the maximizing function will not, in
general, obey p(θ) ≥ 0. Subject to this inequality p? (θ) must
satisfy
[1]
{p? (θ) > 0, fKL (θ) = MI} or {p? (θ) = 0, fKL (θ) < MI}
at every θ. With finite data fKL (θ) − MI must be an analytic
function of θ, and therefore must be smooth with a finite
numbers of zeros, corresponding to the atoms of p? (θ) (Figure
1A). See (28, 29, 46) for related arguments for discreteness,
and (41–43) for other approaches.
The number of atoms occurring in p? (θ) increases as the
data improves. For K atoms there is an absolute bound
MI ≤ log K, saturated if they are perfectly distinguishable. In
Figure 2C we observe that the optimal priors instead approach
a line MI → ζ log K, with slope ζ ≈ 0.75. At large L the
length of parameter space is proportional to the number of
distinguishable points, hence MI → log L. Together these
imply K ∼ L1/ζ , and so the average number density of atoms
grows as
ρ0 = K/L ∼ L1/ζ−1 ≈ L1/3 ,
[2]
Repeated
measurements are equivalent to smaller σ (by σ →
√
σ/ m), so we fix m = 1 here. The Fisher metric is gθθ = 1/σ 2 ,
thus L = 1/σ. An optimal prior is shown in Figure 2, and
in Figure 5A along with its implied distribution of expected
data. This p(x) is similar to that implied by Jeffreys prior,
here pJ (θ) = 1.
We calculated p? (θ) numerically in two ways. After discretizing both θ and x, we can use the Blahut–Arimoto (BA)
�
Z
MI = I(X; Θ) =
p
m
,
θ(1−θ)
𝜃, with steps Δ𝜃 = 1/30
◼
◼
algorithm (37, 38). This converges to the global maximum,
which is a discrete distribution: see Figure 2. Alternatively,
using our knowledge that p? (θ) is discrete, we can instead
adjust the positions θa and weights λa of a finite number of
atoms. See the Supplement for more details.
To see analytically why discreteness arises, we write the
mutual information as
thus pJ (θ) = [π θ(1 − θ)]−1 , and
R√
√
proper parameter space length L =
ds2 = π m.
In the extreme case m = 1, the optimal prior is two delta
functions, p? (θ) = 12 δ(θ) + 12 δ(θ − 1), and MI = log 2, exactly
one bit (29, 30, 33). Before an experiment that will only
ever be run once, this places equal weight on both outcomes;
afterwards it records the outcome. As m increases, weight is
moved from the boundary onto interior points, which increase
in number, and ultimately approach the smooth pJ (θ): see
Figures 1 and 2A.
Similar behavior is seen in a second example, in which we
measure one real number x, normally distributed with known
σ about the parameter θ ∈ [0, 1]:
This gives gθθ =
◼
Fig. 2. Convergence of the Blahut–Arimoto algorithm. This is for the oneparameter Gaussian model Eq. (2) with L = 10 (comparable to m = 10 in
Figure 1). On the right θ is discretized into ten times as many points, but pτ (θ)
clearly converges to the same five delta functions.
We begin with some problems with a single bounded parameter, of length L in the Fisher metric. These tractable cases
illustrate the generic behaviour along either short (irrelevant)
or long (relevant, L 1) parameter directions in higherdimensional examples.
Our first example is the Bernoulli problem, in which we
wish to determine the probability θ ∈ [0, 1] that an unfair
coin gives heads, using the data from m trials. It is sufficient
to record the total number of heads x, which occurs with
probability
m!
θx (1 − θ)m−x .
x!(m − x)!
◼
◼
�
One-Parameter Examples
p(x|θ) =
𝜏 = 10 iterations
𝜏 = 100
𝜏 = 104
Exact p⭑(𝜃)
¶
L 1.
[4]
The function fKL (θ) is sometimes called the Bayes risk, as it quantifies how poorly
the prior will perform if θ turns out to be correct. One of the problems equivalent
to maximising the mutual information (40) is the minimax problem for this (see also
Figure 1):
max I(X; Θ) = min max fKL (θ) = min max
p(θ)
p(θ)
θ
R
dθp(θ)DKL [p(x|θ)kq(x)] .
q(x) p(θ)
The distributions we call expected data p(x) are also known as Bayes strategies, i.e.
distributions on X which are the convolution of the likelihood p(x|θ) with some prior
p(θ). The optimal q(x) from this third formulation (with minq(x) . . .) can be shown
to be such a distribution (40).
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
3
B
𝜃
r=
A
0
B
A
p
Bernoulli: L = ⇡ m
r=
=
k2
Gaussian: L = 1/𝜎
k1
C
1
y2
s = 1ê3
s = 1ê7
y1
‡
0.5
ʇ‡
ʇ‡
ʇÊÊÊ
‡‡
‡‡‡‡‡‡‡‡‡ ‡ ‡ ‡‡‡‡‡‡‡‡‡‡‡‡‡‡‡
‡‡ÊÊÊ
‡‡
‡‡
‡‡
‡‡
‡‡‡‡
‡
‡‡
‡
‡‡
‡‡
‡‡
‡‡
‡‡
‡‡‡
Ê
Ê
‡‡
‡‡
ʇÊ
‡Ê‡
ÊÊ
Ê
Ê
d eff
𝜁=1
𝜁 = 3/4
1
0
‡
‡
‡‡‡‡ ‡ ‡
‡‡‡‡
‡‡‡‡
2
Number of atoms K
Ê
ÊÊÊ
‡Ê‡
ʇÊ
ʇ‡
ʇ‡
‡‡‡‡
‡‡‡‡
‡‡‡
‡‡‡
5
Ê
Ê
rá2
Ê
Ê
rá1
rá0
0
2
Ê
Ê
Ê
Bernoulli
Gaussian
Ê
10 20
1ês
Ê
Ê
Ê
Ê
ÊÊ
Ê
s = 1ê19
2
Wr
I(X; ⇥)/ log 2
1
Ê
Ê
Ê Ê
Ê ÊÊ
Ê
ÊÊ
Ê Ê
r=
C
Ê Ê ÊÊ
Ê ÊÊÊÊ
Ê
ÊÊ
Ê Ê
Ê
Ê ÊÊ
s = 1ê50
50 100
1/𝜎
Fig. 3. Behavior of p? (θ) with increasing Fisher length. Panels A and B show
the atoms of p? (θ) for the two one-dimensional models as L is increased (i.e. we
perform more repetitions m or have smaller noise σ ). Panel C shows the scaling of
the mutual A
information (in bits) with the number of atoms K .The dashed line is the
bound MI ≤ log K , and the solid line is the scaling law MI ∼ 3/4 log K .
Thus the proper spacing between atoms shrinks to zero in
the limit of infinite data, i.e. neighboring atoms cease to be
distinguishable.
To derive this scaling law analytically, in a related paper
(51) we consider a field theory for the number density of atoms,
in which the entropy density (omitting numerical factors) is
B
2
S = const. − e−ρ [ρ4 (ρ0 )2 + 1].
From this we find ζ = 3/4, which is consistent with both
examples presented above.
Multi-parameter Example
In the examples above, p? (θ) concentrates weight on the edges
of its allowed domain when data is scarce (i.e. when m is
small or σ is large, hence L is small). We next turn to a multiparameter model in which some parameter combinations are
ill-constrained, and where edges correspond to reduced models.
The physical picture is that we wish to determine the
composition of an unknown radioactive source, from data of xt
Geiger counter clicks at some times t. As parameters we have
the quantities Aµ and decay constants kµ of isotopes µ. The
probability of observing xt should be a Poisson distribution (of
mean yt ) at each time, but we approximate these by Gaussians
to write:‖
p(~x|~
y) ∝
Y
2
e−(xt −yt )
/2σ 2
,
X
yt =
t
‖
Aµ e−kµ t .
[5]
µ
Using a normal distribution of fixed σ here is what allows the metric in Eq. (6) to
be so simple. However the qualitative behavior from the Poisson distribution is very
similar.
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
4
Fig. 4. Parameters and priors for the exponential model (5). Panel A shows the
area of the ~
y plane covered by all decay constants k1 , k2 ≥ 0. Panel B shows the
positions of the delta functions of the optimal prior p? (~
y ) for several values of σ , with
colors indicating the dimensionality r at each point. Panel C shows the proportion of
weight on these dimensionalities.
Fixing σt = σ = const. then brings us to a nonlinear leastsquares model of the type studied by (6, 7). This same model
also arises in other contexts, such as the asymptotic approach
to equilibrium of many dynamical systems (52).
We can see the essential behavior with just two isotopes
in fixed quantities: Aµ = 12 , thus yt = 12 (e−k1 t + e−k2 t ).
Measuring at only two times t1 and t2 , we almost have a
two-dimensional version of Eq. (2), in which the center of the
distribution ~
y = (y1 , y2 ) plays the role of θ above. The mapping between ~k and ~
y is shown in Figure 4A, fixing t2 /t1 = e.
The Fisher information metric is proportional to the ordinary
Euclidean metric for ~
y , but not for ~k:
gµν (~k) =
1 X ∂yt ∂yt
σ2
∂kµ ∂kν
⇐⇒
gst (~
y) =
t
1
δst .
σ2
[6]
Thus Jeffreys prior is a constant on the allowed region of the
~
y plane.
Then we proceed to find the optimum p? (~
y ) for this model,
shown in Figure 4B for various values of σ. When σ is large,
this has delta functions only in two of the corners, allowing
only k1 , k2 = 0 and k1 , k2 = ∞. As σ is decreased, new atoms
appear first along the lower boundary (corresponding to the
one-dimensional model where k1 = k2 ) and then along the
other boundaries. At sufficiently small σ, atoms start filling
in the (two-dimensional) interior.
To show this progression in Figure 4C, we define Ωr as
the total weight on all edges
PD of dimension r, and an effective
dimensionality deff =
r Ωr . This increases smoothly
r=1
from 0 towards D = 2 as the data improves.
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
A
Gaussian model: D = 1
σ = �/��
p(x) from p⭑(𝜃)
�★ (θ)
p(x) from Jeffreys
■
■
■
■
■
Discussion
𝜃
θ
�
B
Exponential
model: D = 2
�
Prior p? (~y)
�★ (�+ )
y+ =
y1 + y2+
p
2
s = 1ê7
σ = �/�
r=0
r=1
■
■
■
■
■
�
to equal decay constants will still have significant weight.
The fact that such edges give human-readable simpler models
(unlike arbitrary submanifolds) was the original motivation
for preferring them in (22), and it is very interesting that our
optimization procedure has the same preference.††
■
0
■
y1 + y2
y+ = p
�+ = (�� +�2� )/ �
�
Fig. 5. Distributions of expected data p(x) from different priors. Panel A is the
one-parameter Gaussian model, with L = 10. Panel B projects the two-parameter
exponential model onto the y1 + y2 direction, for σ = 1/7 where the perpendicular
direction should be irrelevant. The length
is about the same
√ of the relevant direction
p? (~y) of expected
as the one-parameter case: L+ = 7 2. Notice that the distribution
data p(x+ ) from Jeffreys
√prior here is quite different, with almost no weight at the
ends of the range (0 and 2), because this prior still weights the area not the length.
At medium values of σ, the prior p? (~
y ) almost ignores the
width of the parameter
manifold,
and
cares
mostly about its
√
length (L+ = 2/σ along the diagonal). This behavior is very
different to that of Jeffreys prior: in Figure 5B we demonstrate
this by plotting the distributions of data implied by these two
priors. Jeffreys puts almost no weight near the ends of the long
(i.e. stiff, or relevant) parameter’s range, because the (sloppy,
or irrelevant) width happens to be even narrower there than
in the middle. By contrast our effective model puts significant
weight on each end, much like the one-parameter model in
Figure 5A.
The difference between one and two parameters being relevant (in Figure 4B) is very roughly σ = 1/7 to σ = 1/50, a
factor 7 in Fisher length, thus a factor 50 in the number of repetitions m — perhaps the difference between a week’s data and
a year’s. These numbers are artificially small to demonstrate
the appearance of models away from the boundary: more
realistic models often have manifold lengths spread over many
orders of magnitude (5, 8), and thus have some parameters
inaccessible even with centuries of data. To measure these we
need a qualitatively different experiment, justifying a different
effective theory.
The one-dimensional model along the lower edge of Figure
4A is the effective theory with equal decay constants. This
remains true if we allow more parameters k3 , k4 , . . . in Eq. (5),
and p? (~
y ) will still place a similar weight there.∗∗ Measuring xt
also at later times t3 , t4 , . . . will add more thin directions to the
manifold (7), but the one-dimensional boundary corresponding
While the three examples we have studied here are very simple,
they demonstrate a principled way of selecting optimal effective
theories, especially in high-dimensional settings. Following
(45), we may call this rational ignorance.
The prior p? (θ) which encodes this selection is the maximally uninformative prior, in the sense of leaving maximum
headroom for learning from data. But its construction depends on the likelihood function p(x|θ), and thus it contains
knowledge about the experiment through which we are probing
nature. Jeffreys prior pJ (θ) also depends on the experiment,
but more weakly: it is independent of the number of repetitions m, precisely because it is the limit m → ∞ of the
optimal prior (32, 33).
Under either of these prescriptions, performing a second
experiment may necessitate a change in the prior, leading to
a change in the posterior not described by Bayes’ theorem. If
the second experiment is different from the first, then changing to Jeffreys prior for the combined experiment (and then
applying Bayes’ rule just once) will have this effect (56, 57).‡‡
Our prescription differs from Jeffreys in also regarding more
repetitions of an identical experiment as being different. Many
experiments would have much higher resolution if they could
be repeated for all eternity. The fact that they cannot is an
important limit on the accuracy of our knowledge, and our
proposal treats this limitation on the same footing as the rest
of the specification of the experiment.
Keeping m finite is where we differ from earlier work on
prior selection. Bernardo’s reference prior (31) maximizes the
same mutual information, but always in the m → ∞ limit
where it gives a smooth analytically tractable function. Using
I(X; Θ) to quantify what can be learned from an experiment
goes back to Lindley (24). That finite information implies a
discrete distribution was known at least since (26, 27). What
has been overlooked is that this discreteness is useful for
avoiding a problem with Jeffreys prior on the hyper-ribbon
parameter spaces generic in science (5): because it weights the
irrelevant parameter volume, Jeffreys prior has strong dependence on microscopic effects invisible to experiment. The limit
m → ∞ has erased the divide between relevant and irrelevant
parameters, by throwing away the natural length scale on the
parameter manifold. By contrast p? (θ) retains discreteness at
roughly this scale, allowing it to ignore irrelevant directions.
Along a relevant parameter direction this discreteness is no
worse than rounding θ to as many digits as we can hope to
measure, and we showed that in fact the spacing of atoms
decreases faster than our accuracy improves.
††
Edges of the parameter manifold give simpler models not only in the sense of
having fewer parameters, but also in an algorithmic sense. For example, the
Michaelis–Menten model is analytically solvable (53) in a limit which corresponds
to a manifold boundary (54). Stable linear dynamical systems of order n are model
∗∗
boundaries order n + 1 systems (55). Taking some parameter combinations to the
If we have more parameters than measurements then the model must be singular.
extreme can lock spins into Kadanoff blocks (54).
In fact the exponential model of Figure 4 is already slightly singular, since k1 ↔ k2
does not change the data; we could cure this by restricting to k2 ≥ k1 , or by working ‡‡ This view is natural in the objective Bayesian tradition, but see (58) and (59–61) for
with ~
y , to obtain a regular model.
alternative viewpoints.
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
5
Model selection is more often studied not as part of prior
selection, but at the stage of fitting the parameters to data.
From noisy data, one is tempted to fit a model which is more
complicated than reality; avoiding such overfitting improves
predictions. The AIC, BIC (15, 62) and related criteria (19,
20, 63–65) are subleading terms of various measures in the
m → ∞ limit, in which all (nonsingular) parameters of the
true model can be accurately measured. Techniques like MDL,
NML, and cross-validation (63, 66, 67) need not take this limit,
but all are applied after seeing the data. They favor minimally
flexible models close to the data seen, while our procedure
favors one answer which can distinguish as many different
outcomes as possible. It is curious that both approaches can
point towards simplicity. We explore this contrast in more
detail in the Appendix.§§
Being discrete, the prior p? (θ) is very likely to exclude
the true value of the parameter, if such a θtrue ∈ Θ exists.
This is not a flaw: the spirit of effective theory is to focus on
what is relevant for describing the data, deliberately ignoring
microscopic effects which we know to exist (68). Thus the
same effective theory can emerge from different microscopic
physics [as in the universality of critical points describing phase
transitions (69)]. The relevant degrees of freedom are often
quasiparticles [such as the Cooper pairs of superconductivity
(70)] which do not exist in the microscopic theory, but give
a natural and simple description at the scale being observed.
We argued here for such simplicity not on the grounds of the
difficulty of simulating 1023 electrons, nor of human limitations,
but based on the natural measure of information learned.
There is similar simplicity to be found outside of physics.
For example the Michaelis–Menten law for enzyme kinetics
(71) is derived as a limit in which only the ratios of some reaction rates matter, and is useful regardless of the underlying
system. In more complicated systems which we cannot solve
by hand, and for which the symmetries and scaling arguments
used in physics cannot be applied, we hope that our information approach may be useful for identifying the appropriately
detailed theory.
6. M. K. Transtrum, B. B. Machta and J. P. Sethna, Why are Nonlinear Fits
to Data so Challenging?, Phys. Rev. Lett. 104 (2010) 060201
[arXiv:0909.3884].
7. M. K. Transtrum, B. B. Machta and J. P. Sethna, Geometry of nonlinear
least squares with applications to sloppy models and optimization,
Phys. Rev. E83 (2011) 036701 [arXiv:1010.1449].
8. B. B. Machta, R. Chachra, M. K. Transtrum and J. P. Sethna, Parameter
Space Compression Underlies Emergent Theories and Predictive
Models, Science 342 (2013) 604–607 [arXiv:1303.6738].
9. M. K. Transtrum, B. B. Machta, K. S. Brown, B. C. Daniels, C. R. Myers and J. P.
Sethna, Perspective: Sloppiness and emergent theories in physics,
biology, and beyond, J. Chem. Phys. 143 (2015) 010901
[arXiv:1501.07668].
10. T. O’Leary, A. C. Sutton and E. Marder, Computational models in the age
of large datasets, Cur. Op. Neurobio. 32 (2015) 87–94.
11. T. Niksic and D. Vretenar, Sloppy nuclear energy density functionals:
effective model reduction, Phys. Rev. C94 (2016) 024333
[arXiv:1606.08617].
12. D. V. Raman, J. Anderson and A. Papachristodoulou, Delineating
Parameter Unidentifiabilities in Complex Models, arXiv:1607.07705.
13. G. Bohner and G. Venkataraman, Identifiability, reducibility, and
adaptability in allosteric macromolecules, J. Gen. Physiol. 149 (2017)
547–560.
14. A. Raju, B. B. Machta and J. P. Sethna, Information geometry and the
renormalization group, arXiv:1710.05787.
15. H. Akaike, A new look at the statistical model identification, IEEE
Trans. Automat. Contr. 19 (1974) 716–723.
16. N. Sugiura, Further analysts of the data by Akaike’s information
criterion and the finite corrections, Comm. Stat. Th. Meth. 7 (1978)
13–26.
17. V. Balasubramanian, Statistical inference, Occam’s razor, and
statistical mechanics on the space of probability distributions,
Neural Comp. 9 (1997) 349–368.
18. I. J. Myung, V. Balasubramanian and M. A. Pitt, Counting probability
distributions: differential geometry and model selection., PNAS 97
(2000) 11170–11175.
19. D. J. Spiegelhalter, N. G. Best, B. P. Carlin and A. Van Der Linde, Bayesian
measures of model complexity and fit, J. Roy. Stat. Soc. B 64 (2002)
583–639.
20. S. Watanabe, Asymptotic equivalence of Bayes cross validation and
21.
22.
23.
ACKNOWLEDGMENTS.
We thank Vijay Balasubramanian,
William Bialek, Robert de Mello Koch, Peter Grünwald, Jon Machta,
James Sethna, Paul Wiggins, and Ned Wingreen for discussion and
comments. We thank ICTS Bangalore for hospitality.
H.H.M. was supported by NIH grant R01GM107103. M.K.T.
was support by NSF-EPCN 1710727. B.B.M. was supported by a
Lewis-Sigler Fellowship and by NSF PHY 0957573. M.C.A. was
supported by NCN grant 2012/06/A/ST2/00396.
1. L. P. Kadanoff, Scaling laws for Ising models near Tc , Physics 2 (1966)
263–272.
2. K. G. Wilson, Renormalization group and critical phenomena. 1.
Renormalization group and the Kadanoff scaling picture, Phys. Rev.
B4 (1971) 3174–3183.
3. J. L. Cardy, Scaling and renormalization in statistical physics. Cambridge Univ.
Press, 1996.
4. J. J. Waterfall, F. P. Casey, R. N. Gutenkunst, K. S. Brown, C. R. Myers, P. W.
Brouwer and J. P. Sethna, Sloppy-model universality class and the
Vandermonde matrix., Phys. Rev. Lett. 97 (2006) 150601–4.
5. R. N. Gutenkunst, J. J. Waterfall, F. P. Casey, K. S. Brown, C. R. Myers and J. P.
Sethna, Universally Sloppy Parameter Sensitivities in Systems
Biology Models, PLoS Comp. Biol. 3 (2007) e189–8.
§§
Model selection usually starts from a list of models to be compared, in our language
a list of submanifolds of Θ. We can also consider maximising mutual information in
this setting, rather than with an unconstrained function p(θ), and unsurprisingly we
observe a similar preference for highly flexible simpler models. This is also discussed
in the Appendix, at Eq. (S3).
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
6
24.
25.
widely applicable information criterion in singular learning
theory, JMLR 11 (2010) 3571–3594 [arXiv:1004.2316].
P. A. Wiggins and C. H. LaMont, Information-based inference for
singular models and finite sample sizes, arXiv:1506.05855.
M. K. Transtrum and P. Qiu, Model reduction by manifold boundaries,
Phys. Rev. Lett. 113 (2014) 098701–6.
C. E. Shannon, A mathematical theory of communication, Bell Sys.
Tech. J. 27 (1948) 623–656.
D. V. Lindley, On a Measure of the Information Provided by an
Experiment, Ann. Math. Statist. 27 (1956) 986–1005.
A. Rényi, On some basic problems of statistics from the point of
view of information theory, Proc. 5th Berkeley Symp. on Math. Statist.
and Prob. (1967) 531–543. [projecteuclid.org].
26. G. Färber, Die Kanalkapazität allgemeiner Übertragunskanäle bei
begrenztem Signalwertbereich beliebigen Signalübertragungszeiten
sowie beliebiger Störung, Arch. Elektr. Übertr. 21 (1967) 565–574.
27. J. G. Smith, The information capacity of amplitude-and
variance-constrained scalar gaussian channels, Information and
Control 18 (1971) 203–219.
28. S. L. Fix, Rate distortion functions for squared error distortion
measures, Proc. 16th Annu. Allerton Conf. Commun. Control Comput. (1978)
704–711.
29. J. O. Berger, J. M. Bernardo and M. Mendoza, On Priors that Maximize
Expected Information, Recent Developments in Statistics and Their
Applications (1988) 1–20. [www.uv.es/~bernardo/].
30. Z. Zhang, Discrete Noninformative Priors. PhD thesis, Yale University, 1994.
[UMI 9523257].
31. J. M. Bernardo, Reference Posterior Distributions for Bayesian
Inference, J. Roy. Stat. Soc. B 41 (1979) 113–147. [www.uv.es/~bernardo/].
32. B. S. Clarke and A. R. Barron, Jeffreys’ prior is asymptotically least
favorable under entropy risk, J. Stat. Plan. Infer. 41 (1994) 37–60.
33. H. R. Scholl, Shannon optimal priors on independent identically
distributed statistical experiments converge weakly to Jeffreys’
prior, Test 7 (1998) 75–94.
34. H. Jeffreys, An Invariant Form for the Prior Probability in
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
Estimation Problems, Proc. Roy. Soc. A 186 (1946) 453–461.
35. J. E. Kerrich, An Experimental Introduction to the Theory of Probability.
Copenhagen: E Munksgaard, 1946.
36. C. O’Luanaigh, CERN data centre passes 100 petabytes, home.cern
(2013).
37. S. Arimoto, An algorithm for computing the capacity of arbitrary
discrete memoryless channels, IEEE Trans. Inform. Theory 18 (1972)
14–20.
38. R. Blahut, Computation of channel capacity and rate-distortion
functions, IEEE Trans. Inform. Theory 18 (1972) 460–473.
39. K. Rose, A mapping approach to rate-distortion computation and
analysis, IEEE Trans. Inform. Theory 40 (1994) 1939–1952.
40. D. Haussler, A general minimax result for relative entropy, IEEE
Trans. Inform. Theory 43 (1997) 1276–1280.
41. M. N. Ghosh, Uniform Approximation of Minimax Point Estimates,
Ann. Math. Statist. 35 (1964) 1031–1047.
42. G. Casella and W. E. Strawderman, Estimating a Bounded Normal
Mean, Ann. Statist. 9 (1981) 870–878.
43. I. Feldman, Constrained Minimax Estimation of the Mean of the
Normal Distribution with Known Variance, Ann. Statist. 19 (1991)
2259–2265.
44. M. Chen, D. Dey, P. Müller, D. Sun and K. Ye, Frontiers of statistical decision
making and Bayesian analysis. Springer, New York, NY, 2010.
45. C. A. Sims, Rational Inattention: Beyond the Linear-Quadratic
Case, American Economic Review 96 (2006) 158–163.
46. J. Jung, J. Kim, F. Matĕjka and C. A. Sims, Discrete Actions in
Information-Constrained Decision Problems, .
[princeton.edu/~sims/#RIDiscrete].
47. S. Laughlin, A simple coding procedure enhances a neuron’s
information capacity, Z. Naturforsch. 36 c (1981) 910–912.
48. G. Tkačik, C. G. Callan and W. Bialek, Information flow and
optimization in transcriptional regulation, PNAS 105 (2008)
12265–12270 [arXiv:0705.0313].
49. M. D. Petkova, G. Tkačik, W. Bialek, E. F. Wieschaus and T. Gregor, Optimal
decoding of information from a genetic network, arXiv:1612.08084.
50. A. Mayer, V. Balasubramanian, T. Mora and A. M. Walczak, How a
well-adapted immune system is organized, PNAS 112 (2015)
5950–5955 [arXiv:1407.6888].
51. M. C. Abbott and B. B. Machta, An information scaling law: ζ = 3/4,
arXiv:1710.09351.
52. S. H. Strogatz, Nonlinear Dynamics And Chaos. Sarat Book House, 2007.
53. S. Schnell and C. Mendoza, Closed Form Solution for
Time-dependent Enzyme Kinetics, J. Theor. Biol. 187 (1997) 207–212.
54. M. Transtrum, G. Hart and P. Qiu, Information topology identifies
emergent model classes, arXiv:1409.6203.
55. P. E. Paré, A. T. Wilson, M. K. Transtrum and S. C. Warnick, A unified view
model selection, Statistics Surveys 4 (2010) 40–79.
68. P. W. Anderson, More Is Different, Science 177 (1972) 393–396.
69. R. W. Batterman, Philosophical Implications of Kadanoff’s Work on
the Renormalization Group, J. Stat. Phys. 167 (2017) 559–574.
[pitt.edu/~rbatterm/].
70. J. Bardeen, L. N. Cooper and J. R. Schrieffer, Theory of
superconductivity, Phys. Rev. 108 (1957) 1175–1204.
71. L. Michaelis and M. L. Menten, The kinetics of invertin action, FEBS
Letters 587 (2013) 2712–2720. [Translation by T. R. C. Boyde].
72. R. E. Kass and A. E. Raftery, Bayes Factors, J. Amer. Stat. Assoc. 90 (1995)
773.
73. A. Bhadra, J. Datta, N. G. Polson and B. Willard, Default Bayesian
analysis with global-local shrinkage priors, Biometrika 103 (Dec.,
2016) 955–969.
74. D. Simpson, H. Rue, A. Riebler and T. G. Martins, Penalising Model
Component Complexity: A Principled, Practical Approach to
Constructing Priors, Statist. Sci. 32 (2017) 1–28 [arXiv:1403.4630].
75. J. I. Myung, D. J. Navarro and M. A. Pitt, Model selection by normalized
maximum likelihood, Journal of Mathematical Psychology 50 (2006)
167–179.
76. P. D. Grünwald, The Minimum Description Length Principle. MIT Press, 2007.
77. C.-I. Chang and L. D. Davisson, On calculating the capacity of an
infinite-input finite (infinite)-output channel, IEEE Trans. Inform.
Theory 34 (1988) 1004–1010.
78. J. Lafferty and L. Wasserman, Iterative Markov chain Monte Carlo
computation of reference priors and minimax risk, Proc. 17th conf.
Uncert. AI (2001) 293–300 [arXiv:1301.2286].
79. J. Dauwels, Numerical computation of the capacity of continuous
memoryless channels, Proc. 26th Symp. Inf. Th. Benelux (2005).
[www.dauwels.com].
Appendix: Model Selection from Data
The usual discussion of model selection takes place after observing
data x. If we wish to compare some models∗ labeled by d, each
with some prior pd (θ), then one prescription is to choose the model
with the largest p(x). Labelling this explicitly, we write
p(x|d) =
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
pd (θ) > 0 on Θd ⊂ Θ.
[S1]
If the Bayes factor p(x|d)/p(x|d0 ) is larger than one then (absent
any other information) d is preferred over d0 (72).† In the usual
asymptotic limit m → ∞, this idea leads to minimising the Bayesian
information criterion (BIC) (62):
− log p(x|d) ≈ − log p(x|θ̂d ) +
pp. 1989–1994, IEEE, 2015.
56. N. Lewis, Combining independent Bayesian posteriors into a
(2000) 1244–1255.
59. T. Seidenfeld, Why I am not an objective Bayesian; some
reflections prompted by Rosenkrantz, Theor Decis 11 (1979) 413–440.
60. R. E. Kass and L. Wasserman, The selection of prior distributions by
formal rules, J. Amer. Stat. Assoc. 91 (1996) 1343–1370.
61. J. Williamson, Objective Bayesianism, Bayesian conditionalisation
and voluntarism, Synthese 178 (2009) 67–85.
62. G. Schwarz, Estimating the dimension of a model, Ann. Statist. 6
(1978) 461–464.
63. J. Rissanen, Modeling by Shortest Data Description, Automatica 14
(1978) 465–471.
64. C. S. Wallace and D. M. Boulton, An Information Measure for
Classification, The Computer Journal 11 (1968) 185–194.
65. S. Watanabe, A widely applicable Bayesian information criterion,
JMLR 14 (2013) 867–897 [arXiv:1208.6338].
66. P. D. Grünwald, I. J. Myung and M. A. Pitt, Advances in Minimum Description
Length: Theory and Applications. 2009.
67. S. Arlot and A. Celisse, A survey of cross-validation procedures for
dθ p(x|θ) pd (θ),
Θd
of Balanced Truncation and Singular Perturbation
Approximations, in Proceedings of the American Control Conference,
confidence distribution, with application to estimating climate
sensitivity, J. Stat. Plan. Infer. (2017) to appear.
57. N. Lewis, Modification of Bayesian Updating where Continuous
Parameters have Differing Relationships with New and Existing
Data, arXiv:1308.2791.
58. D. Poole and A. E. Raftery, Inference for Deterministic Simulation
Models: The Bayesian Melding Approach, J. Amer. Stat. Assoc. 95
Z
where − log p(x|θ̂d ) =
1 2
χ
2
=
1
2
Pm
i=1
d
log m + O(m0 )
2
[xi − yi (θ̂d )]2 /σ 2 ∼ O(m),
and θ̂d is a maximum likelihood estimator for x, constrained to the
appropriate subspace:
θ̂d (x) = argmax p(x|θ).
θ∈Θd
The term d log m penalises models with more parameters, even
though they can usually fit the data more closely. Despite the name
this procedure is not very Bayesian: one chooses the effective model
(and hence the prior) after seeing the data, rather than simply
updating according to Bayes theorem.‡
∗
†
The word model unfortunately means several things in the literature. We mean
parameter space Θd always equipped with a likelihood function p(x|θ), and usually
with a prior pd (θ). When this is a subspace of some larger model ΘD (whose
likelihood function agrees, but whose prior may be unrelated) then we term the
smaller one an effective model, or a reduced model, although we do not always write
the adjective. The optimal prior p? (θ) defines an effective model in this sense. Its
support will typically be on several boundaries of ΘD . If the boundaries of ΘD (of
all dimensions) are regarded as a canonical list of reduced models, then p? (θ) is
seldom a sub-model of any one of them.
R
If one of the priors is improper, say
dθ pd (θ) = ∞, then p(x|d) will also be
infinite. In this sense the Bayes factor behaves worse than the posterior p(θ|x),
which can still be finite.
‡
Terms penalising more complex models can be translated into shrinkage priors, which
concentrate weight near to simpler models (73). Perhaps the shrinkage priors closest
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
7
Related prescriptions can be derived from minimum description length (MDL) ideas. To allow reconstruction of the data we
transmit both the fitted parameters and the residual errors, and
minimising the (compressed) length of this transmission drives a
tradeoff between error and complexity (63, 64, 66). A convenient
version of this goes by the name of normalized maximum likeihood
(NML) (75, 76), and chooses the model d which maximizes
pNML
(x) =
d
Bayes Factor Comparison
A
1
0
0
0
1
x1
Full model,
pJ(𝜃)
0
x1
Optimal
p⭑(𝜃)
1
NML
1
x2
1
0
0
0
1
x1
Full model,
pJ(𝜃)
0
x1
Three edges,
separately
Zd =
,
Zd
Z
dx0 p x0 |θ̂d (x0 ) .
[S2]
Θd
• Figure S1A compares three models: the complete model (with
Jeffreys prior), the optimal model described by discrete prior
p? (~
y ), and an even simpler model with weight only on the
three vertices ~
y = (0, 0), ( 21 , 12 ), (1, 1).
• Figure S1B instead compares the complete model to three
different one-parameter models (along the three boundaries
of the allowed region of the ~
y plane) and a zero-parameter
model (one point ~
y , in no particularly special place). In terms
of decay rates the three lines are limits k1 = k2 , k1 = 0 and
k2 = ∞.
Corners
only
Bayes
B
This is not Bayesian in origin, and does not depend on the prior on
each effective model d, only its support Θd . The function pNML
(x)
d
is not expected data in the sense of p(x) — it is not the convolution
of the likelihood with any prior.§ In the asymptotic limit pNML
(x)
d
approaches p(x) from Jeffreys prior, and this criterion agrees with
BIC (63), but away from this limit they differ.
In Figure S1 we apply these two prescriptions to the exponential
example treated in the text. At each ~
x ∈ X we indicate which
of a list of models is preferred.¶ Figure S2 instead draws the
distributions being used.
NML Comparison
x2
1
p x|θ̂d (x)
1
One point
Fig. S1. Model selection from data point x. Each large panel here shows which
of a list of effective models is preferred after observing data x ∈ X . On the left the
criterion is maximizing Eq. (S1), on the right the criterion is Eq. (S2). We study the
same exponential model considered above, with σ = 1/10. Panel A compares our
optimal effective model with prior p? (θ) to the full model (with Jeffeys prior) and to an
even simpler model whose prior is just three delta functions. (These are drawn in the
legend below). Panel B compares the full model to three different one-dimensional
models, each allowing only one edge of Θ (with a uniform prior along this, i.e. the
one-dimensional Jeffreys prior) and also to a trivial model (just one point), again with
colors as indicated just below.
Different effective models are preferred for different values of data x.
At a given point x, if several models contain the same θ̂(x) then the
simplest among them is preferred, which in the NML case means
precisely the one with the smallest denominator Zd . In fact a trivial
model consisting of just one point Θ0 = θ̂(x) would always be
preferred if it were among those considered — there is no automatic
preference for models which can produce a wide range of possible
data.
By contrast our prior selection approach aims to be able to
distinguish as many possible outcomes in X as possible. Applied
to the same list of models as in Figure S1, this gives the following
fixed scores (base e):
Ifull = 1.296,
and
I? = 1.630,
Icorners = 1.098
Iupper = 0.852,
Ilower-left = 0.845,
Ilower-right = 1.418,
Ione-point = 0.
[S3]
to this paper’s are the penalised complexity priors of (74). Those are also reparameterization invariant, and also concentrate weight on a subspace of Θ, often a boundary.
However both the subspace (or base model) and the degree of concentration (scaling
parameter) are chosen by hand, rather than being deduced from p(x|θ).
§
This relevant optimization problem here can be described as minimizing worst-case
expected regret, written (75)
NML
pd
p(x|θ̂d (x))
,
q(x)
= argmin max log
x
q
θ̂d (x) ∈ Θd .
Perhaps the closest formulation of our maximum mutual information problem is that
our p? (x), the distribution on X not the prior, can be found as follows (40):
p? = argmin max
q∈B
θ
R
X
dx p(x|θ) log
p(x|θ)
q(x)
where q(x) is constrained to be a Bayes strategy i.e. to arise from some prior p? (θ).
Note the absence of θ̂d (x) and the presence of an integral over X , corresponding
to the fact thats that this maximization takes place without being given a subspace
Θd nor seeing data x. The resulting distributions on X are also different, as drawn in
Figure S2. If plotted on Figure 5B, pNML
(x) from the full model would be somewhere
2
between the two expected data p(x) lines there. But it is not a comparable object; its
purpose is model comparison as in Figure S1.
¶
Recall that ~
x is ~
y corrupted by Gaussian noise, and ~
y is constrained to the area show
in Figure 4A because it arises from decay rates kµ via Eq. (5). We may regard
either yt or kµ as being the parameters, generically θ .
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
8
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
B
A
pNML
(x)
d
x2
p(x|d)
■
■
■
■
�
■
x1
�
x
x2
pNML
(x)
d
■
�
d=
x
Full model, pJ(𝜃)
Optimal p⭑(𝜃)
Corners only
x1
Fig. S2. Distributions over X . Panel A shows p(x|d) Eq. (S1) and pNML
(x)
d
Eq. (S2) for the one-parameter Gaussian example Eq. (2) with σ = 1/10. The
three models being compared are the full model (with a flat prior), the simpler model
defined by p? (θ), and a model with just the endpoints of the line. Under the Bayes
factor comparison, the p? (θ) model would never be preferred here. Panel B draws
pNML
(x) for the two-parameter exponential model, Eq. (5), for the complete model,
d
the effective model defined by the support of p? (θ), and an even simpler model
allowing only three points — the same three models as compared in Figure S1A.
Notice that pNML
(x) is always a constant on the allowed region x ∈ y(Θd ).
d
By definition p? (θ) has the highest score. In second place comes the
line along the lower edge (corresponding to k1 = k2 ). The shorter
lines are strongly disfavored, because they cover a much smaller
range of possible data.
Algorithms
The standard algorithm for maximizing channel capacity (of discrete
memoryless channels) was written down independently by Blahut
(38) and Arimoto (37). This aspect of rate-distortion theory is
mathematically the same as the problem we consider, of maximizing
mutual information by choosing the prior. The algorithm starts
with p0 (θ) = const., and then at each time step updates this by
pτ +1 (θ) =
1 fKL (θ)
e
pτ (θ)
Zτ
0
[S4]
where Zτ =
dθ0 efKL (θ ) pτ (θ0 ) maintains normalization, and
fKL (θ) = DKL [p(x|θ)kp(x)] is computed with pτ (θ). Since this
R
𝜏 = 106 iterations
�
x2
■
■
𝜏 = 1000
Fig. S3. Convergence of the BA algorithm (S4) for the exponential model. This
shows p? (~
y ) for the case σ = 1/50, with ~
y discretized on a grid of spacing 1/100
in the bulk and 1/200 along the boundaries of the allowed region.
x1
■
■
𝜏 = 100
Mattingly, Transtrum, Abbott, Machta
|
Rational Ignorance
is a convex optimization problem, the algorithm is guaranteed to
converge to the global maximum. This makes it a good tool to see
discreteness emerging.
Figures 2 and S3 show the progress of this algorithm for the
one- and two-dimensional models in the text. We stress that the
number and positions of the peaks which form are unchanged when
the discretization of θ is made much finer. Notice also that the
convergence to delta functions happens much sooner near to the
boundaries than in the interior. The convergence to the correct
mutual information I(X; Θ), and towards the optimum distribution
on data space p(x), happens much faster than the convergence to
the correct number of delta functions.
Because θ must be discretized for this procedure, it is poorly
suited to high-dimensional parameter spaces. However once we
know that p? (θ) is discrete it is natural to consider algorithms
exploiting this. With K atoms, we can adjust their positions θ~a
and weights λa using gradients
∂MI
= λa
∂θaµ
Z
dx
~
~
p(x|θ)
∂p(x|θ)
log
µ
∂θ
p(x)
~ θ
~a
θ=
∂MI
= fKL (θ~a ) − 1.
∂λa
[S5]
Figures 1, 2A, 5 and the square plot points in Figure 4 were generated this way. This optimization is not a convex problem (there is
some tendency to place two atoms on top of each other, and thus
use too few points of support) but it can often find the optimum
solution. We can confirm this by calculating fKL (θ) everywhere
— any points for which this is larger than its value at the atoms
indicates that we do not have the optimal solution, and should add
an atom.
Monte Carlo algorithms for this problem have been investigated
in the literature, see (77, 78) and especially (79). (Incidentally, we
observe that (77)’s table 1 contains a version of scaling law Eq. (4),
with ζ ≈ 1/2. No attempt was made there to use the optimal
number of atoms, only to calculate the channel capacity to sufficient
accuracy.)
PNAS
|
7 February, 2018
|
vol. XXX
|
no. XX
|
9
| 7cs.IT
|
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
1
Convergence of the Z-Bus Method for Three-Phase
Distribution Load-Flow with ZIP Loads
arXiv:1605.08511v3 [math.OC] 9 May 2017
Mohammadhafez Bazrafshan, Student Member, IEEE, and Nikolaos Gatsis, Member, IEEE
Abstract—This paper derives a set of sufficient conditions
guaranteeing that the load-flow problem in unbalanced threephase distribution networks with wye and delta ZIP loads has
a unique solution over a region that can be explicitly calculated
from the network parameters. It is also proved that the wellknown Z-Bus iterative method is a contraction over the defined
region, and hence converges to the unique solution.
Index Terms—Three-phase distribution load-flow, Z-Bus
method, ZIP load, contraction mapping, power flow
I. I NTRODUCTION
E
XISTENCE of three-, two-, and one-phase transmission
lines with high R/X ratios in distribution networks is
predominant. Therefore, the Z-Bus iterative method is preferred over classical Newton-Raphson methods for distribution
load-flow [1]. This paper deals with the Z-Bus method when
applied to practical distribution networks including three-phase
wye and delta loads with constant-power, constant-current, and
constant-impedance portions (ZIP loads), as well as potentially
untransposed lines. Conditions on the network components and
loads guaranteeing the existence of a unique load-flow solution
and the convergence of the Z-Bus method are derived.
The availability of conditions guaranteeing power flow
solution existence has several applications in power system
planning and operations. In planning studies, such as conventional and distributed generation planning, conditions for
solution existence serve as assurance that desirable operating conditions are feasible under future load demands. For
example, such conditions provide network-specific guides on
whether a certain lateral will have valid load-flow solutions
over the course of years, as the loads increase [2], [3].
In operations, conditions for solution existence allow power
engineers to determine whether the set of power injections and
flows respect security constraints of power system equipment
[4]. Furthermore, the availability of solvability regions in
terms of nodal voltages renders improved initial estimates
for numerical algorithms and accelerates the online analysis
required for real-time monitoring and control applications [5].
Specifically in regards to distribution networks, their
weakly-meshed structure limits the number of feasible voltage
profiles. This characteristic implies that it is harder to meet
different demand patterns in distribution networks [6]. In this
case, explicit sufficient conditions for solution existence that
Manuscript received May 26, 2016; revised November 22, 2016 and March
1, 2017; accepted April 26, 2017. The authors are with the Department of
Electrical and Computer Engineering, The University of Texas at San Antonio. E-mails: {mohammadhafez.bazrafshan, nikolaos.gatsis}@utsa.edu. This
material is based upon work supported by the National Science Foundation
under Grant No. CCF-1421583.
take into account various load types—that is, wye and delta
ZIP loads—prove to be very useful.
The literature on existence of solutions to the power flow
equations is extensive; see e.g., [3] for a review on general
power networks, and recent works in [5], [7], [8] geared toward
distribution networks. Here, the prior art most closely related
to the present work is reviewed.
The single-phase equivalent of a power distribution network
is considered in [5]. Power flow equations are formulated so
that injection currents can be expressed as a fixed point of a
quadratic map parametrized by constant-power loads and the
network admittance matrix. Sufficient conditions on the loads
and the network admittance matrix are derived to guarantee
existence of a power flow solution.
The work in [7] highlights that the injection currents can be
expressed by a family of fixed-point quadratic maps similarly
to [5] but parametrized by a diagonal matrix. An algorithm
is proposed for finding the diagonal matrix that expands the
regions given in [5] where a power flow solution exists.
Convergence of the Z-Bus method interpreted as a fixedpoint iteration is considered in [8] and [9] for singlephase distribution networks. Specifically, explicit conditions
on constant-power loads and the network admittance matrix
are presented in [8] ensuring convergence of the Z-Bus method
and existence of a unique solution for distribution load-flow
under balanced assumptions. Interestingly, it is numerically
verified in [9] that when distributed generation units are modeled as constant-power nodes, the voltage iterates produced by
the Z-Bus method are indeed contracting.
Single-phase power system modeling is valid under balanced three-phase operation, which is a reasonably accurate assumption for transmission networks. In distribution networks,
there is a prevalence of untransposed three-, two- and onephase transmission lines with unbalanced loading. There is
a gap in the literature, since theoretical results on solution
existence and convergence of load-flow methods overlook the
underlying unbalanced nature of distribution systems.
This paper considers general unbalanced distribution networks that include transmission lines with one, two, or three
available phases as well as wye and delta ZIP loads, in contrast
to [5], [7], [8], which only consider single-phase networks
with constant-power loads. Sufficient conditions for the Z-Bus
method to be a contraction over a region explicitly defined in
terms of the loads and the bus admittance matrix are derived.
A consequence of the contraction is that power-flow equations
are guaranteed to have a unique solution in the defined region.
In order to prove that the Z-Bus method is a contraction
mapping, a candidate region of contraction is defined and
the self-mapping and contraction properties are proved. These
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
steps were followed in [8] for single-phase networks with
constant-power loads. However, the extension to general unbalanced three-phase networks with wye and delta ZIP loads
and missing phases adds a level of complexity that necessitates
the development of novel methods to prove the convergence
of the Z-Bus method.
The major contribution of this paper is to derive sufficient
convergence conditions that handle the complexities arising
from various combinations of load types and phasings in
practical three-phase distribution networks. The conditions
are expressed directly in terms of the network loads and
admittance matrix, and guarantee the existence of a unique
solution to the power flow equations. It becomes immediately
apparent from these conditions that different load types contribute differently to the network solvability. Moreover, the
generality of the network model renders results that are readily
applicable to practical distribution networks without simplifying assumptions. To this end, the conditions are validated on
several IEEE distribution test feeders.
In order to prove the aforementioned conditions, this paper
develops a set of tools that advance the power flow analysis
via contraction methods. These tools are particularly tailored
to handle wye and delta connections, ZIP loads, and missing
phases, and are expected to be useful to other researchers
working in this area. In addition, this paper constructs interesting and non-trivial examples of networks where the Z-Bus
method exhibits oscillatory behavior, and the conditions are
not satisfied. One example amounts to a two-node network
where the non-slack bus has constant-current and impedance
loads, but a constant-power injection (e.g., from a renewable
energy source). The second example features a nontrivial
unbalanced network with phase couplings in its lines.
Paper Organization: The network and the ZIP load models
with wye and delta configurations are introduced in Section
II. The Z-Bus method is briefly reviewed in Section III.
Section IV presents the main theorem, which contains the
sufficient conditions for the Z-bus method to be a contraction.
The theorem is verified numerically in Section V. Section VII
provides pointers to future work, and Appendices A and B
include a detailed proof of the main result.
Notation: For a vector u, uj denotes the j-th element, and
diag(u) is a square matrix with elements of u on the main
diagonal. The notation v = {vn }n∈N is used to denote a
vector constructed from vertically stacking vectors vn for n ∈
N . For a matrix Z, Z•k denotes the k-th column, Zj• denotes
the j-th row, Zjk denotes the element in the j-th row and k-th
column; and [Z•k ]k∈NY denotes a matrix constructed from the
columns of Z in the set NY . The infinity norm for a vector u
and a P
matrix Z is defined as kuk∞ = maxj |uj | and kZk∞ =
maxj k |Zjk |, respectively.
II. N ETWORK
AND LOAD MODEL
Motivated by the various three-phase connections in IEEE
distribution test feeders, a detailed network model is presented
in this section.
2
A. Three-phase network modeling preliminaries
A power distribution network is represented by a graph
(N , E), where N := {1, 2, . . . , N } ∪ {S} is the set of nodes
and E ⊂ N × N := {(n, m)|n, m ∈ N } is the set of
edges. Node S is considered the slack bus, i.e., the point of
interconnection to the transmission network. Furthermore, let
N be partitioned as N = NY ∪ N∆ ∪ {S} where NY and N∆
collect nodes with wye and delta connections, respectively.
The set of available phases at node n is denoted by Ωn ⊆
{a, b, c}. For wye nodes, i.e., n ∈ NY , Ωn may have one,
two, or three phases. For delta nodes, i.e., n ∈ N∆ , Ωn has at
least two phases. For future use, define r(φ) as the right shift
of phase φ as follows:
r(a) = b, r(b) = c, r(c) = a.
(1)
If |Ωn | = 3, then we have that r(φ) ∈ Ωn for all phases
φ ∈ Ωn . If |Ωn | = 2, then there exists only one phase φ ∈ Ωn
such that r(φ) ∈ Ωn . For example, if Ωn = {a, c}, only phase
c has the property that r(c) = a ∈ Ωn .
For every node n ∈ N , the complex line to neutral voltage
vector is denoted by vn := {vnφ }φ∈Ωn ∈ C|Ωn | . The slack
2π
2π
bus voltage is fixed at vS = {1, e−j 3 , ej 3 }. The line to
neutral voltages are collected in a vector v = {vn }n∈N \{S} .
′
Moreover, we use a vector eφφ
∈ C|Ωn | with entries of 1, -1,
n
′
φφ′ T
and possibly 0, such that (en ) vn = vnφ − vnφ gives the line
′
to line voltages between phases φ and φ . It is easy to see that
′
T
φ′ φ T
(eφφ
n ) vn = −(en ) vn .
further an index set J := {1, . . . , J} where J =
PDefine
N
|Ω
n | is the length of vector v, and each j ∈ J is
n=1
a linear index corresponding to a particular pair (n, φ) with
n ∈ N \ {S} and φ ∈ Ωn . In this case, we write the
operation j = Lin(n, φ). The operation Node[Lin(n, φ)] =
Node[j] = n relates the index j ∈ J to the corresponding
node n. Define the set Jn := {j | Node[j] = n} as the set of
all linear indices j that correspond to node n.SFinally, set J
is partitioned
as J = JY ∪ J∆ , where JY = n∈NY Jn and
S
J∆ = n∈N∆ Jn .
B. Three-phase load models
For a load at node n, the complex vector of net current
injections is denoted by in = {iφn }φ∈Ωn and we define i =
{in }n∈N \{S} . Due to existence of loads, the nodal net current
injection i is a function of nodal voltages v. This dependence is
denoted by in (vn ). According to the ZIP load model, in (vn )
is
iPQn =
φcomposed of currents from constant-power loads
iPQn φ∈Ωn , constant-current loads iIn = iφIn φ∈Ωn , and
constant-impedance loads iZn = iφZn φ∈Ωn . For n ∈ N \{S}
and φ ∈ Ωn we have that
iφn (vn ) = iφPQn (vn ) + iφIn (vn ) + iφZn (vn )
(2)
where functions iφPQn (vn ), iφIn (vn ), and iφZn (vn ) are defined
in (3) for wye connections and in (4) for delta connections.
Specifically, for wye connections (n ∈ NY and φ ∈ Ωn ),
the ZIP load components are
iφPQn (vn ) = −(sφLn /vnφ )∗
(3a)
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
iφZn (vn ) =
vnφ
iφ ,
φ Ln
|vn |
φ
−yL
vφ ,
n n
iφIn (vn ) = −
3
(3b)
(3c)
where sφLn is the nominal power in the constant-power portion
of the ZIP model, iφLn is the nominal current for the constantφ
current portion of the ZIP model, and yL
is the nominal
n
admittance in the constant-impedance portion of the ZIP model
at node n ∈ NY and phase φ ∈ Ωn .
For delta connections (n ∈ N∆ and φ ∈ Ωn ), the ZIP load
components are
!∗
φφ′
X
sL
φ
n
(4a)
iPQn (vn ) = −
φφ′ T
(e
)
v
n
n
′
φ ∈Ωn \{φ}
′
iφIn (vn )
iφZn (vn )
′
′
=−
=−
X
φ′ ∈Ωn \{φ}
X
′
iφφ
Ln
′
T
(eφφ
n ) vn
′
T
|(eφφ
n ) vn |
′
φφ
T
yL
(eφφ
n ) vn ,
n
(4b)
(4c)
φ′ ∈Ωn \{φ}
where the dependence of injected currents on nodal voltages
is shown explicitly.
Using (2), i(v) can be decomposed into three parts:
i(v) = iPQ (v) + iI (v) + iZ (v),
(7)
where iPQ (v) = {iPQn }n∈N \{S} , iI (v) = {iIn }n∈N \{S} ,
iZ (v) = {iZn }n∈N \{S} . Moreover, (3c) and (4c) reveal that
iZ (v) can be written as a linear function of v as follows:
iZ (v) = −YL v,
where YL ∈ C
J×J
(8)
has entries given by
φ
YL (j, j) = yL
, if j = Lin(n, φ), n ∈ NY ,
n
X
φφ′
, if , j = Lin(n, φ), n ∈ N∆ ,
yL
YL (j, j) =
n
φ′ ∈Ωn \{φ}
′
φφ
YL (j, k) = −yL
, j = Lin(n, φ), k = Lin(n, φ′ ), n ∈ N∆ ,
n
and the remaining entries of YL are all zero.
Substituting (7) and (8) in (6) and after some manipulation
the following fixed-point equation for v is obtained:
′
φφ
φφ
where sφφ
Ln , iLn , and yLn are respectively the nominal power,
current, and admittance in the ZIP model for nodes n ∈ N∆
and over phases
φ, φ′ ′ ∈ Ωn′ . For′ n ∈ N∆ and
φ, φ′′ ∈ Ωn , we
φφ′
φ φ φφ
φφ
φφ′
φφ
have that sLn = sLn , iLn = iLn , and yLn = yL
.
n
Notice that the constant-current loads in (3b) and (4b)
rightfully adhere to the proper definition given in [10], where
it is emphasized that the power varies directly with voltage
magnitude. To see this, one can calculate the apparent power
consumption for the load model in (3b) which gives sφn =
vnφ [(vnφ /|vnφ |)iφLn ]∗ = |vnφ |(iφLn )∗ , i.e., the apparent power is
only a function of the magnitude of the voltage |vnφ |. It is
easy to see that in this case, the load current magnitude is
|iLn | and the load power factor is
Re[iφ
Ln ]
|iφ
Ln |
, both of which are
constant values, concluding that these models are in line with
the definition in [11, pp. 315] as well. It should be noted that
(3b) and (4b) are different than the constant-current-phasor
model employed in [12].
For future use, if k = Lin(n, φ) where n ∈ NY and φ ∈
Ωn , we define skL = sφLn and ikL = iφLn . For k = Lin(n, φ)
where n ∈ N∆ and φ ∈ Ωn , if r(φ) ∈ Ωn , define skL =
φr(φ)
φr(φ)
φr(φ)
. If r(φ) ∈
/ Ωn , then
sLn , ikL = iLn , and ek = en
′
′
k
Ωn = {φ, φ }, and define sL = 0, ikL = 0, and ek = eφφ
n .
v = Z [iPQ (v) + iI (v)] + w,
−1
where Z = (Y + YL )
(6)
(11)
Equation (10) lends itself to the well-known Z-Bus method
for three-phase networks, outlined in e.g., [1]:
v[t + 1] = Z [iPQ (v[t]) + iI (v[t])] + w,
(12)
where v[t] is the value of v at iteration t. Notice that (12) is an
application of Picard’s iteration for solving nonlinear systems
of equations [14, pp. 182].
IV. S UFFICIENT CONDITIONS FOR CONVERGENCE OF THE
Z-B US METHOD TO A UNIQUE SOLUTION
In this section, sufficient conditions are presented such that
the iteration of the Z-Bus method (12) is contracting, i.e., the
iteration converges geometrically to a unique solution of (10).
To present these conditions, it is beneficial to make a change
of variables in (12). Specifically, let u = Λ−1 v[t] where Λ ∈
CJ×J is an invertible diagonal matrix with entries λj . Matrix
Λ will serve as a design parameter which can be leveraged
to potentially expand the convergence region of the Z-Bus
method. Parameterizing v by u and multiplying both sides of
(12) by Λ−1 yield the following iteration for u:
u[t + 1] = T(u[t]),
where iS is the complex current injection of the slack bus.
Matrices Y, YNS , YSN , and YSS are formed by concatenating
the admittance matrices of transmission lines, transformers and
voltage regulators, given e.g., in [13]. It follows from (5) that
i(v) = Yv + YNS vS ,
, and
w = −ZYNS vS .
III. T HE Z-B US M ETHOD
For a three-phase network, the multidimensional Ohm’s law
is given by
v
Y
YNS
i
,
(5)
=
YSN YSS vS
iS
(10)
J
(13)
J
where T(u) : C → C is the following mapping
T(u) = Λ−1 Z [iPQ (Λu) + iI (Λu)] + Λ−1 w,
(14)
and the Z-Bus method of (12) can be equivalently written as
v[t + 1] = ΛT(Λ−1 v[t]).
(15)
Since the iterations in (13) and (15) have a one-to-one correspondence, convergence of (13) to a fixed point ufp = T(ufp )
is equivalent to the convergence of (15) to a respective solution
vfp = Λufp of (10).
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
4
In what follows, conditions such that the mapping T(u)
in (14) is a contraction over a region D ⊆ CJ are derived.
As a consequence, T(u) has a unique fixed point over D, and
iterations (13) converge geometrically to the unique fixed point
[15, Prop. 3.1.1]. The definition of contraction is as follows:
Definition. For a closed set D ⊆ CJ , a mapping T(u) is a
contraction on D if the following two properties hold:
1) Self-mapping property: u ∈ D ⇒ T(u) ∈ D.
2) Contraction property: u, ũ ∈ D ⇒ kT(u) − T(ũ)k ≤
αku − ũk,
where k.k is some norm and α ∈ [0, 1) is a constant.
kΛ−1 ZY diag(sY )diag(wY )−1 k∞
1 − Rλ̄/w
−1 ∆
kΛ Z diag(s∆ )diag(w∆ )−1 k∞
+
1 − 2Rλ̄/ρ
+ kΛ−1 ZY diag(iY )k∞ + kΛ−1 Z∆ diag(i∆ )k∞ ≤ R (C3)
kΛ−1 ZY diag(sY )ΛY [diag(wY )−1 ]2 k∞
(1 − Rλ̄/w)2
−1 ∆
2kΛ Z diag(s∆ )Λ∆ [diag(w∆ )−1 ]2 k∞
+
(1 − 2Rλ̄/ρ)2
2kΛ−1 ZY diag(iY )ΛY diag(wY )−1 k∞
1 − Rλ̄/w
4kΛ−1 Z∆ diag(i∆ )Λ∆ diag(w∆ )−1 k∞
+
≤ α < 1.
1 − 2Rλ̄/ρ
+
In regards to the previous definition, any number α ∈ [0, 1)
˜
u)k
uniformly for all
that upperbounds the ratio kT(u)−T(
ku−ũk
u, ũ ∈ D (u 6= ũ), is called contraction modulus.
In order to state the conditions for (14) to be a contraction
map, the following quantities are defined:
(C4)
sY = {skL }k∈JY ∈ C|JY | , iY = {ikL }k∈JY ∈ C|JY | , (16b)
wY = {wk }k∈JY ∈ C|JY | , ZY = [Z•k ]k∈JY CJ×|JY | , (16c)
If conditions (C1)–(C4) hold for some R > 0, then T(u) is
a contraction mapping and has as a consequence a unique fixed
point ufp in the ball defined by DR . In other words, the power
flow equations have a unique solution in DR . The value of R
can be interpreted as the infinity norm of the voltage difference
from the no-load power flow solution scaled by Λ. The latter
is obtained as u = Λ−1 w upon setting iPQ = iI = 0 in (14).
w∆ = {eTk wNode[k] }k∈J∆ ∈ C|J∆ | ,
Moreover, the contraction property of T(u) implies the
following relationships for successive iterates of the Z-Bus
method upon initialization with u[0] ∈ DR :
w = min |wk |, λ̄ = max |λk |, ρ = min {|eTk wNode[k] |}, (16a)
k∈J
k∈J
k∈J∆
ΛY = diag({λk }k∈JY ) ∈ C|JY |×|JY | ,
s∆ = {skL }k∈J∆ ∈ C|J∆ | , i∆ = {ikL }k∈J∆ ∈ C|J∆ | ,
(16d)
(16e)
J×|J∆ |
(16g)
∆
Z =
[Z∆
•lk ]k∈J∆
∈C
,
(16f)
where lk ∈ {1, . . . , |J∆ |} is the order of k in J∆ , and
Z•k − Z•k′ , if φ, r(φ) ∈ Ωn ,
Z∆
=
•lk
0J ,
if φ ∈ Ωn , r(φ) ∈
/ Ωn ,
′
with k = Lin(n, φ), k = Lin(n, r(φ)),
∆
Λ = diag({maxl∈JNode[k] |λl |}k∈J∆ ) ∈ C|J∆ |×|J∆ | . (16h)
In (16), the notation wn = {wk }k∈Jn relates the indices in
w to corresponding ones in v. Respectively for wye and delta
connections, quantities sY and s∆ collect all constant-power
loads, iY and i∆ collect all constant-current loads, ZY and
Z∆ collect corresponding columns of the impedance matrix
Z, and wY , w∆ collect the voltages induced by the slack bus.
Matrix ΛY contains the elements of the design matrix Λ that
correspond to wye nodes and matrix Λ∆ selects the maximum
of the per node entries of the design matrix Λ for delta nodes.
Theorem 1 establishes the convergence of the Z-Bus method
in three-phase distribution networks with ZIP loads, and is
proved in Appendix B using intermediate results in Appendix A.
Theorem 1. Define the closed ball DR := {u ∈ CJ | ku −
Λ−1 wk∞ ≤ R} where R > 0. Then T(u) in (14) is a
contraction mapping on DR with contraction modulus α if
the following four conditions hold:
1 − Rλ̄/w > 0,
(C1)
1 − 2Rλ̄/ρ > 0,
(C2)
ku[t + 1] − ufp k∞ ≤ αku[t] − ufp k∞ ,
ku[t + 1] − u[t]k∞ ≤ αku[t] − u[t − 1]k∞ .
(17a)
(17b)
Equation (17a) implies that the iterates in (13) converge to ufp
with a convergence rate that is upperbounded by α. Equation
(17b) can be used to numerically evaluate the convergence
ku[t+1]−u[t]k∞
. We will see in
rate by computing the ratio ku[t]−u[t−1]k
∞
Section V for the IEEE test feeders that the aforementioned
ratio turns out to be much smaller than the estimate provided
by the left-hand side of (C4).
Theorem 1 on T(u) implies certain convergence properties
for iteration (12), as stated next.
Theorem 2. If conditions (C1)–(C4) hold, then by initializing
v[0] ∈ D′R := {v|kΛ−1 (v−w)k∞ ≤ R}, the Z-Bus iterations
in (12) remain within D′R and converge to a unique solution.
Moreover, the convergence is R-linear with rate α [i.e., the
sequence generated by (12) is dominated by the geometric
sequence Bαt where B is a non-negative scalar].
Proof. The self-mapping property of T(u) and the one-toone correspondence between v and u guarantee that the
iterates v[t] remain within D′R . Regarding the convergence
rate of (12), note that the contraction property of T(u) implies
(cf. [15, Prop. 3.1.1])
ku[t] − ufp k∞ ≤ αt ku[0] − ufp k∞ .
Since we have u = Λ
−1
(18)
v, it follows from (18) that
kΛ−1 (v[t] − vfp )k∞ ≤ αt kΛ−1 (v[0] − vfp )k∞ . (19)
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
5
From the definition of the infinity norm it follows that
1
(vj [t] − vjfp )
λj
1
= max
|vj [t] − vjfp |
j∈J
λj
1
≥ max |vj [t] − vjfp |
λ̄ j∈J
1
(20)
⇒kΛ−1 (v[t] − vfp )k∞ ≥ kv[t] − vfp k∞ .
λ̄
Moreover, we have that
kΛ−1 (v[t] − vfp )k∞ = max
j∈J
1
(vj [0] − vjfp )
λj
1
(vj [0] − vjfp )
≤ max
j∈J mink∈J |λk |
1
≤
max |vj [0] − vjfp |
mink∈J |λk | j∈J
1
kv[0] − vfp k∞ . (21)
⇒kΛ−1 (v[0] − vfp )k∞ ≤
mink∈J |λk |
kΛ−1 (v[0] − vfp )k∞ = max
j∈J
Combining (20), (21), and (19) yields
kv[t] − vfp k∞ ≤ Bαt ,
(22)
λ̄
where B = minj∈J |λj | kv[0] − vfp k∞ ≥ 0. Since the quantity
kv[t] − vfp k∞ is upperbounded by a geometric sequence,
the rate of convergence is (informally) said to be geometric;
formally, the convergence is R-linear with rate α [16, pp.
619].
Conditions (C1)–(C4) can be expressed via inequalities
involving polynomials of degree at most four in the radius
of the contraction region. The coefficients of the polynomials
can be readily written in terms of the network parameters
and loads. Quartic polynomials and their roots have been
completely characterized, and can be routinely computed in
terms of the polynomial coefficients—see for example [17].
The convergence region can thus be easily computed. Significant analytical advantages can also be afforded by the
presented conditions, as the convergence region can potentially
be expressed explicitly in terms of the network loads and
parameters. An alternative set of nonlinear non-polynomial
conditions for convergence of the Z-Bus method is presented
in [18].
Conditions (C1)–(C4) can also be graphically solved, as
explained next. First, the quantities in (16) must be computed
which depend only on the feeder data. The set of R values
satisfying each condition can simply be plotted on the real
line. The intersection of the feasible regions in these plots
gives a range for R. In the next section, the valid range
of R satisfying conditions (C1)–(C4) is computed for IEEE
distribution test feeders, and the convergence of the Z-Bus
method is numerically illustrated.
V. N UMERICAL
VERIFICATION ON
IEEE FEEDERS
In this section, the range of R where conditions (C1)–(C4)
are satisfied is given for IEEE-13, IEEE-37, and IEEE-123
distribution test feeders [19]. Furthermore, the self-mapping
property as well as the convergence of the Z-Bus method are
numerically confirmed for iterations on voltages in (12). This
is because the Z-Bus method is routinely implemented on the
variable v as opposed to u.
The feeder data are available online [20]. We model
transformer nodal admittances according to [1]. However, to
avoid singularities in the bus admittance matrix, delta-delta
transformers are modeled using a modification suggested in
[21] which entails connecting a small shunt admittance to the
ground. For voltage regulators, the models are derived from
[11, Ch. 7].1
To numerically verify conditions (C1)–(C4), the design
parameter Λ is selected to be Λ = diag(w). This choice
allows for the center of DR to be the vector of all one’s. This
practically means that the no-load solution for scaled voltages
is u = Λ−1 w = 1.
For the IEEE-13 test feeder, conditions (C1)–(C4) yield the
following inequalities for R:
1 − 1.043R > 0; 1 − 1.203R > 0
0.041
0.127
+
+ 0.032 + +0.01 ≤ R
1 − 1.043R 1 − 1.203R
0.127
0.048
+
(1 − 1.043R)2 (1 − 1.203R)2
0.024
0.064
+
< 1.
+
1 − 1.043R 1 − 1.203R
For the IEEE-37 test feeder, the conditions are
1 − 1.037R > 0; 1 − 1.198R > 0
0.037
+ 0.018 ≤ R
1 − 1.198R
0.042
0.043
+
< 1.
(1 − 1.198R)2 1 − 1.198R
(23a)
(23b)
(23c)
(24a)
(24b)
(24c)
Finally, for the IEEE-123 feeder, the conditions are
1 − 1.078R > 0; 1 − 1.238R > 0
(25a)
0.001
0.129
+
+ 0.030 + 0.012 ≤ R
(25b)
1 − 1.078R 1 − 1.238R
0.001
0.129
+
2
(1 − 1.0777R)
(1 − 1.238R)2
0.060
0.027
+
+
< 1. (25c)
1 − 1.078R 1 − 1.238R
For each feeder, the subsets of the real line where conditions
(C1)–(C4) are satisfied are depicted in Fig. 1, by solving
(23), (24), and (25). The intersection of these regions for the
IEEE-13, IEEE-37 and IEEE-123 feeders is [Rmin , Rmax ] =
[0.29, 0.48], [Rmin , Rmax ] = [0.06, 0.64], and [Rmin , Rmax ] =
[0.22, 0.54], respectively. Both Rmin and Rmax reveal important information about the Z-Bus convergence. Specifically,
larger Rmax guarantees convergence when the initialization is
far from the unique solution. Smaller Rmin guarantees that the
unique solution is close to the center of the ball. The latter
implies that the deviation of the load-flow voltage solution
from the no-load solution is tightly bounded. Notice that these
regions are dependent on the design matrix Λ = diag(w).
1 The MATLAB scripts that compute the convergence regions for the IEEE
test feeders and run the Z-Bus method are provided online at the following
page: https://github.com/hafezbazrafshan/contraction-mapping.
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
(C1)
(C2)
(C3)
6
vn [0] = vS
vn [0] = wn
vn [0] = wn − Rmin wn
vn [0] = wn − Rmax wn
Rmin
Rmax
(C4)
0.5
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
kΛ−1 (v[t] − w)k∞
0
R
(a)
(C1)
(C2)
(C3)
(C4)
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
R
(b)
0
(C1)
(C2)
(C3)
1
2
3
4
Iteration t
(C4)
Demonstration of the self-mapping property of the Z-Bus
iterations in (12) for IEEE-123 test feeder. As long as the Z-Bus
method is initialized with v[0] ∈ D′R , where R satisfies conditions
(C1)–(C4), successive iterations v[t] remain within D′R .
Fig. 3.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
R
(c)
Fig. 1. Regions on the real line where conditions (C1)–(C4) hold
for (a) IEEE-13, (b) IEEE-37, and (c) IEEE-123 test feeders. The
presence of a mark indicates that the condition is satisfied.
vn [0] = vS
vn [0] = wn
vn [0] = wn − Rmin wn
vn [0] = wn − Rmax wn
103
101
kv[t + 1] − v[t]k∞
Lower bound of contraction modulus α
1
IEEE-13
IEEE-37
IEEE-123
0.9
0.8
10−1
10−3
0.7
0.6
10−5
0.5
0.4
10−7
0.3
0
1
2
3
4
Iteration t
0.2
0.1
Fig. 4. Convergence of the Z-Bus iterations in (12) for IEEE-123 test
0
0
0.1
0.2
0.3
0.4
0.5
0.6
R
Fig. 2. Lower bound of contraction modulus α versus R in the range for
which conditions (C1)–(C4) are satisfied.
For example, setting Λ = IJ for the IEEE-13 feeder yields
Rmax = 0.60 and also shifts the center of the ball from 1
to w. A study on the impact of Λ on the contracting region
remains for future work.
Figure 2 depicts the lower bound of the contraction modulus
α evaluated from (C4) as a function of the feasible R for
the three IEEE feeders. It is revealed in Fig. 2 that a more
localized initialization (i.e., smaller R) yields smaller α, with
the smallest α occurring at Rmin . For the IEEE-13, IEEE37, and IEEE-123 feeders, the smallest α is respectively
0.50, 0.09, and 0.34. Upon initializing within DRmin , the
convergence rate is guaranteed to be at most as much as the
aforementioned values.
The self-mapping property of the Z-Bus iterations (12) is
demonstrated in Fig. 3 for the IEEE-123 test feeder. It is
shown that when the initialization v[0] is within D′R , where
R ∈ [Rmin , Rmax ], the sequence v[t] produced by the ZBus method in (15) remains within D′R . In Fig. 3, a case of
special interest is the graph corresponding to the initialization
with the flat voltage profile, that is, when vn [0] = vS , where
feeder. It is observed that the Z-Bus method converges to the unique
solution very fast—within 4 iterations.
π
π
vS = {1, e−j2 3 , ej2 3 } is the voltage at the slack bus. The
graph shows that the typical initialization with the flat voltage
profile is within the contracting region; and upon initializing
with vn [0] = vS ∈ D′R for all nodes, the iterates remain
within D′R .
The distance between two successive iterates of the Z-Bus
method for the IEEE-123 test feeder is plotted in Fig. 4.
It is shown that the distance kv[t + 1] − v[t]k∞ practically
converges to zero within 4 iterations, implying that the fixedpoint solution is obtained. For the sequences of Fig. 4, the
empirical convergence rate is also numerically calculated using
(17b) and is depicted in Fig. 5. All the numerical convergence
rates are smaller than 0.34, which is the smallest theoretically
obtained contraction modulus for the IEEE-123 network.
The merits of this work is that it characterizes certain
regions with unique power flow solution in three-phase distribution networks and that it guarantees convergence of the
Z-Bus method. Working with networks that satisfy solvability
conditions can provide a degree of assurance that the numerical
algorithm—the Z-Bus method in this case—will converge to
a solution.
Further research is required to obtain conditions that ac-
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
7
0.12
vn [0] = vS
vn [0] = wn
vn [0] = wn − Rmin wn
vn [0] = wn − Rmax wn
kΛ−1 (v[t+1]−v[t])k∞
kΛ−1 (v[t]−v[t−1])k∞
0.1
0.08
0.06
0.04
1
2
3
4
Iteration t
Fig. 5. Numerical convergence rate of the Z-Bus iterations in (12) for
the iterations in Fig. 4. The numerical convergence rate at iteration t
is calculated using (17b). Notice that these convergence rates are all
smaller than 0.34, which is the tightest upper bound given in Fig. 2
for IEEE-123.
curately characterize the largest region with a unique power
flow solution. As was previously alluded to, it is possible to
leverage the design parameter Λ for this purpose. By setting
Λ = diag(w), the conditions are satisfied for the practical
IEEE test networks; developing comprehensive methods for
the selection of parameter Λ are left for future work.
It should be noted at this point that this work derives
sufficient conditions for convergence. Generally speaking, sufficient conditions are conservative by nature. That is, the existence of loads and network parameters that do not satisfy the
conditions yet yield convergence is not precluded. It is clear
from conditions (C3) and (C4) that by increasing the constantpower and constant-current loads, there will eventually be no
R > 0 that ensures contraction.
As an example, in the IEEE-123 test feeder, after 50% of
load increase, no R > 0 satisfies condition (C3), however the
Z-Bus converges to a solution within 5 iterations. Therefore,
the ultimate goal would be to provide a set of conditions that
divide the voltage space to two regions, namely, one where
there is convergence and another one of non-convergence.
However, this is still an open research question. The next
section provides two network examples where the Z-Bus
method fails to converge in cases where conditions (C1)–(C4)
are not satisfied yet a solution may exist.
VI. N UMERICAL
EXAMPLES OF NON - CONVERGENCE
In this section, we provide two example networks where the
Z-Bus method does not converge. The first example (Section
VI-A) depicts a simple two-node network for which an analytical solution is found; however, convergence conditions are not
satisfied and the Z-Bus method oscillates. The second example
(Section VI-B) features a non-trivial unbalanced network with
constant-power loads where the convergence occurs only when
the loads are reduced by a certain percentage. The convergence
conditions also are satisfied for a close percentage, which
reveals that they are relatively tight in this example.
A. Balanced two-node network with power injection
Consider a two-node network where the slack bus S is connected to a secondary bus through an ideal phase-decoupled
three-phase line with real-valued series impedance yt . The
secondary bus contains real-valued ZIP components sL , iL ,
and yL on each phase, where iL , yL > 0 but sL < 0, that is,
the constant-power portion of the ZIP component is considered
to be a power injection. In this case, we have that sL = sL 1,
iL = iL 1, YL = yL I where 1 ∈ R3 is a vector of all one’s and
I ∈ R3×3 is the 3 × 3 identity matrix. Moreover, the network
admittance matrix is Y = −YNS = yt I. Denote the voltage
vector at the non-slack bus with v = {v a , v b , v c } and let the
2π
2π
slack bus voltage be fixed at vS = {1, e−j 3 , ej 3 }.
Due to the fact that the network is balanced and the values
of loads are real, it turns out that a solution of the form
2π
2π
v a = v b ej 3 = v c e−j 3 = v exists where v is a real number.
To demonstrate this, we write the Ohm’s law for phase a as
follows:
v
sL
− iL
= (yt + yL )v − yt .
(26)
−
v
|v|
Assuming that we are only interested in non-negative voltages v, that is, when |v| = v, the following quadratic equation
can be obtained for v:
(yt + yL )v 2 + (iL − yt )v + sL = 0.
The solution to (27) is explicitly given as
√
yt − i L
∆
v=
±
,
2(yt + yL ) 2(yt + yL )
(27)
(28)
with ∆ = (yt − iL )2 − 4(yt + yL )sL . Keep in mind that (28)
is not necessarily the only admissible voltage solution for this
network since we have only limited the search to v ∈ R≥0 .
In order to simplify the ensuing computations, let us select
yt = yL = 21 pu, and set the current injection iL = 21 pu.
Substituting these values into (28), the real non-negative
voltage solution v parameterized by sL is thus computed
√
v = −sL .
(29)
It is clear that for values sL √
< 0, the network admits at least
one voltage solution vsol = −sL vS .
At this point, it is our intention to show that even though
voltage solutions exist in this example network, the Z-Bus
method may not converge and the convergence conditions
(C1)–(C4) are not satisfied.
1
We first compute Z = (Y +YL )−1 = yt +y
I = I and w =
L
1
2 vS . By selecting Λ = I, convergence conditions (C1)–(C4)
for this network yield the following inequalities parameterized
by sL :
1 − 2R > 0,
1
2|sL |
+ ≤ R,
1 − 2R 2
4|sL |
2
+
< 1,
(1 − 2R)2
1 − 2R
(30a)
(30b)
(30c)
where in (30), due to the absence of delta-connected nodes,
(C2) is not present.
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
2
8
v[0] = vs
v[0] = 2vsol
v[0] = 0.5vsol
v[0] = vsol + 0.1vS
1.8
101
kv[t + 1] − v[t]k∞
1.6
102
kv[t]k∞
1.4
1.2
1
0.8
0.6
100
10−1
10−2
10−3
10−4
0.4
10−5
0.2
10−6
0
0
2
4
6
8
0
10
Iterations of the Z-Bus method performed on the two-node
test network while setting sL = −0.5 pu. Since Z-Bus convergence
conditions in (30) are not feasible for any sL , the convergence of the
Z-Bus method is not guaranteed, even when there is knowledge that
solution (29) exists.
Inequality (30b) can be transformed to the quadratic inequality 2R2 − 2R + 12 + 2|sL | ≤ 0, which does not have
a feasible solution in R for any sL < 0. However, existence
of power solutions for sL < 0 is easily perceived by (29).
Figure 6 plots the infinity norm of the iterates of the Z-Bus
method, that is, kv[t]k∞ , when sL = −0.5 pu. Figure 6 is
representative of the fact that the Z-Bus method does not
converge to the solution (29).
B. Unbalanced three-node network
2
4
6
8
10
12
14
16
18
20
Iteration t
Iteration t
Fig. 6.
θ=1
θ = 0.5
θ = 0.3
θ = 0.2
θ = 0.12
θ = 0.11
Iterations of the Z-Bus method performed on the threenode test network for several values of parameter θ. When the
value of θ is such that Z-Bus convergence conditions (33) are not
feasible, the convergence of the Z-Bus is not guaranteed. The black
line corresponding to θ = 0.12 depicts a very long oscillation.
As θ approaches θsol = 0.1060, the Z-Bus method converges.
sol
When
, a√ unique solution exists within DR where R ∈
√ θ ≤ θ
1− 1−9.4360θ 1+ 1−9.4360θ
[
,
].
2
2
Fig. 7.
the following inequalities in R parameterized by θ:
1−R>0⇒R<1
(33a)
2.359θ
≤ R ⇒ R2 − R + 2.359θ ≤ 0
(33b)
1−R
2.359θ
< 1 ⇒ R2 − 2R + 1 − 2.359θ > 0. (33c)
(1 − R)2
It is not hard to see that (33) has a valid solution for R only
sol
√ Z-Bus is
Consider a three-node network with the set of nodes if θ ≤ θ = 0.1060. The range of √R for convergent
1− 1−9.4360θ 1+ 1−9.4360θ
,
].
N = {1, 2, S} and the set of edges {(1, S), (1, 2)}. The series then given as [Rmin , Rmax ] = [
2
2
The Z-Bus method is run on this example network for several
admittances for edges (1, S) and (1, 2) are as follows:
representative values of θ using the initialization v[0] = w.
0.077 − j5.33 0.01 − j0.09 0.02 − j0.08
The corresponding Z-Bus iterations are shown in Fig. 7. It is
0.087 − j8 0.03 − j0.07 , (31a) revealed in Fig. 7, that for almost all percentages θ higher than
Y1S = 0.01 − j0.09
0.02 − j0.08 0.03 − j0.07 0.07 − j1.5
θsol the Z-Bus method fails to converge. However, for θ ≤ θsol ,
0.056 − j8.66
0
0.02 − j0.07
the Z-Bus iterations are guaranteed
to converge
√
√ to the unique
1+ 1−9.4360θ
0
0.02 − j4.8 0.03 − j0.05 , (31b) solution in DR where R ∈ [ 1− 1−9.4360θ
Y12 =
,
].
2
2
0.02 − j0.07 0.03 − j0.05 0.03 − j3.8
where all values are per unit. The load buses {1, 2}
only contain
values of
constant-power loads with per unit
T
and sL2 =
sL1 = θ 0.7 + j1.5 0.8 + j1.5 0.8 + j2.5
T
θ 0.6 + j2.5 0.6 + j0.5 0.3 + j0.5 where θ ∈ (0, 1] is
a parameter that is going to be varied to obtain convergence
threshold for the Z-Bus method.
Our intention is to find θ such that the convergence of the
Z-Bus method is guaranteed and numerically investigate how
conservative the theoretical bound is. First, network parameters
are computed:
Y1S + Y12 −Y12
Y=
(32a)
−Y12
Y12
−Y1S
v
YNS =
(32b)
, w= S .
O
vS
Next, to find θ such that Z-Bus convergence is guaranteed,
by setting Λ = I, the convergence conditions (C1)–(C4) yield
VII. C ONCLUDING R EMARKS
A set of sufficient conditions was derived that guarantee
the convergence of the Z-Bus method to a unique solution of
the power flow problem in three-phase distribution networks
with ZIP loads. It was numerically demonstrated that the IEEE
test feeders with wye and delta ZIP loads satisfy the derived
conditions. Characterizing the largest regions with unique
power flow solution is an open research question. Investigating
the effect of the design parameter Λ to this end is the subject
of future work.
A PPENDIX A
U SEFUL I NTERMEDIATE R ESULTS
This Appendix proves intermediate results on the product
Z[iPQ +iI ] that shows up in (12), and an inequality on complex
numbers, which will be later used.
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
A. Rearranging
P
k∈J∆
9
Zjk iPQ (v)k
B. Rearranging
∆
Lemma 1. Consider the matrix Z defined in (16g). For n ∈
N∆ , φ ∈ Ωn , k = Lin(n, φ) ∈ J∆ , j ∈ J , and indices
lk ∈ {1, . . . , |J∆ |}, the following equality holds:
k ∗
X
X
sL
Z∆
Zjk iPQ (v)k = −
.
(34)
jlk
Tv
e
k n
k∈J
k∈J
∆
∆
Proof. Using (4a) for constant-PQ loads with delta connections, we can write
X
Zjk iPQ (v)k =
k∈J∆
−
X
n∈N∆
′∗
X
ZjLin(n,φ)
φ∈Ωn
sφφ
Ln
X
′
T ∗
(eφφ
n ) vn
φ′ ∈Ωn /{φ}
.
(35)
Consider the term in parenthesis in (35). Since n ∈ N∆ , then
|Ωn | ≥ 2, i.e., node n has at least two available phases. If
|Ωn | = 2, then without loss of generality we can assume
Ωn = {φ, φ′ } where φ′ = r(φ) and r(φ′ ) ∈
/ Ωn . The term
in parenthesis in (35) can be written as follows:
′∗
ZjLin(n,φ)
=
=
sφφ
Ln
′
T ∗
(eφφ
n ) vn
′
+ ZjLin(n,φ′ )
sφLnφ
∗
(36)
where k ′ = Lin(n, r(φ)), and equality (a) results from the
fact that Z∆
/ Ωn [cf. (16g)].
jlk is zero when φ ∈ Ωn and r(φ) ∈
If node n has three phases then Ωn = {a, b, c} and r(a) = b,
r(b) = c, and r(c) = a. The term in parenthesis in (35) equals:
∗
∗
sac
sab
ZjLin(n,a) abLnT ∗ + ZjLin(n,a) acLnT ∗
(en ) vn
(en ) vn
∗
∗
sbc
sba
+ ZjLin(n,b) baLnT ∗ + ZjLin(n,b) bcLTn ∗
(en ) vn
(en ) vn
∗
∗
sca
scb
+ ZjLin(n,c) caLnT ∗ + ZjLin(n,c) cbLTn ∗
(en ) vn
(en ) vn
∗
ab∗
sL n
sab
Ln
= ZjLin(n,a) ab T ∗ − ZjLin(n,b) ba T ∗
(en ) vn
(en ) vn
∗
bc∗
sbc
sL n
Ln
+ ZjLin(n,b) bc T ∗ − ZjLin(n,c) bc T ∗
(en ) vn
(en ) vn
∗
∗
sca
sca
Ln
Ln
+ ZjLin(n,c) ca T ∗ − ZjLin(n,a) ca T ∗
(en ) vn
(en ) vn
∗
X
sφr(φ)
Ln
=
ZjLin(n,φ) − ZjLin(n,r(φ))
φr(φ) T ∗
(en
) vn
φ∈Ωn
∗
∗
k
k
X
X
s
sL
[Zjk − Zjk′ ] TL ∗ =
=
Z∆
jlk T ∗ .
ek vn
ek vn
k∈J
k∈J
n
k∈J∆
Zjk iI (v)k
Lemma 2. Consider the matrix Z∆ defined in (16g). For n ∈
N∆ , φ ∈ Ωn , k = Lin(n, φ) ∈ J∆ , j ∈ J , and indices
lk ∈ {1, . . . , |J∆ |}, the following equality holds:
X
X
ikL eTk vn
Z∆
.
(38)
Zjk iI (v)k =
jlk
|eTk vn |
k∈J
k∈J
∆
n
′∗
Proof. Follow Appendix A-A with the terms
∗
sk
L
∗
eT
v
k n
replaced by
′
φφ′ T
iφφ
Ln (en ) vn
′
T
|(eφφ
n ) vn |
and
T
ik
L ek vn
|eT
k vn |
sφφ
Ln
′
T ∗
(eφφ
n ) vn
and
, respectively.
C. A useful inequality for complex numbers
Lemma 3. For two complex numbers x, y 6= 0 we have that
y
|x − y|
x
−
.
≤2
|x| |y|
|x|
(39)
Proof. The following hold:
x
y
y
y
y
x
≤
+
−
−
−
|x| |y|
|x| |x|
|x| |y|
|y| − |x|
|x − y|
|x − y|
=
+ |y|
≤2
. (40)
|x|
|x||y|
|x|
Notice that (40) holds since ||y| − |x|| ≤ |x − y|.
′
(eφn φ )T vn∗
φr(φ)∗
φr(φ)∗
sL n
sL n
ZjLin(n,φ) φr(φ)
− ZjLin(n,r(φ)) φr(φ)
(en
)T vn∗
(en
)T vn∗
∗
∗
∗
sk
skL (a) X ∆ skL
Z
[Zjk − Zjk′ ] TL ∗ = Z∆
=
jlk T ∗ ,
jlk T ∗
ek vn
ek vn
ek vn
k∈Jn
P
A PPENDIX B
P ROOF OF T HEOREM 1
In this section, the proof of Theorem 1 is given. First,
to guarantee that for all u ∈ DR the corresponding lineto-neutral and line-to-line voltage magnitudes are positive,
conditions (C1) and (C2) are derived in Appendices B-A and
B-B respectively. Condition (C3) ensures the self-mapping
property of T(u) and is derived in Appendix B-C. Appendix
B-D derives condition (C4), which guarantees the contraction
property of T(u).
A. Bounds on line to neutral voltages
This section proves two results. First, a lower bound on all
line to neutral voltages is derived in (42) which will be later
used in proving the self-mapping and contraction properties.
Second, condition (C1) is proved.
Due to condition u ∈ DR , the following holds for k ∈ J :
wk
wk
| ≤ R ⇒ |uk −
| ≤ R,
max |uk −
k∈J
λk
λk
⇒ |λk uk − wk | ≤ |λk |R ⇒ |vk − wk | ≤ R|λk |,
⇒ |wk | − |vk | ≤ R|λk | ⇒ |wk | − R|λk | ≤ |vk |
⇒ |wk |(1 − R|λk |/|wk |) ≤ |vk |.
(37)
n
For |Ωn | = 2 or |Ωn | = 3, substituting the appropriate term
from (36) or (37) for the parenthesis in (35) yields (34).
(41)
(42)
Equation (42) provides a lower bound for the magnitude of
line to neutral voltages in terms of R and w.
The proof of (C1) is based on (42). In (42), in the term in
parenthesis, we replace |wk | by its minimum w [defined in
(16a)], and replace |λk | by its maximum λ̄ [also defined in
(16a)] to obtain a lower bound for the left-hand side of (42)
which is common for all k ∈ J :
|wk |(1 − R|λk |/|wk |) ≥ |wk |(1 − Rλ̄/w).
(43)
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
10
To ensure that for all u ∈ DR , the magnitude of the line to
neutral voltages is positive, the right-hand side of (43) must
be greater than zero, yielding condition (C1).
The right-hand side of (49) is the sum of four terms. An
upper bound for each term is provided next.
For the first term in (49), the following holds:
skL ∗
)
vk
k∈JY
k∈JY
" ∗ ∗ ∗ #
X
skL
skL
skL
=
Zjk
−
+
vk
wk
wk
B. Bounds on line to line voltages
Similar to Appendix B-A, this section proves two results.
First, we find a lower bound for line to line voltages in delta
connections. Next we derive condition (C2).
For n ∈ N∆ , any line to line voltage can be written as
eTk vn , where k = Lin(n, φ) and φ ∈ Ωn . Using the triangle
inequality, we find the following:
|eTk vn | = |eTk (vn − wn ) + eTk wn |
≥
|eTk wn |
−
|eTk (vn
X
≤
≤
(44)
Due to Hölder’s inequality, for n = Node[k] we have that
=
|eTk (vn − wn )| ≤ kek k1 kvn − wn k∞ ,
(a)
l∈Jn
⇒ |eTk (vn − wn )| ≤ 2 max{|λl ||ul −
l∈Jn
wl
|}
λl
(b)
⇒ |eTk (vn − wn )| ≤ 2R max{|λl |},
l∈Jn
(45)
where (a) holds because kek k1 = 2 for k ∈ J , and (b) holds
due to the assumption u ∈ DR . Using (45) in (44) we find:
|eTk vn |
≥
|eTk wn |
− 2R max{|λl |}
(46)
l∈Jn
≥ |eTk wn |(1 − 2R max{|λl |}/|eTk wn |).
l∈Jn
l∈Jn
(48)
To guarantee that for all u ∈ DR , the magnitude of line-to-line
voltages is positive, (48) must be greater than zero, yielding
condition (C2).
C. Self-mapping property
To obtain a sufficient condition for the self-mapping property, it is assumed that u ∈ DR , and an upper bound for
the term kT(u) − Λ−1 wk∞ is sought. The j-th entry of
kT(u) − Λ−1 wk∞ can be upper-bounded as follows:
≤
1
|λj |
+
X
k∈JY
wj
1
wj
wj
|=
Zj• [iPQ (v) + iI (v)] +
−
λj
λj
λj
λj
X
X
Zjk iPQ (v)k
Zjk iPQ (v)k +
|Zjk |
k∈JY
X
X
k∈J∆
Zjk iI (v)
!
.
k∈JY
|skL |R|λk |
(|wk | − R|λk |) |wk |
|Zjk ||skL |
(b)
|wk | − R|λk |
≤
≤
X
|Z∆
jlk |
X
|Z∆
jlk |
k∈J∆
k∈J∆
≤
X
k∈JY
X
k∈JY
|Zjk |
|skL |
|wk |
|Zjk ||skL |
,
|wk |(1 − Rλ̄/w)
(50)
X
|skL ||eTk (wn − vn )|
skL
|Z∆
+
jlk |
T
T
T
|ek vn ||ek wn |
e k wn
k∈J
∆
|skL |2R max{|λl |}
l∈Jn
(|eTk wn | − 2R max{|λl |})|eTk wn |
l∈Jn
skL
|Z∆
|
+
jlk
eTk wn
k∈J∆
k
X |Z∆
2R maxl∈Jn {|λl |}
jlk ||sL |
≤
+
1
|eTk wn | |eTk wn | − 2R maxl∈Jn {|λl |}
k∈J
∆
≤
(b)
≤
k
|Z∆
jlk ||sL |
X
|eTk wn |(1 − 2R maxl∈Jn {|λl |}/|eTk wn |)
X
|eTk wn |(1 − 2Rλ̄/ρ)
k∈J∆
k
|Z∆
jlk ||sL |
k∈J∆
,
(51)
where n = Node[k], index lk is defined in (16g), inequality
(a) is due to (45) and (46), and inequality (b) is due to (48).
The upper bound for the third term in (49), is given by
X
iI (v) =
k∈JY
X
k∈JY
Zjk
X
vk k
iL ≤
|Zjk ||ikL |.
|vk |
(52)
k∈JY
Lemma 2 is invoked to provide an upper bound for the
fourth term in (49):
X
Zjk iI (v)k =
X
Zjk iI (v)k ≤
k∈J∆
(49)
+
X
k∈J∆
k∈JY
Zjk iI (v) +
X
k∈JY
X
|sk |
|skL ||wk − vk |
|Zjk | L
+
|vk ||wk |
|wk |
∆
(a)
|eTk wn |(1 − 2R max{|λl |}/|eTk wn |)
|T(u)j −
|Zjk |
where (a) is due to (41), and (b) is due to (42) and (43).
For the second term in (49), we use Lemma 1, and we write
k ∗
X
X
sL
∆
Zjk iPQ (v)k =
Zjlk
Tv
e
k n
k∈J∆
k∈J∆
"
k ∗ #
∗
∗
X
skL
sL
skL
∆
−
+
Zjlk
=
T
T
T
ek vn
e k wn
ek vn
k∈J
(47)
Equation (47) provides a lower bound for the magnitude of
line to line voltages in terms of R and w.
The proof of (C2) is based on (47). In (47), in the term
in parenthesis, we replace the term |eTk wn | by its minimum
ρ [given in (16a)], and replace |λk | by its maximum λ̄ [also
given in (16a)] to obtain a lower bound for the line to line
voltages which is common for all k ∈ J∆ and n = Node[k]:
≥ |eTk wn |(1 − 2Rλ̄/ρ).
X
k∈JY
⇒ |eTk (vn − wn )| ≤ 2 max{|vl − wl |}
Zjk (
k∈JY
(a)
− wn )|
X
Zjk iPQ (v)k =
⇒
k∈J∆
X
Z∆
jlk
k∈J∆
X
k∈J∆
eTk vn k
i
|eTk vn | L
k
|Z∆
jlk ||iL |.
(53)
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
11
Using equations (50), (51), (52), and (53) in (49) yields
"
X
|Zjk ||skL |
1
wj
|≤
|T(u)j −
λj
|λj |
|wk |(1 − Rλ̄/w)
k∈J
≤
≤
X
+
k∈J∆
X
+
k∈JY
k
|Z∆
jlk ||sL |
|eTk wNode[k] |(1 − 2Rλ̄/ρ)
|Zjk ||ikL | +
X
k∈J∆
#
k
|Z∆
jlk ||iL | .
≤
(54)
The self-mapping property holds if the right-hand side of (54)
is upper-bounded by R for all j ∈ J , which is ensured by
(C3).
D. Contraction property
To obtain a sufficient condition for the contraction property,
we seek to find an upper-bound for the term kT(u)−T(ũ)k∞
proportional to ku − ũk∞ . Setting v = Λu and ṽ = Λũ, the
j-th entry of T(u) − T(ũ) is upper-bounded as follows:
|T(u)j − T(ũ)j |
1
1
=
Zj• [iPQ (v) + iI (v)] −
Zj• [iPQ (ṽ) + iI (ṽ)]
λj
λj
X
1
Zjk [iPQ (v)k − iPQ (ṽ)k ]
≤
|λj |
X
Zjk [iPQ (v)k − iPQ (ṽ)k ]
X
Zjk [iI (v)k − iI (ṽ)k ]
X
Zjk [iI (v)k − iI (ṽ)k ]
k∈J∆
+
k∈JY
+
k∈J∆
!
k∈JY
(c)
≤
(b)
≤
X
k∈JY
k∈JY
|λk uk − λk ũk |
(|wk | − R|λk |)2
|wk
− ũk∞ ,
(56)
where inequality (a) is due to (41) and (b) is due to (43).
Using Lemma 1 for the second term in (55) yields:
X
Zjk [iPQ (v)k − iPQ (ṽ)k ]
X
Z∆
jlk
k∈J∆
=
k∈J∆
X
|eTk wn |2 (1 − 2Rλ̄/ρ)2
k
2|Z∆
jlk ||sL |
k∈J∆
X
Zjk [iI (v)k − iI (ṽ)k ]
X
Zjk
max{|λl |}kun − ũn k∞
l∈Jn
max{|λl |}ku − ũk∞ , (57)
l∈Jn
vk k
ṽk k
i −
i
|vk | L |ṽk | L
X 2|Zjk ||ik |
(b) X
2|Zjk ||ikL |
L
|vk − ṽk | ≤
|vk − ṽk |
|vk |
|wk | − R|λk |
k∈JY
k∈JY
X
k∈JY
(55)
|Zjk ||skL |
|Zjk ||skL |
|λk |ku
|2 (1 − Rλ̄/w)2
|eTk wn |2 (1 − 2Rλ̄/ρ)2
− 2R maxl∈Jn {|λl |}/|eTk wn |)2
k
2|Z∆
jlk ||sL |
k∈J∆
k∈JY
≤
k
X
X |Zjk ||sk |
sL ∗
skL ∗
L
Zjk ( ) − ( )
=
≤
|vk − ṽk |
vk
ṽk
|vk ||ṽk |
≤
X
k
2|Z∆
jlk ||sL |kvn − ṽn k∞
k∈JY
k∈JY
(a)
|eTk wn |2 (1
where n = Node[k], (a) is due to Hölder’s inequality and
(46), and (b) is due to (48).
For the third term in (55) we have that:
(a)
The expression in (55) is a sum of four terms. In what
follows, we find an upper bound for each term.
For the first term we have that
X
Zjk [iPQ (v)k − iPQ (ṽ)k ]
X
X
2|Zjk ||ikL |
|wk |(1 − Rλ̄/w)
|λk |ku − ũk∞ ,
X
Zjk [iI (v)k − iI (ṽ)k ]
X
Z∆
jlk
k∈J∆
=
k∈J∆
(a)
≤
(b)
≤
(c)
≤
≤
eTk vn k
eT ṽn k
iL − Tk
i
T
|ek vn |
|ek ṽn | L
k
X 2|Z∆
jlk ||iL |
k∈J∆
|eTk vn |
skL
eTk vn
∗
−
skL
eTk ṽn
∗ #
|eTk vn − eTk ṽn |
k
X 2|Z∆
jlk ||iL |kek k1 kvn − ṽn k∞
k∈J∆
|eTk wn | − 2R maxl∈Jn {|λl |}
X
k
4|Z∆
jlk ||iL |
max{|λl |}kun
T
|ek wn |(1 − 2Rλ̄/ρ) l∈Jn
X
|eTk wn |(1 − 2Rλ̄/ρ)
k∈J∆
k∈J∆
k
4|Z∆
jlk ||iL |
− ũn k∞
max{|λl |}ku − ũk∞ , (59)
l∈Jn
where n = Node[k], (a) is due to Lemma 3, (b) is due to (46)
and Hölder’s inequality, and (c) is due to (48).
Using (56), (57), (58), and (59) in (55) yields:
"
X
|Zjk ||skL ||λk |
1
|T(u)j − T(ũ)j | ≤
|λj |
|wk |2 (1 − Rλ̄/w)2
k∈J
Y
"
(58)
where inequality (a) is due to Lemma 3, (b) is due to (41), and
(c) is due to (43) and the fact that kuk − ũk k ≤ ku − ũk∞ .
For the fourth term, Lemma 2 is used to obtain:
.
k∈JY
(|eTk wn | − 2R maxl∈Jn {|λl |})2
|eTk (vn − ṽn )|
k
|Z∆
jlk ||sL |kek k1 kvn − ṽn k∞
k∈J∆
=
k∈JY
+
≤
X
k∈J∆
(b)
≤
|(eTk vn )||(eTk ṽn )|
k∈J∆
(a)
Y
k
|Z∆
jlk ||sL |
X
+
k
X 2|Z∆
jlk ||sL | maxl∈JNode[k] {|λl |}
k∈J∆
|eTk wNode[k] |2 (1 − 2Rλ̄/ρ)2
IEEE TRANSACTIONS ON POWER SYSTEMS (ACCEPTED)
+
X
k∈JY
+
2|Zjk ||ikL ||λk |
|wk |(1 − Rλ̄/w)
k
X 4|Z∆
jlk ||iL | maxl∈JNode[k] {|λl |}
k∈J∆
12
|eTk wNode[k] |(1 − 2Rλ̄/ρ)
#
Mohammadhafez Bazrafshan received the B.S. degree from Iran University
of Science and Technology in 2012, and the M.S. degree in 2014 from the
University of Texas at San Antonio both in electrical engineering, where he
is currently a PhD student with research interests in smart power grids.
ku − ũk∞
(60)
For T(u) to be a contraction, the coefficient multiplying ku −
ũk∞ must be less than 1 for all j ∈ J , which is ensured by
(C4).
R EFERENCES
[1] T. H. Chen, M. S. Chen, K. J. Hwang, P. Kotas, and E. A. Chebli,
“Distribution system power flow analysis-a rigid approach,” IEEE Trans.
Power Del., vol. 6, no. 3, pp. 1146–1152, July 1991.
[2] B. C. Lesieutre, P. W. Sauer, and M. A. Pai, “Existence of solutions
for the network/load equations in power systems,” IEEE Trans. Circuits
Syst. I: Fundam. Theory Appl., vol. 46, no. 8, pp. 1003–1011, Aug 1999.
[3] D. K. Molzahn, B. C. Lesieutre, and C. L. DeMarco, “A sufficient
condition for power flow insolvability with applications to voltage
stability margins,” IEEE Trans. Power Syst., vol. 28, no. 3, pp. 2592–
2601, Aug. 2013.
[4] F. Wu and S. Kumagai, “Steady-state security regions of power systems,”
IEEE Trans. Circuits Syst., vol. 29, no. 11, pp. 703–711, Nov 1982.
[5] S. Bolognani and S. Zampieri, “On the existence and linear approximation of the power flow solution in power distribution networks,” IEEE
Trans. Power Syst., vol. 31, no. 1, pp. 163–172, Jan. 2016.
[6] H. D. Chiang and M. E. Baran, “On the existence and uniqueness of
load flow solution for radial distribution power networks,” IEEE Trans.
Circuits Syst., vol. 37, no. 3, pp. 410–416, Mar 1990.
[7] S. Yu, H. D. Nguyen, and K. S. Turitsyn, “Simple certificate of
solvability of power flow equations for distribution systems,” in Proc.
IEEE PES General Meeting, July 2015, pp. 1–5.
[8] C. Wang, A. Bernstein, J. Y. L. Boudec, and M. Paolone, “Explicit conditions on existence and uniqueness of load-flow solutions in distribution
networks,” IEEE Trans. Smart Grid, 2016.
[9] T. Q. Zhao, H. D. Chiang, and K. Koyanagi, “Convergence analysis of
implicit Z-bus power flow method for general distribution networks with
distributed generators,” IET Generation, Transmission & Distribution,
vol. 10, no. 2, pp. 412–420, Feb. 2016.
[10] “Load representation for dynamic performance analysis [of power systems],” IEEE Trans. Power Syst., vol. 8, no. 2, pp. 472–482, May 1993.
[11] W. H. Kersting, Distribution System Modeling and Analysis, 3rd ed.
CRC Press, 2002.
[12] A. Garces, “A linear three-phase load flow for power distribution
systems,” IEEE Trans. Power Syst., vol. 31, no. 1, pp. 827–828, Jan.
2016.
[13] T. H. Chen, M. S. Chen, T. Inoue, P. Kotas, and E. A. Chebli, “Threephase cogenerator and transformer models for distribution system analysis,” IEEE Trans. Power Del., vol. 6, no. 4, pp. 1671–1681, Oct. 1991.
[14] J. M. Ortega and W. C. Rheinboldt, Iterative Solution of Nonlinear
Equations in Several Variables. Academic Press, Inc., 1970.
[15] D. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation:
Numerical Methods. Massachussetts, USA: Athena Scientific, 1989.
[16] J. Nocedal and S. J. Wright, Numerical Optimization. New York, USA:
Springer, 2006.
[17] R. W. D. Nickalls, “The quartic equation: Invariants and euler’s solution
revealed,” The Mathematical Gazette, vol. 93, no. 526, pp. 66–75,
2009. [Online]. Available: http://www.jstor.org/stable/40378672
[18] M. Bazrafshan and N. Gatsis, “On the solution of the three-phase loadflow in distribution networks,” in 50th Asilomar Conference on Signals,
Systems, and Computers, Nov. 2016.
[19] W. H. Kersting, “Radial distribution test feeders,” in Proc. IEEE PES
Winter Meeting, vol. 2, 2001, pp. 908–912 vol.2.
[20] “Distribution Test Feeders,” https : //ewh.ieee.org/soc/pes/dsacom/
testfeeders/, accessed: 2016-10-12.
[21] M. J. Gorman and J. J. Grainger, “Transformer modelling for distribution
system studies part II: Addition of models to YBUS and ZBUS,” IEEE
Trans. Power Del., vol. 7, no. 2, pp. 575–580, April 1992.
Nikolaos Gatsis received the Diploma degree in Electrical and Computer
Engineering from the University of Patras, Greece, in 2005 with honors.
He completed his graduate studies at the University of Minnesota, where
he received the M.Sc. degree in Electrical Engineering in 2010, and the
Ph.D. degree in Electrical Engineering with minor in Mathematics in 2012.
He is currently an Assistant Professor with the Department of Electrical and
Computer Engineering at the University of Texas at San Antonio. His research
interests lie in the areas of smart power grids, renewable energy management,
communication networks, and cyberphysical systems, with an emphasis on
optimal resource management. Prof. Gatsis co-organized symposia in the
area of Smart Grids in IEEE GlobalSIP 2015 and IEEE GlobalSIP 2016.
He also served as a Technical Program Committee member for symposia
in IEEE SmartGridComm 2013, IEEE SmartGridComm 2014, and IEEE
SmartGridComm 2015.
| 3cs.SY
|
A Study of Quasi-Gorenstein Rings
arXiv:1508.04597v2 [math.AC] 4 Sep 2016
EHSAN TAVANFAR and MASSOUD TOUSI
A BSTRACT. In this paper several quasi-Gorenstein counterparts to some known properties of Gorenstein rings are given. We, furthermore, give an explicit description of the
attached prime ideals of certain local cohomology modules.
1 Introduction
Throughout this article (R, m) is a commutative Noetherian local ring of dimension d with
identity where m denotes the unique maximal ideal of R. Furthermore, M always stands for
a finitely generated d ′ -dimensional R-module.
Following [2], we say that R is a quasi-Gorenstein ring precisely when Hdm (R) ∼
= E(R/m)
or, equivalently, R has a canonical module which is a rank one free module. In the geometric vein, a normal projective variety X is quasi-Gorenstein if the canonical divisor KX of X is
Cartier. Indeed, a Cohen-Macaulay quasi-Gorenstein ring is Gorenstein. According to [21]
(c.f. [38]), roughly speaking, quasi-Gorenstein rings arise from the theory of linkage. That is,
loosely speaking, they are residue rings of a Gorenstein ring modulo an ideal which is linked
to an unmixed almost complete intersection. Hence there are so many of them. From anb satisfies the S 2 condition then the trivial
other perspective, if R has a canonical module and R
extension of R by its canonical module is quasi-Gorenstein (see [2]). This, again, shows the
ubiquity of quasi-Gorenstein rings. When R is a complete normal domain with a canonical ideal ωR , the first author of the present paper, in [43, Remark 3.2.] and [43, Theorem
L
3.5.(i)], endowed R ωR with another R-algebra structure, by which it is a quasi-Gorenstein
domain (Note that the trivial extension is never domain). It is also worthwhile to mention
that by, [13, Lemma (2.4)], the class of quasi-Gorenstein rings contains the class of unique
factorization domains with canonical module. Such non-Cohen-Macaulay unique factorization domains have their origin in the invariant theory, see [5]. Rees Algebras provides us
with another important source of quasi-Gorenstein rings (see, e.g., [28] and [22]). In the light
of [41, Theorem 6.1.], there are (even) isolated rational singularity Q-Gorenstein rings with
non-Cohen-Macaulay quasi-Gorenstein cyclic cover. In [27, Theorem, page 336] the author
2010 Mathematics Subject Classification. 13H10, 13D45.
Key words and phrases. Canonical module, G-dimension, Gorenstein rings, limit closure, local cohomology, quasi-Gorenstein rings.
This research was in part supported by a grant from IPM (No. 93130211).
1
2
A S TUDY OF QUASI -G ORENSTEIN R INGS
gives a nice description of non-Cohen-Macaulay quasi-Gorenstein Fano 3-folds1 (over C)
with isolated non-rational loci, as the projective cone defined by an ample invertible sheaf
L on an Abelian surface. In [23] the authors give examples of quasi-Gorenstein Buchsbaum
affine semigroup rings of any admissible depth. It is noteworthy to point out that affine semigroup rings of [23] are not the only place where the non-Cohen-Macaulay quasi-Gorenstein
Buchsbaum rings come from. Another source of such rings is the Segre product of two hypersurfaces of a-invariant zero. For example, over an algebraically closed field k, the Segre
product ring,
¡
¢ ¡
¢
R := k[a, b, c]/(a 3 + b 3 + c 3 ) # k[x, y, z, w]/(x 4 + y 4 + z 4 + w 4 ) ,
is a quasi-Gorenstein Buchsbaum normal domain of dimension 4 and depth 2 (see, [19, Theorem (4.1.5)], [19, Proposition (4.2.2)], [19, Theorem (4.3.1)] and [32, Theorem A]). We end
this part of the introduction by stating that, in the light of Kunz’s [29, Proposition 1.1], the
class of almost complete intersections and the class of quasi-Gorenstein rings are disjoint.
A vast amount of research has been devoted to studying the class of Gorenstein rings.
Compared with Gorenstein rings, our understanding of the behaviour and properties of
quasi-Gorenstein rings is limited. In this paper we aim to increase our knowledge about
quasi-Gorenstein rings. For example, in Section 3 we deal with some natural questions concerning the interaction between quasi-Gorensteinness and regular elements. Namely, we
prove the following fact:
Theorem 1.1. (See Proposition 3.1 and Corollary 3.4(i)) If R is quasi-Gorenstein and x is a regular element of R, then R/xR is quasi-Gorenstein if and only if R/xR satisfies the S 2 condition
S
and this is equivalent to the assertion that x ∉
¡
¢ p.
p∈AttR Hdm−1 (R)
Accordingly, the main obstruction here is the failure of S 2-condition. Hence, the next natural question is that whether the S 2 -ification of R/xR is, always, quasi-Gorenstein provided R
is quasi-Gorenstein and R/xR has a canonical module? We settle this question affirmatively.
Theorem 1.2. (See Corollary 3.4(ii)) If R is quasi-Gorenstein and x is a regular element of R
such that R/xR has a canonical module, then the S 2 -ification of R/xR is quasi-Gorenstein.
Perhaps it is worth to stress the following application of Theorem 1.1 to the homological conjectures. In [43, Remark 3.6.(i)], applying both of two characterizations of quasiGorensteinness of R/xR, given in Theorem 1.1, as well as the first author’s reduction of some
homological conjectures to normal complete quasi-Gorenstein domains in [43], together
with Ochiai and Shimomoto’s [34], it is deduced that the validity of the Hochster’s Canonical
Element Conjecture in a dense open subset (in some sense) settles the Canonical Element
Conjecture in general.
Another immediate question is the deformation property of quasi-Gorenstein rings. In
[43, Proposition 3.4.], the first author of the present paper, shows that the Ulrich’s deformation of certain Gorenstein rings to unique factorization domains, developed in [44], has a
1
This means a 3-dimensional normal projective variety X such that its anti-canonical divisor, −KX , is an
ample Cartier divisor.
E HSAN TAVANFAR
AND
3
M ASSOUD T OUSI
quasi-Gorenstein counterpart. Although, at the time of writing the paper, we do not know
whether the quasi-Gorensteinness deforms in general, but we show that the following variant of deformation of quasi-Gorensteinness, which we call it as analytic deformation2, holds
true.
Theorem 1.3. (See Theorem 3.7) Suppose that x is a regular element of R such that R/x n R is
quasi-Gorenstein for each n ∈ N. Then R is quasi-Gorenstein.
In order to deduce Theorem 1.1 and Theorem 1.2, we used a special case of the following
fact which is the main result of Section 2.
Theorem 1.4. (See Theorem 2.9(ii)) Let M be a formally equidimensional module satisfying
the Serre-condition S 2 . Assume, additionally, that M has a canonical module ωM 3 . Set,
n := inf{i : i ≥ 1 and Hd−i
m (M) 6= 0}.
If the formal fibers of R satisfy the S n+2 -condition4 , then
¡
¢
¡
¢
(ω
)
AttR Hd−n
= n + 1 and htωM p ≥ n + 2}.
(M)
=
{
p
∈
Spec(R)
:
Depth
M
p
m
Rp
The section 2. of the paper is, also, devoted to reminding some definitions and wellknown facts as well as giving some remarks and lemmas which are required throughout the
paper.
By, e.g. [15], a Cohen-Macaulay local ring is Gorenstein if and only if Hdm (R) has finite
injective dimension. In [12], the authors introduce the notion of Gorenstein injective modules as a generalization of the concept of injective modules. Gorenstein injective modules
and, its related invariant, the Gorenstein injective dimension are studied by many authors.
For example, in [45, Theorem 2.6] the author shows that if R is a Cohen-Macaulay complete
local ring and Hdm (R) is Gorenstein injective then R is Gorenstein. Subsequently, in [46], the
authors relax the complete assumption and prove that even if Hdm (R) has finite Gorenstein
injective dimension, then the Cohen-Macaulay ring R is Gorenstein. On the other hand, for
the injective dimension case, there exists
counterpart. That is, R is a
¡ da quasi-Gorenstein
¢
quasi-Gorenstein ring if and only if IdR Hm (R) < ∞ ([1, Theorem 3]). This encouraged us
to investigate the Gorenstein injective version of the Aoyama’s theorem which is, indeed, a
theorem whenever R is Cohen-Macaulay. We proved the following fact which also recovers
the Cohen-Macaulay case of [46].
Theorem 1.5. (See Theorem 5.7) Assume that Hdm (R) has finite Gorenstein injective dimension.
If, furthermore,
DepthRb (ωRb )
Ext1Rb (ωRb , ωRb ) = . . . = Ext b
R
then R is quasi-Gorenstein.
2
(ωRb , ωRb ) = 0,
As far as we know, our paper is the first place where such a variant of deformation is studied. This is remarkable, because the ordinary deformation problem of quasi-Gorensteinness is mysterious.
3
See, Definition and Remark 2.2 for the definition of the canonical module of an R-module.
4
e.g., if R is a homomorphic image of a Gorenstein ring.
4
A S TUDY OF QUASI -G ORENSTEIN R INGS
The proof of Theorem 1.5 shows that the vanishing of Ext1R (ω, ωR ) and Ext2R (ωR , ωR ), in
the following sense, is sufficient for our purpose .
Question 1.6. (See Question 4.7) Assume that R has a canonical module ωR and that,
GdimR (ωR ) < ∞.
Then do we have, Ext1R (ωR , ωR ) = Ext2R (ωR , ωR ) = 0?
In fact an affirmative answer to the above question, which is proposed in Section 4, shows
that R is quasi-Gorenstein if and only if Hdm (R) has finite Gorenstein injective dimension.
In [4], the famous theorem of Bass gives a criterion for Gorensteinness in terms of the
irreducibility of all of the parameter ideals. Indeed, there exists a non-Gorenstein ring with
an irreducible parameter ideal. But, by virtue of [30], there is an invariant ℓd (R) of R such
that R is Gorenstein if and only if some parameter ideal contained in mℓd (R) is irreducible.
In Section 5, we prove the following quasi-Gorenstein version of [30] (see Section 5 for the
definition of the limit closure).
b is unmixed and the limit
Theorem 1.7. (See Theorem 5.7) R is quasi-Gorenstein if and only if R
closure of each parameter ideal (of some parameter ideal contained in mℓd (R) ) is irreducible.
It is worth pointing out that [30, Proposition 2.3.], [14], and [35, 3.2 Theorem] also show
that the limit closure of parameter ideals of non-Cohen-Macaulay rings can be considered as
a counterpart to the parameter ideals of Cohen-Macaulay rings. We would like to stress that
Theorem 1.7 may have the following computational application. According to Theorem 1.7
for verifying the quasi-Gorensteinness of a ring R, using
system, we can find a sys¡ Macaulay2
¢
lim
lim
tem of parameters x of R and then check whether, µ {x}lim
:
m
=
µ({x}
R
R ) + 1, wherein {x}R
denotes the limit closure of x. 5 . This can be, considerably, faster than computing the free
resolutions, especially when the defining ideal of R has too many generators or generators
with too many summands. It is noteworthy to mention, here, that, e.g., the limit closure of
any system of parameters x of a Buchsbaum local ring R is just the ideal, (x12 , . . . , xd2 ) : x1 · · · xd
(see, [17, Lemma (4.6).]).
2 Attached Prime ideals of Local Cohomology modules
Reminder 2.1. Here we aim to review some, probably known, definitions and facts.
(i) Throughout the paper, we use the notation Assht to denote the set,
AsshtR (M) := {p ∈ AssR (M) : dim(R/p) = d ′ }.
5
M is said to be equidimensional (respectively, unmixed) if Assht(M) = minAssR (M) (respectively, Assht(M) = AssR (M)). M is called formally equidimensional (respectively, forb 6 is an equidimensional (respectively, unmixed) R-module.
b
mally unmixed) if M
Here, the notation µ stands for the minimal number of generators. Recall that a primary ideal a of (R, m) is
irreducible if and only if µ(a : m) = µ(a) + 1.
6
b stands for the completion of M with respect to the m-adic topology.
Here the notation M
E HSAN TAVANFAR
AND
5
M ASSOUD T OUSI
(ii) For each ideal a of R we set,
¡
¢
htM a := htR/0:R M (a + 0 :R M)/0 :R M .
Note that, dimRp Mp = htM p, for each p ∈ SuppR (M). Thus,
htM a = inf{htM p : p ∈ Var(a)
\
SuppR M} = inf{dimRp Mp : p ∈ Var(a)
\
SuppR M}.
(iii) We say that M satisfies the Serre-condition S n precisely when,
DepthRp (Mp ) ≥ min{n, dimRp (Mp )},
for each p ∈ Spec(R). In fact, the following statements are equivalent.
(a) M satisfies S n .
(b) grade(a, M) ≥ min{n, htM a} for each ideal a of R.
(c) if a is an ideal generated by an M-regular sequence of length j ≤ n −1 then for each
p ∈ AssR (M/aM) we have htM (p) = j .
¡
¢
(iv) We use the notation, M∨ , to denote the Matlis dual HomR M, E(R/m) of an R-module
M. The canonical module for R, denoted by ωR , is a finitely generated R-module such
∼ d
that, ω∨
R = Hm (R). The module ωR , if exists, is an d -dimensional R-module satisfying
the Serre-condition S 2 and is unique up to isomorphism ([6, 12.1.6 Theorem], [6, 12.1.9
Proposition(i)] [6, 12.1.18 Theorem(i)]). So R is a quasi-Gorenstein ring if and only if
R has a canonical module which is a rank one free R-module. In particular a quasiGorenstein ring is S 2 and unmixed ([3, Lemma 1.1]). A fortiori, by [2, (1.8)] we have,
0 :R ωR =
\
q,
(1)
q
where the intersection runs over the primary components q of the zero ideal satisfying
dim(R/q) = d . This, again, implies the unmixedness of quasi-Gorenstein rings. If R is
a homomorphic image of a Gorenstein local ring S of dimension n then by virtue of
[6, 11.2.6 Local Duality Theorem ] we have, ωR := ExtSn−d (R, S).
(v) (See [39, Lemma 1.9.]) Assume that R is a homomorphic image of a Gorenstein local
ring S of dimension n. For each l ∈ Z set, ωlM := ExtSn−l (M, S). By [6, 11.2.6 Local Duality
Theorem] there are functorial isomorphisms
Hlm (N) ∼
= ExtSn−l (N, S)∨ ,
(2)
6
A S TUDY OF QUASI -G ORENSTEIN R INGS
M
for all finitely generated R-modules N. The R-module ωdim
is said to be the canoniM
cal module of M and is denoted by ωM . The R-module ωM satisfies S 2 and AssR (ωM ) =
AsshtR (M). In particular, dim(ωM ) = dim(M) (see [39, Lemma 1.9.(c)]).
dim(M)−dim(R/p)
Let p ∈ SuppR (ωM ). Then, by [39, Lemma 1.9(a)], (ωM )p ∼
= ωMp
. It follows,
dim(M)−dim(R/p)
from (2), that HpRp
(Mp ) 6= 0. Consequently, dim(R/p) + dim(Mp ) = dim(M).
In particular, we get (ωM )p = ωMp .
The argument of the preceding paragraph in conjunction with [39, Lemma 1.9(b)]
yields,
SuppR (ωM ) = {p ∈ SuppR (M) : dim(Mp ) + dim(R/p) = dim(M)}.
Similarly, as in part (iv) of the preceding reminder, the definition of the canonical module
of a finitely generated R-module can be extended to all local rings.
Definition and Remark 2.2. Let M be a finitely generated d ′ -dimensional R-module. Then
∼ d′
we say that a finitely generated R-module ωM is the canonical module for M if ω∨
M = Hm (M).
A similar argument as in the proof of [6, 12.1.6 Theorem] shows that the canonical module
N
∼
b
ωM of M, if exists, is unique up to isomorphism. Note that, ωM
b = ωM
R R. By Reminder 2.1(v),
b
M always has a canonical module. In Lemma 2.4 we prove some counterparts to the properties
of ωM stated in the part (v) of the preceding reminder. Note that, if R has a canonical module
ωR then, by [6, 6.1.10 # Exercise], for each d -dimensional R-module N we have HomR (N, ωR ) ∼
=
ωN .
Remark 2.3. Let M be a (finitely generated) d ′ -dimensional R-module. By virtue of the Cob = dim(S)
hen’s structure theorem there exists a Gorenstein ring S such that dim(R) = dim(R)
b
and R is a homomorphic image of S.
(a) Let x ∈ R\ZR (M). Then from the exact sequence,
x
0 → M → M → M/xM → 0,
x
′
′
∆
′
′
x
′
we get the exact sequence, Hdm −1 (M) → Hdm −1 (M) → Hdm −1 (M/xM) → Hdm (M) → Hdm (M), which
yields the exact sequence,
′
′
′
0 → Hdm −1 (M)/xHdm −1 (M) → Hdm −1 (M/xM) → 0 :Hd ′ (M) x → 0.
m
(3)
Applying the Matlis dual functor to the exact sequence (3) we obtain the exact sequence,
ι
0 → ωM
b M
b → C → 0,
b → ωM/x
b /xωM
¡ ′
¢∨
′
where, C = Hdm −1 (M)/xHdm −1 (M) . By [6, 11.2.6 Local Duality Theorem],
O
¢
¢
¡ ′
¡ d ′ −1
d−1 b ∨
b
b R,
b E(R/
b mR)
b
b
= HomRb HdmRb−1 (M)
R/x
C∼
= HmRb (M)/xH
b (M)
mR
b
R
¢ ´
∼
b ∨
b R,
b Hd −1 (M)
= HomRb R/x
b
mR
³
¡
∼
x.
= 0 :Extd −d ′ +1 (M,S)
b
S
′
(4)
E HSAN TAVANFAR
AND
7
M ASSOUD T OUSI
(b) Now assume, furthermore, that R is formally equidimensional. Set, M = R. Then
b p) = dim(R),
b for each p ∈ Spec(R).
b Set,
[31, page 250, Lemma 2] implies that, ht(p) + dim(R/
b
b
R = S/a. Let, p = P/a ∈ Spec(R) where P ∈ Spec(S). We have,
b − dim(R/
b p) = htb (p) = dim(R
b p ).
dim(S P ) = htS P = dim(S) − dim(S/P) = dim(R)
R
¡ dim(Rb p )−1
¢
b S)p ∼
b S)P ∼
b p, S P) ∼
b p ), E(R
b p /pR
b p ) . ConTherefore, Ext1S (R,
(R
= Ext1S (R,
= Ext1S P (R
= HomRb p H b
pRp
sequently,
³
¢´
¡
¢
¡ dim(Rb p )−1
∼
b
b
b
b
b
R
/x
R
,
Hom
(
R
),
E(
R
/
p
R
)
Cp ∼
0
:
x
Hom
H
1
=
p
p
bp
p
p
p
b
bp
R
Ext (R,S)
R
p=
b
pRp
S
¢
b
b p )−1
dim(R
∼
b p ), E(R
b p /pR
b p) .
b p )/xHdim(Rp )−1 (R
(R
= HomRb p H b
b
pRp
pRp
¡
We summarize the above observation as follows.
(c) If in addition R is presumed to be formally equidimensional then, setting M = R, the
b p )−1
b
dim(R
b p )/xHdim(Rp )−1 (R
b p ), for each
module C in the exact sequence (4) is locally dual to, H
(R
bp
pR
b
p ∈ Spec(R).
bp
pR
In the following lemma we show that some of the properties of the canonical module,
which are well-known in the case where R is a homomorphic image of a Gorenstein ring,
hold in general.
Lemma 2.4. Let M has a canonical module. The following statements hold.
(i) AssR (ωM ) = AsshtR (M). In particular, dim(ωM ) = dim(M) and SuppR (ωM ) ⊆ SuppR (M).
(ii) SuppR (ωM ) = {p ∈ SuppR (M) : htM p + dim(R/p) = d ′ }.
(iii) For each ideal a of R we have htωM a ≥ htM a. But, htωM p = htM p for each p ∈ Supp(ωM ).
(iv) ωM satisfies the Serre condition S 2 .
N
T
b for
Proof. (i) By, [31, Theorem 23.2], we can deduce that AssR N = {q R; q ∈ AssRb (N R R)}
each finitely generated R-module N. Thus, [39, Lemma 1.9.(c)] implies that,
AssR (ωM ) = {q
= {q
= {q
\
\
\
R : q ∈ AssRb (ωM
O
R
R : q ∈ AssRb (ωM
b )}
b
R : q ∈ Asshtb (M}.
b
R)}
R
T
b = AsshtR (M). Let, p ∈ AsshtR (M).
Thus it is enough to prove that, {q R : q ∈ AsshtRb (M)}
b
b
b
b pR).
b So we have dim(M)
b =
There exists q ∈ Spec(R) such that pR ⊆ q and dim(R/q) = dim(R/
b pR)
b = dim(R/
b q). Since, q ∈ min(pR),
b so [31, Theorem 9.5.] imdim(M) = dim(R/p) = dim(R/
T
b as required.
b
plies that q R = p. Thus, q ∈ SuppRb (M). Consequently q ∈ AsshtRb (M),
T
b
b and p = q R. We have p ∈ AssR (M) because q ∈ Ass b (M).
Conversely, let q ∈ AsshtRb (M)
R
b = dimR (M).
b pR)
b ≥ dim(R/
b q) = dimb (M)
Also, dimR (M) ≥ dim(R/p) = dim(R/
R
8
A S TUDY OF QUASI -G ORENSTEIN R INGS
(ii) Let N be a finitely generated R-module and p ∈ SuppR (N). Then we can choose q ∈
T
b such that dim(R/
b q) = dim(R/
b pR)
b = dim(R/p). Therefore, q R = p and so R
b q is a
min(pR)
N
N
N
∼
b
b
b
faithfully flat extension of Rp . Since, (N R R) = Np Rp Rq so we have q ∈ SuppRb (N R R).
Due to the formula [7, Theorem A.11] we have,
O
¢
¡ O
b q = dim(Np
b q ) = dim(Np ) + dim(R
b q /pR
b q ) = dim(Np ).
R)
R
dim (N
R
Rp
Now, our claim follows from the above observation in conjunction with the fact that ωM
b =
N
b
ωM R R.
(iii) Assume that R is complete. Then using Reminder
2.1(v)
we have (ωM )p ∼
= ωMp for
¡
¢
each p ∈ SuppR (ωM ). It turns out that, htωM p = dimRp (ωM )p = dimRp (Mp ) = htM p, for each
p ∈ SuppR (ωM ). It follows that,
\
htωM a = inf{htωM p : p ∈ Var(a) SuppR (ωM )}
\
= inf{htM p : p ∈ Var(a) SuppR (ωM )}
\
≥ inf{htM p : p ∈ Var(a) SuppR (M)}
= htM a.
But,
¡
¡
¢
¢
b + 0 :b ω b )/0 :b ω b = htω aR.
b
htωM a = htR/0:R ωM (a + 0 :R ωM )/0 :R ωM = htR/0:
b b ω b (aR
b
R M
R M
M
R
M
b Hence the statement follows from the preceding inequalBy the same token, htM a = htM
b aR.
ity. Clearly, since SuppR (ωM ) ⊆ SuppR (M), so in fact the inequality, htωM p ≥ htM p, is an equality for each p ∈ SuppR (ωM ).
N
∼
b
(iv) The statement follows from the fact that ωM
b = ωM
R R satisfies S 2 .
The second part of the following lemma, which is a special case of the first part, will be
used several times throughout the remainder of the paper. It is worth pointing out that one
implication of the second part of the following lemma can be deduced by [3, Lemma 2.1], in
the case where d ′ = d .
Lemma 2.5. The following statements hold.
(i) Let p ∈ SuppR (M) such that htM p ≥ 3. Suppose, furthermore, that p * ZR (M). If p ∈
¡ ′
¢
¡
¢
b ω b = 2.
Att Hdm −1 (M) , then grade pR,
M
¡ ′
¢
(ii) (c.f. Remark 2.7) Suppose that DepthR (M) ≥ 1 and dim(M) ≥ 3. Then, m ∈ Att Hdm −1 (M)
if and only if DepthRb (ωM
b ) = 2.
Proof. We first deal with the case where R¡is complete.
¢ Let x ∈ p\ZR (M). Then, by Reminder
d ′ −1
2.1(v), x ∈ p\ZR (ωM ). The fact that p ∈ AttR Hm (M) in conjunction with [6, 7.2.5. Exercise]
¡ ′
¢
′
implies that, p = 0 :R Hdm −1 (M)/A for some submodule A of Hdm −1 (M). Hence we get an exact
sequence,
′
′
′
Hdm −1 (M)/xHdm −1 (M) → Hdm −1 (M)/A → 0,
E HSAN TAVANFAR
AND
9
M ASSOUD T OUSI
¡ ′
¢
′
because x belongs to p. Thus, p/xR ∈ AttR/xR Hdm −1 (M)/xHdm −1 (M)
by virtue
³¡ . ′Consequently,
¢∨ ´
d −1
d ′ −1
of [6, 10.2.20 Corollary], we can conclude that, p/xR ∈ AssR/xR Hm (M)/xHm (M) . It
turns out that,
³
¡ ′
¢∨ ´
′
HomR/xR R/p, Hdm −1 (M)/xHdm −1 (M)
6= 0.
(5)
We have, dimRp /xRp (Mp /xMp ) = dimRp (Mp ) − 1 ≥ 2. Therefore Lemma 2.4(iii) yields,
htωM/xM (p/xR) ≥ htM/xM (p/xR) ≥ 2.
In particular, since ωM/xM is S 2 as R/xR-module so using Reminder 2.1(iii) we get,
It turns out that,
¢
¡
gradeR/xR p/xR, ωM/xM ≥ min{2, htωM/xM (p/xR)} ≥ 2.
HomR/xR (R/p, ωM/xM ) = Ext1R/xR (R/p, ωM/xM ) = 0
(6)
Now, using (5) and (6), in light of the exact sequence (4) of Remark 2.3, we can deduce that,
Ext1R/xR (R/p, ωM /xωM ) 6= 0.
Consequently, gradeR/xR (p/xR, ωM /xωM ) ≤ 1, i.e. gradeR (p, ωM ) ≤ 2. On the other hand,
htωM p ≥ htM p ≥ 3 which implies that, gradeR (p, ωM ) ≥ min{2, htωM p} ≥ 2. So the statement
holds in this case.
¢
¡ ′
b with
In the general case, in light of [6, 11.2.10 Exercise], there exists q ∈ AttRb Hd b−1 (M)
mR
N
N
T
b and 3 ≤ htM p ≤ ht b q. Then the above
b q * Z b (M R R)
p = q R. We have, q ∈ SuppRb (M R R),
M
R
b ω b ) ≤ 2. This completes the
argument shows that gradeRb (q, ωM
b ) = 2 and thence gradeR
b (pR,
M
b ≥ ht b (pR)
b = htM (p) ≥ 3 and ω b
proof because by Lemma 2.4(iii) and (iv) we have, htω (pR)
b
M
M
M
is S 2 .
(ii) We may and we do assume that R is complete. One implication follows from the
preceding part. Note that,
³¡ ′
¢∨ ´
d −1
d ′ −1
m/xR ∈ AssR/xR Hm (M)/xHm (M) ,
³
¡ ′
¢∨ ´
′
6= 0. Hence we can conclude the reif and only if, HomR/xR R/m, Hdm −1 (M)/xHdm −1 (M)
verse implication, also, by tracing back the proof of part (i).
Remark 2.6. It is, perhaps, worth pointing out that the reverse of the first part of the preceding lemma does not hold in general. For example, assume that R is a quasi-Gorenstein
Buchsbaum non-Cohen-Macaulay ring of dimension 4 and depth 2 (see [23, Lemma 2.2.]
or the Segre product ring given in the introduction). Let p ∈ Spec(R) with ht(p) = 3. R is
S 2 and htR (p) ≥ 3, hence by Reminder 2.1(iii)¡we have¢2 ≤ gradeR (p, R) ≤ Depth(R) = 2. So,
d−1
b R)
b = grade (p, R) = 2. But p ∉ AttR Hd−1
gradeRb (pR,
m (R) , because Hm (R) has finite length.
R
Also, it seems that we can not deduce Lemma 2.5(i) from Lemma 2.5(ii) by localization, because we shall need the further assumption, htM p + dim(R/p) = dim(M), which is not assumed.
10
A S TUDY OF QUASI -G ORENSTEIN R INGS
Remark 2.7. This remark is due to Raheleh Jafari. In fact, what follows inspired us to prove
L
Mp . Define the map
the preceding lemma, as a slight generalization. Set, M0 :=
p∈SuppR (M),
htM p=0
0
∂−1
M : M → M such that its projection onto Mp is the natural localization map for each p ∈
Supp(M) with htM p = 0. Assume,
inductively,
that Mi and ∂iM−1 : Mi −1 → Mi are constructed.
¡
¢
L
Coker(∂iM−1 ) p , and ∂iM : Mi → Mi +1 to be the following map. For
Then set, Mi +1 :=
p∈SuppR (M),
htM p=i +1
¡
¢
m ∈ Mi and p ∈ SuppR (M) with htM (p) = i + 1, the component of ∂iM (m) in Coker(∂iM−1 ) p is
m/1, where, − : Mi → Coker(∂iM−1 ) is the natural map. Then, we get the complex,
∂−1
M
∂0M
∂iM−2
∂iM−1
∂iM+1
C (M) := 0 → M → M0 → M1 . . . → Mi −1 → Mi → . . . ,
which is called the Cousin complex of M with respect to the hight filtration. The Cousin
complex is introduced and investigated firstly in [40] as the Commutative Algebraic analogue
of the Cousin complex introduced by A. Grothendieck and R. Hartshorne [20, Chapter IV].
i
Denote by H M
the i -th cohomology of the Cousin complex C (M).
Now assume, additionally, that M is S 2 and formally equidimensional and the formal
d ′ −2
fibers of R are Cohen-Macaulay. Then by virtue of [10, Theorem 2.4] we have H M
6= 0 if and
2
only if H b (ωM
b ) 6= 0. Moreover in spite of [11, Lemma 2.2.(iii)] and [10, Theorem 2.1.] we have
m
¡ R d ′ −1
¢
d ′ −2
m ∈ AttR HR (M) if and only if H M
6= 0. This, again, proves Lemma 2.5(ii) in the case
where M is S 2 and formally equidimensional and formal fibers of R are Cohen-Macaulay.
The following lemma is used in the proof of the second part of Theorem 2.9.
Lemma 2.8. Assume that M is not Cohen-Macaulay and that M has a canonical module ωM .
Put,
s := max{i : i < d ′ and Him (M) 6= 0}.
¡ s
¢
If DepthRb Hm
(M)∨ = 0, then,
(
d ′,
s =0
.
Depth(ωM ) =
′
d − s + 1, s > 0
Proof. It suffices to prove the claim in the complete case, so we assume that R is complete.
Set, R′ := R/0 :R M. Recall that since, 0 :R M ⊆ 0 :R ωM , so both of M and ωM are R′ -modules.
Furthermore, ωM is also
module of M over R′ . In addition, the numbers s,
¡ sthe canonical
¢
∨
Depth(ωM ) and Depth Hm (M) are independent of taking the base ring to be either R or R′ .
Thus, without loss of generality, we can presume additionally that dim(R) = dim(M) and that
R also has a canonical module (since R is assumed to be complete). So the statement follows
from [3, Lemma 2.1].
In the following theorem we give an explicit description of the attached prime ideals of
some of local cohomology modules, namely the first non-zero local cohomology supported
at m starting from the point dim(M) − 1. We also, as an application, reprove [39, Theorem
1.15] (In the case where R is not necessarily homomorphic image of a Gorenstein ring and
E HSAN TAVANFAR
AND
11
M ASSOUD T OUSI
our proof is quite different). Clearly we may replace ωR with R in the following theorem
provided R is quasi-Gorenstein. Furthermore, the first part of the next theorem shall be used
in the proof of the Corollary 3.4(i).
Theorem 2.9. Assume that M is formally equidimensional and S 2 and that M has a canonical
module ωM . Then the following statements hold.
(i) Suppose that the formal fibers of R are S 3 7 . Then we have,
¡
¡ ′
¢
¢
AttR Hdm −1 (M) = {p ∈ Spec(R) : DepthRp (ωM )p = 2, and, htωM p ≥ 3}.
(ii) Assume that, n ≥ 2 and that formal fibers of R satisfy the S n+1 condition. Then, ωM
satisfies the S n -condition, if and only if, Him (M) = 0, for each d ′ − n + 2 ≤ i ≤ d ′ − 1. In
this case,
¡
¢
¡ ′
¢
Att Hdm −n+1 (M) = {p ∈ Spec(R) : DepthRp (ωM )p = n and , htωM p ≥ n + 1}.
b is S 2 too.
Proof. Note that, Since M and formal fibers of R satisfies the S 2 -condition, so M
b is also S 1 so, minb (M)
b = Ass b (M).
b Consequently, in view of the our hypothesis, M
b
Since M
R
R
b Then there exists p ∈ Asshtb (M)
b such that p ⊆ q. By Lemma
is unmixed. Let q ∈ SuppRb (M).
R
2.4(i), we have p ∈ AssRb (ωM
b ). Hence, q ∈ SuppR
b (ωM
b ). So the identity,
b q) = d ′ ,
htM
b q + dim(R/
(7)
holds due to Lemma 2.4(ii).
(i) If d ′ ≤ 2 then M is a Cohen-Macaulay R-module and the statement
follows
from Lemma
¡
¢
2.4(iv). So we suppose that d ′ ≥ 3. Let p ∈ Spec(R) such that Depth (ωM )p = 2 and htωM p ≥
3. Then p ∈ SuppR (ωM ) and in view of Lemma 2.4(iii), we have htM p = htωM p ≥ 3. Let,
T
b Then, q R = p. Using [7, Proposition 1.2.16.(a)] and the fact that R
b q is a
q ∈ minRb (pR).
b q ), DepthR (Mp ) =
faithfully flat extension of Rp we can deduce that dimRp (Mp ) = dimRb q (M
p
¡
¢
¡
¢
b
DepthRb q (Mq ) and DepthRb q (ωM
b )q = DepthRp (ωM )p . On the other hand, in light of Reminder 2.1(v) we get, (ω b )q ∼
= ω b and thence by Lemma 2.5(ii),
M
Mq
¡
¢
b q )−1
dim(M
b q) .
(
M
bq
qR
b q ∈ Attb H
qR
Rq
¡ ht p−1
¢
Thus [6, 11.2.10 Exercise (iii)] yields pRp ∈ AttRp HpRMp (Mp ) . So the statement follows from
Lemma 2.4(ii) and [6, 11.2.11 Exercise]. ¡
¢
′
Now, conversely, assume that p ∈ AttR Hdm −1 (M) . By [6, 11.2.10 Exercise (iii)] there ex¢
¡ d ′ −1
b such that q lies over p and q ∈ Attb H
b . Hence, the identity (7) in conjuncists q ∈ R
(
M)
R
b
mR
¡
¢
b
b q ∈ Attb Hdim(Mq )−1 (M
b q ) . This, and equality
tion with [6, 11.2.8 Exercise(i)] implies that qR
R
b
qR q
7
Note that all formal fibers of a homomorphic image of a Cohen-Macaulay ring are Cohen-Macaulay
12
A S TUDY OF QUASI -G ORENSTEIN R INGS
b q ∉ Ass b (M
b q ), because M
b is unmixed. Next, since M
b is S 2 and M
b q is not Cohen7, yields qR
Rq
b q ) ≥ 3. Therefore we can invoke Lemma 2.5(ii) to deMacaulay, so we deduce that dimRb q (M
¢
¡
b
b
duce that Depth (ωM
b )q = 2. Hence, Depth(Rq /pRq ) ≤ 2, by [7, Proposition 1.2.16] (Note that,
N
∼
∼
b
b q /pR
b q is S 3 we can deduce that,
(ωM )p Rp Rq = (ωM
b )q = ωM
b q ). But then using the fact that R
b q /pR
b q is Cohen-Macaulay. So, it follows that htM p ≥ 3, otherwise Mp and thence
the fiber R
b
Mq is Cohen-Macaulay which is a contradiction. But then we get,
¡
¢
¡
¢
2 ≤ min{2, htωM p} ≤ DepthRp (ωM )p ≤ DepthRb q (ωM
b )q = 2.
(ii) We, firstly, argue by induction on d ′ . ωM is S 2 (see Lemma 2.4(iv)). If d ′ ≤ 2, then both
of M and ωM are Cohen-Macaulay and there is nothing to prove. So we assume that d ′ ≥ 3.
If n = 2, then the statement follows from the preceding part. Suppose that n > 2 and the
statement has been proved for smaller values of n.
Presume that ωM is S n . Since ωM satisfies the Serre-condition S n−1 also, so our inductive
hypothesis implies that Him (M) = 0 for each d ′ − n + 3 ≤ i ≤ d ′ − 1 and that,
¢
¡
¡ ′
¢
Att Hdm −n+2 (M) = {p ∈ Spec(R) : DepthRp (ωM )p = n − 1 and , htωM p ≥ n}.
But ωM satisfies S n -condition, so the set on the right hand side of the above identity is
d ′ −n+2
the¡ empty set. ¢Consequently, we get Hm
(M) = 0, as claimed. It remains to describe,
′
Att Hdm −n+1 (M) . In the case where d ′ ≤ n we have ωM is Cohen-Macaulay and d ′ − n + 1 ≤
1. Since, M is S 2 and d ′ ≥ 3, DepthR (M) ≥ 2 and consequently¡ H0m (M) = H¢1m (M) = 0. So
′
we may assume that d ′
n, in particular d ′
3. Let, p ∈ Att Hdm −n+1 (M) . We, firstly,
¢
¡ d ′ −n+1
b such that q lies over p
deal with the case that p 6= m. There exists q ∈ AttRb H b
(M)
mR
(see [6, 11.2.10 Exercise(iii)]). Then [6, 11.2.8. Exercise(i)] together with the identity (7)
b q )−n+1
¡ dimb (M
¢
b q ∈ Att b H Rq
b q ) . So, our inductive assumption on dimenimplies that qR
(M
qR q
b
¡ qR q ¢
¡
¢
b q ) ≥ n + 1. Therefore,
sion implies that DepthRb q (ωM
b )q = n and dimR
b q (ωM
b )q = dimR
b q (M
¡
¢
b q /pR
b q ) ≤ n and whence the fiber R
b q /pR
b q is Cohen-Macaulay. If dimRp (ωM )p ≤ n,
Depth(R
then (ωM )p is Cohen-Macaulay (Recall that ωM is assumed
to¢ be S n ) and hence (ωM
b )q is
¡
Cohen-Macaulay which is a contradiction. Hence dimRp (ωM )p ≥ n + 1. So,
¡
¡
¢
¢
¡
¢
n ≤ min{n, dimRp (ωM )p } ≤ DepthRp (ωM )p ≤ DepthRb q (ωM
b )q = n.
Thus our statement holds in this case where p 6= m. Suppose that p = m. Since dim(ωM ) =
dim(M) ≥ n + 1 so it is enough to show that DepthR (ωM ) = n. Let x be an M-regular el′
∼
ement. Since, Hdm −1 (M) = 0, so the exact sequence (4) in Remark 2.3 yields, ωM
b =
b /xωM
¡
¢
¢
¡
′
i
b .
b = 0 for each d ′ − n + 2 ≤ i ≤ d ′ − 1 and mR
b ∈ Attb Hd −n+1 (M)
ωM/x
b M
b . We have, H b M
R
b
mR
mR
b →M
b → M/x
b M
b → 0 a standard argument shows that,
So using
the
exact
sequence,
0
→
M
¢
¡
¢
¡
b M)
b . Conb M
b = 0 for each d ′ − n + 2 ≤ i ≤ d ′ − 2 and that mR
b ∈ Attb Hd ′ −n+1 (M/x
Hi M/x
R
b
mR
b
mR
sequently, by Lemma 2.8 we get DepthRb (ωM/x
b ) = n − 1, i.e.
b /xωM
b (ωM
b M
b ) = n − 1, i.e. DepthR
)
=
n.
DepthR (ωM ) = DepthRb (ωM
b
¡
¢
For the reverse inclusion, let p ∈ Spec(R) with DepthRp (ωM )p = n and htωM p ≥ n + 1. Let
¡
¢
b So by [7, Proposition 1.2.16] and [7, Theorem A.11] we have, Depthb (ω b )q =
q ∈ minb (pR).
R
Rq
M
E HSAN TAVANFAR
AND
13
M ASSOUD T OUSI
element in p. By, Lemma 2.4(i), x is an ωM -regular
n and htωMb q ≥ n +1. Let ¡x be an M-regular
¢
∼ b /xω b ).
element. Thus, DepthRb q (ωM/x
)
q ≥ n (Recall that ωM/x
=
n
−
1
and htωM/x
b M
b q
b M
b = ωM
b M
b
M
Now, since ωM/x
b M
b satisfies the S n−1 -condition, so in view of our inductive hypothesis we get
¢
¡ ′
b = 0, so we may deduce that q ∈
b M)
b . Consequently, since Hd ′ −n+2 (M)
q ∈ AttRb Hd b−n+1 (M/x
m
R
¢
¡ d ′ −n+1
¡ d ′ −n+1 m¢Rb
b which yields p ∈ AttR Hm
Attb H
(M)
(M) due to [31, Theorem 9.5.] and [6, 11.2.10
R
b
mR
Exercise(iii)].
Now, conversely, assume that, Him (M) = 0, for each d ′ −n +2 ≤ i ≤ d ′ −1. Then, the inductive assumption shows that,
¡ ′
¢
¡
¢
Att Hdm −(n−1)+1 (M) = {p ∈ Spec(R) : DepthRp (ωM )p = n − 1 and , htωM p ≥ n} = ;,
and that ωM satisfies
p ∈ Spec(R) such
¡
¢S n−1 . Suppose by way of contradiction that¡there exists
¢
that DepthRp (ωM )p
min{n, htωM p}. Also, we have, DepthRp (ωM )p ≥ min{n − 1, htωM p}.
¡
¢
¡ ′
¢
These facts yield htωM p ≥ n and DepthRp (ωM )p = n − 1. But then p ∈ Att Hdm −(n−1)+1 (M)
which is a contradiction.
3 Quasi-Gorenstein rings and regular elements
S
¡
If x ∈ R\Z(R) then [24, Lemma 5.1] states that ωR/xωR ∼
= ωR/xR if and only if x ∉
¢ p.
p∈AttR Hdm−1 (R)
But, as the proof8 of the Lemma shows us, we strongly believe that the author means the following observation:
The map, ι, of the exact sequence (4) in Remark 2.3 is an isomorphism if and only if x ∉
S
¡
¢ p. In fact the proof does not deal with the following question:
p∈AttR Hdm−1 (R)
Whether the map, ι, of the exact sequence (4) in Remark 2.3 is an isomorphism whenever
ωR /xωR ∼
= ωR/xR ?
At least in the case where ωR = R, as the following lemma shows us, the above question
has a positive answer.
Proposition 3.1. Suppose that R is a quasi-Gorenstein ring. Assume that x ∈ m is a regular
S
element of R. Then R/xR is a quasi-Gorenstein ring if and only if x ∉
¡
¢ p.
p∈AttR Hdm−1 (R)
Proof. Set M = R in the exact sequence (4) in Remark 2.3. Suppose that, x ∉
¡
S
¢ p.
p∈At t Hdm−1 (R)
Then C = 0 in the exact sequence (4). Thus, ι, is an isomorphism and whence ωR/xR ∼
=
ωR /xωR ∼
= R/xR.
Now, conversely, assume that R/xR is quasi-Gorenstein. It is harmless to assume that
S
R is complete. Namely the assertion, x ∉
¡
¢ q, is equivalent to the assertion that
b
q∈AttRb Hd −1
(R)
b
mR
d−1 b
b
(R) = 0. But according to [6, Flat Base Change Theorem 4.3.2] and faithfully
(R)/xH
Hd−1
b
b
mR
8
mR
The proof is given in the paragraph before the statement of the lemma.
14
A S TUDY OF QUASI -G ORENSTEIN R INGS
d−1
b the latter statement holds if and only if Hd−1
flatness of R
m (R)/xHm (R) = 0, that is to say,
S
x∉
¡
¢ q. On the other hand it is a well-known fact that a commutative local ring is
q∈Att Hdm−1 (R)
quasi-Gorenstein if and only if its completion is so.
¡
¢
d−1
∼
Suppose by way of contradiction that x ∈ p for some p ∈ Att Hd−1
m (R) . Since Hm (R/xR) =
ER/xR (R/m) ∼
= 0 :Hdm (R) x, the exact sequence (3) in Remark 2.3 gives us an exact sequence,
f
d−1
0 → Hd−1
m (R)/xHm (R) → ER/xR (R/m) → ER/xR (R/m) → 0.
b Dualizing
Using [6, 10.2.11 Theorem] we can deduce that f = sidR/xR E(R/m) for some s ∈ R.
we obtain the exact sequence,
¡
¢∨
sidR/xR
d−1
0 → R/xR −→ R/xR → Hd−1
(R)/xH
(R)
→ 0.
m
m
(8)
¡ d−1 ¢
Since p ∈ Att
so a similar
argument as in the proof of Lemma 2.5(i) shows that
R
¡ Hm (R) , d−1
¢
p/xR ∈ AttR/xR Hd−1
(R)/xH
(R)
.
Consequently,
the exact
m
m
¡
¢ sequence (8) together with [6,
10.2.20 Corollary] implies that p/xR ∈ AssR/xR R/(sR + xR) . Since R/xR is quasi-Gorenstein
so by Reminder 2.1(iv), R/xR satisfies the Serre
condition
¡
¢ S 2 . Consequently,
¡
¢ in view of Reminder 2.1(iii), the fact that, p/xR ∈ AssR/xR R/(sR+xR) yields htR/xR p/xR = 1, i.e. htR (p) =
2. Thus, dim(R/p) = d − 2, by virtue of Reminder 2.1(iv) and [31, page 250, Lemma 2]. Hence
¡ dim(R/p)+1 ¢
¡
¢
we have, p ∈ Att Hm
(R) . Consequently [6, 11.2.8 Exercise] yields pRp ∈ Att H1pRp (Rp ) .
On the other hand, R is S 2 and htR (p) = 2, thence Rp is Cohen-Macaulay and H1pRp (Rp ) = 0.
¡
¢
This contradicts with pRp ∈ Att H1pRp (Rp ) .
The following lemma shall be used in the proof of the corollary 3.4(ii).
Lemma 3.2. Assume that S is an R-algebra which is finitely generated as R-module (S is, indeed, semi-local but not necessarily local). Then,
¡
¢
HomR S, E(R/m) ∼
=
M
E(S/n).
n∈Max(S)
Proof. Recall that S is a semi-local ring (see e.g., [31, page 69, 9.3.] and [31, page ¡66, Lemma
¢
2]). Taking into account an epimorphism, Rn → S we get a monomorphism HomR S, E(R/m) →
n
¡
¢
L
E(R/m). It follows that HomR S, E(R/m) is an Artinian R-module and so it is an Artinian
i =1
¡
¢
¡
¢
S-module. But HomR S, E(R/m) is also an injective S-module. Therefore HomR S, E(R/m) ∼
=
L
µ(n)
E(S/n)
(see [7, Theorem 3.2.8]). We aim to prove that µ(n) = 1 for each n ∈ Max(S).
n∈Max(S)
Let, n ∈ Max(S). Since S/n is a finite extension of R/m so there exists u ∈ N such that S/n ∼
=
E HSAN TAVANFAR
u
L
i =1
AND
15
M ASSOUD T OUSI
R/m. We have,
µ(
n)u
M
i =1
R/m ∼
=
µ(
n)
M
i =1
¡
¢
S/n ∼
= HomS S/n, HomR (S, E(R/m)
¡
¢
∼
= HomR S/n, E(R/m)
u
¡M
¢
∼
R/m, E(R/m)
= HomR
i =1
∼
=
u
M
R/m.
i =1
It turns out that µ(n) = 1 as required.
The subsequent lemma is required for the next corollary.
Lemma 3.3. Let S be an R-algebra which is finitely generated as R-module. The following
statements hold.
N
N
(i) Assume that, the natural map ϕ : Hdm (R) → Hdm (R) R S, given by α 7→ α 1 is an Risomorphism. Then given an S-module N and a map f : HdmS (S) → HdmS (N) we have f is
an R-homomorphism if and only if it is an S-homomorphism.
(ii) Suppose that R has a canonical module ωR . For each d -dimensional finitely generated
S-module N the well-known R-isomorphism,
¡
¢
d
HomR HomR (N, ωR ), E(R/m) ∼
= HmS (N),
is an S-isomorphism too.
Proof. (i) For each finite S-module N by representing Hdm (R), (respectively HdmS (N)) as the
last cohomology of the Čech complex C• (x, R) (respectively, C• (xS, N)) for some system of
parameters x for R, it is easily seen that the natural R-isomorphism,
O
N → HdmS (N),
ψN : Hdm (R)
R
N
N
which takes an element, [r /xα ] R n ∈ Hdm (R) R N, to, [r n/xα ], is an S-isomorphism too. According to our hypothesis, we have the R-isomorphism, ψS ◦ ϕ : Hdm (R) → HdmS (S). Therefore,
³
O
¡
¢ Hom(ψS ,id)
¢
¡
¡
¢´
∼
HomS HdmS (S), HdmS (N)
HomS Hdm (R)
S, HdmS (N) ∼
=
= HomR Hdm (R), HomS S, HdmS (N)
R
−1
−1
¡
¢ Hom(ϕ ◦ψS
∼
∼
= HomR Hdm (R), HdmS (N)
=
,id)
¡
¢
HomR HdmS (S), HdmS (N) .
Let us to denote the composition
above
¢ isomorphisms with λ. It is easily seen that
¡ d of the
d
λ( f ) = f for each f ∈ HomS HmS (S), HmS (N) . This fact together with the surjectivity of λ
proves the claim.
16
A S TUDY OF QUASI -G ORENSTEIN R INGS
(ii) Let N be a finitely generated S-module. As we have seen just in the previous part there
N
exists a natural S-isomorphism, ψN : Hdm (R) R N → HdmS (N). Hence we have the following
chain of S-isomorphisms,
O
¡
¢
¡
¢O
HomR HomR (N, ωR ), E(R/m) ∼
N∼
N∼
= HdmS (N).
= HomR ωR , E(R/m)
= Hdm (R)
R
R
(The reader could easily verify that the first isomorphism is an S-isomorphisms too)
Corollary 3.4. Assume that R is a quasi-Gorenstein ring and x ∈ m is a regular element. Then
the following statements hold.
(i) A necessary and sufficient condition for R/xR to be quasi-Gorenstein is that R/xR satisfies the S 2 condition.
(ii) If R/xR has a canonical module then the S 2 -ification of R/xR is quasi-Gorenstein (up to
localization).
b is quasi-Gorenstein and hence S 2 . Hence the
Proof. (i) Since, R is quasi-Gorenstein so R
formal fibers of R are S 2 . Therefore the formal fibers of R/xR are S 2 too. It turns out that
b R
b is S 2 . Therefore, the statement follows from Theorem
R/xR satisfies S 2 precisely when R/x
3.5.
(ii) Let us denote the S 2 -ification HomR/xR (ωR/xR , ωR/xR ) of R/xR by S. Recall that, S is
∼ L E(S/n).
semi-local. Let mS be the Jacobson radical of S. We aim to prove that, Hd−1
mS (S) =
n∈Max(S)
Then the statement follows after localization. In order to accomplish this, we show that
∼ d−1
Hd−1
mS (S) = HmS (ωR/xR ) as S-modules. On the other hand using Lemma 3.2 we get R/xRisomorphisms,
¡
¢
¡
¢
L
E(S/n) ∼
= HomR/xR S, ER/xR (R/m) = HomR/xR HomR/xR (ωR/xR , ωR/xR ), ER/xR (R/m) ∼
=
n∈Max(S)
Hd−1
mS (ωR/xR ),
∼L
which is an S-isomorphism too by Lemma 3.3(ii). So, Hd−1
mS (S) = n∈Max(S) E(S/n), as was
to be proved.
Let us to denote the natural ring homomorphism R/xR → S, defined by the rule y 7→
yidωR/xR , with g . Let x be a system of parameters for R/xR whose image in S, also denoted
by x, is a system of parameters for S. In the following we will prove that not only Hd−1
m (g ) is
d−1
d−1
∼
an isomorphism (see (10)) but also Hm (R/xR) = Hm (ωR /xωR ) (see (11)). It turns out that
N
∼ d−1
Hd−1
(R/xR) R S →
the isomorphism, ΨS : Hd−1
m (S) = Hm (ωR/xR ) as R/xR-module. Consider N
m
N
n
d−1
n
Hd−1
m (S), which maps an element, [a/(x 1 . . . x d−1 ) ] R s ∈ Hm (R/xR) R S to [as/(x 1 . . . x d−1 ) ].
If we take into account the map ϕ of preceding lemma, then an easy verification shows that
d−1
∼ d−1
Hd−1
m (g ) = ψS ◦ ϕ. This implies that ϕ is bijective. Hence Hm (S) = Hm (ωR/xR ) as S-module
by virtue of Lemma 3.3(i). This gives us our desired isomorphism,
∼ d−1
Hd−1
mS (S) = HmS (ωR/xR )
of the preceding paragraph and completes our proof.
E HSAN TAVANFAR
AND
17
M ASSOUD T OUSI
Since R is quasi-Gorenstein so Lemma 2.4(ii) implies that ht(p) + dim(R/p) = dim(R) for
each p ∈ Spec(R). Furthermore R satisfies S 2 . Thus Reminder 2.1(iii) yields ht(p) = 1 for each
p ∈ Ass(R/xR). It follows that dim(R/p) = d − 1 for each p ∈ Ass(R/xR), i.e. R/xR is unmixed.
Therefore the identity (1) of Reminder 2.1(iv) shows that, 0 :R/xR ωR/xR = 0. Thus S is a finitely
generated integral extension of R/xR by the natural map g . Since R is quasi-Gorenstein, so
by virtue of Remark 2.3(c) and Reminder 2.1(iv), the exact sequence (4) of Reminder 2.3 gives
f
b R
b → ω b b → C → 0, where,
us an exact sequence, F := 0 → R/x
R/x R
¡ ht(p)−1
¢
b p )/xHht(p)−1 (R
b p ), E b (R
b
b
Cp ∼
(R
= HomRb p H b
Rp p /pRp )
b
pRp
(9)
pRp
g
b On the other hand we have the exact sequence, G := 0 → R/xR → S →
for each p ∈ Spec(R).
′
b yields the exact sequence,
C → 0. Tensoring G with R
gb
b R
b → Hom b (ω b b , ω b b ) → C′
Gb : 0 → R/x
R
R/x R
R/x R
O
R
b → 0.
R
b R
b is unmixed and thence (ω b b )q ∼
A similar argument as above shows that R/x
b R)
b q for
R/x R = ω(R/x
b
b
each q ∈ Spec(R/x R) (see Reminder 2.1(v)). In light of Proposition 3.1 and (9)in conjuncb ∉ Suppb b (C) if and only if
tion with the first part of the corollary we can deduce that p/x R
R/x R
b p /x R
b p satisfies the S 2 condition. On the other hand by [3, Proposition 1.2] we have gb b is
R
p/x R
b p /x R
b p satisfies S 2 . It turns out that,
an isomorphism if and only if R
′
dimR/x
b (C
b (C) = dimR/x
b R
b R
O
R
b = dimR/xR (C′ ).
R)
Since R/xR is S 1 , so R/xR is locally Cohen-Macaulay at each prime ideal of height less
than or equal one, whence by [3, Proposition 1.2] htR/xR (0 :R/xR C′ ) ≥ 2 and thereby,
¡
¢
dimR/xR (C′ ) = dim (R/xR)/(0 :R/xR C′ ) ≤ dim(R/xR) − ht(0 :R/xR C′ ) ≤ d − 3.
Thus using the exact sequence G we conclude that
∼ d−1
Hd−1
m (S) = Hm (R/xR).
(10)
b b ∼ d−1 b b ) similarly. This
Since dim(C) = dim(C′ ), using F , we deduce that Hd−1
m (R/x R) = Hm (ωR/x
R
implies that,
d−1
∼ d−1 b b ∼ d−1 b b ) ∼
Hd−1
(11)
m (R/xR) = Hm (R/x R) = Hm (ωR/x
R = Hm (ωR/xR ).
The first part of the preceding corollary is a special case of the following theorem.
b satisfies the S 2 -condition and that x is a regular element of R.
Theorem 3.5. Suppose that R
∼
Then, ωRb /xωRb = ωR/x
b R
b if and only if ωR
b /xωR
b satisfies the Serre-condition S 2 .
18
A S TUDY OF QUASI -G ORENSTEIN R INGS
Proof. We prove that ωR/x
b R
b is isomorphic to ωR
b /xωR
b provided the latter is S 2 . The converse is
obvious by By Remark 2.1(v). Suppose, to the contrary, that the map, ¡ι, in the exact sequence
¢∨
(4) of Remark 2.3, for M := R, is not isomorphism, i.e., coker( ι) = C = Hd−1
(R)/xHd−1
(R) 6=
m
m
S
S
0. It follows that, x ∈
¡
¢ p, and thence, x ∈
¡
¢ p. So, by Theorem 2.9(i),
p∈Att Hdm−1 (R)
b
p∈Att Hdm−1 (R)
¡
¢
b containing x, such that Depth b (ω b )q = 2 but htω q ≥
there exists a prime ideal q ∈ Spec(R),
b
Rq
R
R
¡
¢
3. Consequently, DepthRb q (ωRb /xωRb )q = 1 and htωRb /xωRb q ≥ 2, which violates the S 2 -property
of ωRb /xωRb .
The first author of the paper, in [43, Proposition 3.4.], shows that the Ulrich’s deformation
of certain Gorenstein rings to unique factorization domains, developed in [44], has a quasiGorenstein counterpart. We will mention the statement of this deformation theorem, for the
sake of completeness of the paper.
Theorem 3.6. (see, [43, Proposition 3.4.] and [44, Proposition 1]) Suppose that P is a regular
ring and R := P/a is a quasi-Gorenstein ring. Assume, furthermore, that R is locally complete
intersection at codimension ≤ 1. Then there exists a unique factorization domain S (which is
of finite type over P) and a regular sequence y of S such that R ∼
= S/(y).
In spite of the above deformation theorem, at the of writing this paper, it is not clear for us
whether the quasi-Gorenstein property deforms. However, in view of the following theorem,
the quasi-Gorenstein property adheres a variant of deformation which we call it as analytic
deformation. The following theorem will be used in the proof of Theorem 4.5.
Theorem 3.7. Suppose that x ∈ R\Z(R) and R/x n R is quasi-Gorenstein for each n ∈ N. Then R
is quasi-Gorenstein.
Proof. Our proof reduces to the complete case, so there exists a Gorenstein local ring (S, n)
such that dim(S) = d and R is a homomorphic image of S. The commutative diagrams,
xn
0 −−−−→ R −−−−→
°
°
°
x n+1
R −−−−→ R/x n R −−−−→ 0
x
, n+1 R)
y
y f n :=(r +x n R7→r x+x
(12)
0 −−−−→ R −−−−→ R −−−−→ R/x n+1 R −−−−→ 0
leads to the commutative diagrams,
0 −−−−→ HomS (R, S)/x n HomS (R, S) −−−−→ Ext1S (R/x n R, S) −−−−→ 0 :Ext1 (R,S) x n
S
x
x
x
Ext1 ( f ,id )
x
S n S
−−−−→ 0
,
0 −−−−→ HomS (R, S)/x n+1 HomS (R, S) −−−−→ Ext1S (R/x n+1 R, S) −−−−→ 0 :Ext1 (R,S) x n+1 −−−−→ 0
S
(13)
for each n ∈ N. Taking the inverse limit we get the exact sequence,
0 → lim ωR /x n ωR → lim Ext1S (R/x n R, S) → lim 0 :Ext1 (R,S) x n .
←−
n∈N
←−
n∈N
←−
n∈N
S
(14)
E HSAN TAVANFAR
AND
19
M ASSOUD T OUSI
But, ωR ∼
= lim ωR /x n ωR , by [31, Theorem 8.7] and [31, page 63, 8.2.]. By our hypothesis we
←−
have,
n∈N
O
¡
¢
¡ d−1
¢
n
n
∼
Ext1S (R/x n R, S) ∼
R/x n R, ES (S/n)
= HomS Hd−1
n (R/x R), ES (S/n) = HomS Hm (R/x R)
³
´
¡
n
n
∼
(R/x
R),
Hom
R/x
R,
E
(S/
n
)
= HomR/x n R Hd−1
S
S
m
¡ d−1
¢
∼
= HomR/x n R Hm (R/x n R), ER/x n R (R/m)
∼
= R/x n R,
R/x n R
for each n ∈ N. Hence the inverse system,
¡
¢
lim Ext1S (R/x n R, S), Ext1S ( f n , idS ) n∈N ,
←−
n∈N
is isomorphic to an inverse system (R/x n R, γn : R/x n+1 R → R/x n R)n∈N . For each n ∈ N there
exists αn ∈ R such that γn (r + x n+1 R) = αn r + x n R. We claim that all but finitely many of
α′i s are invertible. In fact, one can easily verified that if our claim is wrong then for an arbiT
(m/x n R)l = 0 for each n ∈ N.
trary element (r n + x n R)n∈N ∈ lim R/x n R, we get r n + x n R ∈
←−
n
n∈N
l ∈N
Consequently lim R/x R = 0 and thereby ωR = 0 by the exact sequence (14) which is a con←−
n∈N
tradiction. Hence, without loss of generality we can assume that all of α′i s are invertible.
n+1
Modify the diagram (12) so that x n+1 is replaced by α−1
in the lower row and the vertical
n x
−1
−1
maps f n and xi d R are replaced by αn f n and αn x, respectively. This yields a modification of
1
the diagram (13) in which Ext1S ( f n , idS ) is replaced by α−1
n ExtS ( f n , idS ) and, by a slight abuse
of notation, xidExt1 (R,S) is replaced by α−1
n xidExt1S (R,S) . The result is an exact sequence analS
ogous to the exact sequence (14) in which the second non-zero module is isomorphic to
lim R/x n R ∼
= R.
←−
n∈N
Moreover, lim 0 :Ext1 (R,S) x n = 0, by the following fact. The ascending chain of modules,
←−
n∈Nn
S
{0 :Ext1 (R,S) x n }n∈N , stabilizes eventually, so that the multiplication by a fixed sufficiently large
S
power of x is zero. This concludes our proof.
4 Quasi-Gorenstein rings and the G-dimension of the canonical module
Definition 4.1. [9] A finitely generated R-module N is said to be a totally reflexive R-module if
it satisfies the following conditions.
¡
¢
(i) N is reflexive, i.e. the natural evaluation map, ηN : N → HomR HomR (N, R), R , is an
isomorphism.
¡
¢
(ii) ExtiR (N, R) = ExtiR HomR (N, R), R = 0 for each i ≥ 1.
20
A S TUDY OF QUASI -G ORENSTEIN R INGS
Definition 4.2. [9] An (augmented) G-resolution of a finitely generated R-module N is an exact
sequence, · · · → Gi → Gi −1 → · · · → G0 → N → 0, where Gi is totally reflexive. Moreover, the
G-dimension of N, denoted by G-dimR N, is the least integer n ≥ 0 such that there exists a Gresolution of N with Gi = 0 for each i > n (If there does not exist such a resolution then we say
that G-dimR N = ∞).
Lemma 4.3. Suppose that (R, m) is a zero depth local ring and M is a totally reflexive Rmodule. If M is indecomposable and 0 :R M = 0, then M ∼
= R. In particular, if HomR (M, M) ∼
= R,
∼
then M = R.
Proof. We assume that M is not free and we look for a contradiction. In particular, M has no
non-zero free direct summand. Let P• be a free resolution of M and F• be a free resolution of
M∗ . Then according to the totally reflexiveness of M we get a complete resolution,
A
. . . → P2 → P1 → P0 → F∗0 → F∗1 → F∗2 → . . . ,
of M where im(A) = im(M → F∗0 ). If the matrix A has some entries in R\m, then there exists
an element x := (r 1 , . . . , r i −1 , 1, r i +1 , . . . , r l ) ∈ im(A) = im(M →¡ F∗0 ). Let¢π : F∗0 → R be the projection map onto the i -th component.
Then,
π⌉im(A) ∈ HomR im(A), R , so O im(A) (x) = R, where
¡
¢
O im(A) (x) = { f (x) : f ∈ HomR im(A), R }. Therefore by [7, Lemma 9.5.1.], M ∼
= im(M → F∗0 ) =
im(A), has a non-zero free direct summand which is a contradiction. It turns out that entries
of A are elements of m9 . This immediately yields, 0 :R m ⊆ 0 :R im(A) = 0 :R im(M → F∗0 ) = 0 :R
M = 0, which contradicts with the fact that depth(R) = 0. It follows that M is free, as claimed.
If, HomR (M, M) ∼
= R, then 0 :R M ⊆ 0 :R HomR (M, M) = 0 :R R = 0. Furthermore, M is indecomposable, otherwise, R ∼
= HomR (M, M) would be decomposable which is a contradiction.
Remark 4.4. Let M be an R-module such that HomR (M, M) ∼
= R and Ext1R (M, M) = 0. Let
x1 ∈ R\Z(M). Then HomR/x1 R (M/x1 M, M/x1 M) ∼
= R/x1 R. If in addition, Ext2R (M, M) = 0, and
x1 , x2 is an M-regular sequence such that x1 ∉ Z(R) then by [31, page 140, Lemma 2.(i)]],
Ext1R/x1 R (M/x1 M, M/x1 M) ∼
= Ext2R (M/x1 M, M). On the other hand from the vanishing of the
modules Ext1R (M, M) and Ext2R (M, M) we can deduce that Ext2R (M/x1 M, M) = 0 and whence
Ext1R/x1 R (M/x1 M, M/x1 M) = 0. Thus, again, by the same argument as above we obtain,
HomR/¡(x
1 ,x2
³
´
¡
¢
¢ M/¡(x , x )M¢, M/¡(x , x )M¢ ∼
R/
(x
,
x
)R
.
=
1
2
1
2
1
2
)R
Theorem 4.5. Assume that R has a canonical module ωR and that Gdim(ωR ) < ∞. Then R is
quasi-Gorenstien provided ExtiR (ωR , ωR ) = 0 for each 1 ≤ i ≤ Depth(ωR ).
Proof. We induct on dim(R). In the case where dim(R) = 0 the statement follows from Lemma
4.3 and the Auslander-Bridger formula [9, Theorem 1.25.]. So we assume that dim(R) > 0 and
9
In general, the minimal free resolutions of M and M∗ for some totally reflexive R-module M, similarly as
above, give us a matrix A such that its entries may lie in R\m. For example, set M := R. Then it is easily seen that
the given matrix A will be the matrix [1].
E HSAN TAVANFAR
AND
21
M ASSOUD T OUSI
the statement has been proved for smaller values of dim(R). Applying the Auslander-Bridger
formula [9, Theorem 1.25.] we get,
Depth(Rp ) = Depth(ωRp ) + Gdim(ωRp ) ≥ DepthR (ωRp ) ≥ min{2, dim(Rp )},
for each p ∈ SuppR (ωR ). In particular, by virtue of [3, Proposition 1.2] we can conclude that R
is S 2 and thence HomR (ωR , ωR ) ∼
= R.
Let, Depth(ωR ) = 2 (respectively, Depth(ωR ) = 1). Since Depth(R) ≥ Depth(ωR ) so there
exists a regular sequence x := x1 , x2 (resp. x := x1 ) on R. By [2, 1.10], x is also a regular sequence on ωR . The above remark in conjunction with our hypothesis implies that
¡
¢
HomR/(x) ωR /xωR , ωR /xωR ∼
= R/(x).
So,
\
SuppR/x (ωR /xωR )
¡
¢
= AssR/xR (HomR/(x)R (ωR /xωR , ωR /xωR )
¡
¢
= AssR/xR R/xR .
m/(x) ∈ AssR/xR (ωR /xωR ) = AssR/xR (ωR /xωR )
¡
¢
Consequently, Depth R/xR = 0. Thus, Lemma 4.3 together with [8, Corollary 1.4.6], implies
that ωR /xωR ∼
= R/(x). Thus, µ(ωR ) = µ(ωR /xωR ) = 1. Also by [3, Lemma 1.1] together with (1)
in Reminder 2.1, we have 0 :R ωR = 0. Hence ωR ∼
= R.
Now we deal with the case where Depth(ω¡R ) ≥ 3. We
that R is complete.
¢ may assume
¡
¢
S
S
In this case, Lemma 2.5(ii) implies that m ∉ Att Hd−1
(R)
.
Let,
x
∈
R\
Z(R)
(
m
¡
¢ p) .
p∈Att Hdm−1 (R)
Then the exact sequence (4) in Remark 2.3 shows that ωR /xωR is the canonical module of
R/xR. Thus using [8, Corollary 1.4.6] we can conclude that GdimR/xR (ωR/xR ) < ∞. Our hypothesis in conjunction with [31, page 140, Lemma 2.(i)] implies that ExtiR/xR (ωR/xR , ωR/xR ) =
0 for each 1 ≤ i ≤ DepthR/xR (ωR/xR ). Therefore using the inductive assumption we can deduce that R/xR is quasi-Gorenstein. By the same token, R/x n R is quasi-Gorenstein for each
n ∈ N. Thus, by virtue of Theorem 3.7 we conclude that R is quasi-Gorenstein.
Remark 4.6. At this time we do not know whether the finiteness of the G-dimension of ωR
implies that Ext1R (ωR , ωR ) = Ext2R (ωR , ωR ) = 0. Note that in light of [1, Theorem 3], this is, indeed, the case when ProjdimR (ωR ) < ∞. We stress that, by arguing as in the proof of Theorem
4.5, an affirmative answer to this question shows that R is quasi-Gorenstein precisely when
Gdime(ωR ) < ∞ (c.f. the hypothesis of Theorem 4.5). Hence, it is perhaps worthwhile to
propose the following questions.
Question 4.7. Suppose that (the local ring) R has a canonical module with finite G-dimension.
Then, do we have Ext1R (ωR , ωR ) = Ext2R (ωR , ωR ) = 0?
Question 4.8. Assume that (the local ring) R has a canonical module with finite G-dimension.
Is R a quasi-Gorenstein ring?
We stress again that an affirmative answer to the former question would answer the latter
question positively.
The following remark sheds some light on the question of vanishing of Ext1R (ωR , ωR ) and
Ext2R (ωR , ωR ) which is related to the question 4.8.
22
A S TUDY OF QUASI -G ORENSTEIN R INGS
Remark 4.9. The following statement holds.
(i) Suppose that R is a homomorphic image of a Gorenstein local ring S with dim(R) =
dim(S). Then we can invoke the Grothendieck spectral sequence,
¢ i ,j
¡
i +j
j
ExtiR ωR , ExtS (R, S) ⇒ ExtS (ωR , S),
and [36, Theorem 10.33 (Cohomology Five-Term Exact Sequence)] to obtain the exact
sequence,
2
∨
d−1
Hd−2
m (ωR ) → ExtR (ωR , ωR ) → Hm (R)
O
R
1
∨
ωR → Hd−1
m (ωR ) → ExtR (ωR , ωR ) → 0. (15)
(ii) It is, perhaps, worthwhile to give an example of a ring satisfying S 2 but Ext1R (ωR , ωR ) 6= 0.
By virtue of [16,
( Theorem (1.1)] there exists a 5-dimensional Quasi-Buchsbaum ring
Him (R) = 0,
i 6= 2, 3
R such that,
, and it is a homomorphic image of a Gorenstein
i
H (R) ∼
= R/m, i = 2, 3
m
ring S. In particular since R is Cohen-Macaulay at punctured spectrum and Depth(R) =
2 so R satisfies the S 2-condition. Hence by [3, Remark 1.4] Hdm (ωR ) ∼
= E(R/m) and thence
R = ωωR . But it is not S 3 . Thus in view of Theorem 2.9(ii) in conjunction with the fact
that R = ωωR we can deduce that H4m (ωR ) 6= 0. Therefore the exact sequence (15) yields
Ext1R (ωR , ωR ) 6= 0.
(iii) In spite of the argument of the previous part, Ext1R (ωR , ωR ) = 0 whenever R and the formal fibers of R satisfy the S 3 -condition. Namely, under this condition we can pass to
b of R to use the exact sequence (15). Since, R = ωω so our hypoththe completion R
R
esis in conjunction with Theorem 2.9(ii) implies that Hd−1
(ω
)
=
0.
Hence
our claim
R
m
follows from the exact sequence (15).
(iv) Assume that R and formal fibers of R satisfies S 4 , DepthR (ωR ) ≥ 3 and Ext2R (ωR , ωR ) has
finite length. We have Ext1R (ωR , ωR ) = 0 by the previous part. We claim, furthermore,
that Ext2R (ωR , ωR ) = 0. We assume, again, that R is complete. Since R = ωωR we can
d−1
use Theorem 2.9(ii) to conclude that Hd−2
m (ωR ) = Hm (ωR ) = 0. Therefore the exact
sequence (15) yields
O
d−1
Ext2R (ωR , ωR )∨ ∼
ωR .
= Hm (R)
R
In addition by Reminder 2.1(iv) and [3, Lemma 1.1] we have, Supp(ωR ) = Spec(R). So,
³
³¡
O ¢∨ ´
O ¢
¡ d−1
¢´
¡
d−1
d−1
∨
ωR
ωR = AssR Hm (R)
AttR Hm (R)
= AssR HomR ωR , Hm (R)
R
R
¢\
¡
¡
¢
∨
= AssR Hd−1
Supp(ωR ) = AttR Hd−1
m (R)
m (R) .
¡ d−1
¢
N
Hence in view of the above
identity
and
Lemma
2.5(ii)
we
have,
m
∉
Att
ω
H
(R)
.
R
R
R
m
¢
¡
Now the fact that AssR Ext2R (ωR , ωR ) ⊆ {m} implies that Ext2R (ωR , ωR ) = 0.
In fact, for the vanishing of the second cohomology module Ext2R (ωR , ωR ) we imposed
several additional assumptions. Perhaps, this shows that the vanishing of Ext2R (ωR , ωR )
is considerably more complicated than the vanishing of Ext1R (ωR , ωR ).
E HSAN TAVANFAR
AND
23
M ASSOUD T OUSI
5 Several Characterizations of Quasi-Gorenstein Rings
Definition 5.1. Let M be an R-module and x := x1 , .¡. . , x v be a sequence of elements
¢ of R. FolS
t+1
t+1
t
t
lowing [25], by limit closure of x on M we mean,
(x1 , . . . , x v )M :M x1 . . . x v and we will
t≥0
denote it by {x}lim
M .
Remark 5.2. Let M be an R-module and x := x1 , . . . , x v be a sequence of elements of R. The
following statements hold.
¢
S ¡ i t+i
(i) It is easily verified that the submodule, {xi }lim
(x1 , . . . , x vi t+i )M :M x1i t . . . x vi t , of M
M =
t≥0
¢
S ¡ i +t
coincides with
(x1 , . . . , x vi +t )M :M x1t . . . x vt for each i ∈ N.
t≥0
(ii) For each i , j ∈ N with i ≤ j , we denote the the multiplication map,
j −i
j −i
. . . x v : M/(xi )M → M/(x j )M,
¡
¢
¡
¢
by ϕi ,j . So we have the direct system M/(xi )M, ϕi ,j i ,j ∈N whose direct limit lim M/(xi )M is
x1
−→
v
the top local cohomology H(x
(M) (c.f.
1 ,...,x v )
i ∈N
[6, 5.2.9. Theorem]).
(iii) By part (i), the kernel of the natural R-homomorphism λi : M/(xi )M → lim M/(xl )M,
−→
i.e. the module,
³S¡
t≥0
(x1i +t , . . . , x vi +t )M :M
l ∈N
¢´ i
t
t
i lim
x1 . . . x v /(x )M, is the submodule {x }M /(xi )M of
M/(xi )M. In particular, we have the induced R-monomorphism,
¡
¢
l
∼ v
λ˜i : M/{xi }lim
M → lim M/(x )M = H(x1 ,...,x v ) (M).
−→
l ∈N
¡
¢
(iv) It is easily verified that, similar to part (ii), we have a direct system M/{xi }lim
,
ψ
i
,j
M
i ,j ∈N
j lim
for which the R-homomorphism ψi ,j : M/{xi }lim
M → M/{x }M is, again, the multiplication
j −i
map by x1
j −i
. . . x v . Furthermore, for each i , j ∈ N the following diagram is commutative,
M/{xi }lim
M
λ˜i
¡
// lim M/(xl )M
−→
▼▼▼
▼▼▼ l ∈N
▼▼
ψi , j ▼▼▼
▼&&
OO ˜
λj
¢
M/{x j }lim
M
The injectivity of λ˜i together with above commutative diagram implies that ψi ,j is an Rmonomorphism for each i , j ∈ N. In addition, we have the natural isomorphism,
v
l
∼
lim M/{xi }lim
M → H(x1 ,...,x v ) (M) = lim M/(x )M.
−→
i ∈N
−→
l ∈N
10
(v) By [35, 3.2 Theorem] {x}lim
M = xM provided x is a regular sequence on M .
10
It is worth pointing out that Remark 5.2(v) remains true if we drop both of the assumptions for the ring R to
be Noetherian and local.
24
A S TUDY OF QUASI -G ORENSTEIN R INGS
(vi) ([30, Proposition 2.3.]) Suppose that M is finitely generated and {x}lim
M = xM. Then x 1 , . . . , x v
is a regular sequence on M.
Remark 5.3. Let M be an R-module. Then,
(i) [7, Definition 1.2.18] We use the notation Soc(M) to denote the socle of M, i.e. the Rmodule HomR (R/m, M) ∼
= 0 :M m which is isomorphic to the sum of the simple submodules
of M. Furthermore the length of M as R-module is denoted by λR (M).
(ii) Assume that R has a canonical module ωR . Then,
³
¡ d
¢´∨
N
N
N
∼
∼
b
Soc Hm (R)
= (ωR R R)
= Hdm (R)∨ R (R/m) ∼
R (R/m) = (ωR /mωR ).
(iii) In view of (ii) we have,
¶
¶
µ³
µ³
¡
¢´∨
¡
¢ ´∨
= λR Soc Hdm (R)
µ(ωR ) = VdimR/m (ωR /mωR ) = VdimR/m Soc Hdm (R)
³
³
¡
¢´
¡
¢´
= λR Soc Hdm (R) = VdimR/m Soc Hdm (R) .
Consequently,
³
³
¢´
¡
¡
¢´
d b
d
b
b
(
R)
=
Vdim
Hom
Hom
R/
m
R,
H
µ(ωRb ) = VdimR/
R/
m
,
H
(R)
.
R/m
b
R
b mR
b
m
R
b
mR
Lemma 5.4. Let A be an Artinian R-module. It is well-know that since A is an m-torsion Rb
module so A has a natural R-module
structure (see [6, 10.2.9 Remark.]). We have injdimR (A) =
injdimRb (A).
Proof. Let E • be a minimal injective resolution of A. Recall that, as A is an Artinian module,
E i is a finite copy of ER (R/m) for each i ≥ 0 and that we can endow both of A and E(R/m)
b
b mR)
b and E • is a minimal R-injective
b
with an R-module
structure such that ER (R/m) ∼
= ERb (R/
resolution of A too. Thus injdimR (A) = injdimRb (A).
Definition 5.5. [9] An R-module N is called Gorenstein injective (Gorenstein flat) if there exists an exact complex I of injective R-modules (exact complex F of flat R-modules) such that
N
Ker(I0 → I1 )] ∼
= N (Coker(F1 → F0 ) ∼
= N) and such that HomR (E, I) (E R F) is exact for every
injective R-module E.
Definition 5.6. [9] An (augmented) Gorenstein injective (respectively, Gorenstein flat) resolution of M is an exact sequence, 0 → M → B0 → B1 → · · · → Bi → · · · (respectively, · · · → Fi →
· · · → F1 → F0 → M → 0), where each Bi (respectively, each Fi ) is Gorenstein injective (respectively, Gorenstein flat). The Gorenstein injective (respectively, Gorenstein flat) dimension of N,
denoted by GinjdimR N (respectively, GfdimR N), is the least integer n ≥ 0 such that there exists
a Gorenstein injective resolution of N (respectively, Gorenstein flat resolution of N) such that
Bi = 0 (respectively, Fi = 0) for each i > n (If there does not exist such a resolution then we say
that GinjdimR N = ∞ (respectively, GfdimR N = ∞) ).
E HSAN TAVANFAR
AND
25
M ASSOUD T OUSI
According to the parts (v) and (vi) of Remark 5.2 we can deduce that R is Cohen-Macaulay
if and only if the assertion {x}lim
R = xR holds for some (equivalently for every) system of parameters x of R. Also recall that a Cohen-Macaulay ring is unmixed. Therefore the next theorem is a generalization of the fact that a Cohen-Macaulay ring is Gorenstein if and only if
each (some) parameter ideal is irreducible.
Theorem 5.7. The following statements are equivalent.
¡
¢
(i) injdim Hdm (R) < ∞.
¡
¢
(ii) GinjdimR Hdm (R) < ∞ and Extib (ωRb , ωRb ) = 0, for each 1 ≤ i ≤ DepthRb (ωRb ).
R
(iii) R is quasi-Gorenstein.
b is unmixed and the top local cohomology module of R with respect to m has one di(iv) R
mensional socle.
b is unmixed and {x}lim is irreducible for every system of parameters x := x1 , . . . , xd of R.
(v) R
R
b is unmixed and there exists a system of parameter x := x1 , . . . , xd of R such that {xn }lim
(vi) R
R
is irreducible for each n ∈ N.
b is unmixed and {x}lim is an irreducible ideal for some system of parameters x contained
(vii) R
R
in a high enough power of m.
Proof. (i)⇔(iii): One implication is clear. For the reverse, recall that Hdm (R) is an Artinian Rb with its induced R-module
b
structure. So by Lemma 5.4 we
module and that Hdm (R) ∼
= Hd b (R)
mR
can assume
that
R
is
complete
and
whence
R
has
a
canonical
module ωR . But the finiteness
¡ d
¢
of injdim Hm (R) implies that fdR (ωR ) < ∞. Therefore by [31, Theorem 7.10.] PdR (ωR ) =
fdR (ωR ) < ∞. Hence ωR is¡ free by¢ virtue of [1, Theorem 3].
¢
¡
b <
(ii)⇔(iii): If GinjdimR Hdm (R) < ∞ then [37, Lemma 3.5.] implies that GinjdimRb Hd b (R)
mR
∞. Hence using [9, Theorem 4.16.] and [9, Proposition 4.24] we conclude that GdimRb (ωRb ) <
∞. Thus the statement follows from Theorem 4.5. The reverse is straightforward.
b is quasi-Gorenstein. So by Reminder 2.1(iv) R
b is an unmixed ring. Since
(iii)⇒(iv): R
³
¡
¢
¡
¢
¡
¢´
d
∼
∼
HomR R/m, H (R) = HomR R/m, E(R/m) = R/m, we have VdimR/m Soc Hd (R) = 1.
m
(iv)⇒(v): By
m
Remark 5.2(iii) there is an R-monomorphism R/{xi }lim
R
³
→ lim R/(xn ) ∼
= Hdm (R)
−→
n∈N
¡
¢´
for each i ∈ N. So the fact that VdimR/m Soc Hdm (R) = 1 implies that either R/{xi }lim
R is zero
i lim
i lim
or it has one dimensional socle. Let, {x }R 6= R. Then since R/{x }R is an Artinian ring with
¡
¢
i lim
VdimR/m HomR (R/m, R/{xi }lim
R ) = 1, by [31, Theorem 18.1.] R/{x }R is a Gorenstein Artinian
i lim
ring and consequently the zero ideal of R/{xi }lim
R is irreducible, i.e. {x }R is an irreducible
11
i lim
ideal of R. Also if, {xi }lim
R = R , then {x }R is an irreducible ideal of R.
(v)⇒(vi): It is obvious.
11
Here, it is worth to mention the following. The Hochster’s Monomial Conjecture states that for each local
ring R and each system of parameters y of R we have {y}lim
R ( R. Hence in the case where either both of R and
R/m have the same characteristic or dim(R) ≤ 3, we already know that {x}lim
R ( R.
26
A S TUDY OF QUASI -G ORENSTEIN R INGS
(vi)⇒(iii): By our hypothesis, for sufficiently large n, R/{xn }lim
R is a Gorenstein Artinian
¡
¢
n lim
ring and consequently R/{x }R has one-dimensional socle. On the other hand Soc Hdm (R)
is a non-zero finitely generated R-module and,
¡
¢
n lim
∼
Soc Hdm (R) = Soc(lim R/{xn }lim
R ) = lim Soc(R/{x }R ).
−→
−→
n∈N
n∈N
It turns out that, Soc(R/{xn }lim
) → Soc(Hdm (R)) is onto for some m ∈ N and each n ≥ m and
¢ R
¡ d
thereby VdimR/m Hm (R) = 1. Hence ωRb is cyclic by Remark 5.3(iii). On the other hand our
unmixedness hypothesis together with the identity (1) of Reminder 2.1 yields, 0 :Rb ωRb = 0, i.e.
b
ωb ∼
= R.
R
(vii)⇒(iv): Let ℓd (R) ∈ N be as in [30, Definition 2.4.]. Suppose that x is a system of parameters in mℓd (R) such that {x}lim
Precisely
2.5.],
R is an irreducible ideal.
¢
¡
¢
¡ as in [30, Proposition
[18, Lemma 3.12] implies that the natural map Soc R/(x) → Soc lim R/(xn ) is onto. By Re−→
n∈N
n
mark 5.2(iv), there exists a natural isomorphism lim R/{xn }lim
R → lim R/(x ) which yields the
−→
−→
following commutative diagram,
¡
¢
Soc R/(x)
n∈N
n∈N
−−−−→
y
¢
¡
Soc lim R/(xn )
−→
n∈N
x
∼
=
n lim
Soc(R/{x}lim
R ) −−−−→ Soc(lim R/{x }R ).
−→
n∈N
¡
¢
¡ d
¢
n lim ∼
R/{x
}
)
→
Soc
lim
This implies that Soc(R/{x}lim
Soc
H
(R)
is an epimorphism too
=
m
R
R
−→
n∈N
¡
¢
and whence Socdim Hdm (R) = 1.
(v)⇒(vii): It is obvious.
It is worth pointing out that the equivalence (iv)⇔(iii) in the preceding theorem may be
considered as a counterpart to the well-known fact that R is a Gorenstein ring if and only if R
is a Cohen-Macaulay ring of type 1. Also the equivalence (vii)⇔(iii) is a generalization of the
main result of [30, Theorem 2.7.], i.e. the following: 12 .
Theorem: ([30] ) There exists an integer ℓ such that R is Gorenstein if and only if some
parameter ideal contained in mℓ is irreducible.
Remark 5.8. Assume that R has a canonical module ωR . By the identity (1) of Reminder 2.1, R
N
b = 0 if and only if R
b is unmixed.
is unmixed if and only if 0 :R ωR = 0 if and only if 0 :Rb (ωR R R)
b
Hence we can replace R with R in all parts of the Theorem 5.7 provided that R has a canonical
module. However, [33, EXAMPLE 2.3] gives us an example of a Noetherian local domain such
that its completion is not unmixed.
12
This is mentioned in the introduction of [30]
E HSAN TAVANFAR
AND
M ASSOUD T OUSI
27
In the following example we construct a ring R such that Hdm (R) has a one dimensional
socle but it is not quasi-Gorenstein. This implies that in Theorem 5.7 the unmixedness conb is necessary.
dition of R
Example 5.9. Let K be a field and consider the power series ring K[[X, Y]]. Then,
R = K[[X, Y]]/(X 5 Y 5 ),
is a Gorenstein ring. Let a = (X 4 Y 3 , X 3 Y 4 )/(X 5 Y 5 ) be an ideal of R. An easy verification shows
that 0 :R a is a principal ideal generated by the image of X 2 Y 2 in R and that a ( 0 :R (0 :R a) =
(X 3 Y 3 )/(X 5 Y 5 ). Note that htR (a) = 0. This implies that 0 :R a ∼
= HomR (R/a, R) is the canonical
module of R/a which is cyclic but 0 :R/a (0 :R a) = 0 :R (0 :R a)/a 6= 0. Therefore the canonical
module of R/a is not free and R/a is not quasi-Gorenstein. However by dualizing
an epimor³
¡ d
¢´
d
Soc
H
(R)
phism R → ωR we get an embedding Hm (R) → E(R/m). Therefore, VdimR/
= 1,
b mR
b
m
although R is not quasi-Gorenstein.
It is well-known that a Cohen-Macaulay ring R is Gorenstein if and only if there exists an
irreducible system of parameters for R. So, in accordance with Theorem 5.7, perhaps it is
b is unmixed and there exist a system
natural to ask whether R is quasi-Gorenstein provided R
of parameters for R whose limit closure is irreducible? The answer is negative.
Example 5.10. Let R = C[[X, Y, Z, T]]/(XY, XT, ZY, ZT). We denote the images of X, Y, Z, T in R
by x, y, z, t , respectively. Then R is a non-Cohen-Macaulay unmixed complete local ring of
dimension 2. One can verify that x + y, z + t is a system of parameters for R and that,
¡
¢
(x + y)2 , (z + t )2 : (x + y)(z + t ) = (x, y, z, t ).
So (x + y, z + t )lim is the unique maximal ideal of R and in particular it is irreducible. But R is
not quasi-Gorenstein as R is a non-Cohen-Macaulay ring of dimension 2.
In the Theorem 5.7 the quasi-Gorensteinness is characterized with the aid of the limit
closure of parameter ideals. We would like to digress momentarily to give another such application of the limit closure. The following proposition will be use in the proof of Theorem
5.12.
n
lim
Proposition 5.11.³ Suppose that there exists n ∈ N such
´ that m ({x}M /xM) = 0, for each s.o.p.
¡
¢
x for M. Then, mn (x1 , . . . , xi )M :M xi +1 /(x1 , . . . , xi )M = 0, for each s.o.p. x for M and for every
0 ≤ i ≤ d ′ − 1.
¡
¢
Proof. Let, m ∈ (x1 , . . . , xi )M :M xi +1 , where 0 ≤ i ≤ d ′ − 1. Then,
u
u
xi +1 m ∈ (x1 , . . . , xi , xiu+1
+1 , x i +2 , . . . , x d ′ )M,
for each u ∈ N. Therefore, x1 . . . xi xiu+1 . . . xdu′ m ∈ (x12 , . . . , xi2 , xi2u
, . . . , xd2u′ )M which yields m ∈
+1
n
u
u
{x1 , . . . , xi , xiu+1 , . . . , xdu′ }lim
M . Thus our hypothesis yields m m ∈ (x 1 , . . . , x i , x i +1 , . . . , x d ′ )M for
each u ∈ N. Now the assertion follows from the Krull’s intersection theorem.
28
A S TUDY OF QUASI -G ORENSTEIN R INGS
Theorem 5.12. The following statements hold.
(i) If R is a generalized Cohen-Macaulay ring, then there exists n ∈ N0 such that, mn ({x}lim
R /xR) =
13
0, for each s.o.p. x for R . The reverse also holds whenever R is a homomorphic image
of a Cohen-Macaulay ring.
(ii) R is a Buchsbaum14 ring if and only if m({x}lim
R /xR) = 0 for each s.o.p. x for R.
¡ lim
¢
n
Proof. If there ¡exists n ∈ N such that
m
{x}
/xR
= 0 for each s.o.p. x for R then by PropoR
¢
sition 5.11, mn (x1 , . . . , xi )R :R xi +1 ⊆ (x1 , . . . , xi )R, for each s.o.p x of R and 0 ≤ i ≤ d − 1. Thus
R is generalized Cohen-Macaulay by [42, Page 260, Proposition 16]. By the same token each
s.o.p. for R is a weak m-sequence of R, i.e. R is a Buchsbaum ring, provided n ≤ 1.
Conversely assume that R is generalized Cohen-Macaulay. Let, mαi Him (R) = 0, for each
d−1
P
d
αi . We show that, m(2 −1)n {x}lim
0 ≤ i ≤ d − 1. Choose an integer n ∈ N0 satisfying, n ≥
R ⊆
i =0
xR, for each s.o.p. x for R. To this aim we induct on d ≥ 1 (for the case where d = 0 the
assertion is obvious). If, d = 1, then for every s.o.p. x of R there exists t ∈ N such that {x}lim
R =
t+1
t
t
x R :R x = (0 :R x ) + xR. So by [7, Corollary 8.1.3.(b)] we are done in this case. Suppose
that the statement is true for smaller values of d . Let x be a s.o.p. for R and r ∈ {x}lim
R , i.e.
j
j
j +1
x1 . . . xd r ∈ (x1
j +1
, . . . , xd ) for some j ∈ N. Then there exists r ′ ∈ R such that,
j
j
j +1
x1 . . . xd−1 r − xd r ′ ∈ (x1
j +1
j
, . . . , xd−1 )R :R xd .
Thus, [7, Corollary 8.1.3.(b)] shows that,
j
j
j +1
j +1
mn (x1 . . . xd−1 r ) ⊆ (x1 , . . . , xd−1 , xd )R.
Therefore, mn r ⊆ {x1 , . . . , xd−1 }lim , where the notation means modulo xd R. We claim that,
R
d−2
¡
¢
P
βi , where βi is the least integer satisfying mβi ⊆ 0 :R Him (R) . But we defer the proof
2n ≥
i =0
of the claim for a while to see what it implies. Using our claim in conjunction with the
d −1
d
inductive hypothesis we can deduce that m(2 −1)2n mn r ⊆ (x 1 , . . . , x d−1 )R. So, m(2 −1)n r ⊆
(x1 , . . . , xd )R, as desired.
Now, we prove our claim. By [7, Corollary 8.1.3.] we have, 0 :R xd ⊆ Γm (R). Hence the
exact sequence, 0 → 0 :R xd → R → R/(0 :R xd ) → 0, shows that Him (R) ∼
= Him (R/0 :R xd ) for
xd
each i ≥ 1. Therefore using the exact sequence, 0 → R/(0 :R xd ) → R → R → 0, we can deduce
that, mαi +αi +1 Him (R) = 0 for each 0 ≤ i ≤ d − 2, provided mαi Him (R) = 0 for each 0 ≤ i ≤ d − 1.
d−2
d−1
P
P
(αi + αi +1 ) ≤ 2n. This proves our claim.
αi ≤ n. so,
Since,
i =0
i =0
Furthermore if R is a Buchsbaum ring then by [42, Proposition 1.17] every s.o.p. for R is
both of an unconditioned strong d -sequence15 and also a weak m-sequence. Hence by virtue
13
See [6, 9.5.7 Exercise] for the definition of Generalized Cohen-Macaulay rings.
See [42, Page 14, Theorem 2 and definition] for the definition of Buchsbaum rings.
15
See [26, Definition 1.2.] for the definition of an unconditional strongly d-sequence.
14
E HSAN TAVANFAR
AND
29
M ASSOUD T OUSI
d ¡
¢
P
of [26, Theorem 3.6.(ii)] we have, {x}lim
(x1 , . . . , c
x j , . . . , xd ) : x j . Since x is a weak mR =
j =1
¡
¢
sequence, so we have, m (x1 , . . . , xbi , . . . , xd ) : xi ⊆ (x1 , . . . , xbi , . . . , xd ). This, immediately gives
us, m{x}lim
R ⊆ xR.
Acknowledgement
We would like to express our deepest gratitude to Raheleh Jafari for many fruitful discussions and helpful suggestions, most notably for Remark 2.7. We are also grateful to Kamran
Divaani-Aazar for valuable comments.
References
[1] Y. Aoyama, On the depth and the projective dimension of the canonical module, Japan. J. Math. (N.S.), 6,
(1980), no. 1, 61-66.
[2]
, Some basic results on canonical modules, J. Math. Kyoto Univ., 23, (1983), no. 1, 85-94.
[3] Y. Aoyama and S. Goto, On the endomorphism ring of the canonical module, J. Math. Kyoto Univ., 25-1,
(1985), 21-30.
[4] H. Bass, On the ubiquity of Gorenstein rings, Math. Z., 82, (1963), 8–28.
[5] M. -J. Bertin, Anneaux d’invariants d’anneaux de polynomes, en caractéristique p, C. R. Acad. Sci. Paris Sér.
A-B 264 (1967), A653–A656.
[6] M. P. Brodmann and R. Y. Sharp, Local Cohomology: An Algebraic Introduction with Geometric Applications, second edition, Cambridge University Press, 2013.
[7] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge University Press, 1993.
[8] L. W. Christensen, Gorenstein dimensions., Berlin: Springer, 2000.
[9] L. W. Christensen and H-Børn Foxby and H. Holm, Beyond totally reflexive modules and back. A survey on
Gorenstein dimensions., New York, NY: Springer, 2011.
[10] M. T. Dibaei, A study of Cousin complexes through the dualizing complexes, Commun. Algebra, 33, (2005),
no. 1, 119–132.
[11] M. T. Dibaei and R. Jafari, Cohen-Macaulay loci of modules, Commun. Algebra, 39, (2011), no. 10, 3681–
3697.
[12] E. E. Enochs and O. M.G. Jenda, Gorenstein injective and projective modules, Math. Z. 220, (1995), no. 4,
611–633.
[13] R. Fossum, H-B Foxby, P. Griffith and I. Reiten, Minimal injective resolutions with applications to dualizing
modules and Gorenstein modules, Publ. Math., Inst. Hautes Étud. Sci., 45, (1975), 193–215.
[14] L. Fouli and C. Huneke, What is a system of parameter?, Proc. Am. Math. Soc., 139, (2011), 2681-2696.
[15] H-B Foxby, Isomorphisms between complexes with applications to the homological theory of modules,
Math. Scand., 40, (1977), 5–19.
[16] S. Goto, A note on quasi-Buchsbaum rings, Proc. Am. Math. Soc., 90, (1984), 511–516.
[17] S. Goto, On the associated graded rings of parameter ideals in Buchsbaum rings, J. Algebra 85 (1983), no. 2,
490–534.
30
A S TUDY OF QUASI -G ORENSTEIN R INGS
[18] S. Goto and H. Sakurai, The equality I2 = QI in Buchsbaum rings, Rend. Sem. Mat. Univ. Padova, 110, (2003),
25-56.
[19] S. Goto and K. Watanabe, On graded rings. I, J. Math. Soc. Japan, 30, (1978), no. 2, 179–213.
[20] R. Hartshorne, Residues and duality. Appendix: Cohomologie à support propre et construction du foncteur
f ! . par P. Deligne., 1966.
[21] H. Hassanzadeh Hafshejani, N. Shirmohammadi and H. Zakeri, A note on quasi-Gorenstein rings, Arch.
Math., 91, (2008), no. 4, 318–322.
[22] W. Heinzer and M. Kim and B. Ulrich, The Cohen-Macaulay and Gorenstein properties of rings associated
to filtrations, Commun. Algebra, 39, (2011), no. 10, 3547–3580.
[23] M. Herrmann and N. Trung, Examples of Buchsbaum quasi-Gorenstein rings, Proc. Am. Math. Soc., 117,
(1993), 619-625.
[24] M. Hochster, Canonical elements in local cohomology modules and the direct summand conjecture, J. Algebra, 84, (1983), 503–553.
[25] C. Huneke, Tight closure, parameter ideals, and geometry, in Six Lectures on Commutative Algebra (J. Elias,
J.M. Giral, R.M. MirÂt’o-Roig, and S. Zarzuela, eds), Progress in Mathematics, vol. 166, Birkhauser Verlag,
Basel, 1998, 187-239.
[26] C. Huneke, M. Katzman, R. Y. Sharp and Y. Yao, Frobenius test exponents for parameter ideals in generalized
Cohen-Macaulay local rings, J. Algebra, 305, (2006), no. 1, 516–539.
[27] S. Ishii, Quasi-Gorenstein Fano 3-folds with isolated non-rational loci., Compos. Math. 77 (1991), no. 3,
335–341.
[28] M. Johnson and B. Ulrich, Serre’s condition Rk for associated graded rings, Proc. Am. Math. Soc., 127, (1999),
no. 9, 2619–2624.
[29] E. Kunz, Almost complete intersections are not Gorenstein rings., J. Algebra 28 (1974), 111–115.
[30] T. Marley, M. W. Rogers, H. Sakurai, Gorenstein rings and irreducible parameter ideals, Proc. Am. Math.
Soc., 136, (2008), no. 1, 49–53.
[31] H. Matsumura, Commutative ring theory. Transl. from the Japanese by M. Reid., 1986 (English).
[32] C. Miyazaki, Graded Buchsbaum algebras and Segre products, Tokyo J. Math., 12, (1989), no. 1, 1–20.
[33] J. Nishimura, A few examples of local rings, I, Kyoto Journal of Mathematics, 52, (2012), 51âĂŞ87.
[34] T. Ochiai and K. Shimomoto, Bertini theorem for normality on local rings in mixed characteristic (applications to characteristic ideals)., Nagoya Math. J. 218 (2015), 125–173.
[35] L. O’Carroll, On the generalized fractions of Sharp and Zakeri, J. London Math. Soc., 28 (2), (1983), 417-427.
[36] J. J. Rotman, An introduction to homological algebra. 2nd ed., 2nd ed., Berlin: Springer, 2009.
[37] R. Sazeedeh, Gorenstein injectivity of the section functor, Forum Math., 22, (2010), no. 6, 1117–1127.
[38] P. Schenzel, A note on almost complete intersections, Seminar D. Eisenbud, B. Singh and W. Vogel, Vol. 2,
Teubner-Texte Math. 48, 49-54 (1982).
[39] Peter Schenzel, On the use of local cohomology in algebra and geometry, Basel: Birkhäuser, 1998.
[40] R.Y. Sharp, The Cousin complex for a module over a commutative Noetherian ring, Math. Z., 112, (1969),
340–356.
[41] Anurag K. Singh, Cyclic covers of rings with rational singularities., Trans. Am. Math. Soc. 355 (2003), no. 3,
1009–1024.
[42] J. Stückrad and W. Vogel, Buchsbaum rings and applications. An interaction between algebra, geometry and
topology, 1986 (English).
E HSAN TAVANFAR
AND
M ASSOUD T OUSI
31
[43] Ehsan Tavanfar, Reduction of certain homological conjectures to excellent unique factorization domains,
arXiv:1607.00025v1 [math.AC].
[44] B. Ulrich, Gorenstein rings as specializations of unique factorization domains, J. Algebra 86 (1984), no. 1,
129–140.
[45] T. Yoshizawa, On Gorenstein injectivity of top local cohomology modules, Proc. Am. Math. Soc., 140, (2012),
no. 6, 1897–1907.
[46] M. Rahro Zargar and H. Zakeri, On injective and Gorenstein injective dimensions of local cohomology modules, arXiv:1204.2394v1 [math.AC] 11 (2012).
E HSAN TAVNAFAR , D EPARTMENT OF M ATHEMATICS , S HAHID B EHESHTI U NIVERSITY, G.C., T EHRAN ,
I RAN .
E-mail address: [email protected]
M. T OUSI , D EPARTMENT OF M ATHEMATICS , S HAHID B EHESHTI U NIVERSITY, G.C., T EHRAN , I RAN .
E-mail address: [email protected]
| 0math.AC
|
arXiv:1712.08665v1 [math.ST] 22 Dec 2017
Quasi-maximum likelihood estimation for
cointegrated solutions of continuous-time
state space models observed at
discrete-time points
Vicky Fasen-Hartmann †‡
Markus Scholz §
In this paper we investigate quasi-maximum likelihood (QML) estimation for the parameters of a cointegrated solution of a continuous-time linear state space model observed at discrete time points. The class of cointegrated solutions of continuous-time linear state space models is equivalent to the class of cointegrated continuous-time ARMA
(MCARMA) processes. The model is not in innovation form. Therefore we have to
construct some pseudo-innovations to be able to define a QML-function. We divide the
parameter vector appropriate in long-run and short-run parameters using a representation for cointegrated solutions of continuous-time linear state space models as a sum of a
Lévy process plus a stationary solution of a linear state space model. Then we establish
the consistency of our estimator in three steps. First, we show the consistency for the
QML estimator of the long-run parameters. In the next step, we calculate its consistency
rate. Finally, we use these results to prove the consistency for the QML estimator of the
short-run parameters. After all we derive the limiting distributions of the estimators. The
long-run parameters are asymptotically mixed normally distributed whereas the short-run
parameters are asymptotically normally distributed.
AMS Subject Classification 2010: Primary: 91G70, 62H12
Secondary: 62M10, 60F05
Keywords: Asymptotically normal, cointegration, (super-)consistency, identifiability, Kalman filter,
MCARMA process, mixed normal distribution, pseudo-innovation, quasi-maximum likelihood estimation, state space model.
1 Introduction
This paper deals with the statistical inference of cointegrated solutions of continuous-time linear state
space models. The source of randomness in our model is a Lévy process, i.e. an Rm -valued stochastic
† Institute
of Stochastics, Englerstraße 2, D-76131 Karlsruhe, Germany, email: [email protected]
by the Deutsche Forschungsgemeinschaft through the research grant FA 809/2-2 is gratefully acknowl-
‡ Financial support
edged.
§ Allianz
Lebensversichung-AG, Reinsburgstraße 19, D-70197 Stuttgart, Germany.
1
process L = (L(t))t≥0 with L(0) = 0m P-a.s., stationary and independent increments and càdlàg sample
paths. A typical example of a Lévy process is a Brownian motion. More details on Lévy processes can
be found, e.g., in the monograph of Sato [31]. For A ∈ MN,N (R), B ∈ MN,m (R) and C ∈ Md,N (R) an
Rd -valued continuous-time linear state space model (A, B,C, L) is defined by the state and observation
equation
dX (t) = AX (t)dt + BdL(t)
Y (t) = CX (t)
for t ≥ 0.
(1.1)
The state vector process X = (X (t))t≥0 is an RN -valued process and the output process Y = (Y (t))t≥0
is Rd -valued. The topic of this paper are cointegrated solutions Y of linear state space models. Cointegrated means that Y is non stationary but has stationary increments and there exist linear combinations
of Y which are stationary. The cointegration space is the space spanned by all vectors β so that β TY
is stationary. Since the driving noise in this model is a Lévy process we allow flexible marginals, in
particular, they can be Gaussian if we use a Brownian motion as Lévy process. The class of cointegrated solutions of linear state space models is huge; they are equal to the class of cointegrated multivariate continuous-time ARMA (MCARMA) processes (see Fasen and Scholz [13]). As the name
suggests, MCARMA processes are the continuous-time versions of the popular and well-established
ARMA processes in discrete-time. Continuous-time models provide the basis for option pricing, asset allocation and term structure theory. The underlying observations of asset prices, exchange rates,
and interest rates are often irregularly spaced, in particular, in the context of high frequently data.
Consequently, one often works with continuous-time models which infer the implied dynamics and
properties of the estimated model at different frequencies from the one used in the estimation.
There is not much known about the statistical inference of cointegrated solutions of continuoustime state space and MCARMA models, respectively which allow a moving average part respectively,
go away from the Gaussian assumption. Most attention is paid to cointegrated Gaussian MCAR(p)
processes, in particular Gaussian MCAR(1) processes which are multivariate Ornstein-Uhlenbeck
processes. An algorithms to estimate the structural parameters in a cointegrated Gaussian MCAR(p)
model by maximum-likelihood started already by Harvey and Stock [15, 16], and were further explored in the well-known paper of Bergstrom [4] without analyzing the properties of the estimators.
Frequency domain estimators for a cointegrated Gaussian MCAR(p) model are presented in Chambers and McCrorie [9]. Besides, statistical inference and identification of ergodic continuously observed Gaussian MCAR(1) processes are considered in Kessler and Rahbek [20] and observations at
discrete time points in Kessler and Rahbek [21]. Stockmarr and Jacobsen [36] do analogous things
by observing the process on a discrete time-grid, which is getting finer and finer. There are only few
papers investigating non-Gaussian cointegrated MCARMA processes. For example, Fasen [10] treats
a multiple regression model in continuous-time where the stationary part is a multivariate OrnsteinUhlenbeck process and the process is observed on an equidistant time-grid. The model in Fasen [11]
is similarly, however there the stationary part is an MCARMA process and the process is observed on
a high-frequency time grid.
Sampling Y with distant h > 0 results in Y (h) = (Y (kh))k∈N , a cointegrated solution of a state-space
model in discrete time. QML estimation for cointegrated solutions of state space models in discrete
time were investigated in Aoki [1] and in the unpublished work of Bauer and Wagner [2]. However,
our discrete-time state space model has a different representation and does not fit into their framework
since our noise is only going into the state equation whereas in [1, 2] it is going into both the state and
the observation equation. On this way their model is already in innovation form making calculations
easier. However, their model excludes not only our model but also aggregated linear processes.
2
Although linear state-space models sampled equidistantly belong to the class of ARMA processes
problems arise by identifying the ARMA parameters. Unfortunately there do not exist explicit representations for the ARMA parameters; even stationary solutions of state space models have semiexplicit ARMA representations (see Schlemm and Stelzer [33, Theorem 4.2]). Moreover, in this
representation the innovations are only uncorrelated and not iid. However, most of the well-known
results for cointegrated ARMA models assume an iid noise elsewise even a Gaussian white noise, see
e.g., the monographs of Johansen [18], Lütkepohl [23] and Reinsel [27]. Hence, it is at the state of
the art more or less impossible to use the ARMA structure of Y (h) for estimation of identifiable model
parameters. Therefore we need a different attempt which incorporates as well the identifiability of the
model parameters; our ansatz is QML estimation.
Without any transformation of the state space model given in (1.1) we do not see directly if there
exists a cointegrated solution not to mention the cointegration space. In the case of a minimal statespace model the eigenvalues of A determine if there exists a stationary or a cointegrated solution. If
A has eigenvalue 0 and it has the same geometric and algebraic multiplicity, and all other eigenvalues
have negative real parts then there exists a cointegrated solution Y . It is well-known in econometrics
that in a cointegrated model the asymptotic behavior of the QML estimator for the long-run parameters
and the QML estimator for the short-run parameters differ. Therefore we need a representation for a
cointegrated solution of a linear state space model where we can see clearly the cointegration space
to get a separation in long-run and short-run parameters. If there exists a cointegrated solution of a
continuous-time linear state space model then due to Fasen and Scholz [13, Theorem 3.2] it has the
form
Y (t) = C1 Z +C1 B1 L(t) +Yst (t),
t ≥ 0,
(1.2)
where B1 ∈ Mc,m (R) and C1 ∈ Md,c (R) have rank c ≤ d. The process Yst is a stationary solution of a
state space model driven by the Lévy process L and Z is a c-dimensional random vector. The process
Y is obviously cointegrated with cointegration space C1⊥ , the space orthogonal to C1 , if Var(L(1)) is
non-singular. Because then Y is non-stationary with stationary increments and C1⊥ TY = C1⊥ TYst is
stationary but C1TY is in any component non-stationary. The probabilistic properties of Y are analyzed
in detail in Fasen and Scholz [13] and lay the groundwork for the present paper.
The aim of this paper is to investigate QML estimators for C1 , B1 and the parameters of the stationary process Yst from the discrete-time observations Y (h), . . . ,Y (nh) for some fixed h > 0. The
parameters of C1 are the long-run parameters whereas the other parameters are the short-run parameters. Since our process Y (h) is not in innovation form we use the Kalman-filter to calculate the linear
innovations and to construct an error correction form (see Fasen and Scholz [13, Proposition 5.4 and
Theorem 5.7]). However, the linear innovations and the error correction form are of infinite order in
contrast to the usual finite order as, e.g., in [24, 28, 37] for VARMA models and in [1, 2] for discretetime state space models making calculations more involved. Infinite means that we use infinitely
many past values. However, in practice we do not have infinitely many past values of Y (h) for our
QML method so that we have to approximate the innovations without using the past values but as well
not loosing much information. Not only from the practical point of view problems arise by the infinite
order representation but also from the theoretical point of view since Y (t) is only defined for t ≥ 0
excluding the infinite past as well. Since Y (h) is non-stationary it is not obvious to define Y (−kh) for
negative values. Therefore we have to develop a method to construct Y (−kh) for k ∈ N making it
possible to interpret this infinite error correction form appropriate. Indeed the linear innovations are
stationary but in general it is not possible to say anything about the mixing properties; for more details
we refer to [33] in the case of stationary models. Hence, standard limit results for stationary mixing
3
processes can not be applied.
The representation of the innovations motivates the definition of the pseudo-innovations and hence,
the pseudo-Gaussian likelihood function. The term pseudo reflects in the first case that we do not use
the real innovations and in the second case that we do not have a Gaussian model. This approach is
standard for stationary models (cf. [34]) but not so well investigated for non-stationary models. In
our model the pseudo-innovations are non-stationary as well and hence, classical methods for QML
estimation for stationary models do not work, e.g., the convergence of the quasi-maximum-likelihood
function by a law of large numbers and an ergodic theorem, respectively. To avoid the problem of
the non-stationarity in our model we use a stepwise approach to derive asymptotic results. In the
first step, we prove the consistency of the long-run parameter estimator and in the second step its
consistency rate; the long-run parameter estimator is super-consistent. In the third step we are able
prove the consistency of the short-run parameter estimator. However, for the proof we have to divide
the likelihood-function appropriate so that one part depends only on the short-run parameters and is
based on stationary processes. This decomposition is not obvious and presumes as well a splitting
of the pseudo-innovations in a non-stationary and a stationary part depending only on the short-run
parameters. This concept of a stepwise proof goes back to Saikkonen [29, 30] for MLE in Gaussian
cointegrated models. As in Saikkonen [29, 30] we require for the investigation of the asymptotic
behavior of our estimators some kind of stochastic equicontinuity conditions. However, our conditions
are somewhat different to [29, 30] fitting into the context of our model.
The paper is structured on the following way. An introduction into QML estimation for cointegrated continuous-time linear state space models is given in Section 2. First, we state in Section 2.1
the assumptions on our parametric family of cointegrated output processes Y . Then we define the
pseudo-innovations for the QML estimation by the Kalman filter in Section 2.2. Based on the pseudoinnovations we calculate the pseudo-Gaussian log-likelihood function in Section 2.3. In Section 2.4
we introduce some identifiability conditions to get a unique minimum of the likelihood function. The
main results of this paper are given in Section 3 and Section 4. First, we show the consistency of the
QML estimator in Section 3. Finally, we calculate the asymptotic distribution of the QML estimator
in Section 4. The short-run QML estimator is asymptotically normally distributed and mimics the
properties for QML estimators for stationary models whereas the long-run QML estimator is asymp√
totically mixed normally distributed with a convergence rate of n instead of n occurring in stationary
models. We conclude the paper with some comments on the practical applicability of our estimator in
Section 5. Eventually, in Appendix A we present some asymptotic results and stochastic equicontinuity conditions which we use throughout the paper. Because of their technicality and to keep the paper
readable they are moved to the appendix.
Notation
We use as norms the Euclidean norm k·k in Rd and the Frobenius norm k·k for matrices, which is
submultiplicative. 0d×s denotes the zero matrix in Rd×s and Id is the identity matrix in Rd×d . For a
matrix A ∈ Rd×d we denote by AT its transpose, tr(A) its trace, det(A) its determinant, rank A its rank,
λmin (A) its smallest eigenvalue and σmin (A) its smallest singular value. If A is symmetric and positive
1
1
semi-definite we write A 2 for the principal square root, i.e. A 2 is a symmetric, positive semi-definite
1
1
matrix satisfying A 2 A 2 = A. For a matrix A ∈ Rd×s with rank A = s, A⊥ is a d × (d − s)-dimensional
matrix with rank (d − s) satisfying AT A⊥ = 0s×(d−s) and A⊥T A = 0(d−s)×s . The space of all m × ndimensional real-valued matrices is Mm,n (R). For two matrices A ∈ Rd×s and B ∈ Rr×n , we denote
by A ⊗ B the Kronecker product which is an element of Rdr×sn and by vec(A) the operator which
converts the matrix A into a column vector. We write ∂i for the partial derivative operator with respect
4
to the ith coordinate and ∂i, j for the second partial derivative operator with respect to the ith and jth
coordinate. Further, for a matrix function f (ϑ ) in Md,m (R) with ϑ ∈ Rs the gradient with respect to
f (ϑ ))
the parameter vector ϑ is denoted by ∇ϑ f (ϑ ) = ∂ vec(
∈ Rdm×s . Let ξ = (ξk )k∈N and η = (ηk )k∈N
∂ϑT
be d-dimensional stochastic processes then Γξ ,η (l) = Cov(ξ1 , η1+l ) and Γξ (l) = Cov(ξ1 , ξ1+l ), l ∈
p
w
N0 , are the covariance functions. Finally, we denote with −−→ weak convergence and with −−→
convergence in probability. In general C denotes a constant.
2 Step-wise quasi-maximum likelihood estimation
2.1 Parametric model
Let Θ ⊂ Rs , s ∈ N, be a parameter space. We assume that we have a parametric family (Yϑ )ϑ ∈Θ of
solutions of continuous-time cointegrated linear state space models of the form
Yϑ (t) = C1,ϑ Z +C1,ϑ B1,ϑ Lϑ (t) +Yst,ϑ (t),
t ≥ 0,
(2.1)
where (Yst,ϑ (t))t≥0 is a stationary solution of the state-space model
dXst,ϑ (t) = A2,ϑ Xst,ϑ (t)dt + B2,ϑ dLϑ (t),
Yst,ϑ (t) = C2,ϑ Xst,ϑ (t)
(2.2)
with A2,ϑ ∈ MN−c,N−c (R), B1,ϑ ∈ Mc,m (R), B2,ϑ ∈ MN−c,m (R), C1,ϑ ∈ Md,c (R) and C2,ϑ ∈ Md,N−c (R).
Only the covariance matrix ΣϑL of the Lévy process Lϑ is parameterized. The parameter vector of the
underlying process Y is denoted by ϑ 0 , i.e. (A2 , B1 , B2 ,C1 ,C2 , L) = (A2,ϑ 0 , B1,ϑ 0 , B2,ϑ 0 ,C1,ϑ 0 ,C2,ϑ 0 , Lϑ 0 )
where Yst is a stationary solution of the state space model (A2 , B2 ,C2 , L). Throughout the paper we
shortly write (A2,ϑ , B1,ϑ , B2,ϑ ,C1,ϑ ,C2,ϑ , Lϑ ) for the cointegrated state space model with solution Yϑ
as defined in (2.1). To be more precise we have the following assumptions on our model.
Assumption A. For any ϑ ∈ Θ the cointegrated state space model (A2,ϑ , B1,ϑ , B2,ϑ ,C1,ϑ ,C2,ϑ , Lϑ )
satisfies the following conditions:
(A1) The parameter space Θ is a compact subset of Rs .
(A2) The true parameter vector ϑ 0 lies in the interior of the parameter space Θ.
(A3) The Lévy process Lϑ has mean zero and non-singular covariance matrix ΣϑL = E[Lϑ (1)Lϑ (1)T ].
Moreover, there exists a δ > 0 such that EkLϑ (1)k4+δ < ∞ for any ϑ ∈ Θ.
(A4) The eigenvalues of A2,ϑ have strictly negative real parts.
(A5) The triplet (A2,ϑ , B2,ϑ ,C2,ϑ ) is minimal with McMillan degree N − c (see [14, Chapter 4.2] for
the definition of McMillian degree).
(A6) The matrices B1,ϑ and C1,ϑ have full rank c.
(A7) The c-dimensional random vector Z does not depend on ϑ , EkZk2 < ∞ and Z is independent of
Lϑ .
⊥ are three
(A8) The functions ϑ 7→ A2,ϑ , ϑ 7→ Bi,ϑ , ϑ 7→ Ci,ϑ for i ∈ {1, 2}, ϑ 7→ ΣLϑ and ϑ1 7→ C1,
ϑ1
times continuously differentiable.
5
T T
(A9) Aϑ := diag(0, A2,ϑ ), Bϑ := (BT
1,ϑ , B2,ϑ ) , Cϑ := (C1,ϑ ,C2,ϑ ). Moreover, Cϑ has full rank.
(A10) For any λ , λ ′ ∈ σ (Aϑ ) = σ (A2,ϑ ) ∪ {0} and any k ∈ Z\{0}: λ − λ ′ 6= 2π k/h (Kalman-Bertram
criterion).
Remark 2.1.
(i) (A1) and (A2) are standard assumptions for QML estimation.
(ii) Assumption (A3)-(A4) are sufficient assumptions to guarantee that there exists a stationary
solution Yst,ϑ of the state space model (2.2).
(iii) Due to assumption (A5) the state space representation of Yst,ϑ in (2.2) with A2,ϑ ∈ MN−c,N−c (R),
B2,ϑ ∈ Md−c,m (R) and C2,ϑ ∈ Md,N−c (R) is unique up to a change of basis.
(iv) We require that c respectively the cointegration rank r = d − c is known in advance to be able to
estimate the model adequately. In reality, it is necessary to estimate first the cointegration rank
r and obtain from this c = d − r.
(v) Using the notation in (A9) it is possible to show that Yϑ is the solution of the state space model
(Aϑ , Bϑ ,Cϑ , Lϑ ). Furthermore, on account of (A5) and (A6) the state space model (Aϑ , Bϑ ,Cϑ )
is minimal with McMillian degree N (see [13, Lemma 2.4]) and hence, as well unique up to a
(h)
(h)
change of basis. That in combination with (A10) is sufficient that Yϑ := (Yϑ (k))k∈N0 :=
(Yϑ (kh))k∈N0 is a solution of a discrete-time state space model with McMillian degree N as
well.
⊥ is not uniquely defined. However, we assume that we have an unique map ϑ 7→
(vi) Usually C1,
1
ϑ1
⊥
C1,ϑ1 which is as well three times differentiable..
Furthermore, we assume that the parameter space Θ is a product space of the form Θ = Θ1 ×Θ2 with
Θ1 ⊂ Rs1 and Θ2 ⊂ Rs2 , and ϑ = (ϑ1T , ϑ2T )T is a s-dimensional parameter vector where ϑ1 ∈ Θ1 and
ϑ2 ∈ Θ2 . The idea is that ϑ1 is the s1 -dimensional vector of long-run parameters modelling the cointegration space and hence, responsible for the cointegration of Yϑ whereas ϑ2 is the s2 -dimensional
vector of short-run parameters which has no influence on the cointegration of the model. Since the
matrix C1,ϑ is responsible for the cointegration property (cf. [13, Theorem 3.2]) we parameterize C1,ϑ
with the sub-vector ϑ1 and use for all the other matrices ϑ2 . In summary we parameterize the matrices
with the following sub-vectors (A2,ϑ2 , B1,ϑ2 , B2,ϑ2 ,C1,ϑ1 ,C2,ϑ2 , Lϑ2 ) for (ϑ1 , ϑ2 ) ∈ Θ1 × Θ2 = Θ.
2.2 Linear and pseudo-innovations
In this section, we define the pseudo-innovations which are essential to define the QML function.
Sampling at distance h > 0 maps the class of continuous-time state space models to discrete-time state
space models. These class of state space models are not in innovation form and hence, we use a result
from Fasen and Scholz [13] to calculate the linear innovations by the Kalman filter. The Kalman
filter constructs the linear innovations as εϑ∗ (k) = Yϑ (kh) − Pk−1Yϑ (kk) where Pk is the orthogonal
projection onto span{Yϑ (lh) : −∞ < l ≤ k}. Thus, εϑ∗ (k) is orthogonal to the Hilbert space generated
by span{Yϑ (lh), −∞ < l < k} where the closure is taken in the Hilbert space of square-integrable
random variables with inner product (Z1 , Z2 ) 7→ E(Z1T Z2 ). In our setting the linear innovations are as
follows.
6
(h)
Definition 2.2. Let Ωϑ be the unique solution of the discrete-time algebraic Riccati equation
(h)
(h)
(h)
T
(h)
Ωϑ =eAϑ h Ωϑ eAϑ h − eAϑ h Ωϑ CϑT Cϑ Ωϑ CϑT
where
(h)
Σϑ
(h)
=
Z h
0
B1,ϑ ΣϑL BT1,ϑ
T
and Kϑ = eAϑ h Ωϑ CϑT Cϑ Ωϑ CϑT
novations
of
εϑ∗
= (εϑ∗ (k))k∈N
eA2,ϑ u B2,ϑ ΣϑL BT1,ϑ
T
B1,ϑ ΣLϑ BT2,ϑ eA2,ϑ u eA2,ϑ u B2,ϑ ΣϑL BT2,ϑ eA2,ϑ u
(h)
(h)
−1
(h)
(h) T
Cϑ Ωϑ eAϑ h + Σϑ ,
(h)
of Yϑ
−1
!
du,
be the steady-state Kalman gain matrix. Then the linear in(h)
:= (Yϑ (k))k∈N := (Yϑ (kh))k∈N are the unique stationary solution
(h)
(h) (h)
Xϑ∗ (k) = (eAϑ h − Kϑ Cϑ )Xϑ∗ (k − 1) + Kϑ Yϑ (k − 1),
(h)
εϑ∗ (k) = Yϑ (k) −Cϑ Xϑ∗ (k).
(h)
(h)
Moreover, Vϑ = E(εϑ∗ (1)εϑ∗ (1)T ) = Cϑ Ωϑ CϑT is the prediction covariance matrix of the Kalman
filter.
More details on the well-definedness of this definition can be found in Fasen and Scholz [13]. In our
case of a cointegrated process Y the process (Xϑ∗ (k))k∈N is non-stationary in contrast to the stationary
process (εϑ∗ (k))k∈N . We obtain recursively
(h)
k−1
(h)
(h) (h)
(h)
εϑ∗ (k) = Yϑ (k) −Cϑ (eAϑ h − Kϑ Cϑ )k−1 Xϑ∗ (1) − ∑ Cϑ (eAϑ h − Kϑ Cϑ ) j−1 Kϑ Yϑ (k − j).
j=1
(h)
Since all eigenvalues of (eAϑ h − Kϑ Cϑ ) lie inside the unit circle (see Scholz [35, Lemma 4.6.7]) the
j−1 (h) j
(h)
function L(z, ϑ ) = Id −Cϑ ∑∞j=1 eAϑ h − Kϑ Cϑ
Kϑ z for z ∈ C is well-defined and due to Fasen
and Scholz [13, Lemma 5.6] has the representation as
⊥T
L(z, ϑ ) = −α (ϑ )C1,
ϑ z + k(z, ϑ )(1 − z)
(h)
(h)
Aϑ h −K C )i K
for some linear filter k(z, ϑ ) = Id − ∑∞j=1 K j (ϑ )z j with K j (ϑ ) = ∑∞
ϑ
i= j Cϑ (e
ϑ
ϑ in Md,d (R)
and a matrix α (ϑ ) ∈ Md,d−c (R) with full rank d − c. This representation of L(z, ϑ ) helps us to
choose the initial condition in the Kalman recursion appropriate so that the linear innovations are
really stationary. Therefore,
it is important to know that the stationary process Yst,ϑ can be deRt
fst,ϑ (t − s) dLϑ (s), t ∈ R, with fst,ϑ (u) = C2,ϑ eA2,ϑ u B2,ϑ 1[0,∞) (u) and
fined on R as Yst,ϑ (t) = −∞
Lϑ (−t−) for t < 0 with
the Levy process (Lϑ (t))t∈R is defined on the negative real-line as Lϑ (t) = e
(h)
e
an independent copy (Lϑ (t))t≥0 of (Lϑ (t))t≥0 . Then we have an adequate definition of ∆Yϑ (k) :=
R kh
(h)
(h)
(h)
Yϑ (k) − Yϑ (k − 1) for negative values as well as ∆Yϑ (k) = −∞
f∆,ϑ (kh − s) dLϑ (s), k ∈ Z, with
f∆,ϑ (u) = fst,ϑ (u) − fst,ϑ (u − h) +C1,ϑ B1,ϑ 1[0,h) (u). As notation we use B for the backshift operator
(h)
(h)
satisfying BYϑ (k) = Yϑ (k − 1).
Lemma 2.3. Let Assumption A hold. Then
(h)
(h)
εϑ∗ (k) = −Π(ϑ )Yϑ (k − 1) + k(B, ϑ )∆Yϑ (k),
7
k ∈ N,
(h)
(h)
(h)
∞
⊥ T and k(B, ϑ )∆Y
where Π(ϑ ) = α (ϑ )C1,
ϑ
ϑ (k) = ∆Yϑ (k) − ∑ j=1 K j (ϑ )∆Yϑ (k − j). The matrix
sequence (K j (ϑ )) j∈N is uniformly exponentially bounded, i.e. there exist positive constants c and
0 < ρ < 1 such that supϑ ∈Θ kK j (ϑ )k ≤ cρ j , j ∈ N.
Proof. It remains to show that (K j (ϑ )) j∈N is uniformly exponentially bounded. The proof follows in
(h)
the same line as Schlemm and Stelzer [34, Lemma 2.6] using that all eigenvalues of (eAϑ h − Kϑ Cϑ )
lie inside the unit circle (cf. Scholz [35, Lemma 4.6.7]).
(h)
(h)
Due to Π(ϑ )Yϑ (k − 1) = Π(ϑ )Yϑ ,st (k − 1) we receive
(h)
(h)
εϑ∗ (k) = −Π(ϑ )Yϑ ,st (k − 1) + k(B, ϑ )∆Yϑ (k).
(h)
From this representation we see nicely that (εϑ∗ (k))k∈N is indeed a stationary process. Defining Yϑ
on the negative integers as
k−1
(h)
(h)
(h)
Yϑ (−k) = C1,ϑ Z +Yst,ϑ (0) − ∑ ∆Yϑ (− j) = C1,ϑ Z + Lϑ (−kh) +Yst,ϑ (−k)
j=0
(h)
for k ∈ N0 ,
(h) (h)
we use as initial condition in the Kalman recursion Xϑ∗ (1) = ∑∞j=0 (eAϑ h − Kϑ Cϑ ) j Kϑ Yϑ (− j) so
that
(h)
(h)
(h) (h)
εϑ∗ (k) = Yϑ (k) − ∑∞j=1 Cϑ (eAϑ h − Kϑ Cϑ ) j−1 Kϑ Yϑ (k − j).
Definition 2.4. The pseudo-innovations are defined for k ∈ N as
(h)
(h)
(h)
εk (ϑ ) = −Π(ϑ )Yk−1 + k(B, ϑ )∆Yk
∞
(h)
(h)
(h) (h)
= Yk − ∑ Cϑ (eAϑ h − Kϑ Cϑ ) j−1 Kϑ Yk− j .
j=1
(h)
For ϑ = ϑ 0 the pseudo-innovations (εk (ϑ 0 ))k∈N are the linear-innovations (εϑ∗ 0 (k))k∈N .
Lemma 2.5. Let Assumption A hold.
(h)
(h)
(a) The matrix functions Π(ϑ ), k(z, ϑ ), Vϑ and (Vϑ )−1 are Lipschitz continuous on Θ and three
times partial differentiable.
(h)
(b) supϑ ∈Θ k(Vϑ )−1 k ≤ C.
(h)
(c) infϑ ∈Θ σmin ((Vϑ )−1 ) > 0.
Proof. (a) is a consequence of Assumption A and [35, Lemma 5.9.3]. However, [35, Lemma 5.9.3]
shows only the twice continuous differentiability but the proof of the existence of the third partial
differential is analog.
(b) follows from (a) and the compactness of Θ.
(h)
(c) Due to [35, Lemma 5.9.1] the matrix (Vϑ )−1 is non-singular. Thus, we can conclude the statement
from (a) and the compactness of Θ.
Thus, the pseudo innovations are three times differentiable and we receive an analog version of
Lemma 2.3.
8
Lemma 2.6. Let Assumption A hold and let u, v ∈ {1, . . . , s}. Then the following results hold.
(i) The matrix sequence (∂v K j (ϑ )) j∈N in Md,d (R) is uniformly exponentially bounded such that
(h)
(h)
(h)
∂v εk (ϑ ) = −∂v Π(ϑ )TYk−1 − ∑∞j=1 ∂v K j (ϑ )∆Yk− j .
(ii) The matrix sequence (∂u,v K j (ϑ )) j∈N in Md,d (R) is uniformly exponentially bounded such that
(h)
(h)
(h)
∂u,v εk (ϑ ) = −∂u,v Π(ϑ )TYk−1 − ∑∞j=1 ∂u,v K j (ϑ )∆Yk− j .
Proof. The proof is analog to Schlemm and Stelzer [34, Lemma 2.11] and recall the representation
given in Lemma 2.3 where (K j (ϑ )) j∈N is an uniformly exponentially bounded matrix sequence.
Finally, we end this subsection with some probabilistic properties of the pseudo-innovations. For
reasons of brevity, we write ∂i1 := ∂ ϑ∂ 1i for the partial derivatives with respect to the ith -component of
the long-run parameter vector ϑ1 ∈ Θ1 , i ∈ {1, . . . , s1 }, and similarly ∂ jst := ∂ ϑ∂ 2 j for the partial deriva-
tives with respect to the jth -component of the short-run parameter vector ϑ2 ∈ Θ2 , j ∈ {1, . . . , s2 }.
Analogously we define ∂i,1j and ∂i,stj , respectively for the second partial derivatives.
Lemma 2.7. Let Assumption A hold and i, j ∈ {1, . . . , s2 }. Then
(h)
(h)
(h)
(a) (εk (ϑ 0 )T , ∂ jst εk (ϑ 0 )T , ∂i,stj εk (ϑ 0 )T )k∈N is a stationary and ergodic sequence.
(h)
(h)
(b) Ekεk (ϑ 0 )k4 < ∞,
(h)
Ek∂ jst εk (ϑ 0 )k4 < ∞
(h)
(c) E(∂ist εk (ϑ 0 )εk (ϑ 0 )T ) = 0d×d
(h)
(h)
and
(h)
Ek∂i,stj εk (ϑ 0 )k4 < ∞.
and
(h)
(h)
E(∂i,stj εk (ϑ 0 )εk (ϑ 0 )T ) = 0d×d .
(h)
(d) E(εk (ϑ 0 )εk (ϑ 0 )T ) = Vϑ0 .
Proof.
(a) Representation (1.2) yields
(h)
(h)
(h)
Π(ϑ 0 )Yk
⊥
= α (ϑ10 , ϑ20 )(C1,
)TYk = Π(ϑ 0 )Yst,k ,
ϑ10
⊥ T (h)
(h)
) Yst,k ,
∂ist Π(ϑ 0 )Yk = ∂ist α (ϑ 0 ) (C1,
ϑ10
(h)
(h)
⊥
∂i,stj Π(ϑ 0 )Yk = ∂i,stj α (ϑ 0 ) (C1,
)TYst,k ,
ϑ0
1
and hence,
(h)
(h)
(h)
εk (ϑ 0 ) = −Π(ϑ 0 )Yst,k−1 + k(B, ϑ )∆Yk ,
(h)
(h)
∞
(h)
∂ist εk (ϑ 0 ) = −∂ist Π(ϑ 0 )TYst,k−1 − ∑ ∂ist Kl (ϑ 0 )∆Yk−l ,
(h)
(h)
l=1
∞
(2.3)
(h)
∂i,stj εk (ϑ 0 ) = −∂i,stj Π(ϑ 0 )TYst,k−1 − ∑ ∂i,stj Kl (ϑ 0 )∆Yk−l .
l=1
These are obviously stationary processes. Fasen and Scholz [13, Proposition 5.8] state already that
(h)
(εk (ϑ 0 ))k∈N is ergodic with finite second moments. The same arguments lead to the ergodicity of
(h)
(h)
(h)
(εk (ϑ 0 )T , ∂ jst εk (ϑ 0 )T , ∂i,stj εk (ϑ 0 )T )k∈N .
(h)
(b) The finite fourth moment of (εk (ϑ 0 ))k∈N and its partial derivatives are consequences of their
9
series representation (2.3) with uniformly exponentially bounded coefficient matrices and the finite
(h)
(h)
fourth moment of Yst,k and ∆Yk due to Assumption (A3) and [25, Proposition 3.30].
(h)
(h)
(c) A consequence of (2.3) is that both ∂ist εk (ϑ 0 ) and ∂i,stj εk (ϑ 0 ) are elements of the Hilbert
(h)
(h)
space generated by {Yl , −∞ < l < k}. But εk (ϑ 0 ) is orthogonal to the Hilbert space generated by
(h)
{Yl , −∞ < l < k} so that the statements follow.
(d) is a conclusion of the construction of the linear innovations by the Kalman filter.
2.3 Quasi-maximum likelihood estimation
We estimate the model parameters via an adapted quasi-maximum likelihood estimation method. Minus two over n times the logarithm of the pseudo-Gaussian likelihood function is given by
(h)
Ln (ϑ ) =
i
1 n h
(h)
(h) −1 (h)
(h)
T
π
+
log
detV
+
ε
ϑ
)
V
ε
ϑ
)
.
d
log
2
(
(
∑
ϑ
ϑ
k
k
n k=1
(h)
The pseudo-innovations εk (ϑ ) are constructed by the infinite past {Y (h) (l) : −∞ < l < k}. However,
(h)
(h)
the infinite past is not known, we only have the finite observations Y1 , . . . ,Yn . Therefore, we have to
(h)
approximate the pseudo-innovations and the likelihood-function. For a starting value Xb1 (ϑ ), which
is usually a deterministic constant, we define recursively the approximate pseudo-innovations as
(h)
(h)
(h) (h)
(h)
Xbk (ϑ ) = (eAϑ h − Kϑ Cϑ )Xbk−1 (ϑ ) + Kϑ Yk−1 ,
(h)
(h)
(h)
b
εk (ϑ ) = Yk −Cϑ Xbk (ϑ ),
and the approximate likelihood-function as
i
n h
(h)
(h) −1 (h)
T
cn(h) (ϑ ) = 1 ∑ d log 2π + log detV (h) + b
b
(
(
ε
ϑ
)
V
ε
ϑ
)
.
L
k
k
ϑ
ϑ
n k=1
Then the QML estimator
T bT T
cn(h) (ϑ )
ϑbn := (ϑbn,1
, ϑn,2 ) := argminϑ ∈Θ L
cn(h) (ϑ ). The estimator
is defined as the minimizer of the pseudo-Gaussian log-likelihood function L
ϑbn,1 estimates the long-run parameter ϑ1 and the estimator ϑbn,2 estimates the short-run parameter ϑ2 .
cn(h) (ϑ ) or Ln(h) (ϑ ) as a conclusion
However, for our asymptotic results it doesn’t matter if we use L
of the next proposition.
Assumption B. For every u, v ∈ {1, . . . , s} we assume that
(h)
(h)
2
2
b
b
E sup kX1 (ϑ )k < ∞, E sup k∂u X1 (ϑ )k < ∞
ϑ ∈Θ
ϑ ∈Θ
(h)
and Xb1 (ϑ ) is independent of (Lϑ (t))t≥0 .
and
(h)
E sup k∂u,v Xb1 (ϑ )k2
ϑ ∈Θ
<∞
Proposition 2.8. Let Assumption A and B hold. Moreover, let γ < 1 and u, v ∈ {1, . . . , s}. Then
p
cn(h) (ϑ ) − Ln(h) (ϑ )| −−
(a) nγ supϑ ∈Θ |L
→ 0,
10
cn(h) (ϑ ) − ∂u Ln(h) (ϑ )| −−p→ 0,
(b) nγ supϑ ∈Θ |∂u L
cn(h) (ϑ ) − ∂u,v Ln(h) (ϑ )| −−p→ 0.
(c) nγ supϑ ∈Θ |∂u,v L
The proof of this proposition is similarly to the proof of [33, Lemma 2.7 and Lemma 2.15]. However,
(h)
(h)
they are some essential differences since in their paper (Yk )k∈N and (εk (ϑ ))k∈N are stationary
sequences where in our setup they are non-stationary. Furthermore, we require different convergence
rates. A detailed proof can be found in Appendix B.
We split now the pseudo-innovation sequence based on the decomposition ϑ = (ϑ1T , ϑ2T )T so that
one part is stationary and depends only ϑ2 :
(h)
(h)
(h)
εk (ϑ ) = εk,1 (ϑ ) + εk,2 (ϑ ),
(h)
(h)
(h)
(2.4)
εk,1 (ϑ ) := − Π(ϑ1 , ϑ2 ) − Π(ϑ10 , ϑ2 ) Yk−1 + k(B, ϑ1 , ϑ2 ) − k(B, ϑ10 , ϑ2 ) ∆Yk
where
(h)
(h)
(h)
(h)
εk,2 (ϑ ) := εk,2 (ϑ2 ) = −Π(ϑ10 , ϑ2 )Yk−1 + k(B, ϑ10 , ϑ2 )∆Yk .
and
Due to similar calculations as in (2.3) we receive that the process
(h)
(h)
(h)
εk,2 (ϑ2 ) = −Π(ϑ10 , ϑ2 )Yst,k−1 + k(B, ϑ10 , ϑ2 )∆Yk ,
(2.5)
k ∈ N,
(h)
is indeed stationary. Moreover, εk,1 (ϑ10 , ϑ2 ) = 0 for any ϑ2 ∈ Θ2 and k ∈ N. Finally, we separate the
(h)
log-likelihood function Ln (ϑ ) in
(h)
(h)
(h)
Ln (ϑ ) = Ln,1 (ϑ ) + Ln,2 (ϑ2 ),
where
(h)
(h)
(h)
Ln,1 (ϑ ) := Ln (ϑ1 , ϑ2 ) − Ln (ϑ10 , ϑ2 )
(h)
(h)
= log detVϑ − log detVϑ 0 ,ϑ +
1
2
1 n (h)
(h) −1 (h)
εk,1 (ϑ )T Vϑ
εk,1 (ϑ )
∑
n k=1
2
1 n (h)
(h) −1 (h)
(h)
(h) −1 (h)
T
ε
ϑ
)
V
ε
ϑ
εk,2 (ϑ2 )T Vϑ
εk,2 (ϑ2 )
(
(
)
+
2
∑
∑
k,1
k,2
ϑ
n k=1
n k=1
1 n (h)
(h) −1 (h)
εk,2 (ϑ2 ),
− ∑ εk,2 (ϑ2 )T Vϑ 0 ,ϑ
1 2
n k=1
1 n (h)
(h)
(h)
(h)
(h) −1 (h)
εk,2 (ϑ2 ).
Ln,2 (ϑ2 ) := Ln (ϑ10 , ϑ2 ) = d log 2π + log detVϑ 0 ,ϑ + ∑ εk,2 (ϑ2 )T Vϑ 0 ,ϑ
1 2
1 2
n k=1
n
+
(h)
(h)
Obviously, Ln,2 (ϑ2 ) depends only on the short-run parameters whereas Ln,1 (ϑ ) depends on all
parameters. Furthermore, we have the following relations:
(h)
Ln,1 (ϑ10 , ϑ2 ) = 0
(h)
and
(h)
(h)
Ln (ϑ10 , ϑ2 ) = Ln,2 (ϑ2 ) for any ϑ2 ∈ Θ2 .
(h)
(2.6)
This immediately implies Ln (ϑ 0 ) = Ln,2 (ϑ20 ). In the remaining of the paper we will see that the
(h)
asymptotic properties of ϑbn,1 are determined by L (ϑ ) whereas the asymptotic properties of ϑbn,2
n,1
11
(h)
(h)
are completely determined by Ln,2 (ϑ2 ). Since Ln,2 (ϑ2 ) is based only on stationary processes it
is not surprising that ϑbn,2 exhibits the same asymptotic properties as QML estimators for stationary
processes.
2.4 Identifiability
In order to properly estimate our model, we need a unique minimum of the likelihood function and
therefore we need some identifiability criteria for the family of stochastic processes (Yϑ , ϑ ∈ Θ). The
first assumption guarantees the uniqueness of the long-run parameter ϑ10 .
⊥T C k ≥ ckϑ − ϑ 0 k for ϑ ∈ Θ.
Assumption C. There exist a constant c > 0 so that kC1,
1
1
ϑ1 1
Remark 2.9.
⊥T C k has a zero in ϑ 0 but not that kC ⊥T C k =
(i) Without Assumption C we have only that kC1,
ϑ1 1
1
1,ϑ1 1 6 0
0
⊥T
0
for ϑ1 6= ϑ1 . In particular, C1,ϑ1 C1 6= 0(d−c)×c for ϑ1 6= ϑ1 implies that the space spanned by C1
and C1,ϑ1 are not the same.
⊥T and C ⊥T C = 0
⊥T
(ii) Due to the Lipschitz-continuity of C1,
(d−c)×c the upper bound kC1,ϑ1 C1 k ≤
ϑ1
1,ϑ 0 1
1
Ckϑ1 − ϑ10 k is valid as well.
⊥T C B k > 0 for ϑ 0 6= ϑ since
(iii) Note that Assumption C implies that kΠ(ϑ )C1 B1 k = kα (ϑ )C1,
1
1
ϑ1 1 1
(h)
α (ϑ ) and B1 have full rank, and thus, the process εk,1 (ϑ ) k∈N is indeed non-stationary for all
long-run parameters ϑ1 6= ϑ10 .
(iv) The matrix function α (ϑ ) is continuous and has full column rank d − c so that necessarily
infϑ ∈Θ σmin (α (ϑ )) > 0. Applying Bernstein [5, Corollary 9.6.7] gives for some C > 0
⊥T
0
kΠ(ϑ )C1 k ≥ inf {σmin (α (ϑ ))} kC1,
ϑ1 C1 k ≥ Ckϑ1 − ϑ1 k.
ϑ ∈Θ
The next assumption guarantees the uniqueness of the short-run parameter ϑ20 .
Assumption D. For any ϑ20 6= ϑ2 ∈ Θ2 there exists a z ∈ C such that either
h
h
i−1 (h)
i−1 (h)
(h)
(h)
K
or Vϑ 0 ,ϑ 6= V (h) .
Kϑ 0 ,ϑ 6= C IN − eAh − K (h)C z
Cϑ 0 ,ϑ2 IN − eAϑ2 h − Kϑ 0 ,ϑ Cϑ 0 ,ϑ2 z
1
1
2
1
1
2
1
2
(h)
Lemma 2.10. Let Assumption A and D hold. The function L2 : Θ2 → R defined by
(h) −1 (h)
(h)
(h)
(h)
ε1,2 (ϑ2 )
L2 (ϑ2 ) := d log(2π ) + log detVϑ 0 ,ϑ + E ε1,2 (ϑ2 )T Vϑ 0 ,ϑ
1
2
1
2
(2.7)
has a unique global minimum at ϑ20 .
Proof. The proof is analogous to the proof of Lemma 2.10 in Schlemm and Stelzer [34].
(h)
Without the additional Assumption D we obtain only that L2 (ϑ2 ) has a minimum in ϑ20 but not the
uniqueness.
Due to Fasen and Scholz [13, Theorem 3.1] a canonical form for cointegrated state space processes
already exists and can be used to construct a model class satisfying Assumption C and Assumption D.
Further details are presented in [12].
12
3 Consistency of the QML estimator
In order to show the consistency of the QML estimator, we follow the ideas of Saikkonen [30]. Thus,
we prove the consistency in three steps. In the first step, we prove the consistency of the long-run
QML estimator ϑbn,1 and next we determine its consistency rate. Thirdly, we prove the consistency of
the short-run QML estimator ϑbn,2 by making use of the consistency rate of the long-run QML estimator. Throughout the rest of this paper we assume that Assumption A - D always hold. Furthermore,
we denote by (W (r))0≤r≤1 = ((W1 (r)T ,W2 (r)T ,W3 (r)T )T )0≤r≤1 a (2d + m)-dimensional Brownian
motion with covariance matrix
!
T
Z h
ΣL
ΣL eA2 u
duψ (1)T
(3.1)
ΣW = ψ (1)
Tu
A
u
A
u
T
A
2
2
2
e B 2 ΣL e B 2 ΣL B 2 e
0
where
C1 B1
C2
,
ψ0 := 0d×m
C2
Im×m 0m×N−c
∞
ψ (z) =
∑ ψ jz j,
z ∈ C,
j=0
0d×m C2 (eA2 h j − eA2 h( j−1) )
, j ≥ 1,
ψ j = 0d×m
C2 eA2 h j
Im×m
0m×N−c
(3.2)
(Wi (r))0≤r≤1 , i = 1, 2, are d-dimensional Brownian motions and (W3 (r))0≤r≤1 is an m-dimensional
Brownian motion.
3.1 Consistency of the long-run QML estimator
To show the consistency for the long-run parameter it suffices to show the following theorem where
B(ϑ10 , δ ) := {ϑ1 ∈ Θ1 : kϑ1 − ϑ10 k ≤ δ } denotes the closed ball with radius δ around ϑ10 , and
B(ϑ10 , δ ) := Θ1 \B(ϑ10 , δ ) denotes its complement.
Theorem 3.1. For any δ > 0 we have
lim P
inf
n→∞
cn(h) (ϑ 0 ) > 0 = 1.
cn(h) (ϑ ) − L
L
ϑ ∈B(ϑ10 ,δ )×Θ2
Corollary 3.2. In particular, ϑbn,1 − ϑ10 = o p (1).
3.1.1 Proof of Theorem 3.1
The following lemmata are important for the proof of the theorem.
(h)
Lemma 3.3. Let L(h) := (Lk )k∈Z := (L(kh))k∈Z and define
(h)
Ln,1,1 (ϑ ) :=
(h)
i
1 n h
(h) −1
(h) T
(h)
Vϑ
Π(ϑ )C1 B1 Lk−1 ,
Π(ϑ )C1 B1 Lk−1
∑
n k=1
(h)
(h)
Ln,1,2 (ϑ ) := Ln,1 (ϑ ) − Ln,1,1 (ϑ ).
(h)
Then |Ln,1,2 (ϑ )| ≤ Ckϑ1 − ϑ10 kUn for ϑ ∈ Θ with Un = 1 + Vn + Wn = O p (1), and Un and Vn are
defined as in Proposition A.3.
13
Proof. Define
(h)
(h)
(h)
εk,1,1 (ϑ ) := − Π(ϑ1 , ϑ2 ) − Π(ϑ10 , ϑ2 ) C1 B1 Lk−1 = −Π(ϑ1 , ϑ2 )C1 B1 Lk−1 ,
(h)
(h)
(h)
εk,1,2 (ϑ ) := εk,1 (ϑ ) − εk,1,1 (ϑ )
(h)
(h)
= − Π(ϑ1 , ϑ2 ) − Π(ϑ10 , ϑ2 ) Yst,k−1 + k(B, ϑ1 , ϑ2 ) − k(B, ϑ10 , ϑ2 ) ∆Yk .
(h)
(h)
(h)
(h)
Then (εk,1,2 (ϑ ))k∈N is a stationary sequence and εk,1 (ϑ ) = εk,1,1 (ϑ ) + εk,1,2 (ϑ ). First, note that
(h)
(h)
(h)
Ln,1,2 (ϑ ) = log detVϑ1 ,ϑ2 − log detVϑ 0 ,ϑ +
1
2
2 n (h)
(h) −1 (h)
(h)
εk,1,1 (ϑ )T Vϑ
[εk,2 (ϑ2 ) + εk,1,2 (ϑ )]
∑
n k=1
1 n (h)
(h)
(h)
(h) −1
εk,1,2 (ϑ )T Vϑ
[2εk,2 (ϑ2 ) + εk,1,2 (ϑ )]
∑
n k=1
1 n (h)
(h) −1
(h)
(h) −1
εk,2 (ϑ2 ).
− Vϑ 0 ,ϑ
+ ∑ εk,2 (ϑ2 )T Vϑ
1 2
n k=1
+
In the following we use Bernstein [5, (2.2.27) and Corollary 9.3.9] to get the upper bound
1 n (h)
(h)
(h) −1 (h)
[εk,2 (ϑ ) + εk,1,2 (ϑ )]
εk,1,1 (ϑ )T Vϑ
∑
n k=1
1 n (h)
(h) −1 (h)
(h)
[εk,2 (ϑ ) + εk,1,2 (ϑ )]
tr εk,1,1 (ϑ )T Vϑ
∑
n k=1
!
n
(h) −1 1
(h)
(h)
(h)
= tr Vϑ
∑ [εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,1 (ϑ )T
n k=1
=
(h) −1
≤ k(Vϑ
k
1 n (h)
(h)
(h)
∑ [εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,1 (ϑ )T .
n k=1
Similarly we find upper bounds for the other terms. Moreover, due to Lemma 2.5 (b)
(h)
(h)
(h)
|Ln,1,2 (ϑ )| ≤ | log detVϑ1 ,ϑ2 − log detVϑ 0 ,ϑ | +C
1
2
1 n (h)
(h)
(h)
[εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,1 (ϑ )T
∑
n k=1
+C
1 n
(h)
(h)
(h)
[2εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,2 (ϑ )T
∑
n k=1
+C
Vϑ1 ,ϑ2
(h)
−1
(h)
− Vϑ 0 ,ϑ
1
2
−1
1 n (h)
(h)
∑ εk,2 (ϑ2 )εk,2 (ϑ2 )T .
n k=1
Since Vϑ−1 and log detVϑ are Lipschitz continuous by Lemma 2.5(a), we obtain
(h)
|Ln,1,2 (ϑ )| ≤ C kϑ1 − ϑ10 k +
1 n (h)
(h)
(h)
∑ [εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,1 (ϑ )T
n k=1
14
(3.3)
1 n
1 n (h)
(h)
(h)
(h)
(h)
+
[2εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,2 (ϑ )T + kϑ1 − ϑ10 k
∑
∑ εk,2 (ϑ2 )εk,2 (ϑ2 )T
n k=1
n k=1
!
.
Moreover, Π(ϑ ) is Lipschitz continuous as well (see Lemma 2.5(a)) and the sequence of matrix functions (K j (ϑ )) j∈N and (∇ϑ K j (ϑ )) j∈N are exponentially bounded (see Lemma 2.3 and Lemma 2.6).
(h)
Due to (A.4) and εk,1,1 (ϑ10 , ϑ2 ) = 0 we receive
1 n (h)
(h)
(h)
∑ [εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,1 (ϑ )T ≤ Ckϑ1 − ϑ10 kVn
n k=1
(3.4)
(h)
and due to (A.6) and εk,1,2 (ϑ10 , ϑ2 ) = 0
1 n
(h)
(h)
(h)
[2εk,2 (ϑ ) + εk,1,2 (ϑ )]εk,1,2 (ϑ )T ≤ Ckϑ1 − ϑ10 kWn .
∑
n k=1
(3.5)
1 n (h)
(h)
∑ εk,2 (ϑ2 )εk,2 (ϑ2 )T ≤ CWn
n k=1
(3.6)
Finally,
(h)
as well. Then (3.3)-(3.6) result in the upper bound |Ln,1,2 (ϑ )| ≤ Ckϑ1 − ϑ10 k(1 +Vn +Wn ). A direct
consequence of Proposition A.3 is Un = 1 +Vn +Wn = O p (1).
Lemma 3.4.
(h)
(h)
p
(a) supϑ2 ∈Θ2 |Ln,2 (ϑ2 ) − L2 (ϑ2 )| −→ 0 as n → ∞.
(b)
(h)
1
n Ln,1 (ϑ )
R
(h)
−−
→ 01 k(Vϑ )−1/2 Π(ϑ )C1 B1W3 (r)k2 dr and the convergence holds in the space of
continuous functions on Θ with the supremum norm.
w
Proof.
(a) is a consequence of Proposition A.1(a) and the continuous mapping theorem.
(b) First, supϑ ∈Θ | n1 Ln,1,2 (ϑ )| = o p (1) due to Lemma 3.3 and Θ compact. Second, a conclusion of
Proposition A.1(b) and the continuous mapping theorem is that
!
!
1 n (h) (h)T
1 (h)
(h) −1/2
(h) −1/2
T
T T
Π(ϑ )C1 B1
(ϑ ) = tr Vϑ
L
∑ Lk−1Lk−1 B1 C1 Π(ϑ ) Vϑ
n n,1,1
n2 k=1
Z 1
w
(h) −1/2
(h) −1/2
T
T
T T
Π(ϑ )C1 B1
−
−
→ tr Vϑ
W3 (r)W3 (r) dr B1 C1 Π(ϑ ) Vϑ
=
Z 1
0
0
(h)
k(Vϑ )−1/2 Π(ϑ )C1 B1W3 (r)k2 dr,
and the convergence holds in the space of continuous functions on Θ with the supremum norm due
(h)
to the continuity of Π(ϑ ) and (Vϑ )−1 (cf. Lemma 2.5). In the first and in the last equality we used
Bernstein [5, 2.2.27] which allows us to permutate matrices in the trace.
15
Proof of Theorem 3.1. On the one hand, due to Proposition 2.8
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
≥
=
cn(h) (ϑ 0 )
cn(h) (ϑ ) − L
L
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
inf
(h)
(h)
cn(h) (ϑ ) − Ln(h) (ϑ )|
(Ln (ϑ ) − Ln (ϑ 0 )) − 2 sup |L
ϑ ∈B(ϑ10 ,δ )×Θ2
ϑ ∈Θ
(h)
(h)
Ln (ϑ ) − Ln (ϑ 0 ) + o p (1).
On the other hand, due to Lemma 2.10 and Lemma 3.4(a)
(h)
(h)
inf Ln,2 (ϑ2 ) − Ln,2 (ϑ20 )
ϑ2 ∈Θ2
≤
(h)
(h)
(h)
(h)
(h)
(h)
sup |Ln,2 (ϑ2 ) − L2 (ϑ2 )| + inf L2 (ϑ2 ) − L2 (ϑ20 ) + |L2 (ϑ20 ) − Ln,2 (ϑ20 )| = o p (1).
ϑ2 ∈Θ2
ϑ2 ∈Θ2
Using (2.6) and the above results we receive
cn(h) (ϑ 0 ) ≥
cn(h) (ϑ ) − L
L
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
≥
=
(h)
(h)
Ln (ϑ ) − Ln (ϑ 0 ) + o p (1)
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
(h)
(h)
(h)
Ln,1 (ϑ ) + inf Ln,2 (ϑ2 ) − Ln,2 (ϑ20 ) + o p (1)
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
ϑ ∈Θ2
(h)
Ln,1 (ϑ ) + o p (1).
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
Hence, it suffices to show that for any c > 0
lim P
inf
(h)
Ln,1 (ϑ )
ϑ ∈B(ϑ10 ,δ )×Θ2
n→∞
> c = 1.
(3.7)
An application of Lemma 3.4(b) and the continuous mapping theorem yield
1 (h)
w
Ln,1 (ϑ ) −
inf
−
→
0
n
ϑ ∈B(ϑ1 ,δ )×Θ2
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
Z 1
0
(h)
k(Vϑ )−1/2 Π(ϑ )C1 B1W3 (r)k2 dr.
(3.8)
Due to Bernstein [5, Corollary 9.6.7]
Z 1
0
(h)
(h)
k(Vϑ )−1/2 Π(ϑ )C1 B1W3 (r)k2 dr ≥ σmin ((Vϑ )−1 )
Z 1
0
kΠ(ϑ )C1 B1W3 (r)k2 dr.
(3.9)
Moreover,
Z 1
0
kΠ(ϑ )C1 B1W3 (r)k2 dr
Z 1
tr [B1W3 (r)]T [Π(ϑ )C1 ]T [Π(ϑ )C1 ][B1W3 (r)] dr
0
Z 1
T
T
= tr [Π(ϑ )C1 ] [Π(ϑ )C1 ] [B1W3 (r)][B1W3 (r)] dr
=
0
where
we used Bernstein [5, 2.2.27] to permutate the matrices in the trace. The random matrix
R1
T
[B
W
1 3 (r)] [B1W3 (r)] dr is P-a.s. positive definite since B1 and the covariance matrix of W3 have
0
full rank. Hence, there exists an m × m-dimensional symmetric positive random matrix W ∗ with
16
R1
T
0 [B1W3 (r)][B1W3 (r)] dr
Z 1
0
= W ∗W ∗T . Then we obtain similarly as above with Bernstein [5, 2.2.27]
kΠ(ϑ )C1 B1W3 (r)k2 dr = tr [W ∗ ]T [Π(ϑ )C1 ]T [Π(ϑ )C1 ]W ∗ = kΠ(ϑ )C1W ∗ k2 .
Again an application of Bernstein [5, Corollary 9.6.7] and (3.9) yields
Z 1
0
(h)
(h)
k(Vϑ )−1/2 Π(ϑ )C1 B1W3 (r)k2 dr ≥ σmin ((Vϑ )−1 )
(h)
Z 1
0
kΠ(ϑ )C1 B1W3 (r)k2 dr
(3.10)
= σmin ((Vϑ )−1 )kΠ(ϑ )C1W ∗ k2
(h)
≥ σmin ((Vϑ )−1 )σmin W ∗W ∗T kΠ(ϑ )C1 k2
Z 1
(h) −1
T
B1W3 (r)[B1W3 (r)] dr kΠ(ϑ )C1 k2 .
= σmin ((Vϑ ) )σmin
0
R
R1
1
T
T
T
T
Since B1 0 W3 (r)W3 (r) drB1 is P-a.s. positive definite σmin B1 0 W3 (r)W3 (r) drB1 > 0 P-a.s. On
(h)
the one hand, infϑ ∈B(ϑ 0 ,δ )×Θ2 σmin ((Vϑ )−1 ) > 0 due Lemma 2.5. On the other hand, Assumption C
1
(cf. Remark 2.9) implies that infϑ ∈B(ϑ 0 ,δ )×Θ2 kΠ(ϑ )C1 k2 > c2 δ 2 > 0. To conclude
1
inf
ϑ ∈B(ϑ10 ,δ )×Θ2
Z 1
0
(h)
k(Vϑ )−1/2 Π(ϑ )C1 B1W3 (r)k2 dr > 0
(h)
P-a.s.
p
which finally gives with (3.8) that infϑ ∈B(ϑ 0 ,δ )×Θ2 Ln,1 (ϑ ) −→ ∞ and thus, (3.7) is proven.
1
3.2 Super-consistency of the long-run QML estimator
From the previous section we already know that the QML estimator ϑbn,1 for the long-run parameter
is consistent. In the following we will calculate its consistency rate. For 0 ≤ γ < 1 define the set
for n ∈ N
(3.11)
Nn,γ (ϑ10 , δ ) := ϑ1 ∈ Θ1 : kϑ1 − ϑ10 k ≤ δ n−γ
and N n,γ (ϑ10 , δ ) := Θ1 \Nn,γ (ϑ10 , δ ) as its complement.
Theorem 3.5. Let 0 ≤ γ < 1. For any δ > 0 we have
(h)
(h)
0
c
c
Ln (ϑ ) − Ln (ϑ ) > 0 = 1.
lim P
inf
n→∞
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
Corollary 3.6. In particular, ϑbn,1 − ϑ10 = o p (n−γ ) for 0 ≤ γ < 1.
3.2.1 Proof of Theorem 3.5
The proof uses the next lemma.
Lemma 3.7. Let the notation of Lemma 3.3 hold. Then
(h)
(h)
(h)
(h)
(a) Ln,1,1 (ϑ ) ≥ Cσmin ((Vϑ )−1 )kϑ1 − ϑ10 k2 σmin 1n ∑nk=1 B1 Lk−1 [B1 Lk−1 ]T .
(h)
(h)
(h)
(h)
(b) Ln,1,1 (ϑ ) ≤ Ck(Vϑ )−1 kkϑ1 − ϑ10 k2 tr 1n ∑nk=1 B1 Lk−1 [B1 Lk−1 ]T .
17
Proof.
(a) Several applications of Bernstein [5, Corollary 9.6.7] give similarly as in (3.10)
(h)
Ln,1,1 (ϑ ) =
≥
1 n
(h)
(h)
∑ k(Vϑ )−1/2Π(ϑ )C1 B1Lk−1k2
n k=1
(h)
σmin ((Vϑ )−1 )σmin
!
1 n
(h) T
(h)
∑ B1Lk−1[B1Lk−1] kΠ(ϑ )C1 k2 .
n k=1
An application of Assumption C (cf. Remark 2.9) yields (a).
(b) The submultiplicativity of the norm gives
(h)
Ln,1,1 (ϑ ) =
1 n
(h)
(h)
∑ k(Vϑ )−1/2 Π(ϑ )C1 B1Lk−1k2
n k=1
(h)
≤ k(Vϑ )−1/2 k2 kΠ(ϑ )C1 k2
=
1 n
(h)
∑ kB1 Lk−1k2
n k=1
!
1 n
(h) T
(h)
∑ [B1Lk−1][B1Lk−1] .
n k=1
(h)
k(Vϑ )−1/2 k2 kΠ(ϑ )C1 k2 tr
In the last line we applied Bernstein [5, 2.2.27]. Due to Π(ϑ10 , ϑ2 )C1 = 0d×c we have
(h)
Ln,1,1 (ϑ )
≤
!
1 n
(h) T
(h)
∑ B1Lk−1 [B1Lk−1] .
n k=1
(h)
k(Vϑ )−1 kkΠ(ϑ )C1 − Π(ϑ10 , ϑ2 )C1 k2 tr
Finally, the Lipschitz continuity of Π(ϑ ) and hence of Π(ϑ )C1 yield the statement.
Proof of Theorem 3.5. Due to Proposition 2.8 the lower bound
inf
cn(h) (ϑ 0 ) ≥
cn(h) (ϑ ) − L
n L
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
≥
inf
(h)
(h)
n Ln (ϑ ) − Ln (ϑ 0 ) + o p (1)
inf
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
(h)
(h)
(h)
nLn,1 (ϑ ) + inf n Ln,2 (ϑ2 ) − Ln,2 (ϑ20 ) + o p (1)
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
ϑ ∈Θ2
(h)
holds. We investigate now the second term. Note that Ln,2 (ϑ2 ) depends only on the short-run pa(h)
rameter. Therefore, we take the infeasible estimator ϑbst := arg min
L (ϑ2 ) for the short-run
ϑ2 ∈Θ2
n,2
(h)
n,2
parameter ϑ20 minimizing Ln,2 (ϑ2 ). For this reason, we can interpret this as a „classical“ stationary
(h)
estimation problem. Applying a Taylor-expansion of nLn,2 around ϑ20 yields
√
√
(h)
(h)
(h)
st
st
n · Ln,2 (ϑbn,2
n∇ϑ2 Ln,2 (ϑ n,2 ) · n(ϑbn,2
) − Ln,2 (ϑ20 ) =
− ϑ20 )
st − ϑ 0 k. Since √n∇ L (h) (ϑ
for an appropriate intermediate value ϑ n,2 ∈ Θ2 with kϑ n,2 − ϑ20 k ≤ kϑbn,2
ϑ2 n,2
n,2 )
2
√ bst
0
and n(ϑn,2 − ϑ2 ) are asymptotically normally distributed (these are special and easier calculations
(h)
st ) − L (h) (ϑ 0 ) = O (1). Finally,
as in Section 4.2) we can conclude n · Ln,2 (ϑbn,2
p
2
n,2
inf
cn(h) (ϑ 0 ) ≥
cn(h) (ϑ ) − L
n· L
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
18
inf
(h)
n · Ln,1 (ϑ ) + O p (1).
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
Thus, if we can show that
p
(h)
sup
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
(3.12)
nLn,1 (ϑ ) −−→ ∞,
then for any c > 0
lim P
n→∞
(h)
(h)
0
c
c
Ln (ϑ ) − Ln (ϑ ) > 0
inf
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
≥ lim P
n→∞
inf
ϑ ∈N n,γ (ϑ10 ,δ )×Θ2
(h)
(h)
0
c
c
n Ln (ϑ ) − Ln (ϑ ) > c = 1.
Before we prove (3.12) we first note that due to Equation (3.7) we only have to consider the set
M n,γ (ϑ10 , δ1 ) := N n,γ (ϑ10 , δ1 ) ∩ B(ϑ10 , δ1 ) ⊆ Θ1 ∩ B(ϑ10 , δ1 )
(h)
for n large enough instead of the whole set N n,γ (ϑ10 , δ1 ) in the infimum. Note that infϑ ∈Θ σmin ((Vϑ )−1 ) >
0 by Lemma 2.5. Then Lemma 3.3 and Lemma 3.7 give the lower bound
(h)
(h)
(h)
Ln,1 (ϑ ) ≥ Ln,1,1 (ϑ ) − |Ln,1,2 (ϑ )|
≥
≥ Cnkϑ1 − ϑ10 k2 σmin
|
Finally,
inf
ϑ ∈M n,γ (ϑ10 ,δ )×Θ2
!
1 n
(h) T
(h)
∑ B1Lk−1[B1Lk−1] −Ckϑ1 − ϑ10 kUn
n k=1
!
!
1
1 n
(h) T
(h)
∑ B1Lk−1[B1Lk−1] − nkϑ1 − ϑ 0k Un .
n2 k=1
1
{z
}
(h)
Cσmin ((Vϑ )−1 )kϑ1 − ϑ10 k2 σmin
(h)
nLn,1 (ϑ )
=:Zn (ϑ )
inf
≥
ϑ ∈Mn,γ (ϑ10 ,δ )×Θ2
≥ Cn2−2γ
inf
Cn2 kϑ1 − ϑ10 k2
ϑ ∈M n,γ (ϑ10 ,δ )×Θ2
!
Zn (ϑ )
Zn (ϑ ).
Due to Proposition A.1(b) and Lemma 3.3 we receive
Z 1
w
T
T
inf
→ σmin B1
Zn (ϑ ) −−
W3 (r)W3 (r) drB1 > 0
ϑ ∈M n,γ (ϑ10 ,δ )×Θ2
inf
ϑ ∈M n,γ (ϑ10 ,δ )×Θ2
!
P-a.s.
0
(h)
p
Thus, finally supϑ ∈Mn,γ (ϑ 0 ,δ )×Θ2 nLn,1 (ϑ ) −−→ ∞ for 0 ≤ γ < 1.
1
3.3 Consistency of the short-run QML estimator
Next, we consider the consistency of the short-run parameter estimator ϑbn,2 with the help of the order
of consistency of the long-run parameter estimator ϑbn,1 which we determined in Corollary 3.6. We
show a sufficient condition given by the next theorem. Therefore define for δ > 0 the set B(ϑ20 , δ ) :=
{ϑ2 ∈ Θ2 : kϑ2 − ϑ20 k ≤ δ } as closed ball with radius δ around ϑ20 and B(ϑ20 , δ ) := Θ2 \B(ϑ20 , δ ) as
19
its complement.
Theorem 3.8. Then for any δ > 0 we have
(h)
(h)
0
c
c
lim P
inf
Ln (ϑ ) − Ln (ϑ ) > 0 = 1.
ϑ ∈Θ1 ×B(ϑ20 ,δ )
n→∞
Corollary 3.9. In particular, ϑbn,2 − ϑ20 = o p (1).
3.3.1 Proof of Theorem 3.8
Again we prove some auxiliary results before we state the proof of the theorem.
Lemma 3.10. For
1
2
< γ < 1, δ1 > 0 and ε > 0 we have
(h)
lim P
sup
|Ln,1 (ϑ )| ≤ ε = 1.
n→∞
ϑ ∈Nn,γ (ϑ10 ,δ1 )×Θ2
Proof. Due to Lemma 3.3 and Lemma 3.7 we have the upper bound
(h)
(h)
(h)
|Ln,1 (ϑ )| ≤ |Ln,1,1 (ϑ )| + |Ln,1,2 (ϑ )|
≤
(h)
Ck(Vϑ )−1 kkϑ1 − ϑ10 k2 tr
!
1 n
(h)
(h) T
∑ B1Lk−1[B1Lk−1] +Ckϑ1 − ϑ10 kUn.
n k=1
Then, using Lemma 2.5 results in
sup
ϑ ∈Nn,γ (ϑ10 ,δ1 )×Θ2
(h)
Ln,1 (ϑ )
1−2γ
≤ C δ1 n
tr
!
1 n
(h) T
(h)
∑ B1Lk−1[B1Lk−1 ] +Cδ1n−γ Un .
n2 k=1
(3.13)
Since Un = O p (1) by Lemma 3.3 and
!
Z1
1 n
w
(h)
(h) T
T
T
tr
∑ B1Lk−1[B1Lk−1] −−→ tr B1 0 W3 (r)W3 (r) drB1 > 0 P-a.s.
n2 k=1
by Proposition A.1(b) and the continuous mapping theorem, the right hand side of (3.13) converges
to 0 in probability if 12 < γ < 1. This proves the lemma.
Lemma 3.11. For any δ > 0 and ε > 0 we have
(h)
(h)
0
lim P
inf Ln,2 (ϑ2 ) − Ln,2 (ϑ2 ) > ε = 1.
ϑ2 ∈B(ϑ20 ,δ )
n→∞
Proof. We have
(h)
(h)
Ln,2 (ϑ2 ) − Ln,2 (ϑ20 )
inf
ϑ2 ∈B(ϑ20 ,δ )
≥
inf
ϑ2 ∈B(ϑ20 ,δ )
(h)
(h)
Ln,2 (ϑ2 ) − L2 (ϑ2 ) +
20
inf
ϑ2 ∈B(ϑ20 ,δ )
(h)
(h)
−Ln,2 (ϑ20 ) + L2 (ϑ20 )
+
inf
ϑ2 ∈B(ϑ20 ,δ )
(h)
(h)
L2 (ϑ2 ) − L2 (ϑ20 ) .
On the one hand, the first two terms converge to zero in probability due to Lemma 3.4(a) and the con
(h)
(h)
(h)
tinuous mapping theorem. On the other hand, infϑ2 ∈B(ϑ 0 ,δ ) L2 (ϑ2 ) − L2 (ϑ20 ) > 0 since L2 (ϑ2 )
2
has a unique minimum in ϑ20 by Lemma 2.10.
Proof of Theorem 3.8. Let us assume that 12 < γ < 1. Apparently, the parameter subspace Θ1 is
the union of Θ1 = N n,γ (ϑ10 , δ1 ) ∪ Nn,γ (ϑ10 , δ1 ) and thus, we have already shown Theorem 3.8 for
the set N n,γ (ϑ10 , δ1 ) × B(ϑ20 , δ ) instead of Θ1 × B(ϑ20 , δ ) in Theorem 3.5. It remains to investigate
Nn,γ (ϑ10 , δ1 ) × B(ϑ20 , δ ). For any δ1 > 0 we obtain by Proposition 2.8
(h)
(h)
0
c
c
Ln (ϑ ) − Ln (ϑ ) > 0
lim P
inf
n→∞
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ )
= lim P
n→∞
≥ lim P
n→∞
≥ lim P
n→∞
inf
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ )
(h)
(h)
Ln (ϑ ) − Ln (ϑ 0 )
(h)
Ln,1 (ϑ ) + inf
inf
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ )
ϑ2 ∈B(ϑ20 ,δ )
(h)
sup
|Ln,1 (ϑ )|
0
0
ϑ ∈Nn,γ (ϑ1 ,δ1 )×B(ϑ2 ,δ )
≤ε;
>0
(h)
(h)
Ln,2 (ϑ2 ) − Ln,2 (ϑ20 )
inf
ϑ2 ∈B(ϑ20 ,δ )
>0
(h)
(h)
Ln,2 (ϑ2 ) − Ln,2 (ϑ20 ) >
ε .
Then a consequence of Lemma 3.10 and Lemma 3.11 is
(h)
(h)
0
c
c
Ln (ϑ ) − Ln (ϑ ) > 0 ≥ 1
lim P
inf
n→∞
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ )
which proves in combination with Theorem 3.5 the claim.
4 Asymptotic distributions of the QML estimator
The aim of this section is to derive the asymptotic distributions of the long-run parameter estimator
ϑbn,1 and the short-run parameter estimator ϑbn,2 . These two estimators have a different asymptotic
behavior and a different convergence rate. On the one hand, we prove the asymptotic normality of
the short-run QML estimator and on the other hand, we show that the long-run QML estimator is
asymptotically mixed normally distributed.
4.1 Asymptotic distribution of the long-run parameter estimator
We derive in this section the asymptotic distribution of the long-run QML estimator ϑbn,1 . From
Corollary 3.6 we already know that ϑbn,1 − ϑ10 = o p (n−γ ), for 0 ≤ γ < 1. Since the true parameter
ϑ 0 = ((ϑ10 )T , (ϑ20 )T )T is an element of the interior of the compact parameter space Θ = Θ1 × Θ2
due to Assumption A, the estimator ϑbn,1 is at some point also an element of the interior of Θ1 with
probability one. Because the parametrization is assumed to be threefold continuously differentiable,
T ,ϑ
bT )T via the first order condition ∇ϑ L
cn(h) (ϑbn,1 , ϑbn,2 ) = 0s .
we can find the minimizing ϑbn = (ϑbn,1
1
1
n,2
21
We apply a Taylor-expansion of the score vector around the point (ϑ10 , ϑbn,2 ) resulting in
c(h) (ϑ , ϑbn,2 )n(ϑbn,1 − ϑ 0 ),
cn(h) (ϑ 0 , ϑbn,2 ) + n−1 ∇2 L
0s1 = ∇ϑ1 L
ϑ1
n
n,1
1
1
(h)
(4.1)
c (ϑ , ϑbn,2 ) denotes the matrix whose ith row, i = 1, . . . , s1 , is equal to the ith row of
where ∇ϑ2 1 L
n
n,1
(h)
i
2
c
b
In the case
∇ϑ1 Ln (ϑ n,1 , ϑn,2 ) with ϑ in,1 ∈ Θ1 such that kϑ in,1 − ϑ10 k ≤ kϑbn,1 − ϑ10 k.
c(h) (ϑ , ϑbn,2 ) is non-singular we receive
∇ϑ2 1 L
n
n,1
−1
c(h) (ϑ , ϑbn,2 )
cn(h) (ϑ 0 , ϑbn,2 ).
n(ϑbn,1 − ϑ10 ) = − n−1 ∇2ϑ1 L
∇ ϑ1 L
n,1
1
n
cn(h) (ϑ ) and the HesThus, we have to consider the asymptotic behavior of the score vector ∇ϑ1 L
cn(h) (ϑ ). Based on Proposition 2.8 it is sufficient to consider ∇2 Ln(h) (ϑ ) and
sian matrix ∇2ϑ1 L
ϑ1
(h)
∇ϑ1 Ln (ϑ ), respectively.
4.1.1 Asymptotic behavior of the score vector
First, we show the convergence of the gradient with respect to the long-run parameter ϑ1 . For this,
we consider the partial derivatives with respect to the ith -component of the parameter vector ϑ , i =
1, . . . , s1 , of the log-likelihood function. These partial derivatives are given due to the differentiation
rules for matrix functions (see, e.g., Lütkepohl [23, Appendix A.13]) by
(h)
∂i Ln (ϑ )
= tr
+
1 n (h) −1 (h)
(h)
(h) −1
(h)
T
− ∑ tr Vϑ
εk (ϑ )εk (ϑ ) Vϑ
∂iVϑ
n k=1
(h) −1 (h)
(h)
∂i εk (ϑ )T Vϑ
εk (ϑ ).
(4.2)
(h)
(h) −1
∂iVϑ
Vϑ
2 n
∑
n k=1
Proposition 4.1. The score vector with respect to the long-run parameters ϑ1 satisfies
(h)
(1)
w
∇ϑ1 Ln (ϑ 0 ) −−→ J1 (ϑ 0 ) := J1 (ϑ 0 )
(s1 )
···
J1
where
(i)
J1 (ϑ 0 )
= 2 tr
+ 2 tr
(h) −1
Vϑ 0
"
−∂i1 Π(ϑ 0 ), 0d×d
(h) −1
Vϑ 0
Z
0
1
#
#
T
W (r) dW (r)
∞
T
(ϑ 0 ) ,
k(1, ϑ 0 )
−Π(ϑ 0 )
!#
Γ∂ 1 k(B,ϑ 0 )∆Y (h) ,ε (h) (ϑ 0 ) (0) + ∑ Γ−∂ 1 Π(ϑ 0 )∆Y (h) ,ε (h) (ϑ 0 ) ( j)
i
j=1
i
and (W # (r))0≤r≤1 = ((W1 (r)T ,W2 (r)T )T )0≤r≤1 with the notation of Proposition A.1.
Proof. Equation (4.2) implies for i = 1, . . . , s1 that
n
(h) −1 1 (h)
(h)
(h)
(h) −1
(h) −1 1
(h)
0 (h)
0 T
)
)
∂i1 Ln (ϑ 0 ) = tr Vϑ 0
∂i Vϑ 0 − tr Vϑ 0
∂i1Vϑ 0 Vϑ 0
ε
ϑ
ε
ϑ
(
(
∑ k
k
n k=1
22
+ 2 · tr
(h) −1 1
Vϑ 0
n
n
∑
k=1
(h)
(h)
∂i1 εk (ϑ 0 ) εk (ϑ 0 )T
=: In,1 + In,2 + In,3 .
(h)
(h)
(h)
Note that the second term In,2 converges due to the ergodicity of (εk (ϑ 0 ))k∈N , E(εk (ϑ 0 )εk (ϑ 0 )T ) =
(h)
Vϑ 0 (see Lemma 2.7(a,e)) and Birkhoff’s Ergodic Theorem (cf. Bradley [7, 2.3 Ergodic Theorem])
so that
(h) −1 (h)
a.s.
(h) −1
(h) −1 1 (h)
1 (h)
In,2 −−→ − tr Vϑ 0
∂i Vϑ 0 Vϑ 0
∂i Vϑ 0 .
Vϑ 0 = − tr Vϑ 0
a.s.
Hence, In,1 + In,2 −−→ 0. Thus, it only remains to show the convergence of the last term In,3 . We obtain
with Proposition A.1(a,c) and the continuous mapping theorem
(h)
1 n
(h)
∂i1 εk (ϑ 0 ) εk (ϑ 0 )T
∑
n k=1
(h)
(h)
1 n
1 n
(h) T
(h)
1
0
0
∂
ϑ
ϑ
[
[ ∂i1 Π(ϑ 0 ) Yk−1 ][Π(ϑ 0 )Yst,k−1 ]T
][k(B,
]
+
Π(
)
Y
)∆Y
∑
∑
i
k−1
k
n k=1
n k=1
(h) (h)
1 n
+ ∑ [ ∂i1 k(B, ϑ 0 ) ∆Yk ]εk (ϑ 0 )T
n k=1
∞
Z 1
w
−−→ − ∂i1 Π(ϑ 0 )
W1 (r)dW1 (r)T k(1, ϑ 0 )T − ∑ Γ∂i1 Π(ϑ 0 )∆Y (h) ,k(B,ϑ 0 )∆Y (h) ( j)
=−
0
Z 1
+ ∂i1 Π(ϑ 0 )
0
j=1
∞
W1 (r) dW2 (r)T Π(ϑ 0 )T + ∑ Γ∂ 1 Π(ϑ 0 )∆Y (h) ,Π(ϑ 0 )Y (h) ( j)
j=1
st
i
(4.3)
+ Γ∂ 1 k(B,ϑ 0 )∆Y (h) ,ε (h) (ϑ 0 ) (0).
i
w
(i)
Then the continuous mapping theorem results in In,3 −−→ J1 (ϑ 0 ) which concludes the proof.
4.1.2 Asymptotic behavior of the Hessian matrix
(h)
The second partial derivatives of the log-likelihood function Ln (ϑ ) are given by
(h)
(h)
(h)
(h) −1
(h) −1 2 (h)
(h) −1
∂i, j Ln (ϑ ) = tr Vϑ
∂i, jVϑ − Vϑ
∂iVϑ
∂ jVϑ
Vϑ
1 n (h) −1 (h)
(h)
(h) −1 2 (h)
− ∑ tr Vϑ
εk (ϑ )εk (ϑ )T Vϑ
∂i, jVϑ
n k=1
−1 (h)
1 n (h) −1
(h) −1
(h)
(h)
(h)
∂ jVϑ
∂iVϑ
V (h) εk (ϑ )εk (ϑ )T Vϑ
+ ∑ tr Vϑ
n k=1
1 n (h) −1 (h)
(h)
(h)
(h)
(h) −1
(h) −1
εk (ϑ )εk (ϑ )T Vϑ
∂ jVϑ
∂iVϑ
Vϑ
+ ∑ tr Vϑ
n k=1
(h) −1 (h)
1 n (h) −1
(h)
(h)
∂ j εk (ϑ )εk (ϑ )T Vϑ
∂iVϑ
− ∑ tr Vϑ
n k=1
2 n
(h)
(h) −1 (h)
+ ∑ ∂i, j εk (ϑ )T Vϑ
εk (ϑ )
n k=1
23
(h) −1
2 n (h) −1 (h)
(h)
(h)
tr Vϑ
εk (ϑ ) ∂i εk (ϑ )T Vϑ
∂ jVϑ
∑
n k=1
2 n (h)
(h) −1
(h)
∂ j εk (ϑ )
+ ∑ ∂i εk (ϑ )T Vϑ
n k=1
−
8
=:
∑ In, j .
(4.4)
j=1
Since the Hessian matrix should be asymptotically positive definite we need an additional assumption.
⊥T C
Assumption E. The ((d − c)c × s1 )- dimensional gradient matrix ∇ϑ1 C1,
is of full column
1
ϑ1
rank s1 .
The asymptotic distribution of the Hessian matrix is given in the next proposition.
Proposition 4.2. Let Assumption E additionally hold. Define the (s1 × s1 )-dimensional random matrix Z1 (ϑ 0 ) as
Z 1
(h) −1 1
0
0
T
1
0 T
∂i Π(ϑ ) W1 (r)W1 (r) dr ∂ j Π(ϑ )
[Z1 (ϑ )]i, j := 2 · tr Vϑ 0
0
for i, j = 1, . . . , s1 . Then Z1 (ϑ 0 ) is almost surely positive definite and
(h)
w
n−1 ∇ϑ2 1 Ln (ϑ 0 ) −−→ Z1 (ϑ 0 ).
Proof. First, we prove the asymptotic behavior of the score vector and then, in the next step, that the
limit is almost surely positive definite.
Step 1: The first term n1 In,1 in (4.4) converges to zero due to the additional normalizing rate of n−1 .
Due to Proposition A.1 (a,c) we have for j = 2, . . . , 7 that In, j = O p (1) (see exemplarily (4.3) for In,5 )
and hence, 1n ∑7j=2 In, j converges in probability to zero. To summarize
(h)
n−1 ∂i,1j Ln (ϑ 0 )
1
= In,8 + o p (1) = 2 · tr
n
(h) −1 1
Vϑ 0
n2
n
∑
(h)
(h)
∂i1 εk (ϑ 0 )∂ j1 εk (ϑ 0 )T
k=1
+ o p (1).
Due to Lemma 2.6 and Proposition A.1 (a,c) we receive
(h)
n−1 ∂i,1j Ln (ϑ 0 )
= 2 · tr
(h) −1 1
Vϑ 0
n2
n
∑
(h) (h)T
∂i1 Π(ϑ 0 )Yk−1Yk−1 ∂ j1 Π(ϑ 0 )T
k=1
+ o p (1).
Then Proposition A.1(b) and the continuous mapping theorem result in
(h)
n−1 ∂i,1j Ln (ϑ 0 )
w
−−→ 2 · tr
(h) −1 1
∂i Π(ϑ 0 )
Vϑ 0
Z 1
0
T
W1 (r)W1 (r) dr
T
∂ j1 Π(ϑ 0 )
.
In particular, we have also the joint convergence of the partial derivatives.
R
Step 2: Note that we can take W1 = C1 B1W3 and define M := B1 01 W3 (r)W3 (r)T drBT
1 , which is a
P-a.s. positive definite c × c matrix. We apply the Cholesky decomposition M = M∗ M∗T . By using
24
properties of the vec operator and the Kronecker product (see [5, Chapter 7.1]) we have
T
(h) − 1
(h) − 1
Vϑ 0 2 ∂i1 Π(ϑ 0 )C1 M Vϑ 0 2 ∂ j1 Π(ϑ 0 )C1
T
(h) − 12
(h) − 1
⊥T
0 1 ⊥T
α
(
ϑ
∂
vec
V
C
M
C
M
= 2vec Vϑ 0 2 α (ϑ 0 )∂i1C1,
)
C
j 1,ϑ10 1 ∗
ϑ10 1 ∗
ϑ0
T
(h) −1
⊥T
0 T
0
1 ⊥T
)
V
)
vec
C
= 2vec ∂i1C1,
M
⊗
.
C
α
(
ϑ
α
(
ϑ
∂
C
0
0
1
1
0
j 1,ϑ
ϑ
ϑ
[Z1 (ϑ 0 )]i, j = 2 tr
1
1
(4.5)
(h) −1
(h) −1
α (ϑ 0 ) = rank(M) · rank α (ϑ 0 )T Vϑ 0
α (ϑ 0 ) due
Furthermore, rank M ⊗ α (ϑ 0 )T Vϑ 0
(h) −1
to [5, Fact 7.4.23] and thus, M ⊗ α (ϑ 0 )T Vϑ 0
α (ϑ 0 ) has full rank c · (d − c) a.s. Now, if we
consider the Hessian matrix Z1 (ϑ 0 ), we have
T
(h) −1
0 T
0
⊥T
⊥T
)
V
)
∇
C
Z1 (ϑ 0 ) = 2∇ϑ1 C1,
M
⊗
.
C
C
α
(
ϑ
α
(
ϑ
0
0
ϑ
1
1
0
1
1,ϑ
ϑ
ϑ
1
1
⊥T C
is of full column rank
Due to Assumption E the ((d − c)c × s1 )-dimensional matrix ∇ϑ1 C1,
1
0
ϑ1
and hence, the product has full rank s1 . Therefore, we have the positive definiteness almost surely.
4.1.3 Asymptotic mixed normality of the long-run QML estimator
We are able now to show the weak convergence of the long-run QML estimator and thus, we have one
main result.
Theorem 4.3. Let Assumption E additionally hold. Then we have as n → ∞
w
n(ϑbn,1 − ϑ10 ) −−→ −Z1 (ϑ 0 )−1 · J1 (ϑ 0 )
where J1 (ϑ 0 ) is defined as in Proposition 4.1 and Z1 (ϑ 0 ) as in Proposition 4.2, respectively.
Proof. From (4.1) we know that
cn(h) (ϑ 0 , ϑbn,2 ) + n−1 ∇2 L
c(h) (ϑ , ϑbn,2 )n(ϑbn,1 − ϑ 0 ).
0s1 = ∇ϑ1 L
1
ϑ1
n
n,1
1
(4.6)
(h)
In Proposition 4.1 we already derived the asymptotic behavior of the score vector ∇ϑ1 Ln (ϑ10 , ϑ20 )
(h)
and in Proposition 4.2 the asymptotic behavior of the Hessian matrix n−1 ∇2ϑ1 Ln (ϑ10 , ϑ20 ). However,
(h)
for the proof of Theorem 4.3 we require now the asymptotic behavior of ∇ϑ1 Ln (ϑ10 , ϑbn,2 ) and
n−1 ∇2 L (h) (ϑ , ϑbn,2 ). Therefore we use a kind of stochastic equicontinuity condition.
ϑ1
n
n,1
Lemma 4.4. For every ε > 0 and every η > 0, there exists an integer n(ε , η ) and real numbers
δ1 , δ2 > 0 such that for 21 < γ < 1,
(h)
(h)
(a) P supϑ2 ∈B(ϑ 0 ,δ2 ) k∇ϑ1 Ln (ϑ10 , ϑ2 ) − ∇ϑ1 Ln (ϑ10 , ϑ20 )k > ε ≤ η for n ≥ n(ε , η ),
2
(h)
(h)
(b) P supϑ ∈Nn,γ (ϑ 0 ,δ1 )×B(ϑ 0,δ2 ) kn−1 ∇2ϑ1 Ln (ϑ )−n−1 ∇2ϑ1 Ln (ϑ 0 )k > ε ≤ η
1
2
25
for n ≥ n(ε , η ).
Proof.
(h)
(h)
(a) Note that on the one hand, ∇ϑ1 Ln,1,1 (ϑ10 , ϑ2 ) = 0 since εn,1,1 (ϑ10 , ϑ2 ) = 0 and on the other hand,
(h)
∇ϑ1 Ln,2 (ϑ2 ) = 0 clearly. Hence,
(h)
sup
ϑ2 ∈B(ϑ20 ,δ2 )
(h)
sup
=
(h)
k∇ϑ1 Ln (ϑ10 , ϑ2 ) − ∇ϑ1 Ln (ϑ10 , ϑ20 )k
ϑ2 ∈B(ϑ20 ,δ2 )
(h)
k∇ϑ1 Ln,1,2 (ϑ10 , ϑ2 ) − ∇ϑ1 Ln,1,2 (ϑ10 , ϑ20 )k.
We can conclude with similar calculations as in Lemma 3.3 applying (A.4) and (A.6) that
(h)
sup
ϑ2 ∈B(ϑ20 ,δ2 )
(h)
k∇ϑ1 Ln,1,2 (ϑ10 , ϑ2 ) − ∇ϑ1 Ln,1,2 (ϑ10 , ϑ20 )k ≤
sup
ϑ2 ∈B(ϑ20 ,δ2 )
Ckϑ2 − ϑ20 kUn ≤ Cδ2Un .
Since Un = O p (1) due to Lemma 3.3 we obtain the statement.
(h)
(b) Due to ∇2ϑ1 Ln,2 (ϑ2 ) = 0 we have
(h)
sup
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ2 )
≤
(h)
kn−1 ∇2ϑ1 Ln (ϑ ) − n−1 ∇2ϑ1 Ln (ϑ 0 )k
(h)
sup
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ2 )
(h)
sup
+
(h)
kn−1 ∇2ϑ1 Ln,1,1 (ϑ ) − n−1 ∇ϑ2 1 Ln,1,1 (ϑ 0 )k
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ2 )
(h)
kn−1 ∇2ϑ1 Ln,1,2 (ϑ ) − n−1 ∇2ϑ1 Ln,1,2 (ϑ 0 )k.
Then the first term is bounded by (A.2) and the second term by (A.4) and (A.6), respectively. Hence,
≤
sup
ϑ ∈Nn,γ (ϑ10 ,δ1 )×B(ϑ20 ,δ2 )
≤ C δ2
1
1 n (h) (h)
Ckϑ − ϑ 0 k 2 ∑ Lk−1 [Lk−1 ]T + Ckϑ − ϑ 0 kUn
n k=1
n
!
1
1 n (h) (h) T
Lk−1 [Lk−1 ] + Cδ2Un .
∑
2
n k=1
n
Since Un = O p (1) due to Lemma 3.3 and
statement (b) follows.
1
n2
(h)
(h)
∑nk=1 Lk−1 [Lk−1 ]T = O p (1) due to Proposition A.1(b),
cn(h) (ϑ 0 , ϑbn,2 ) to J1 (ϑ 0 ) follows then by Proposition 2.8, ProposiThe weak convergence of ∇ϑ1 L
1
tion 4.1 and Lemma 4.4(a). Due to Proposition 2.8, Proposition 4.2 and Lemma 4.4(b) we have that
c(h) (ϑ , ϑbn,2 ) converges weakly to the random matrix Z1 (ϑ 0 ). In particular, Proposition A.1
n−1 ∇ϑ2 1 L
n,1
n
also guarantees the joint convergence of both terms. Finally, the almost sure positive definiteness of
Z1 (ϑ 0 ) allows us to take the inverse and reorder (4.1) so that
−1
w
c(h) (ϑ , ϑbn,2 )
cn(h) (ϑ 0 , ϑbn,2 ) −−
→ −Z1 (ϑ 0 )−1 · J1 (ϑ 0 ).
∇ ϑ1 L
n(ϑbn,1 − ϑ10 ) = − n−1 ∇2ϑ1 L
1
n
n,1
26
4.2 Asymptotic distribution of the short-run parameter estimator
Lastly, we derive the asymptotic normality of the short-run QML estimator ϑbn,2 which we prove by
using a Taylor-expansion of the QML-function similarly as in Section 4.1. Before we start the proof
(h)
(h)
we want to derive some mixing properties of the process (Yst,k , ∆Yk )k∈Z because this will be used
throughout this section.
(h)
(h)
Lemma 4.5. The process (Yst,k , ∆Yk )k∈Z is strongly mixing with mixing coefficients α∆Y (h) ,Y (h) (m) ≤
st
Cρ m for some 0 < ρ < 1. In particular, for any δ > 0,
∞
∑ α∆Y
m=1
δ
(h) ,Y (h)
st
(m) 2+δ < ∞.
(h)
Proof. Due to Equation (2.2) the process Yst has the state space representation
(h)
(h)
(h)
(h)
Xst,k = eA2 h Xst,k−1 +
with
Yst,k = C2 Xst,k
Z kh
(k−1)h
eA2 (kh−u) B2 dLu ,
k ∈ N.
(h)
Masuda [26, Theorem 4.3] proved that (Xst,k )k∈N is β -mixing with an exponentially rate since
(h)
(h)
EkXst,k k2 < ∞. Having Ek∆Lk k2 < ∞ in mind as well we can conclude on the same way as in [26,
Theorem 4.3] that the Markov process
!
! Z
(h)
(h)
kh
∆Lk−1
∆Lk
0m×m
0m×(N−c)
Im
dLu .
=
+
(h)
(h)
0(N−c)×m
eA2 h
eA2 (kh−u) B2
Xst,k
Xst,k−1
(k−1)h
is β -mixing with mixing coefficients β∆L(h) ,X (h) (m) ≤ Cρ1m for some 0 < ρ1 < 1. Hence, it is as well
α -mixing with mixing coefficient α∆L(h) ,X (h) (m) ≤ β∆L(h) ,X (h) (m) ≤ Cρ1m . Finally, it is obvious of the
definition of α -mixing that
(h)
∆Yk
(h)
Yst,k
!
=
C1 B1 C2
−C2
0d×m C2 0d×N−c
(h)
∆Lk
(h)
Xst,k
(h)
Xst,k−1
is α -mixing with α∆Y (h) ,Y (h) (m) ≤ α∆L(h) ,X (h) (m − 1) ≤ Cρ1m−1 .
st
4.2.1 Asymptotic behavior of the score vector
First, we prove that the partial derivatives have finite variance.
(h)
Lemma 4.6. For any ϑ2 ∈ Θ2 , n ∈ N, and i = 1, . . . , s2 the random variable E|∂ist Ln (ϑ 0 )|2 < ∞.
Proof. We have due to Lemma 2.7 (b) and the Cauchy-Schwarz inequality
(h)
(h) −1 (h)
εk (ϑ 0 )εk (ϑ 0 )T
Vϑ 0
(h) −1 st (h)
∂i Vϑ 0
(h) −1 (h) 0
(h)
εk (ϑ )
+ 2 · ∂ist εk (ϑ 0 )T Vϑ 0
21
(h)
(h)
(h)
≤ C · Ekεk (ϑ 0 )k4 +C · Ekεk (ϑ 0 )k4 Ek∂ist εk (ϑ 0 )k4 < ∞,
E − tr
Vϑ 0
so that the statement follows with (4.2).
27
2
Now we can prove the convergence of the covariance matrix of the score vector where we plug in the
true parameter.
(h)
(h) −1 (h)
(h)
εk,2 (ϑ2 ) that
Lemma 4.7. We have for all ϑ2 ∈ Θ2 and ℓk,2 (ϑ2 ) = εk,2 (ϑ2 )T Vϑ 0 ,ϑ
1
h
2
i
(h)
(h)
(h)
lim Var ∇ϑ2 Ln (ϑ 0 ) = ∑ Cov ∂ist ℓ1,2 (ϑ20 ), ∂ jst ℓ1+l,2 (ϑ20 )
n→∞
l∈Z
1≤i, j≤s2
=: I(ϑ20 ).
(4.7)
Proof. We can derive the result in a similar way as in Schlemm and Stelzer [34, Lemma 2.14]. Hence,
we only sketch the proof to show the differences. A detailed proof can be found in Scholz [35, Section
5.9]. It is sufficient to show that for all ϑ2 ∈ Θ2 and all i, j = 1, . . . , s2 the sequence
(h)
[Var ∇ϑ2 Ln (ϑ 0 ) ]i, j = n−1
n
n
∑ ∑
k1 =1 k2 =1
(h)
(h)
(i, j)
Cov ∂ist ℓk1 ,2 (ϑ20 ), ∂ jst ℓk2 ,2 (ϑ20 ) =: In (ϑ20 )
(4.8)
converges as n → ∞. By the representation of the partial derivatives in (4.2) and (2.3) the sequence
(h) −1 (h) 0
(h)
(h) −1 st (h)
(h) −1 (h)
(h)
(h)
∂ist ℓk,2 (ϑ20 ) = − tr Vϑ 0
εk,2 (ϑ20 )εk,2 (ϑ20 )T Vϑ 0
∂i Vϑ 0 + ∂ist εk,2 (ϑ20 )T Vϑ 0
εk,2 (ϑ2 )
(h) −1 (h) 0
(h) −1 (h)
(h)
(h) −1 st (h)
(h)
= − tr Vϑ 0
εk (ϑ 0 )εk (ϑ 0 )T Vϑ 0
∂i Vϑ 0 + ∂ist εk (ϑ 0 )T Vϑ 0
εk (ϑ )
is stationary and the covariance in (4.8) depends only on the difference l = k1 − k2 . If we can show
that
st (h)
0
st (h)
0
(
(
ℓ
),
ℓ
)
|<∞
(4.9)
|
Cov
∂
ϑ
∂
ϑ
∑
i 1,2 2
j 1+l,2 2
l∈Z
then the Dominated Convergence Theorem implies
(i, j)
In
(ϑ20 )
n−1
=
n
∑
(h)
(h)
(n − |l|) Cov ∂ist ℓ1,2 (ϑ20 ), ∂ jst ℓ1+l,2 (ϑ20 )
l=−n
n→∞
−−−→
st (h)
0
st (h)
0
∑ Cov ∂i ℓ1,2(ϑ2 ), ∂ j ℓ1+l,2 (ϑ2 ) .
l∈Z
Due to Lemma 4.5 and the uniformly exponentially bound of (K j (ϑ )) and (∂ist K j (ϑ )) finding the
dominant goes in the same vein as in the proof of [34, Lemma 2.14].
In the following we derive the convergence of the score vector with respect to the short-run parameters
by a truncation argument.
Proposition 4.8. For the gradient with respect to the short-run parameters the asymptotic behavior
√
w
(h)
n · ∇ϑ2 Ln (ϑ 0 ) −−→ N (0, I(ϑ20 ))
holds, where I(ϑ20 ) is the asymptotic covariance matrix given in (4.7).
(h)
Proof. First, we realize that representation (4.2) and Lemma 2.7 (c,d) result in E ∇ϑ2 Ln (ϑ 0 ) =
0s2 . Due to (2.3) we can rewrite (4.2) for m ∈ N as
(h)
∂ist Ln (ϑ 0 ) =
1 n
1 n (i)
(i)
(i)
(i)
Ym,k − EYm,k + ∑ Zm,k − EZm,k ,
∑
n k=1
n k=1
28
(4.10)
(i)
where Ym,k := tr
+
(h) −1 st (h)
∂i Vϑ 0
m
∑
ι1 =0
m
+
−
ι2 =0
m
∑
tr
∑
tr
ι1 ,ι2 =0
+2·
(i)
(h) −1
(h)
Vϑ 0
(h) −1 st (h)
∂i Vϑ 0
(h)T
(h) −1 st (h)
∂i Vϑ 0
(h)T
(h)
Π(ϑ 0 )Yst,k−1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
(h) −1
Vϑ 0
(h)
(h) −1 st (h)
∂i Vϑ 0
(h)T
∑
tr
∑
ι2 =0
m
∑
tr
ι1 ,ι2 =0
(i)
(i)
Kι1 (ϑ 0 )∆Yk−ι1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
ι1 =0
m
m
−2·
(h) −1
Vϑ 0
− tr
(h) −1 (h)T
(h)
Yst,k−1 Π(ϑ 0 )T
∂ist Π(ϑ 0 )Yst,k−1 Vϑ 0
(h)
(h) −1 (h)T
Yst,k−1 Π(ϑ 0 )T
tr ∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
+ 2 · tr
−2·
Π(ϑ 0 )Yst,k−1Yst,k−1 Π(ϑ 0 )T Vϑ 0
(h) −1
(h) (h)T
(h) −1 st (h)
Kι1 (ϑ 0 )∆Yk−ι1 Yst,k−1 Π(ϑ 0 )T Vϑ 0
tr Vϑ 0
∂i Vϑ 0
Vϑ 0
(h)T
(h)
(h) −1
∂ist Π(ϑ 0 )Yst,k−1 Vϑ 0
∆Yk−ι2 Kι2 (ϑ 0 )T
(h)
(h) −1
(h)T
∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
∆Yk−ι2 Kι2 (ϑ 0 )T ,
Zm,k :=Vm,k +Um,k ,
∞
(i)
(h) −1
(h) (h)T
(h) −1 st (h)
and Vm,k := ∑ tr Vϑ 0
∂i Vϑ 0
Kι1 (ϑ 0 )∆Yk−ι1 Yst,k−1 Π(ϑ 0 )T Vϑ 0
ι1 =m+1
∞
−
ι1 =m+1 ι2 =0
∞
+2·
(i)
Um,k :=
∞
∑ ∑ tr
−2·
∑
∑
tr
(h) −1
∑ ∑
+2·
tr
tr
ι2 =m+1
∞
∞
∑ ∑
(h)
(h) −1 st (h)
∂i Vϑ 0
(h)T
Kι1 (ϑ 0 )∆Yk−ι1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
(h)T
(h)
(h) −1 st (h)
∂i Vϑ 0
Π(ϑ )Yst,k−1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
(h) −1
Vϑ 0
(h)T
(h)
(h) −1
∆Yk−ι2 Kι2 (ϑ 0 )T ,
∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
(h)
(h)T
(h) −1 st (h)
∂i Vϑ 0
Kι1 (ϑ 0 )∆Yk−ι1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
(h)
(h) −1
(h)T
∂ist Π(ϑ 0 )Yst,k−1 Vϑ 0
∆Yk−ι2 Kι2 (ϑ 0 )T
tr
ι1 =0 ι2 =m+1
(1) T
Vϑ 0
∞
∑
(h) −1
Vϑ 0
(h)
(h) −1 (h)T
∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
Yst,k−1 Π(ϑ 0 )T
ι1 =m+1 ι2 =0
ι1 =0 ι2 =m+1
m
−2·
∑ ∑ tr
ι2 =m+1
∞
−
tr
ι1 =m+1
∞
∞
∞
(h) −1
(h)T
(h)
∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
∆Yk−ι2 Kι2 (ϑ 0 )T .
(1) T
(s ) T
(s ) T
We define Ym,k := (Ym,k , . . . ,Ym,k2 )T as well as Zm,k := (Zm,k , . . . , Zm,k2 )T and use a truncation
argument analogous to Schlemm and Stelzer [34, Lemma 2.16]. The main difference to [34] is that in
(i)
(i)
(i)
our case the definition of Ym,k , Vm,k and Um,k is more complex including the two stochastic processes
(h)
∆Y (h) , Yst and additional summands. We show the claim in three steps.
(h)
Step 1: The process Ym,k depends only on m + 1 past values of ∆Y (h) and Yst . Hence, it inherits the
29
(h)
strong mixing property of (∆Y (h) ,Yst ) and satisfies αYm (l) ≤ α∆Y (h) ,Y (h) (max{0, l − m − 1}). Thus,
st
δ /(2+δ )
∑∞
l=1 (αYm (l))
< ∞. Using the Cramér-Wold device and the univariate
by Lemma 4.5 we have
central limit theorem of Ibragimov [17, Theorem 1.7] for strongly mixing random variables we obtain
1 n
w
√ ∑ (Ym,k − EYm,k ) −−→ N (0s2 , Im (ϑ20 ))
n k=1
as n → ∞ where Im (ϑ20 ) := ∑l∈Z Cov(Ym,1 , Ym,1+l ). Next, we need to show that
m→∞
Im (ϑ20 ) −−−→ I(ϑ20 ).
(4.11)
(h)
(h)
( j)
(i)
Therefore, we first prove that Cov Ym,k ,Ym,k+l → Cov ∂ist ℓ1,2 (ϑ 0 ), ∂ jst ℓ1+l,2 (ϑ 0 ) as m → ∞. Note
that the bilinearity property of the covariance operator implies
(i) ( j)
(h)
(h)
| Cov Ym,k ,Ym,k+l − Cov ∂ist ℓk,2 (ϑ 0 ), ∂ jst ℓk+l,2 (ϑ 0 ) |
(i)
( j)
(h)
(i)
(h)
(h)
= | Cov Ym,k ,Ym,k+l − ∂ jst ℓk+l,2 (ϑ 0 ) + Cov Ym,k − ∂ist ℓk,2 (ϑ 0 ), ∂ jst ℓk+l,2 (ϑ 0 ) |
1/2
(i) 1/2
( j)
(h)
≤ Var Ym,1
Var Ym,1 − ∂ jst ℓ1,2 (ϑ 0 )
1/2
1/2
(i)
(h)
(h)
+ Var Ym,1 − ∂ist ℓ1,2 (ϑ 0 )
Var ∂ jst ℓl,2 (ϑ 0 )
,
(4.12)
where
(i)
(h)
Ym,k − ∂ist ℓk,2 (ϑ 0 ) = −
∞
∑
tr
∑
tr
ι1 =m+1
∞
−
ι2 =m+1
+
∑
(h) −1
Vϑ 0
tr
ι1 ,ι2
max{ι1 ,ι2 }>m
∞
+2·
+2·
−2·
∑
tr
∑
tr
ι1 =m+1
∞
ι2 =m+1
∑
(h) −1
Vϑ 0
ι1 ,ι2
max{ι1 ,ι2 }>m
(h)
(h) −1 st (h)
∂i Vϑ 0
(h)T
Kι1 (ϑ 0 )∆Yk−ι1 Yst,k−1 Π(ϑ 0 )T Vϑ 0
(h)
(h)T
(h) −1 st (h)
∂i Vϑ 0
Π(ϑ 0 )Yst,k−1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
(h) −1
Vϑ 0
(h)
(h)T
(h) −1 st (h)
∂i Vϑ 0
Kι1 (ϑ 0 )∆Yk−ι1 ∆Yk−ι2 Kι2 (ϑ 0 )T Vϑ 0
(h)
(h) −1 (h)T
Yst,k−1 Π(ϑ 0 )T
∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
(h) −1
(h)
(h)T
∂ist Π(ϑ 0 )Yst,k−1 Vϑ 0
∆Yk−ι2 Kι2 (ϑ 0 )T
tr
(h)
(h) −1
(h)T
∂ist Kι1 (ϑ 0 )∆Yk−ι1 Vϑ 0
∆Yk−ι2 Kι2 (ϑ 0 )T .
We obtain with the Cauchy-Schwarz inequality, the exponentially decreasing coefficients (Kl (ϑ 0 ))l∈N
(h)
and the finite 4th-moment of Yst and ∆Y (h) due to Assumption A that for some 0 < ρ < 1,
(i)
(h)
Var Ym,1 − ∂ist ℓ1,2 (ϑ 0 ) ≤ Cρ m .
(h)
(i)
(i)
Moreover, by the proof of Lemma 4.6 we have Var ∂ jst ℓ1,2 (ϑ 0 ) < ∞ and then, Var Ym,1 ) ≤ 2E Ym,1 −
2
2
(h)
(h)
∂ist ℓ1,2 (ϑ 0 ) + 2E ∂ jst ℓ1,2 (ϑ 0 ) < ∞ as well. Thus, (4.12) converges uniformly in l at an exponential
30
rate to zero as m → ∞ and
(h)
(h)
( j) m→∞
(i)
Cov Ym,k ,Ym,k+l −−−→ Cov ∂ist ℓ1,2 (ϑ 0 ), ∂ jst ℓ1+l,2 (ϑ 0 ) .
Then the same arguments as in [34, Lemma 2.16] guarantee that there exists a dominant (cf. Scholz
[35, Section 5.9]) so that dominated convergence results in (4.11) (cf. proof of Lemma 4.7).
Step 2: In this step, we show that √1n ∑nk=1 (Zm,k − EZm,k ) is asymptotically negligible. We have
1 n
tr Var √ ∑ Zm,k
n k=1
!
1 n
≤ 2 · tr Var √ ∑ Um,k
n k=1
!
1 n
+ 2 · tr Var √ ∑ Vm,k
n k=1
!
,
(4.13)
where Um,k and Vm,k are defined in the same vein as Zm,k . Since both terms can be treated similarly
we consider only the first one
!
s2
∞
1 n
1 n
( j)
(i)
′
tr Var √ ∑ Um,k
)
≤
tr
Cov(U
,
U
=
m,k
m,k
∑ ∑ | Cov(Um,1 ,Um,1+l )|. (4.14)
n k=1
n k,k∑
′ =1
i, j=1 l=−∞
With the same arguments as in [34, Lemma 2.16] we obtain independent of i and j the upper bound
∞
(i)
( j)
2m
∑ | Cov(Um,1 ,Um,1+l )| ≤
l=0
(i)
∞
( j)
∑ | Cov(Um,1 ,Um,1+l )| +
l=0
≤ Cρ m
∑
l=2m+1
(i)
( j)
| Cov(Um,1 ,Um,1+l )|
δ
δ +2
m + ∑ α∆Y (h) ,Y (h) (l)
,
∞
st
l=0
which implies tr Var √1n ∑nk=1 Um,k ≤ C(m +C)ρ m due to (4.14). With the same ideas one obtains
an equivalent bound for tr Var √1n ∑nk=1 Vm,k
and thus, we have with (4.13) that
1 n
tr Var √ ∑ Zm,k
n k=1
!
≤ C(m +C)ρ m .
(4.15)
Step 3: With the multivariate Chebyshev inequality (cf. Schlemm [32, Lemma 3.19]) and (4.15) from
Step 2 we obtain for every ε > 0 that
√
1 n
(h)
0
n∇ϑ2 Ln (ϑ ) − √ ∑ [Ym,k − EYm,k ] > ε
lim lim sup P
m→∞ n→∞
n k=1
!
s2
1 n
s2
≤ lim lim sup 2 tr Var √ ∑ Zm,k
≤ lim 2 C(m +C)ρ m = 0.
m→∞ n→∞ ε
m→∞
n k=1
ε
All in all, the results of Step 1 and Step 3 combined with Brockwell and Davis [8, Proposition 6.3.9]
yield the asymptotic normality in Proposition 4.8.
4.2.2 Asymptotic behavior of the Hessian matrix
We require an additional assumption for the Hessian matrix with respect to the short-run parameters
(h)
to be positive definite. Therefore, we need some notation. We write shortly Fϑ := eAϑ h − Kϑ Cϑ .
31
The function is similar to the function in Schlemm and Stelzer [34, Assumption C11]. However, we
define Fϑ different since we do not have a moving average representation of Y (h) with respect to the
innovations ε (h) . Hence, we have to adapt the criterion in [34] and define the function
h
iT
ih
(h) T
j T
T
T
I ⊗ Kϑ ⊗Cϑ (vec IN ) (vec Fϑ ) . . . (vec Fϑ )
.
ψϑ , j := j+1
(4.16)
(h)
vec Vϑ
Assumption F. Let there exist a positive index j0 such that the [( j0 + 2)d 2 × s2 ] matrix ∇ϑ2 ψϑ 0 , j0 has
rank s2 .
Proposition 4.9. Let Assumption F additionally hold. Then
p
(h)
∇2ϑ2 Ln (ϑ 0 ) −−→ Zst (ϑ 0 ),
where the (s2 × s2 )-dimensional matrix Zst (ϑ 0 ) is given by
(h) −1 st (h) 0
(h)
[Zst (ϑ 0 )]i, j = 2E ∂ist ε1 (ϑ 0 )T Vϑ 0
∂ j ε1 (ϑ )
1
(h) −1 st (h)
(h) − 1
(h) −
(h)
∂ j Vϑ 0 Vϑ 0 2 .
+ tr Vϑ 0 2 ∂ist Vϑ 0 Vϑ 0
Moreover, the limiting matrix Zst (ϑ 0 ) is almost surely a non-singular deterministic matrix.
Proof. We proceed as in the proof of Proposition 4.2.
(h)
(h)
(h)
Step 1: Since (∂i,stj εk (ϑ 0 ), ∂ jst εk (ϑ 0 ), εk (ϑ 0 ))k∈N is a stationary and ergodic sequence with finite
absolute moment (see Lemma 2.7(a)) we obtain with Birkhoff’s Ergodic Theorem
p
(h)
(h) −1 st (h)
(h)
(h)
(h) −1
(h) −1
∂i,stj Ln (ϑ 0 ) −−→ tr Vϑ 0
∂i, jVϑ 0 − Vϑ 0
∂ist Vϑ 0 Vϑ 0
∂ jst Vϑ 0
i
h
(h) −1 st (h)
(h)
(h)
(h) −1
− tr Vϑ 0
∂i, jVϑ 0
E ε1 (ϑ 0 )ε1 (ϑ 0 )T Vϑ 0
h
i
(h) −1
(h) −1 st (h)
(h)
(h)
(h)
(h) −1
∂ jst Vϑ 0 Vϑ 0
∂i Vϑ 0
E ε1 (ϑ 0 )ε1 (ϑ 0 )T Vϑ 0
+ tr Vϑ 0
h
i
(h) −1
(h)
(h)
(h) −1 st (h)
(h)
(h) −1
∂ jst Vϑ 0 Vϑ 0
∂i Vϑ 0
+ tr Vϑ 0
E ε1 (ϑ 0 )ε1 (ϑ 0 )T Vϑ 0
h
i
(h)
(h)
(h) −1 st (h)
(h) −1
E ∂ jst ε1 (ϑ 0 )ε1 (ϑ 0 )T Vϑ 0
− tr Vϑ 0
∂i Vϑ 0
h
i
(h) −1
(h)
(h)
+ 2 · tr Vϑ 0
E ε1 (ϑ 0 ) ∂i,stj ε1 (ϑ 0 )T
h
i
(h) −1
(h) −1 st (h)
(h)
(h)
∂ j Vϑ 0
− 2 · tr Vϑ 0
E ε1 (ϑ 0 )∂ist ε1 (ϑ 0 )T Vϑ 0
i
h
(h)
(h)
(h) −1
+ 2 · E ∂ist ε1 (ϑ 0 )T Vϑ 0
∂ jst ε1 (ϑ 0 ) .
Combining this with Lemma 2.7 (c,d) results in
h
(h) −1 st (h) 0 i
p
(h)
(h)
(h)
(h) −1
(h) −1 st (h)
∂i,stj Ln (ϑ 0 ) −−→ tr Vϑ 0
∂ist Vϑ 0 Vϑ 0
∂ j Vϑ 0 + 2 · E ∂ist ε1 (ϑ 0 )T Vϑ 0
∂ j ε1 (ϑ ) .
Step 2: Next we check that Zst (ϑ 0 ) is positive definite with probability one which we show by contradiction similarly to the corresponding proofs in Schlemm and Stelzer [34, Lemma 3.22] or Boubacar
32
and Francq [6, Lemma 4], respectively. From Step 2 we know that
p
(h)
where
and
∇2ϑ2 Ln (ϑ 0 ) −−→ Zst (ϑ 0 ) = Zst,1 (ϑ 0 ) + Zst,2 (ϑ 0 ),
h
i
(h) −1
(h)
(h)
∂ jst ε1 (ϑ 0 )
Zst,1 (ϑ 0 ) := 2 · E ∂ist ε1 (ϑ 0 )T Vϑ 0
1≤i, j≤s2
h
1
1 i
−
−1
−
(h)
(h)
(h)
(h)
(h)
Zst,2 (ϑ 0 ) := tr Vϑ 0 2 ∂ist Vϑ 0 Vϑ 0
∂ jst Vϑ 0 Vϑ 0 2
(4.17)
1≤i, j≤s2
.
We can factorize Zst,2 (ϑ 0 ) in the following way
Zst,2 (ϑ 0 ) = a1 . . . as2
T
a1 . . . as2
with
ai :=
(h) − 21
Vϑ 0
(h) − 21
⊗ Vϑ 0
(h)
vec ∂ist Vϑ 0
and thus, Zst,2 (ϑ 0 ) is obviously positive semi-definite. Similarly, Zst,1 (ϑ 0 ) is positive semi-definite.
It remains to check that for any c ∈ Rs2 \ {0s2 } we have cT Zst,i (ϑ 0 )c > 0 for at least one i ∈ {1, 2}.
We assume for the sake of contradiction that there exists a vector c ∈ Rs2 \{0s2 } such that
cT Zst,1 (ϑ 0 )c + cT Zst,2 (ϑ 0 )c = 0.
(✸)
In order to be zero, each summand cT Zst,1 (ϑ 0 )c and cT Zst,2 (ϑ 0 )c must be zero, since Zst,1 (ϑ 0 ) as
well as Zst,2 (ϑ 0 ) are positive semi-definite. But cT Zst,1 (ϑ 0 )c = 0 is only possible if
i
h
∞
(h)
(h) (h)
c P-a.s.
0d = (∇ϑ2 ε1 (ϑ 0 ))c = − ∑ ∇ϑ2 Cϑ 0 Fϑj−1
0 Kϑ 0 Yk− j
j=1
Rewriting this equation yields
i
i
h
h
∞
(h) (h)
(h) (h)
∇ϑ2 Cϑ 0 Kϑ 0 Yk−1 c = − ∑ ∇ϑ2 Cϑ 0 Fϑj−1
c P-a.s.
0 Kϑ 0 Yk− j
(4.18)
j=2
Hence, for every row i = 1, . . . , d and c = (c1 , . . . , cm )T we obtain
"
#
"
s2
d
∑ ∑
m=1 l=1
(h)
(h)
∂mst (Cϑ 0 Kϑ 0 )i,l Yk−1,l
∞
cm = − ∑
s2
d
∑ ∑
j=2 m=1 l=1
(h)
(h)
∂mst (Cϑ 0 Fϑj−1
0 Kϑ 0 )i,l Yk− j,l
#
cm
P-a.s.
which is equivalent to
i T
i T
h
h
∞
(h)
(h)
(h)
j−1 (h)
T
∇
K
c
=
−
Y
∇ϑ2 eT
K
F
C
e
C
0
0
ϑ
∑ 2 i ϑ ϑ 0 ϑ 0 c Yk− j
i ϑ
k−1
ϑ0
P-a.s.
j=2
i T
h
(h)
(h)
(h)
But then ∇ϑ2 eT
c Yk−1 lies in span{Y j : j ≤ k − 2}. By the definition of the linear innoK
C
0
i ϑ
ϑ0
h
i T
(h)
(h)
(h)
(h)
(h)
vations this is only possible if ∇ϑ2 eT
C
c εk−1 = 0 P-a.s. However, Vϑ 0 = E(εk−1 (εk−1 )T )
K
0
i ϑ
ϑ0
i
h
(h)
is non-singular due to [35, Lemma 5.9.1] so that necessarily ∇ϑ2 eT
i Cϑ 0 Kϑ 0 c = 0d for i = 1, . . . , d.
33
(h)
This is again equivalent to ∇ϑ2 (Cϑ 0 Kϑ 0 )c = 0d 2 . Plugging this in (4.18) gives
i
i
h
h
∞
(h) (h)
(h) (h)
∇ϑ2 Cϑ 0 Fϑ 0 Kϑ 0 Yk−2 c = − ∑ ∇ϑ2 Cϑ 0 Fϑj−1
0 Kϑ 0 Yk− j c.
j=3
(h)
Then we can show similarly ∇ϑ2 (Cϑ 0 Fϑ 0 Kϑ 0 )c = 0d 2 and obtain recursively that
(h)
∇ϑ2 (Cϑ 0 Fϑj 0 Kϑ 0 )c = 0d 2 ,
(4.19)
j ∈ N0 .
On the other hand, we obtain due cT Zst,2 (ϑ 0 )c = 0 under assumption (✸) that
(h)
∇ϑ2 Vϑ 0 c = 0d 2 .
(4.20)
The definition of ψϑ , j in (4.16), (4.19) and (4.20) imply that ∇ϑ2 ψϑ 0 , j c = 0( j+2)d 2 holds for all
j ∈ N, which contradicts Assumption F. Hence, Zst (ϑ 0 ) is almost surely positive definite.
4.2.3 Asymptotic normality of the short-run QML estimator
We conclude this section with the last main result of this paper, namely the asymptotic distribution of
the short-run QML estimator.
Theorem 4.10. Let Assumption F additionally hold. Furthermore, suppose
(h)
(h)
I(ϑ 0 ) = lim Var ∇ϑ2 Ln (ϑ 0 )
and Zst (ϑ 0 ) = lim ∇ϑ2 2 Ln (ϑ 0 ).
n→∞
Then as n → ∞,
√
n→∞
w
n(ϑbn,2 − ϑ20 ) −−→ N (0, Zst (ϑ 0 )−1 I(ϑ 0 )Zst (ϑ 0 )−1 ).
(4.21)
Again we need the following auxiliary result for the proof.
Lemma 4.11. For every ε > 0 and every η > 0, there exists an integer n(ε , η ) and real numbers
δ1 , δ2 > 0 such that for 43 < γ < 1,
√
√
(h)
(h)
(a) P supϑ1 ∈Nn,γ (ϑ 0 ,δ1 ) k n∇ϑ2 Ln (ϑ1 , ϑ20 )− n∇ϑ2 Ln (ϑ10 , ϑ20 )k > ε ≤ η for n ≥ n(ε , η );
1
(h)
(h)
(b) P supϑ ∈Nn,γ (ϑ 0 ,δ1 )×B(ϑ 0,δ2 ) k∇2ϑ2 Ln (ϑ ) − ∇2ϑ2 Ln (ϑ 0 )k > ε ≤ η
1
2
for n ≥ n(ε , η ).
Proof.
(a) We use the upper bound
sup
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
≤
√
√
(h)
(h)
k n∇ϑ2 Ln (ϑ1 , ϑ20 ) − n∇ϑ2 Ln (ϑ10 , ϑ20 )k
sup
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
+
√
√
(h)
(h)
k n∇ϑ2 Ln,1,1 (ϑ1 , ϑ20 ) − n∇ϑ2 Ln,1,1 (ϑ10 , ϑ20 )k
sup
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
√
√
(h)
(h)
k n∇ϑ2 Ln,1,2 (ϑ1 , ϑ20 ) − n∇ϑ2 Ln,1,2 (ϑ10 , ϑ20 )k.
34
(4.22)
Since Π(ϑ 0 )C1 = 0d×c and ∇ϑ2 (Π(ϑ 0 )C1 ) = 0dcs2 we can apply (A.3) and receive
sup
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
≤C
√
√
(h)
(h)
k n∇ϑ2 Ln,1,1 (ϑ1 , ϑ20 ) − n∇ϑ2 Ln,1,1 (ϑ10 , ϑ20 )k
sup
n
3
2
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
≤ Cn
3
2 −2γ
tr
kϑ1 − ϑ10 k2 tr
1 n (h) (h) T
∑ Lk−1 [Lk−1]
n2 k=1
!
1 n (h) (h) T
∑ Lk−1[Lk−1]
n2 k=1
(4.23)
!
p
−−→ 0
(4.24)
(h)
(h)
where we used γ > 3/4 and tr n12 ∑nk=1 Lk−1 [Lk−1 ]T = O p (1) due to Proposition A.1(b). For the
second term we get by (A.4) and (A.6), and similar calculations as in Lemma 3.3 that
sup
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
√
√
(h)
(h)
k n∇ϑ2 Ln,1,2 (ϑ1 , ϑ20 ) − n∇ϑ2 Ln,1,2 (ϑ10 , ϑ20 )k
≤
sup
ϑ1 ∈Nn,γ (ϑ10 ,δ1 )
√
1
p
nCkϑ1 − ϑ10 kUn ≤ Cn 2 −γ Un −−→ 0
(4.25)
due to γ > 3/4 and Un = O p (1). A combination of (4.22)-(4.25) proves (a).
(b) The proof is similar to (a).
Proof of Theorem 4.10. The proof is similar to the proof of Theorem 4.3 using Proposition 4.8, Proposition 4.9, Proposition 2.8 and Lemma 4.11.
5 Conclusion
The main contribution of the present paper is the development of a QML estimation procedure for the
parameters of cointegrated solutions of continuous-time linear state space models sampled equidistantly. To the best of our knowledge it is the first paper estimating the parameters of a cointegrated
output Y of a linear state space and MCARMA model, respectively. We showed that the QML estimator for the long-run parameter is super-consistent and that of the short-run parameter is consistent.
Moreover, the QML estimator for the long-run parameter converges with a n-rate to a mixed normal
√
distribution whereas the short-run parameter converges with a n-rate to a normal distribution.
In this paper we lay the mathematical basis for QML for cointegrated solutions of state-space models. In a separate paper [12] we present a canonical form for the state space model satisfying the
assumptions of this paper, and apply the method to a simulation study. In this simulation study we see
that the estimator works quite well in practice. We decided to split the paper because the introduction
into a canonical form is quite lengthy and would blow up the present paper. Moreover, a drawback of
our estimation procedure is that we assume that the cointegration rank is known in advance which is
not the case in reality. First we have to estimate and test the cointegration rank. For this it is possible
to incorporate some well-known results for estimating and testing the cointegration rank for cointegrated VARMA processes as, e.g., presented in [3, 24, 28, 37]. This will also be considered in [12].
Some parts of [12] can already be found in Scholz [35].
35
A Auxiliary results
A.1 Asymptotic results
For the derivation of the asymptotic behavior of our estimators we require the asymptotic behavior of
the score vector and the Hessian matrix. To obtain these asymptotic results we use the next proposition.
(h)
Proposition A.1. Let Assumption A hold. Furthermore, let (Lk )k∈N0 := (L(kh))k∈N0 be the Lévy
(h)
process sampled at distance h and ∆Lk = L(kh) − L((k − 1)h). Define for n ∈ N, k ∈ N0 ,
(h)
∆Yk
(h)
(h)
ξk := Yst,k
(h)
∆Lk
(h)
Sn :=
and
n
(h)
∑ ξk
.
k=1
∞
i
i
Let L(z, ϑ ) = ∑∞
i=0 Li (ϑ )z and L(z, ϑ ) = ∑i=0 Li (ϑ )z , ϑ ∈ Θ, z ∈ C, where (Li (ϑ ))i∈N0 is a nonstochastic uniformly exponentially bounded continuous matrix sequence in Md,2d+m (R) and similarly (Li (ϑ ))i∈N0 is a nonstochastic uniformly exponentially bounded continuous matrix sequence in
Md,2d+m (R). Moreover, Π(ϑ ) ∈ Md,2d+m (R), Π(ϑ ) ∈ Md,2d+m (R) are continuous matrix functions
(h)
(h)
(h)
as well. We write L(B, ϑ )ξ (h) = (L(B, ϑ )ξk )k∈N0 with L(B, ϑ )ξk = ∑∞
i=0 Li (ϑ )ξk−i and similarly
(h)
T
T
T
T
L(B, ϑ )ξ . Let (W (r))0≤r≤1 = ((W1 (r) ,W2 (r) ,W3 (r) ) )0≤r≤1 be the Brownian motion as defined in the beginning of Section 3. Then
(a) sup
ϑ ∈Θ
i p
h
1 n
(h)
(h) T
(h)
(h) T
ϑ
)
ξ
ϑ
)
ξ
ϑ
)
ξ
ϑ
)
ξ
[L(B,
][L(B,
]
−
E
[L(B,
][L(B,
]
−→ 0
∑
1
1+ j
k
k+ j
n k=1
n
(h)
(h)
w
−
→ Π(ϑ )
(b) n−2 ∑ Π(ϑ )Sk−1 [Sk−1 ]T Π(ϑ )T −
k=1
−1
(c) n
n
∑
(h)
(h)
Π(ϑ )Sk−1 [L(B, ϑ )ξk ]T
k=1
w
−
−
→ Π(ϑ )
for j ∈ N0 ,
Z 1
W (r)W (r)T dr Π(ϑ )T ,
Z 1
W (r)dW (r)T L(1, ϑ )T + ∑ ΓΠ(ϑ )ξ (h) ,L(B,ϑ )ξ (h) ( j).
0
0
∞
j=1
The stated weak convergence results also hold jointly.
Before we state the proof of Proposition A.1 we need some auxiliary results.
Lemma A.2. Let ψ de defined as in (3.2). Then the following statements hold.
(h)
(h)
(a) E(ξk ) = 02d+m and Ekξk k4 < ∞.
(b)
1
n
(h)
p
∑nk=1 ξk −−→ 02d+m
(h)
(h) (h) T
1 n
n ∑k=1 ξk [ξk+l ]
and
p
(h)
(h)
−−→ E(ξ1 [ξ1+l ]T ) =: Γξ (h) (l) for l ∈ N0 .
(h)
T
(c) ∑∞
l=0 Ekξ1 [ξ1+l ] k < ∞.
w
(h)
−−→ (ψ (1)W ∗ (r))0≤r≤1 where (W ∗ (r))0≤r≤1 is a (m + (N − c))-dimensional
(d) √1n S⌊nr⌋
0≤r≤1
Brownian motion with covariance matrix
ΣW ∗ =
Z h
0
T
ΣL
ΣL BT2 eA2 u
T
A
u
A
u
e 2 B2 ΣL e 2 B2 ΣL BT2 eA2 u
36
!
du
and ψ is defined as in (3.2).
(e)
1
n
(h)
(h)
w
∑nk=2 Sk−1 [ξk ]T −−→ ψ (1)
R1
0
W ∗ (r)dW ∗ (r)T ψ (1)T + ∑∞
l=1 Γξ (h) (l).
(h)
Proof. We shortly sketch the proof. The sequence (ξk )k∈N has the MA-representation
(h)
∆Yk
(h)
(h)
Yst,k = ξk =
(h)
∆Lk
k
∑
(h)
(A.1)
ψk− j η j
j=−∞
T
R kh
(h)
(h)
(h) T
T
A2 (kh−u) B dL . Hence, (ξ (h) )
with the iid sequence ηk := ∆L(h)
=
and
R
,
R
2
u
k∈N
(k−1)h e
k
k
k
k
is stationary and ergodic as a measurable map of a stationary ergodic process (see Krengel [22, Theorem 4.3 in Chapter 1]).
(a) is due to Assumption A.
(b) is a direct consequence of Birkhoff’s ergodic theorem.
(h)
(c) follows from Ekηk k4 < ∞, kψ j k ≤ Cρ j for some C > 0, 0 < ρ < 1 and the MA-representation
(A.1).
(d,e) are conclusions of Johansen [19, Theorem B.13] and the MA-representation (A.1).
Proof of Proposition A.1. (a) The proof follows directly by Theorem 4.1 of Saikkonen [29] and the
comment on [29, p.163, line 4] if we can show that Assumption 4.1 and 4.2 of that paper are satisfied.
Since we have uniformly exponentially bounded families of matrix sequences, Saikkonen [29, Assumption 4.1] is obviously satisfied. Saikkonen [29, Assumption 4.2 ] is satisfied due to Lemma A.2.
Note that we have two different coefficient matrices, whereas the results in Saikkonen [29] are
proved for the same coefficient matrix. However, [29, Theorem 4.1] also holds if the coefficient matrices are different as long as each sequence of matrix coefficients satisfies the necessary conditions
as mentioned in the paper of Saikkonen [29, p. 163].
(b,c) Due to Lemma A.2, Saikkonen [29, Assumption 4.3] is satisfied as well. Hence, we can
conclude the weak convergence result from [29, Theorem 4.2(iii)] and [29, Theorem 4.2(iv)], respectively.
A.2 Equicontinuity results
A kind of stochastic equicontinuity condition for the processes given in Proposition A.1 is presented
next.
Proposition A.3. Let the assumption and notation of Proposition A.1 hold. Assume further that
Π(ϑ ), Π(ϑ ) are Lipschitz-continuous and the sequence of matrix functions (∇ϑ (Li (ϑ )))i∈N0 and
(∇ϑ (Li (ϑ )))i∈N0 are uniformly exponentially bounded.
n
(h)
(h)
(a) Define Xn (ϑ ) = ∑ Π(ϑ )Sk−1 [Sk−1 ]T Π(ϑ )T . Then
k=1
kXn (ϑ ) − Xn (ϑ 0 )k ≤ Ckϑ − ϑ 0 k
37
n
(h)
(h)
∑ Sk−1 [Sk−1 ]T
k=1
.
(A.2)
If additionally Π(ϑ 0 ) = 0d×(2d+m) and Π(ϑ 0 ) = 0d×(2d+m) then
kXn (ϑ ) − Xn (ϑ 0 )k ≤ Ckϑ − ϑ 0 k2
n
n
(h)
(h)
∑ Sk−1 [Sk−1 ]T
(A.3)
.
k=1
(h)
(h)
(b) Define Xn (ϑ ) = ∑ Π(ϑ )Sk−1 [L(B, ϑ )ξk ]T . Then
k=1
kXn (ϑ ) − Xn (ϑ 0 )k ≤ Cnkϑ − ϑ 0 kVn
(A.4)
where
Vn =
1 n
1 n (h) (h) T
(h)
(h)
(h)
(h)
Sk−1 [ξk ] + kSn k[kρ (B)kξn k] + ∑ k∆Sk k[kρ (B)kξk k]
∑
n k=1
n k=1
+
i
1 n (h) h
(h) T
Sk−1 L(B, ϑ 0 )ξk
∑
n k=1
(A.5)
= O p (1),
(h)
(h)
kρ (z) = c ∑∞j=0 ρ j z j for some 0 < ρ < 1, c > 0, and kρ (B)kξk k := c ∑∞j=0 ρ j kξk− j k.
n
(h)
(h)
(c) Define Xn (ϑ ) = ∑ [L(B, ϑ )ξk ][L(B, ϑ )ξk ]T . Then there exists a random variable Wn =
k=1
O p (1) so that
kXn (ϑ ) − Xn (ϑ 0 )k ≤ Cnkϑ − ϑ 0 kWn .
(A.6)
In particular, Vn +Wn = O p (1).
Proof.
(a) We have the upper bound
kXn (ϑ ) − Xn (ϑ 0 )k
n
≤
(h)
(h)
∑ [Π(ϑ ) − Π(ϑ 0 )]Sk−1 [Sk−1 ]TΠ(ϑ )T
k=1
n
+
(h)
(h)
∑ Π(ϑ 0 )Sk−1 [Sk−1 ]T[Π(ϑ ) − Π(ϑ 0 )]T
k=1
≤ kΠ(ϑ ) − Π(ϑ 0 )kkΠ(ϑ )k + kΠ(ϑ ) − Π(ϑ 0 )kkΠ(ϑ 0 )k
n
(h)
(h)
∑ Sk−1 [Sk−1 ]T
.
k=1
Since Π(ϑ ) and Π(ϑ ) are Lipschitz continuous, max(kΠ(ϑ ) − Π(ϑ 0 )k, kΠ(ϑ ) − Π(ϑ 0 )k) ≤ Ckϑ −
ϑ 0 k and supϑ ∈Θ max(kΠ(ϑ )k, kΠ(ϑ )k) ≤ C. Thus, (A.2) holds. Moreover, (A.3) is valid because for
Π(ϑ 0 ) = 0d,2d+m , Π(ϑ 0 ) = 0d,2d+m the upper bound
kXn (ϑ ) − Xn (ϑ 0 )k = kXn (ϑ )k ≤ kΠ(ϑ )kkΠ(ϑ )k
n
(h)
(h)
∑ Sk−1 [Sk−1 ]T
k=1
≤ kΠ(ϑ ) − Π(ϑ 0 )kkΠ(ϑ ) − Π(ϑ 0 )k
38
n
(h)
(h)
∑ Sk−1 [Sk−1 ]T
k=1
≤ Ckϑ − ϑ 0 k2
n
(h)
(h)
∑ Sk−1 [Sk−1 ]T
k=1
is valid.
(b) Using a Taylor expansion leads to
vec(L(z, ϑ )) − vec(L(z, ϑ 0 )) =
∞
∑ [vec(L j (ϑ )) − vec(L j (ϑ 0 ))]z j =
j=0
∞
∑ ∇ϑ vec(L∗j (ϑ ( j)))(ϑ − ϑ 0 )z j
j=0
where vec(L∗j (ϑ ( j))) denotes the matrix whose ith row is equal to the ith row of vec(L j (ϑ i ( j)))
with ϑ i ( j) ∈ Θ such that kϑ i ( j) − ϑ 0 k ≤ kϑ − ϑ 0 k for i = 1, . . . , d(2d + m). Due to assumption,
k∇ϑ vec(L∗j (ϑ ( j)))k ≤ Cρ j for j ∈ N0 and some 0 < ρ < 1 so that
kL(z, ϑ ) − L(z, ϑ 0 )k ≤ kρ (|z|)kϑ − ϑ 0 k.
(A.7)
Hence,
(h)
(h)
k L(B, ϑ ) − L(B, ϑ 0 ) ξk k ≤ kϑ − ϑ 0 k[kρ (B)kξk k].
(A.8)
Define L∇ (z, ϑ , ϑ 0 ) := L(z, ϑ ) − L(z, ϑ 0 ) =: ∑∞j=0 L∇j (ϑ , ϑ 0 )z j . Then we apply a Beveridge-Nelson
decomposition (see Saikkonen [29, (9)]) to get
(h)
L∇ (B, ϑ , ϑ 0 )ξk
(h)
= L∇ (1, ϑ , ϑ 0 )ξk + ηk (ϑ , ϑ 0 ) − ηk−1 (ϑ , ϑ 0 )
(h)
∇
0
with ηk (ϑ , ϑ 0 ) := − ∑∞j=0 ∑∞
i= j+1 Li (ϑ , ϑ )ξk− j . Thus,
h
(h) iT
1 n
(h)
Π(ϑ )Sk−1 L(B, ϑ ) − L(B, ϑ 0 ) ξk
∑
n k=1
= Π(ϑ )
1 n (h) (h) T ∇
1 n (h)
1 n (h)
Sk−1 [ξk ] L (1, ϑ , ϑ 0 )T + Π(ϑ ) ∑ Sk−1 [ηk (ϑ , ϑ 0 )]T − Π(ϑ ) ∑ Sk−1 [ηk−1 (ϑ , ϑ 0 )]T
∑
n k=1
n k=1
n k=1
= Π(ϑ )
1 n (h) (h) T ∇
1 n
(h)
(h)
0 T
0 T
]
L
[
S
(1,
)
+
)]
−
(
ξ
ϑ
,
ϑ
Π(
ϑ
)S
[
η
ϑ
,
ϑ
Π(
ϑ
)
n
n
∑ k−1 k
∑ ∆Sk [ηk (ϑ , ϑ 0 )]T .
n k=1
n k=1
Due to (A.7), kL∇ (1, ϑ , ϑ 0 )k ≤ Ckϑ − ϑ 0 k and kL∇j (ϑ , ϑ 0 )k ≤ Ckϑ − ϑ 0 kρ j so that kηk (ϑ , ϑ 0 )k ≤
(h)
kϑ − ϑ 0 k[kρ (B)kξk k] as well. Finally, we receive
(h) iT
1 n (h) h
Sk−1 L(B, ϑ ) − L(B, ϑ 0 ) ξk
(A.9)
∑
n k=1
#
"
n
n
1
1
(h)
(h)
(h)
(h)
(h)
(h)
≤ Ckϑ − ϑ 0 k
∑ Sk−1 [ξk ]T + kSn k[kρ (B)kξn k] + n ∑ k∆Sk k[kρ (B)kξk k] ,
n k=1
k=1
Π(ϑ )
and
[Π(ϑ ) − Π(ϑ 0 )]
i
i
1 n (h) h
1 n (h) h
(h) T
(h) T
≤ Ckϑ − ϑ 0 k
. (A.10)
Sk−1 L(B, ϑ 0 )ξk
Sk−1 L(B, ϑ 0 )ξk
∑
∑
n k=1
n k=1
39
Then (A.9) and (A.10) result in (A.4).
(h)
(h) T
1 n
n ∑k=1 Sk−1 [ξk ]
in the definition of Vn is O p (1) by
(h)
1 n
Lemma A.2. Moreover,
= n ∑k=1 ξk = O p (1) by Birkhoff’s Ergodic Theorem; similarly the
(h)
(h)
1 n
third term n ∑k=1 k∆Sk k[kρ (B)kξk k] is O p (1) by Birkhoff’s Ergodic Theorem. Finally, the last
h
i
n (h)
(h) T
term n1 ∑ Sk−1 L(B, ϑ 0 )ξk
is O p (1) by Proposition A.1(c).
k=1
It remains to prove (A.5). The first term
1 (h)
n Sn
(c) The proof is similarly to the proof of (b).
B Proof of Proposition 2.8
First, we present some auxiliary results for the proof of Proposition 2.8.
(h)
(h)
(h) (h)
Lemma B.1. Let Assumption A and B hold. Define X1 (ϑ ) = ∑∞j=0 (eAϑ h − Kϑ Cϑ ) j Kϑ Y− j . Then
n
o
(h)
(h)
1
E supϑ ∈Θ kb
< ∞.
εk (ϑ )k2
(a) E supϑ ∈Θ kX1 (ϑ )k2 < ∞ and maxk∈N (1+k)
n
o
(h)
(h)
1
< ∞.
εk (ϑ )k2
(b) E supϑ ∈Θ k∂u X1 (ϑ )k2 < ∞ and maxk∈N (1+k)
E supϑ ∈Θ k∂u b
n
o
(h)
(h)
1
(c) E supϑ ∈Θ k∂u,v X1 (ϑ )k2 < ∞ and maxk∈N (1+k)
E supϑ ∈Θ k∂u,v b
< ∞.
εk (ϑ )k2
(h)
Proof. We prove (a) exemplary for (b) and (c). First, remark that EkY j k2 ≤ C(1 + | j|) for j ∈ Z.
(h)
Since all eigenvalues of (eAϑ h − Kϑ Cϑ ) lie inside the unit circle (see Scholz [35, Lemma 4.6.7]) and
all matrix functions are continuous on the compact set Θ and, hence, bounded, we receive for some
(h)
(h)
(h)
0 < ρ < 1 that supϑ ∈Θ keAϑ h − Kϑ Cϑ k ≤ ρ and supϑ ∈Θ kX1 (ϑ )k ≤ C ∑∞j=0 ρ j kY− j k. Thus,
(h)
E sup kX1 (ϑ )k2
ϑ ∈Θ
∞
≤C
(h)
∑ ρ j (EkY− j k2 )1/2
j=0
!2
∞
≤C
∑ ρ j (1 + j)1/2
j=0
!2
< ∞.
Similarly,
k−1
(h)
(h)
(h)
sup kb
εk (ϑ )k ≤ kYk k +Cρ k−1 sup kXb1 (ϑ )k + ∑ Cρ j kYk− j k,
ϑ ∈Θ
such that
E
(h)
sup kb
εk (ϑ )k2
ϑ ∈Θ
ϑ ∈Θ
2 2k−2
2
≤ 3EkYk k + 3C ρ
E
j=1
(h)
sup kXb1 (ϑ )k2
ϑ ∈Θ
(h)
k−1
+3
j=1
≤ C (1 + k) + ρ 2k−2 E sup kXb1 (ϑ )k2 + k
ϑ ∈Θ
∑ Cρ
∞
∑ρj
j=0
Finally, due to Assumption B
max
k∈N
1
(h)
εk (ϑ )k2
E sup kb
(1 + k)
ϑ ∈Θ
(h)
≤ C 1 + E sup kXb1 (ϑ )k2 +
ϑ ∈Θ
40
j
(h)
(EkYk− j k2 )1/2
!2
.
!2
∑ ρ j < ∞.
∞
j=0
!2
Lemma B.2. Let Assumption A and B hold. Furthermore, let u, v ∈ {1, . . . , s}.
(a) Then there exists a positive random variable ξ with Eξ 2 < ∞ and a constant 0 < ρ < 1 so that
(h)
(h)
supϑ ∈Θ kb
εk (ϑ ) − εk (ϑ )k ≤ Cρ k−1 ξ for any k ∈ N.
(b) Then there exists a positive random variable ξ u with E(ξ (u) )2 < ∞ and a constant 0 < ρ < 1 so
(h)
(h)
εk (ϑ ) − ∂u εk (ϑ )k ≤ Cρ k−1 ξ (u) for any k ∈ N.
that supϑ ∈Θ k∂u b
(c) Then there exists a positive random variable ξ (u,v) with E(ξ (u,v) )2 < ∞ and a constant 0 < ρ < 1
(h)
(h)
εk (ϑ ) − ∂u,v εk (ϑ )k ≤ Cρ k−1 ξ (u,v) for any k ∈ N.
so that supϑ ∈Θ k∂u,v b
Proof. (a) We use the representation
(h)
(h)
(h)
(h)
(h)
b
εk (ϑ ) − εk (ϑ ) = Cϑ (eAϑ h − Kϑ Cϑ )k−1 (Xb1 (ϑ ) − X1 (ϑ ))
(h)
(h)
and define ξ = supϑ ∈Θ kXb1 (ϑ )k + supϑ ∈Θ kX1 (ϑ )k. Due to Assumption B and Lemma B.1(a) we
(h)
know that Eξ 2 < ∞. Since all eigenvalues of (eAϑ h − Kϑ Cϑ ) lie inside the unit circle and Cϑ is
bounded as a continuous function on the compact set Θ there exists a C > 0 and 0 < ρ < 1 so that
(h)
supϑ ∈Θ kCϑ (eAϑ h − Kϑ Cϑ )k−1 k ≤ Cρ k−1 and
(h)
(h)
(h)
(h)
(h)
sup kb
εk (ϑ ) − εk (ϑ )k ≤ sup kCϑ (eAϑ h − Kϑ Cϑ )kk−1 sup kXb1 (ϑ ) − X1 (ϑ )k ≤ Cρ k−1 ξ .
ϑ ∈Θ
ϑ ∈Θ
ϑ ∈Θ
(b,c) can be proven similarly.
Proof of Proposition 2.8.
(a) First,
cn(h) (ϑ ) − Ln(h) (ϑ )
L
i
1 n h (h)
(h) −1 (h)
(h) −1 (h)
(h)
(h)
(h)
T
T
b
b
b
=
ε
ϑ
)
−
ε
ϑ
))
V
ε
ϑ
)
−
ε
ϑ
)
V
ε
ϑ
)
−
ε
ϑ
))
.
(
(
(
(
(
(
(
(
∑ k
ϑ
ϑ
k
k
k
k
k
n k=1
Then
(h)
(h)
cn (ϑ ) − Ln (ϑ )|
n sup |L
ϑ ∈Θ
n
(h) −1
(h)
(h)
(h)
(h)
≤ sup k(Vϑ ) k ∑ sup kb
εk (ϑ ) − εk (ϑ )k sup kb
εk (ϑ )k + sup kεk (ϑ )k .
ϑ ∈Θ
k=1 ϑ ∈Θ
ϑ ∈Θ
ϑ ∈Θ
Due to Lemma 2.5 and Lemma B.2(a)
n
≤ Cξ
∑ρ
k=1
k
(h)
(h)
sup kb
εk (ϑ )k + sup kεk (ϑ )k
ϑ ∈Θ
ϑ ∈Θ
with E(ξ 2 ) < ∞. From this and Cauchy-Schwarz inequality we can conclude that
(h)
(h)
c
nE sup |Ln (ϑ ) − Ln (ϑ )|
ϑ ∈Θ
41
2 1/2
≤ C(Eξ )
"
1/2
1/2 #
(h)
(h)
2
2
+ E sup kεk (ϑ )k
.
∑ ρ E sup kbεk (ϑ )k
n
k
k=1
ϑ ∈Θ
ϑ ∈Θ
An application of Lemma B.1(a) yields
∞
≤ C(Eξ 2 )1/2 ∑ ρ k (1 + k)1/2 < ∞.
k=1
cn(h) (ϑ ) − Ln(h) (ϑ )| = O p (1) so that (a) follows. Again (b) and (c) can be
This proves, n supϑ ∈Θ |L
proven similarly.
References
[1] AOKI , M. (1990). State space modeling of time series 2. ed. Springer, Berlin.
[2] BAUER , D. AND WAGNER , M. (2002). Asymptotic properties of pseudo maximum likelihood estimates for multiple frequency I(1) processes. Diskussionsschriften. Universitaet Bern, Departement Volkswirtschaft.
[3] BAUER , D. AND WAGNER , M. (2002). Estimating cointegrated systems using subspace algorithms. J.
Econometrics 111, 47–84.
[4] B ERGSTROM , A. R. (1997). Gaussian estimation of mixed-order continuous-time dynamic models with
unobservable stochastic trends from mixed stock and flow data. Econometric Theory 13, 467–505.
[5] B ERNSTEIN , D. S. (2009). Matrix mathematics: theory, facts, and formulas 2nd ed. Princeton Univ.
Press, Princeton.
[6] B OUBACAR M AÏNASSARA , Y. AND F RANCQ , C. (2011). Estimating structural VARMA models with
uncorrelated but non-independent error terms. J. Multivar. Anal. 102, 496–505.
[7] B RADLEY, R. C. (2007). Introduction to strong mixing conditions vol. 1. Kendrick, Heber City.
[8] B ROCKWELL , P. J. AND DAVIS , R. A. (1998). Time Series: Theory and Methods 2nd ed. Springer Ser.
Statist. Springer, New York.
[9] C HAMBERS , A. AND M C C RORIE , J. R. (2007). Frequency domain estimation of temporally aggregated
Gaussian cointegrated systems. J. Econometrics 136, 1–29.
[10] FASEN , V. (2013). Time series regression on integrated continuous-time processes with heavy and light
tails. Econometric Theory 29, 28–67.
[11] FASEN , V. (2014). Limit theory for high frequency sampled MCARMA models. Adv. Appl. Probab. 46,
846–877.
[12] FASEN -H ARTMANN , V.
process. In preparation.
AND
S CHOLZ , M. (2017). A canonical form for a cointegrated MCARMA
[13] FASEN -H ARTMANN , V. AND S CHOLZ , M. (2017). Cointegrated continuous-time linear state-space and
MCARMA models. Submitted for publication.
[14] H ANNAN , E. AND D EISTLER , M. (2012). The Statistical Theory of Linear Systems. Society for Industrial
and Applied Mathematics, Philadelphia.
[15] H ARVEY, A. C. AND S TOCK , J. H. (1985). The estimation of higher-order continuous time autoregressive
models. Econometric Theory 1, 97–112.
42
[16] H ARVEY, A. C. AND S TOCK , J. H. (1988). Continuous time autoregressive models with common
stochastic trends. J. Econom. Dynam. Control. 12, 365–384.
[17] I BRAGIMOV, I. A. (1962). Some limit theorems for stationary processes. Theory Probab. Appl. 7,
349–382.
[18] J OHANSEN , S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector
autoregressive models. Econometrica 59, 1551–1580.
[19] J OHANSEN , S. (2009). Likelihood-based inference in cointegrated vector autoregressive models. Advanced texts in econometrics. Oxford Univ. Press, Oxford.
[20] K ESSLER , M. AND R AHBEK , A. (2001). Asymptotic likelihood based inference for co-integrated homogenous Gaussian diffusions. Scand. J. Statist. 28, 455–470.
[21] K ESSLER , M. AND R AHBEK , A. (2004). Identification and inference for multivariate cointegrated and
ergodic Gaussian diffusions. Stat. Inference Stoch. Process. 7, 137–151.
[22] K RENGEL , U. (1985). Ergodic theorems: with a supplement. de Gruyter Stud. Math. 6. De Gruyter,
Berlin.
[23] L ÜTKEPOHL , H. (2005). New introduction to multiple time series analysis. Springer, Berlin.
[24] L ÜTKEPOHL , H. AND C LAESSEN , H. (1997). Analysis of cointegrated VARMA processes. J. Econometrics 80, 223–239.
[25] M ARQUARDT, T.
117, 96–120.
AND
S TELZER , R. (2007). Multivariate CARMA processes. Stochastic Process. Appl.
[26] M ASUDA , H. (2004). On multidimensional Ornstein-Uhlenbeck processes driven by a general Lévy
process. Bernoulli 10, 97–120.
[27] R EINSEL , G. C. (1997). Elements of multivariate time series analysis 2. ed. Springer-Verlag, New York.
[28] S AIKKONEN , P. (1992). Estimation and testing of cointegrated systems by an autoregressive approximation. Econometric Theory 8, 1–27.
[29] S AIKKONEN , P. (1993). Continuous weak convergence and stochastic equicontinuity results for integrated
processes with an application to the estimation of a regression model. Econometric Theory 9, 155–188.
[30] S AIKKONEN , P. (1995). Problems with the asymptotic theory of maximum likelihood estimation in
integrated and cointegrated systems. Econometric Theory 11, 888–911.
[31] S ATO , K.-I. (1999). Lévy processes and infinitely divisible distributions. Cambridge Stud. Adv. Math.
68. Cambridge Univ. Press, Cambridge.
[32] S CHLEMM , E. (2011). Estimation of continuous-time ARMA models and random matrices with dependent entries. PhD thesis. Technische Universität München.
[33] S CHLEMM , E. AND S TELZER , R. (2012). Multivariate CARMA processes, continuous-time state space
models and complete regularity of the innovations of the sampled processes. Bernoulli 18, 46–63.
[34] S CHLEMM , E. AND S TELZER , R. (2012). Quasi maximum likelihood estimation for strongly mixing
state space models and multivariate Lévy-driven CARMA processes. Electron. J. Stat. 6, 2185–2234.
[35] S CHOLZ , M. (2016). Estimation of Cointegrated Multivariate Continuous-time Autoregressive Moving
Average Processes. PhD thesis. Karlsruher Institute of Technology (KIT), Karlsruhe.
[36] S TOCKMARR , A. AND JACOBSEN , M. (1994). Gaussian diffusions and autoregressive processes: weak
convergence and statistical inference. Scand. J. Statist. 21, 403–429.
[37] YAP, S. F. AND R EINSEL , G. C. (1995). Estimation and testing for unit roots in a partially nonstationary
vector autoregressive moving average model. J. Amer. Statist. Assoc. 90, 253–267.
43
| 10math.ST
|
arXiv:1411.7631v2 [cs.DS] 17 Nov 2015
Approximate Undirected Maximum Flows in
O(mpolylog(n)) Time
Richard Peng
Georgia Tech
[email protected]
November 18, 2015
Abstract
We give the first O(mpolylog(n)) time algorithms for approximating maximum
flows in undirected graphs and constructing polylog(n)-quality cut-approximating
hierarchical tree decompositions. Our algorithm invokes existing algorithms for
these two problems recursively while gradually incorporating size reductions. These
size reductions are in turn obtained via ultra-sparsifiers, which are key tools in
solvers for symmetric diagonally dominant (SDD) linear systems.
1
Introduction
The problem of finding maximum flows and minimum cuts has been studied extensively
in algorithmic graph theory and combinatorial optimization. It led to important tools
in algorithm design such as augmenting paths [FJF56], blocking flows [Din70, EK72],
dynamic trees [GN80, ST83], dual algorithms [GT86], scaling algorithms [GR98], graph
sparsification [BK96], and electrical flows [CKM+ 11, Ma̧d13]. In its simplest form, the
maximum flow problem asks to route the most flow from a source to a sink while obeying
edge capacities. Its dual, the minimum cut problem, asks for the minimum capacity of
edges whose removal disconnects the sink from the source.
Approximating maximum flows in undirected graphs has received much attention recently due to its tighter interactions with randomized and numerical tools [CKM+ 11,
LRS13, She13a, KLOS14]. Algorithms for this variant have applications in graph partitioning [KRV09, OSVV08, She09], image processing [CMMP13], and as we will describe,
the construction of oblivious routing schemes [RST14].
Recently, algorithms that approximate undirected maximum flows in O(m1+o(1) ǫ−2 )
time were given by Sherman [She13a] and Kelner et. al. [KLOS14]. At the core of these
algorithms are congestion-approximators [Ma̧d10a] and oblivious routing schemes respectively [KLOS14]. Congestion-approximators can be viewed as a small set of representative
1
cuts in the graph, and oblivious routing schemes are more powerful in that they preserve
flows as well as cuts. The runtime of these algorithms stems from both the quality of these
approximators as well as the cost of constructing them. A natural question stemming from
them is to further improve this running time.
Oblivious routing schemes are of independent interest in the study of graph partitioning and routing. Schemes with quality polylog(n) were shown to exist by Räcke [Räc02],
and invoking them would lead to a better running time of mpolylog(n) after preprocessing. However, finding these schemes requires solving an intricate sequence of ratio cut
problems [HHR03, BKR03].
The current best algorithms for approximating ratio cuts are based on invoking (approximate) maximum flows [KRV09, OSVV08, She09]. Following the break-through on
approximate maximum flows, Räcke et al. [RST14] gave a more efficient algorithm for constructing oblivious routing schemes. This result can be viewed as producing a polylog(n)quality oblivious routing scheme by computing maximum flows on graphs of total size
O(mpolylog(n)). This leads to a chick-and-egg situation when approximators and maximum flow algorithms are viewed as black boxes: either gives the other via an overhead of
polylog(n), but to get the calls started we need to invoke routines that run in O(m1+o(1) )
time and produce mo(1) -approximations.
In this paper, we complete this cycle of algorithmic invocations by resolving this
chicken-and-egg situation, leading to improved algorithms to all intermediate problems.
The key observation is that the oblivious routing schemes produced by the Räcke et al.
algorithm have fixed size: producing them via recursive calls does not affect the cost of
invoking them, and any error introduced in the recursion will only show up as a slightly
larger overhead on this fixed size. The main steps of our algorithm on a graph G are:
1. Produce a graph H with size m/polylog(n) that can polylog(n)-approximate G.
2. Construct an approximator for H using the Räcke et. al. algorithm, making more
recursive maximum flow calls.
3. Convert this scheme to one for G, and use it to solve approximate maximum flows.
The size reduction allows us to bound the total size of the maximum flow instances
computed recursively by at most m/2, giving a total size bound of O(m). As H polylog(n)approximates G, the approximator for H returned by the recursive calls is still a polylog(n)quality approximator for G. The fact that its size is O(n) then allows us to bound the
overall cost by O(mpolylog(n)).
This recursive scheme allows us to bypass the more expensive approximators used to
initiate this sequence of algorithmic calls. The total cost in turn reduces from O(m1+o(1) )
to O(mpolylog(n)). Furthermore, these size reductions can be directly obtained via ultrasparsifiers from solvers for linear systems in graph Laplacians [ST14]. This results in
a short pseudocode when the pieces are viewed as black-boxes. Our algorithm is also
analogous to iterative schemes for computing row samples of matrices [LMP13]: the
2
congestion-approximator plays a similar role to the small row sample, and the call structure is analogous to what we use here, with ultra-sparsifiers being the size reductions.
We will introduce the algorithmic tools that we invoke in Section 2, and describe
our algorithm in Section 3. For simplicity, we will limit our presentation to the cut
setting and utilize the oblivious routing schemes as congestion-approximators. As the
oblivious routing construction by Räcke et al. [RST14] also produces embeddings, and size
reductions similar to ultra-sparsifiers were used in the flow based algorithm by Kelner et
al., we believe this scheme can be extended to the flow setting as well.
We will also not optimize for the exponent in log n because further runtime improvements based on this approach are likely. However, major obstacles remain in obtaining
running times of m log5 n or faster:
1. Random sampling based ultra-sparsifiers incur an overhead of log2 n in error.
2. Current oblivious routing constructions are based on top-down divide-and-conquer
with log n levels, each making a sequence of log n maximum flow calls through
rebalancings [RST14]. At present these routines also incur several additional log
factors due to error accumulations over levels of recursion.
3. Oblivious routing schemes incur a distortion of at least log n [Räc08].
4. Producing balanced cuts using maximum flows requires log n maximum flow invocation [OSVV08, She09].
5. The invocation of congestion-approximators to produce approximate maximum flows
requires an iteration count that’s at least quadratic in the distortion, as well as incurring another log n factor overhead. [She13a, KLOS14]. 1
Directly combining these estimates leads to a total cost of about m log11 n. An optimistic
view is that the algorithms using congestion-approximators can depend linearly on the
distortion, and reusing maximum flow calls across the construction scheme leads to recursion on graphs with total size m log n. Even in this case, the overall cost is still about
m log5 n. Therefore, we believe obtaining a running time of O(m log3 n) will require significant improvements to both algorithms that construct oblivious routings and iterative
methods that utilize them.
2
Background
Our presentation follows the notations from [She13a] and [KLOS14]. A flow f meets
demands b if for all vertices v, the total amount of flow enter/leaving v is b v . For edge
capacities u, the congestion of f is the maximum of |f e /u e | over all edges. By a standard
1
We cite the lower cost bounds from [KLOS14] here under the belief that these routines have inherent
connections.
3
reduction via binary search (e.g. Sections 2.2 and 3.1 of [CKM+ 11]), we can focus on the
decision version. For a fixed demand, the problem asks to either route it with congestion
at most 1 + ǫ, or certify via a cut that it cannot be routed with congestion less than 1.
A cut is defined by a subset of vertices S: its demand, b(S), is the total demand of
vertices in S, and its capacity, u(S), is the total capacity of edges leaving S. The ratio
between demand and capacity is a lower bound for the minimum congestion, and the
maxflow-mincut theorem states that the minimum congestion needed to route a demand
is in fact equal to the maximum demand/capacity ratio over all cuts S.
The connections between flows, cuts, congestion, and demand brings us to the notion of
(1 + ǫ)-approximate flow/cut solutions, which will be our standard notion of approximate
solutions. For a demand b and an error ǫ, such a pair consists of a flow and cut whose
congestion and demand/capacity value are within a factor of 1 + ǫ of each other.
We will make extensive use of approximations, and denote them using the ≈κ notation.
For two scalar quantities, x and y, we use x ≈κ y to mean that there exist parameters
γmin and γmax ≤ γmin κ such that γminx ≤ y ≤ γmax x.
2.1
Congestion Approximators
We will use an algorithm by Sherman [She13a] on using congestion-approximators to
compute approximate maximum flows.
Definition 2.1 (Definition 1.1. in [She13b]). An α-congestion-approximator of G is a
matrix R such that for any demand vector b,
kRbk∞ ≈α opt(b)
where opt(b) is the minimum congestion required to route the demands b in G.
Theorem 2.2 (Theorem 1.2. from [She13b]). There is a routine ApproximatorMaxFlow that, given demands b and access to an α-congestion-approximator R, makes
O(α2 log2 nǫ−3 ) iterations and returns an (1 + ǫ)-approximate flow/cut solution for these
demands. Each iteration takes O(m) time, plus computing matrix-vector products involving R and RT .
Räcke et. al. [RST14] showed that these congestion approximators can be efficiently
computed using approximate maximum flow routines. This result can be pharaphrased
as:
Theorem 2.3 (main result of [RST14]). There is a routine CongestionApproximator
that takes a graph G, returns with high probability an O(log4 n)-congestion-approximator
R such that matrix-vector products in R and RT can be performed in O(n) time. Furthermore, this approximator is computed via a series of approximate flow/cut solutions
with error 1/Θ(log3 n) on graphs of sizes m1 , . . . mN such that
N
X
i=1
mi ≤ O(m log4 n),
4
plus an additional running time overhead of O(m log6 n).
This summarizes several aspects of the algorithm for constructing oblivious routings
by Räcke et. al. [RST14]: the fact that the oblivious routing scheme produced gives
a congestion-approximator was observed in the second paragraph of the abstract. The
approximation guarantee is from Theorem 4.1. The invocation costs of R and RT also
follow from oblivious routing scheme being a tree.
The error tolerance in the maximum flow calls of 1/Θ(log3 n) is stated in the abstract
and utilized in the rebalancing step of the proof of Lemma 3.1. Overall the algorithm
performs O(log n) levels of partition based recursion, and the total sizes of graphs at each
level is O(m). Each partition step may adjust the partition O(log n) times using the cutmatching game by Khandekar et. al. [KRV09], which in turn needs O(log2 n) approximate
flow/cut solutions. Combining these bounds gives a total size of O(m log4 n). The running
time overhead comes from applying the O(log2 n) matchings produced in the cut-matching
game to a random vector before routing it using approximate maximum flows.
2.2
Ultra-Sparsifiers and Size Reductions
Ultra-sparsifiers are controlled ways of reducing graphs to tree-like structures. As they
involve pairs of graphs on the same vertex set, we will use scripts to denote the graph
in question in our notations. The following construction can be obtained from [KMP14]
and [AN12].
Theorem 2.4. There is a routine Ultra-Sparsify that takes a graph G = (V, EG , u G )
with n vertices and m edges, and any parameter κ > 1, returns in O(m log n log log n) time
a graph H = (V, EH , u H ) on the same set of vertices with n − 1 + O(m log2 n log log n/κ)
edges such that with high probability we have
u G (S) ≈κ u H (S)
for all subsets of vertices S ⊆ V .
Since minimum cut seeks to minimize u(S)/b(S), u G (S) ≈kappa u H (S) for all S
implies optG (b) ≈κ optH (S), and an α-congestion-approximator for H is also a καcongestion-approximator for G. Note that since we only need to preserve cuts, the
Spielman-Teng construction of ultra-sparsifiers [ST14] with spectral sparsifiers replaced
by cut-sparsifiers [BK96] also gives a similar bound.
These edge reductions are complemented by vertex reductions, which also are crucial
in algorithms using ultra-sparsifiers [ST14, Ma̧d10a, KMP14, Pen13, She13a, KLOS14].
Lemma 2.5 (Lemma 5.8 of [Ma̧d10b], paraphrased). When given a graph H with n
vertices and m = n − 1 + m′ edges, we can produce a graph H ′ = Reduce(H) with O(m′ )
edges such that any α-congestion-approximator RH ′ for H ′ can be converted into an O(α)congestion-approximator R H for H. Furthermore, matrix-vector products involving RH
or RTH can be computed by performing a single matrix-vector product in RH ′ or RTH ′
respectively plus an overhead of O(m).
5
Combining these two steps for edge and vertex reductions gives our key size reduction
routines:
Corollary 2.6. There are routines UltraSparsifyAndReduce and Convert so that
when given a graph G with n vertices and m edges, and an approximation factor κ,
UltraSparsifyAndReduce(G, κ) produces G′ such that with high probability.
1. G′ has at most O(m log2 n log log n/κ) edges, and
2. given an α-congestion-approximator R G′ for G′ , RG = Convert(G, G′ , R′ )
(a) is an O(κα)-congestion-approximator for G, and
(b) matrix-vector products involving R G or RTG can each be performed using one
matrix-vector product involving RG′ or R TG′ plus an overhead of O(m).
3
Recursive Algorithm
Our algorithm recursively calls the two routines for utilizing and constructing congestionapproximators, while reducing sizes using UltraSparsifyAndReduce. Its pseudocode
is given in Figure 1
f = RecursiveApproxMaxFlow(G, ǫ, b)
1. Set κ ← C log6 n log log n for some absolute constant C.
2. G′ ← UltraSparsifyAndReduce(G, κ)
3. RG′ ← CongestionApproximator(G′ ), which in turm makes recursive calls to
RecursiveApproxMaxFlow.
4. RG ← Convert(G, G′ , RG′ )
5. Return ApproximatorMaxFlow(G, RG , ǫ) .
Figure 1: Recursive Algorithmm for Approximate Maximum Flow
We will simplify the analysis by bounding the size reductions at each recursive call
and the overall failure probability of any call.
Lemma 3.1. We have |EG′ | ≤ O(|EG |/(C log4 n)) during each recursive call, and with
high probability all the function invocations terminate correctly,
6
Proof. Corollary 2.6 gives that the size of G′ is at most O(m log2 n log log n/κ), and the
bound follows from the choice of κ = C log6 n log log n.
We can then follow the call structure of this algorithm and accumulate failure probabilities. At each step, Theorem 2.3 gives that the total size of the graphs recursed on
is bounded by O(|EG′ | log4 n) = O(|EG |/C). Therefore, a sufficiently large C means the
total number of recursive calls is bounded by O(m) with high probability. Accumulating
the failure probabilities over these steps using the union bound then gives the overall
success probability.
We remark that to bound the failure probability of each recursive call by 1 − nc , the
routines also need to use n̄, the initial vertex count, instead of the size of the current
instance. This is a situation that occur frequently in analyses of recursive invocations of
Monte-Carlo randomized algorithms [ST14, Pen13, CKM+ 14]. We omit the details here
due to the large number of routines used in black-box manners.
For the rest of this proof we will assume that all black-box invocations terminate
correctly.
Lemma 3.2. On input of a graph with size m, with high probability the cost of the final
call to ApproximatorMaxFlow is at most O(m log32 n log2 log nǫ−3 ).
Proof. The guarantees of CongestionApproximator from Theorem 2.3 gives that
RG′ is an O(log4 n)-congestion-approximator for G′ , and matrix-vector products involving RG′ and RG′ cost O(n). Combining this with Corollary 2.6 gives that RG as returned
by Convert on Line 4 of RecursiveApproxMaxFlow is an O(log10 n log log n)congestion-approximator for G, and the cost of matrix-vector products RG and RTG is
O(m). The running time and error guarantees then follow from Theorem 2.2.
Theorem 3.3. With high probability, RecursiveApproxMaxFlow returns an (1 + ǫ)approximate flow/cut solution in time O(m log32 n log2 log n max{log9 n, ǫ−3 }).
Proof. Theorem 2.3 guarantees that the error parameter in all intermediate calls is at
most ǫ = 1/Θ(log3 n). Let the running time of RecursiveApproxMaxFlow on a
graph with m edges and error ǫ be T (m). We will show by induction, or guess-and-check,
that we can choose C to ensure T (m) ≤ Cm log41 n log2 log n.
The base case of m ≤ log10 n̄, where n̄ is the top level vertex count, follows from
invoking existing approximate maximum flow algorithms [CKM+ 11]. For the inductive
case, let the graphs that we recurse on have sizes m1 . . . mN . Theorem 2.3 and Lemma 3.2
give the following recurrence:
T (m) ≤
N
X
i=1
T (mi ) + O(m log6 n)
+O(m log32 n log2 log nǫ−3 )
≤
N
X
i=1
T (mi ) + O(m log41 n log2 log n).
7
Since all graphs that we compute approximate maximum flows on are subgraphs of
G , Lemma 3.1 allows us to invoke the inductive hypothesis, giving
′
T (m) ≤
N
X
Cmi log41 n log2 log n + O(m log41 n log2 log n).
i=1
The total sizes of the graphs that we recurse on can in turn be bounded via Theorem 2.3
and Corollary 2.6.
N
X
i=1
mi ≤ O(log4 n|EG′ |) ≤ O(m log6 n log2 log n/κ).
Choosing C appropriately then allows us to bound this by m/2, giving a total of
T (m) ≤
C
m log41 n log2 log n + O(m log41 n log2 log n).
2
The inductive hypothesis then follows by picking C to be twice the constant of the trailing
term.
This gives the bound of O(m log41 n log2 log n) when ǫ is set to 1/Θ(log3 n). For the
general case of arbitrary ǫ, Lemma 3.2 gives a bound of O(m log32 n log2 log nǫ−3 ). Note
that the first term is still present, since the second to last call is made on a graph of
size m/2 with ǫ = 1/Θ(log3 n). Summing over both terms then gives the overall runtime
bound.
We remark that the ǫ−3 term arises from a similar dependency in the congestionapproximator based flow routine by Sherman [She13a], which is stated in Theorem 2.2.
Invoking this algorithm in Theorem 2.3 also gives an O(mpolylog(n)) time algorithm
for constructing hierarchical tree decomposition based oblivious routing schemes.
Corollary 3.4. Given an undirected graph G, we can construct in O(m log45 n log2 log n)
time a tree that with high probability corresponds to an O(log4 n)-competitive oblivious
routing scheme .
References
[AN12]
Ittai Abraham and Ofer Neiman. Using petal decompositions to build a low
stretch spanning tree. In Proceedings of the 44th symposium on Theory of
Computing, STOC ’12, pages 395–406, 2012.
[BK96]
András A. Benczúr and David R. Karger. Approximating s-t minimum cuts
in Õ(n2 ) time. In Proceedings of the 28th annual ACM symposium on Theory
of computing, STOC ’96, pages 47–55, 1996.
8
[BKR03]
Marcin Bienkowski, Miroslaw Korzeniowski, and Harald Räcke. A practical
algorithm for constructing oblivious routing schemes. In Proceedings of the
15th annual ACM symposium on Parallel algorithms and architectures, SPAA
’03, pages 24–33, 2003.
[CKM+ 11] Paul Christiano, Jonathan A. Kelner, Aleksander Ma̧dry, Daniel A. Spielman,
and Shang-Hua Teng. Electrical flows, Laplacian systems, and faster approximation of maximum flow in undirected graphs. In Proceedings of the 43rd
annual ACM symposium on Theory of computing, STOC ’11, pages 273–282,
2011.
[CKM+ 14] Michael B. Cohen, Rasmus Kyng, Gary L. Miller, Jakub W. Pachocki,
Richard Peng, Anup B. Rao, and Shen Chen Xu. Solving SDD linear systems
in nearly mlog 1/2 n time. In Proceedings of the 46th Annual ACM Symposium
on Theory of Computing, STOC ’14, pages 343–352, 2014.
[CMMP13] Hui Han Chin, Aleksander Ma̧dry, Gary L. Miller, and Richard Peng. Runtime
guarantees for regression problems. In Proceedings of the 4th conference on
Innovations in Theoretical Computer Science, ITCS ’13, pages 269–282, 2013.
[Din70]
Yefim A. Dinitz. Algorithm for Solution of a Problem of Maximum Flow in a
Network with Power Estimation. Soviet Math Doklady, 11:1277–1280, 1970.
[EK72]
Jack Edmonds and Richard M. Karp. Theoretical improvements in algorithmic efficiency for network flow problems. Journal of the ACM, 19(2):248–264,
April 1972.
[FJF56]
Lester R. Ford Jr and Delbert R. Fulkerson. Maximal flow through a network.
Canadian Journal of Mathematics, 8:399–404, 1956.
[GN80]
Zvi Galil and Amnon Naamad. An O(EV log2 V ) algorithm for the maximal
flow problem. Journal of Computer and System Sciences, 21(2):203–217, 1980.
[GR98]
Andrew V. Goldberg and Satish Rao. Beyond the flow decomposition barrier.
Journal of the ACM, 45:783–797, September 1998.
[GT86]
Andrew V. Goldberg and Robert E. Tarjan. A new approach to the maximum
flow problem. In Proceedings of the 18th annual ACM symposium on Theory
of computing, STOC ’86, pages 136–146, 1986.
[HHR03]
Chris Harrelson, Kirsten Hildrum, and Satish Rao. A polynomial-time tree
decomposition to minimize congestion. In Proceedings of the 15th annual ACM
symposium on Parallel algorithms and architectures, SPAA ‘03, pages 34–43,
2003.
9
[KLOS14] Jonathan A. Kelner, Yin Tat Lee, Lorenzo Orecchia, and Aaron Sidford. An
almost-linear-time algorithm for approximate max flow in undirected graphs,
and its multicommodity generalizations. In Proceedings of the 25th Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA ‘14, pages 217–226,
2014.
[KMP14]
Ioannis Koutis, Gary L. Miller, and Richard Peng. Approaching optimality
for solving SDD linear systems. SIAM Journal on Computing, 43(1):337–354,
2014.
[KRV09]
Rohit Khandekar, Satish Rao, and Umesh Vazirani. Graph partitioning using
single commodity flows. Journal of the ACM, 56(4):19, 2009.
[LMP13]
Mu Li, Gary L. Miller, and Richard Peng. Iterative row sampling. In 54th
Annual IEEE Symposium on Foundations of Computer Science, FOCS ‘13,
pages 127–136, 2013.
[LRS13]
Yin Tat Lee, Satish Rao, and Nikhil Srivastava. A new approach to computing
maximum flows using electrical flows. In Proceedings of the 45th annual ACM
symposium on theory of computing, STOC ’13, pages 755–764, 2013.
[Ma̧d10a]
Aleksander Ma̧dry. Fast approximation algorithms for cut-based problems
in undirected graphs. In 51th Annual IEEE Symposium on Foundations of
Computer Science, FOCS ‘10, pages 245–254, 2010.
[Ma̧d10b]
Aleksander Ma̧dry. Fast approximation algorithms for cut-based problems in
undirected graphs. CoRR, abs/1008.1975, 2010.
[Ma̧d13]
Aleksander Ma̧dry. Navigating central path with electrical flows: From flows
to matchings, and back. In 54th Annual IEEE Symposium on Foundations of
Computer Science, FOCS ‘13, pages 253–262, 2013.
[OSVV08] Lorenzo Orecchia, Leonard J. Schulman, Umesh V. Vazirani, and Nisheeth K.
Vishnoi. On partitioning graphs via single commodity flows. In Proceedings of
the 40th annual ACM symposium on Theory of computing, STOC ’08, pages
461–470, 2008.
[Pen13]
Richard Peng. Algorithm Design Using Spectral Graph Theory. PhD thesis,
Carnegie Mellon University, August 2013. CMU CS Tech Report CMU-CS13-121.
[Räc02]
Harald Räcke. Minimizing congestion in general networks. In Proceedings of
the 43rd Symposium on Foundations of Computer Science, FOCS ‘02, pages
43–52, 2002.
10
[Räc08]
Harald Räcke. Optimal hierarchical decompositions for congestion minimization in networks. In Proceedings of the 40th annual ACM symposium on
Theory of computing, STOC ’08, pages 255–264, 2008.
[RST14]
Harald Räcke, Chintan Shah, and Hanjo Täubig. Computing cut-based hierarchical decompositions in almost linear time. In Proceedings of the 25th
Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ‘14, pages
227–238, 2014.
√
Jonah Sherman. Breaking the multicommodity flow barrier for O( logn)approximations to sparsest cut. In 50th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’09, pages 363–372, 2009.
[She09]
[She13a]
Jonah Sherman. Nearly maximum flows in nearly linear time. In 54th Annual
IEEE Symposium on Foundations of Computer Science, FOCS ‘13, pages
263–269, 2013.
[She13b]
Jonah Sherman. Nearly maximum flows in nearly linear time.
abs/1304.2077, 2013.
[ST83]
Daniel D Sleator and Robert Endre Tarjan. A data structure for dynamic
trees. Journal of Computer and System Sciences, 26(3):362–391, 1983.
[ST14]
Daniel A. Spielman and Shang-Hua Teng. Nearly linear time algorithms for
preconditioning and solving symmetric, diagonally dominant linear systems.
SIAM Journal on Matrix Analysis and Applications, 35(3):835–885, 2014.
11
CoRR,
| 8cs.DS
|
Deterministic Time-Space Tradeoffs for k-SUM
Andrea Lincoln∗1 , Virginia Vassilevska Williams†2 , Joshua R.
Wang‡3 , and R. Ryan Williams§4
1
2
3
arXiv:1605.07285v1 [cs.DS] 24 May 2016
4
Computer Science Department,
[email protected]
Computer Science Department,
[email protected]
Computer Science Department,
[email protected]
Computer Science Department,
[email protected]
Stanford University, USA
Stanford University, USA
Stanford University, USA
Stanford University, USA
Abstract
Given a set of numbers, the k-SUM problem asks for a subset of k numbers that sums to zero.
When the numbers are integers, the time and space complexity of k-SUM is generally studied in
the word-RAM model; when the numbers are reals, the complexity is studied in the real-RAM
model, and space is measured by the number of reals held in memory at any point.
We present a time and space efficient deterministic self-reduction for the k-SUM problem
which holds for both models, and has many interesting consequences. To illustrate:
q
n lg(n)
3-SUM is in deterministic time O(n2 lg lg(n)/ lg(n)) and space O
lg lg(n) . In general, any
polylogarithmic-time improvement over quadratic time for 3-SUM can be converted into an
algorithm with an identical time improvement but low space complexity as well.
√
3-SUM is in deterministic time O(n2 ) and space O( n), derandomizing an algorithm of Wang.
A popular conjecture states that 3-SUM requires n2−o(1) time on the word-RAM. We show
that the 3-SUM Conjecture is in fact equivalent to the (seemingly weaker) conjecture that
every O(n.51 )-space algorithm for 3-SUM requires at least n2−o(1) time on the word-RAM.
√
For k ≥ 4, k-SUM is in deterministic O(nk−2+2/k ) time and O( n) space.
1998 ACM Subject Classification F.2.1 Numerical Algorithms and Problems
Keywords and phrases 3SUM; kSUM; time-space tradeoff; algorithm.
Digital Object Identifier 10.4230/LIPIcs.ICALP.2016.XXX
∗
†
‡
§
Supported by a Stanford Graduate Fellowship.
Supported by NSF Grants CCF-1417238, CCF-1528078 and CCF-1514339, and BSF Grant
BSF:2012338.
Supported by a Stanford Graduate Fellowship.
Supported in part by NSF CCF-1552651. Any opinions, findings and conclusions or recommendations
expressed in this material are those of the authors and do not necessarily reflect the views of the
National Science Foundation.
© Andrea Lincoln, Virginia Vassilevska Williams, Joshua R. Wang and R. Ryan Williams;
licensed under Creative Commons License CC-BY
43rd International Colloquium on Automata, Languages, and Programming (ICALP 2016).
Editors: Ioannis Chatzigiannakis, Michael Mitzenmacher, Yuval Rabani, and Davide Sangiorgi; Article No. XXX;
pp. XXX:1–XXX:16
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XXX:2
Deterministic Time-Space Tradeoffs for k-SUM
1
Introduction
We consider the k-SUM problem: given a list S of n values, determine whether there are
Pk
distinct a1 , . . . , ak ∈ S such that
i=1 ai = 0. This classic problem is a parameterized
version of the Subset Sum problem, which is among Karp’s original NP-Complete problems.1
The brute-force algorithm for k-SUM runs in O(nk ) time, and it is known [22] that an
no(k) time algorithm (where the little-o depends on k) would violate the Exponential Time
Hypothesis [18]. A faster meet-in-the-middle algorithm reduces the k-SUM problem on n
numbers to 2-SUM on O(ndk/2e ) numbers, which can then be solved by sorting and binary
search in O(ndk/2e log n) time. The belief that this meet-in-the-middle approach is essentially
time-optimal is at the heart of many conditional 3-SUM-hardness results in computational
geometry (e.g. [15]) and string matching (e.g. [5, 2]).
The space usage of the meet-in-the-middle approach is prohibitive: the O(n log n) time
solution for 2-SUM uses linear space, which causes the fast k-SUM algorithm to need
Ω(ndk/2e ) space. However, the brute-force algorithm needs only O(k log n) space. This
leads to the natural question: how well can one trade off time and space in solving k-SUM?
Schroeppel and Shamir [23] first studied time-space tradeoff algorithms for Subset Sum.
They showed how to reduce Subset Sum to an instance of k-SUM for any k ≥ 2: split the
elements into k sets of n/k elements each; for each set, compute 2n/k sums corresponding
to the subsets of the set; this forms a k-SUM instance of size 2n/k . Since the k-SUM
instance does not have to be explicitly stored, any time T (N ), space S(N ) algorithm for
k−SUM immediately implies a time T (2n/k ), space S(2n/k ) algorithm for Subset Sum.
Furthermore, Schroeppel and Shamir gave a deterministic Õ(n2 ) time, Õ(n) space algorithm
for 4-SUM, implying a O∗ (2n/2 ) time, O∗ (2n/4 ) space algorithm for Subset Sum.2 They
also generalized the algorithm to provide a smooth time-space tradeoff curve, with extremal
points at O∗ (2n/2 ) time, O∗ (2n/4 ) space and O∗ (2n ) time, O∗ (1) space.
A recent line of work leading up to Austrin et al. [6] has improved this long-standing
tradeoff curve for Subset Sum via randomized algorithms, resulting in a more complex curve.
Wang [25] moved these gains to the k-SUM setting. In particular, for 3-SUM he obtains an
√
Õ(n2 ) time, Õ( n) space Las Vegas algorithm.
Despite the recent progress on the problem, all of the improved algorithms for the general
case of k-SUM have heavily relied on randomization, either utilizing hashes or random prime
moduli. These improvements also all rely heavily on the values in the lists being integers.
For the general case of k-SUM, the previous best deterministic k-SUM results (even for
integer inputs) are the brute-force algorithm, the meet-in-the-middle algorithm, and the
Schroeppel and Shamir 4-SUM algorithm, and simple combinations thereof.
1.1
Our Results
We consider new ways of trading time and space in solving k-SUM, on both integer and real
inputs (on the word-RAM and real-RAM respectively), without the use of randomization.
Our improvements for k-SUM naturally extend to improvements to Subset Sum as well.
Our main result is a deterministic self-reduction for k-SUM. Informally, we show how to
deterministically decompose a list of n numbers into a small collection of shorter lists, such
1
2
Karp’s definition of the Knapsack problem is essentially Subset Sum [19].
The notation Õ suppresses polylogarithmic factors in n, and O∗ suppresses polynomial factors in n.
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
XXX:3
that the k-SUM solution is preserved. This result is shown for k = 3 in Section 4. It is
shown for general k in Section 5.
I Theorem 1. Let g be any integer between 1 and n. k-SUM on n numbers can be reduced
to O(kg k−1 ) instances of k-SUM on n/g numbers. The reduction uses O(ng k−1 ) additional
time and O(n/g) additional words of space.
Theorem 1 has several interesting applications. First, it leads to more efficient k-SUM
algorithms. For example, Gold and Sharir, building on other recent advances, report a
deterministic algorithm for 3-SUM that works in both the word-RAM and real-RAM models
and which runs in time O(n2 lg lg(n)/ lg(n)) [16]. However, this algorithm uses a considerable
amount of space to store a table of permutations. Applying Theorem 1 in multiple ways and
calling their algorithm, we recover the same asymptotic running time but with drastically
better space usage:
I Theorem 2. There
is an O(n2 lg lg(n)/ lg(n)) time deterministic algorithm for 3-SUM
q
n lg(n)
that stores at O( lg lg(n) ) numbers in memory at point. (An analogous statement holds for
3-SUM over the integers.)
Theorem 1 also directly leads to a derandomization of Wang’s space-efficient algorithm
for 3-SUM:
I Theorem 3. For all s ∈ [0, 1/2] there is a deterministic time O(n3−2s ), algorithm which
uses O(ns ) words of space for 3-SUM.
From Theorem 1 we can also derive a more space-efficient algorithm for 4-SUM, and lift
it to a new algorithm for k-SUM:
I Theorem 4. For k ≥ 4, k-SUM is solvable in deterministic O(nk−2+2/(k−3) ) time and
√
O( n) space in terms of words.
A more plausible 3-SUM conjecture.
A rather popular algorithmic conjecture is the 3-SUM Conjecture that 3-SUM on n integers
requires n2−o(1) time on a word-RAM with O(log n) bit words. This conjecture has been
used to derive conditional lower bounds for a variety of problems [15, 5, 2, 20, 3], and appears
to be central to our understanding of lower bounds in low-polynomial time. To refute the
conjecture, one could conceivably construct an algorithm that runs in O(n1.99 ) time, but
utilizes Ω(n1.99 ) space in some clever way. Here we consider a seemingly weaker (and thus
more plausible) conjecture:
I Conjecture 5 (The Small-Space 3-SUM Conjecture). On a word-RAM with O(log n)-bit
words, there exists an > 0 such that every algorithm that solves 3-SUM in O(n1/2+ ) space
must take at least n2−o(1) time.
This conjecture looks weaker than the original 3-SUM Conjecture, because we only have
√
to prove a quadratic-time lower bound for all algorithms that use slightly more than n
space. Proving time lower bounds is generally much easier when space is severely restricted
(e.g. [9, 14, 12, 26, 8]).
Our self-reduction for 3-SUM yields the intriguing consequence that the original 3-SUM
Conjecture is equivalent to the Small-Space 3-SUM conjecture! That is, the non-existence
of a truly subquadratic-time 3-SUM algorithm is equivalent to the non-existence of a truly
subquadratic-time n0.51 -space 3-SUM algorithm, even though the latter appears to be a
more plausible lower bound. We prove:
ICALP 2016
XXX:4
Deterministic Time-Space Tradeoffs for k-SUM
I Theorem 6. If 3-SUM is solvable in time O(n2− ) time, then for every α > 0 there is a
δ > 0 such that 3-SUM is solvable in O(n2−δ ) time and space O(n1/2+α ) in terms of words.
Theorem 6 is interesting, regardless of the veracity of the 3-SUM conjecture. On the
one hand, the theorem reduces the difficulty of proving the 3-SUM Conjecture if it is true,
because we only have to rule out small-space sub-quadratic time algorithms. On the other
hand, the theorem means that refuting the 3-SUM conjecture immediately implies a trulysubquadratic time algorithm for 3-SUM using small space as well, which would be an algorithmic improvement.
2
Preliminaries
2.1
k-SUM and Selection
We will use the following version of the k-SUM problem:
I Definition 7. In the k-SUM problem, we are given an unsorted list L of n values (over Z
Pk
or R) and want to determine if there are a1 , . . . , ak ∈ L such that i=1 ai = 0.
One fundamental case is the 3-SUM problem. Sometimes 3-SUM is presented with three
separate lists, which we denote as 3-SUM’, but the two are reducible to each other in linear
time, and with no impact on space usage.
I Definition 8. In the 3-SUM problem, we are given an unsorted list L of n values and want
to know if there are a, b, c ∈ L such that a + b + c = 0. In the 3-SUM’ problem, we are given
three unsorted lists A, B, and C of values, where |A| = |B| = |C| = n, and want to know if
there are a ∈ A, b ∈ B, c ∈ C such that a + b + c = 0.
As part of our k-SUM algorithms, the classical Selection Problem will also arise:
I Definition 9. In the s-Select problem, we are given an unsorted list L of n values and
a natural number s, and want to determine the sth smallest value in L.
2.2
Computational Model
As standard when discussing sub-linear space algorithms, the input is provided in read-only
memory, and the algorithm works with auxiliary read/write memory which counts towards
its space usage.
Computation on Integers. When the input values are integers, we work in the word-RAM
model of computation: the machine has a word size w, and we assume all input numbers
can be represented with w bits so that they fit in a word. Arithmetic operations (+, −, ∗)
and comparisons on two words are assumed to take O(1) time. Space is counted in terms of
the number of words used.
Computation on Reals. When the input values are real numbers, we work in a natural
real-RAM model of computation, which is often called the comparison-addition model (see,
for example, [21]). Here, the machine has access to registers that can store arbitrary real
numbers; addition of two numbers and comparisons on real numbers take O(1) time. Space
is measured in terms of the number of reals stored.
Time-Space Complexity Notation. We say that k-SUM is solvable in TISP(T (n), S(n))
if k-SUM on lists of length n can be solved by a single algorithm running in deterministic
O(T (n)) time and O(S(n)) space simultaneously on the real-RAM (and if the lists contain
integers, on the word-RAM).
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
2.3
XXX:5
Other Prior Work
Baran, Demaine and Patrascu [7] obtained randomized slightly subquadratic time algorithms
for Integer 3-SUM in the word-RAM. Grønlund and Pettie [17] studied 3-SUM over the reals,
presenting an O(n2 /(log n/ log log n)) time randomized algorithm, as well as a deterministic
algorithm running in O(n2 /(log n/ log log n)2/3 ) time. Recently, Gold and Sharir [16] improved this deterministic running time to O(n2 /(log n/ log log n)). Abboud, Lewi and Williams [1] showed that Integer k-SUM is W[1]-complete under randomized FPT reductions
(and under some plausible derandomization hypotheses, the reductions can be made deterministic). In the linear decision tree model of computation, k-SUM over the reals is known
to require Ω(ndk/2e ) depth k-linear decision trees [13, 4], but the problem can be solved
√
with O(nk/2 log n) depth (2k − 2)-linear decision trees [17]. The randomized decision tree
complexity was improved by Gold and Sharir [16] to O(nk/2 ).
3
Building Blocks
In this section, we describe two tools we use to obtain our main self-reduction lemma for
k-SUM and 3-SUM. The first tool helps us guarantee that we don’t have to generate too
many subproblems in our reduction; the second will allow us to find these subproblems in a
time and space efficient way.
3.1
Domination Lemma
Our deterministic self-reduction for k-SUM will split lists of size n into g sublists of size n/g,
then solve subproblems made up of k-tuples of these sublists. Naively, this would generate
g k subproblems to enumerate all k-tuples. In this section, we show that we only need to
consider O(kg k−1 ) subproblems.
First, we define a partial ordering on k-tuples on [n]k . For t, t0 ∈ [n]k , we say that t ≺ t0
if t[i] < t0 [i] for all i = 1, . . . , k. (Geometrically, the terminology is that t0 dominates t.)
I Lemma 10 (Domination Lemma). Suppose all tuples in a subset S ⊆ [n]k are incomparable
with respect to ≺. Then |S| ≤ knk−1 .
The Domination Lemma can be seen as an extension of a result in [24] (also used in [11]
in a different context) which covers the k = 3 case.
Proof. We will give a cover of all elements in [n]k with few chains under ≺. Then by
Dilworth’s theorem, any set of incomparable elements under ≺ can only have one element
from each chain.
Take any k-tuple t ∈ [n]k such that t[i] = 1 for some i = 1, . . . , k. Letting ` ∈ [n] be the
largest element in t, we define the chain C(t) = {t0 , t1 , . . . , tn−` }, where each tj is given by
tj [i] = t[i] + j for all i = 1, . . . , k. Clearly C(t) forms a chain in [n]k under ≺. Moreover
these chains cover all elements of [n]k : observe that the tuple t appears in the chain C(t0 )
where t0 [i] = t[i] − minj t[j] + 1 for all i = 1, . . . , k.
The number of chains is exactly the number of k-tuples with a 1 in at least one coordinate.
This number is less than k times the number of tuples that have a 1 in dimension i. The
number of tuples with a 1 in dimension i is nk−1 . Thus, the total number of chains is
≤ knk−1 .
J
The Domination Lemma can be applied to show that in any list of numbers, not too
many k-SUM subproblems can have k-SUM solutions. In the following, let g divide n for
ICALP 2016
XXX:6
Deterministic Time-Space Tradeoffs for k-SUM
y
3
2
z
3
1
2
1
0
1
2
3
x
Figure 1 Domination Lemma chains when n = 3 and k = 3. The chain {(1, 1, 1), (2, 2, 2), (3,
3, 3)} is highlighted in red. In three dimensions, the number of chains is roughly proportional to
the surface area of the cube, which is only O(n2 ), despite the fact that there are O(n3 ) points.
Figure 2 A depiction of how L is divided.
simplicity. Given a list L of n numbers divided into g groups of size n/g, a subproblem of L
is simply the union of a k-tuple of groups from L. Note that a subproblem contains at most
kn/g numbers.
I Corollary 11. Given a k-SUM instance L, suppose L is divided into g groups L1 , . . . , Lg
where |Li | = n/g for all i, and for all a ∈ Li and b ∈ Li+1 we have a ≤ b. Then there are
O(k · g k−1 ) subproblems L0 of L such that the smallest k-sum of L0 is less than zero and
the largest k-sum of L0 is greater than zero. Furthermore, if some subproblem of L has its
largest or smallest k-sum equal to 0, then the corresponding k-SUM solution can be found
in O(g k ) time.
Proof. We associate each subproblem of L with a corresponding k-tuple (x1 , . . . , xk ) ∈ [g]k
corresponding to the k sublists (Lx1 , . . . , Lxk ) of L.
Let m[i] be the element in position i · (n/g) when L is in sorted order. Consider any
Pk
Pk
subproblem with i=1 m[xi ] > 0 (smallest k-sum greater than zero) or i=1 m[xi + 1] < 0
(largest k-sum less than zero). We call such a subproblem trivial, since it cannot contain
k-SUM solutions.
Pk
In O(g k ) time, we can determine whether any subproblem has
i=1 m[xi ] = 0, and
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
XXX:7
return the corresponding k-SUM solution if this is the case. Otherwise, we can assume that
Pk
Pk
for each subproblem either it is trivial, or i=1 m[xi ] < 0 < i=1 m[xi + 1].
Consider the set of non-trivial subproblems. Because for all a ∈ Li and b ∈ Li+1 we
have a ≤ b, if for two subproblem k-tuples we have t ≺ t0 , then the smallest k-sum of the
subproblem t0 is at least the largest k-sum of the subproblem t. This implies that at least
one of the two subproblems must be trivial. In other words, the set of nontrivial problems
corresponds to a set of incomparable k-tuples in [g]k . Applying Lemma 10, the number of
nontrivial subproblems is O(kg k ).
J
3.2
Bucket Retrieval and Space-Efficient Selection
A randomized algorithm for k-SUM can partition a list of numbers by choosing a hash
function at random, then loop over the hash function range to partition a given list into
smaller buckets. Given a hash and a bucket number, it is easy to retrieve the contents of
that bucket by scanning the list.
To derandomize this process, we could try to create small “hash” buckets by grouping
the n/g smallest elements together, then the next n/g smallest elements, and so on, without
actually sorting the list. However, retrieving the contents of a bucket may now be difficult
to do with small space: we need to know the smallest and largest elements of a bucket to
retrieve its elements, and we may not be able to store all of these extrema. We require an
efficient algorithm to compute the largest element of a bucket, given the smallest element
and the bucket size.
This problem is equivalent to the selection problem, also known as s-Select, which asks
for the sth smallest element of a list, when we set s = n/g. To reduce from our problem to
s-Select, pretend that every entry less than our smallest element is ∞. (To reduce from
s-Select to our problem, we can pretend our smallest element is −∞.)
The classic median-of-median algorithm can solve s-Select in O(n) time and O(n)
space [10]. Since we care about space usage, we provide an algorithm below which has O(n)
running time, but uses much less space. This algorithm turns out to be optimal for our
purposes, since retrieving the bucket afterwards will already take O(n) time and O(s) space.
I Lemma 12. s-Select can be solved in O(n) time and O(s) space.
Proof. The plan is to scan through the elements of the list, inserting them to a data structure
D which will allow us to track the smallest s elements. We perform n insertions, then query
D to ask for the smallest s elements it contains. To get the claimed algorithm for selection,
we give a data structure can handle these operations in O(1) amortized update time and
O(s) query time, with a data structure using only O(s) space.
One first attempt might be to build a heap of s + 1 elements, which throws away the
largest element whenever it gets full. Since heaps have logarithmic update time and linear
space usage, this results in O(log s) update time, O(s) query time, and O(s) space.
We can improve the update time by batching when we throw out large elements. Suppose
instead we keep an array which can hold up to 2s elements. When the array gets full, we
throw out the largest s elements. To do this, we first compute the (s + 1)th smallest element
in the array. This can be done in O(s) time and O(s) space via the classical median-ofmedians algorithm. We then do a linear scan of the array, and write all elements strictly
less than the median to a new array. To handle ties, we write a copy of the median to the
new array, until it has s elements. When we are given our final query, we again throw out
large elements so that we only have s elements left, and then return those.
ICALP 2016
XXX:8
Deterministic Time-Space Tradeoffs for k-SUM
Updates now take amortized constant time: after s updates, we take O(s) time to clear
out the large elements. The final query takes O(s) time, since we again need to throw out
large elements. The space usage is O(s) since we store up to 2s elements, and running
median-of-medians takes O(s) space. This completes the proof.
J
We will call the above algorithm NextGroup. NextGroup takes as input a value
v, a natural number s, and a list of numbers L, and outputs the next s elements of L in
sorted order after the value v. Other variations on deterministic s-Select algorithms are
mentioned in Appendix A.
Subquadratic 3-SUM implies Subquadratic small-space 3-SUM
4
We will begin by using our building blocks to prove a self reduction for 3-SUM. Then we will
show three intriguing consequences of this self reduction. First, the self reduction can be
used to show a general theorem that takes subquadratic algorithms for 3-SUM and produces
√
subquadratic time algorithms that run in nearly n space. Second, we show that algorithms
for 3-SUM that are subquadratic by polylog factors can be used to obtain 3-SUM algorithms
with the same asymptotic running time and simultaneously small space. Finally, we will
prove that the Small-Space 3-SUM conjecture is equivalent to the 3-SUM conjecture.
4.1
3-SUM Self Reduction
We now proceed to solve 3-SUM using our bucket retrieval subroutine. We will use max S
and min S to refer to the maximum and minimum elements of a list S, respectively.
As anticipated, we split the three arrays into groups of size n/g, and solve 3-SUM on subproblems of this size. Naively there are O(g 3 ) subproblems to solve, but we use Corollary 11
to argue we only get O(g 2 ) subproblems.
I Theorem 13 (3-SUM Self-Reduction Theorem). If 3-SUM is solvable in TISP(T (n), S(n))
then for any g, 3-SUM can be solved in TISP(g 2 (n + T (n/g)), n/g + S(n/g)).
Proof. Consider the following algorithm.
Algorithm 1: 3-SUM Algorithm
Set preva = −∞;
for i ∈ [0, g − 1] do
Set A0 = NextGroup(A, preva, n/g + 1);
Set prevb = −∞;
for j ∈ [0, g − 1] do
Set B 0 = NextGroup(B, prevb, n/g + 1);
Set C 0 = NextGroup(C, − max A0 − max B 0 , n/g + 1);
while min C 0 ≤ − min A0 − min B 0 do
if 3-SUM(A0 , B 0 , C 0 ) returns true then
return true;
Set C 0 = NextGroup(C, max C 0 , n/g + 1);
Set prevb = max B 0 ;
Set preva = max A0 ;
return false;
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
XXX:9
Algorithm 1 is correct because we consider all possible elements of C where the sum of
elements from A0 and B 0 could land, and the choices of A0 and the choices of B 0 cover all
of A and B, respectively. If there are multiple copies of a value in a list we will fail to list
all copies only if it already appeared in a previous sublist. This will not affect correctness
because the value will have already been analyzed.
It’s easy to see that the algorithm calls NextGroup O(g) times for A0 , O(g 2 ) times for
0
B . We claim that we also only call it O(g 2 ) times for C 0 . To show this, we want to apply
Corollary 11. Unfortunately, the groups of C that we extract don’t always line up with our
ideal n/g division; since we start at − max A0 − max B 0 , we may not align at the endpoints
of blocks. Fortunately, we’ve only introduced an extra O(1) possibilities of C 0 for every
(A0 , B 0 ) pair, or O(g 2 ) extras total. Hence we still only make O(g 2 ) calls to NextGroup.
By Lemma 12, these calls will require O(ng 2 ) time and O(n/g) space.
Our algorithm also calls the TISP(T (n), S(n)) algorithm for 3-SUM O(g 2 ) times on
instances of size O(n/g), which requires O(g 2 T (n/g)) time and O(S(n/g)) space.
We have shown Algorithm 1 is correct and has the desired runtime and space usage, so
this completes the proof.
J
4.2
General Theorem for Space Reduction
Our self-reduction for 3-SUM yields the following intriguing consequence: subquadratic-time
algorithms for 3-SUM imply subquadratic-time small-space algorithms for 3-SUM. Plugging
this connection into known 3-SUM algorithms, we can automatically obtain more spaceefficient 3-SUM algorithms for free. From a complexity-theoretic point of view, the consequence is perhaps even more intriguing: it means that the 3-SUM Conjecture is equivalent
to the statement that there is no subquadratic-time n0.51 -space 3-SUM algorithm, even
though the latter appears to be a more plausible lower bound(!).
We begin by stating our generic space reduction theorem.
I Theorem 14 (3-SUM Space Reduction). Suppose 3-SUM is solvable in n2 /f (n) time, where
1 ≤ f (n) ≤ n. Then 3-SUM is solvable by an algorithm running in O(n2 /f (n/g)) time and
O(n/h) space simultaneously, where g(n), h(n) ∈ [1, n] satisfy the relations
r
n
n
n
2
and h +
≤O
.
g(n) ≥ Ω
f (n/g(n))
f (n/(hg(n/h)))
f (n/g(n))
Proof. We will apply our Self-Reduction Theorem for 3-SUM (Theorem 13) in two different
ways. First, we will use the self-reduction (and the constraint on g(n)) to convert our 3SUM algorithm into a linear-space algorithm, with a modest increase in running time (if at
all). Pushing the linear-space algorithm through the self-reduction once more will reduce
the space bound further, without increasing the running time asymptotically (using the
constraint on h(n)).
Let T (n) := n2 /f (n). Set the parameter g(n) ≥ 1 to satisfy
r
n2 /g 2
n
T (n/g) =
= O(n); or, equivalently g = Ω
.
(1)
f (n/g)
f (n/g)
Assuming g satisfies (1), applying the 3-SUM Self-Reduction (Theorem 13) with T (n) = S(n)
and g, we can then solve 3-SUM in
n2
TISP g 2 (n + T (n/g)), n/g + T (n/g) = TISP
,n .
(2)
f (n/g)
ICALP 2016
XXX:10 Deterministic Time-Space Tradeoffs for k-SUM
Now, set new time and space bounds T (n) := n2 /f (n/g(n)), S(n) = n from (2). Then,
applying the 3-SUM Self-Reduction (Theorem 13) with the new T (n), S(n) and some parameter h, we can then solve 3-SUM in TISP(h2 (n + T (n/h)), n/h + S(n/h)) =
n2 /h2
n2
2
TISP h n +
, n/h ⊆ TISP
, n/h ,
f (n/(hg(n/h)))
f (n/g)
by our hypothesis on h.
J
Space-Efficient Fast 3-SUM
4.3
When we apply Theorem 14 directly to known algorithms, we obtain immediate space improvements with negligible loss in running time. Very recently, Gold and Sharir [16] have
given a faster 3-SUM algorithm in the real-RAM model, building on the work of Gronlund
and Pettie [17]:
I Theorem 15 (Gold and Sharir [16]). 3-SUM can be solved in O(n2 lg lg(n)/ lg(n)) time
over the reals and integers.
As discussed in the introduction, their novel approach
uses quite a bit ofspace. Applying
p
n lg(n)/ lg lg(n) , with the same
Theorem 14, we can reduce the space usage to only O
asymptotic running time of Gold and Sharir.
q
lg(n)
n lg(n)
I Corollary 16 (Space-Efficient 3-SUM Algorithm). 3-SUM is in TISP n2 lglg(n)
, lg
lg(n) .
Proof. We shall apply Theorem 14. First, set f (n) := lg(n)/ lg lg(n), so that 3-SUM is
2
solvable in O(nq
/f (n)) time by Theorem
q 15.
lg(n)
Set g(n) := n lg
and h(n) :=
lg(n)
of logarithms, observe that
√
f (n/g) = f (Õ( n)) = Θ(f (n)),
n lg lg(n)
lg(n) .
By our choice of f (n) and basic properties
(3)
and furthermore
√
f (n/(hg(n/h))) = f Õ( n)/Õ(n1/4 ) = Θ(f (n)).
(4)
By (3), we have
s
r
n lg lg(n)
n
g=
≥Ω
, so the first constraint of Theorem 14 is satisfied.
lg(n)
f (n/g)
Moreover, by (4) we have
n
n lg lg(n)
n
h +
=
+
, which is O
f (n/(hg(n/h)))
lg(n)
Θ(f (n))
2
n
f (n/g)
by (3).
Therefore the second constraint of Theorem 14 is also satisfied,
and 3-SUM is solvable by
p
2
an algorithm running in O(n /f (n)) time and O
nf (n) space simultaneously.
J
In general, Theorem 14 provides a generic reduction from faster 3-SUM algorithms to
faster space-efficient 3-SUM algorithms. To illustrate:
I Corollary 17. If 3-SUM is solvable in O(n2 / lga (n)) time for some constant a > 0, then
√
3-SUM is in TISP(n2 / lga (n), n lga/2 (n)).
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
XXX:11
Proof. We apply Theorem 14. By assumption we have 3-SUM in O(n2 /f (n)) time, where
√
√
f (n) = lga n. Set g(n) := n/ lga/2 (n), and h(n) := n/ lga/2 (n). Note that f (n/g(n)) =
Θ(lga (n)) and f (n/(h(n) · g(n/h(n)))) = Θ(lga (n)), similar to Corollary 16. Therefore
g(n) =
√
n/ lg
a/2
(n) ≥ Ω
r
n
f (n/g(n))
and
h2 +
n
≤O
f (n/(hg(n/h)))
n
lga n
≤O
n
f (n/g(n))
.
Hence Theorem 14 applies to these settings of the parameters, and 3-SUM is in O(n2 /f (n/g)) =
√
O(n2 / lga (n)) time and O(n/h) = O( n lga/2 (n)) space.
J
4.4
The 3-SUM Conjecture and Small Space
Finally, we use the Space Reduction Theorem (Theorem 14) to show that the 3-SUM conjecture is false, then it is also false with respect to small-space algorithms.
I Lemma 18. If 3-SUM is in O(n2− ) time for some > 0, then for every α > 0, there is a
δ > 0 such that 3-SUM is solvable in O(n2−δ ) time and O(n1/2+α ) space, simultaneously.
Proof. The proof of Theorem 14 applies the 3-SUM Self Reduction (Theorem 13) twice.
We will basically perform the first part of the proof of Theorem 14, but instead of applying
the second part of the proof, we have to choose a different setting of parameters, focused on
minimizing the space usage instead of preserving running time.
Let T (n) := n2 /f (n) with f (n) = n . We first reduce the space usage of the algorithm
to linear. To this end, set g(n) := n(1−)/(2−) . Then, applying the 3-SUM Self-Reduction
(Theorem 13) with T (n) = S(n) and g(n), we can then solve 3-SUM in
n2
n2
, n = TISP
TISP(g 2 (n + T (n/g)), n/g + T (n/g)) = TISP
,
n
.
f (n/g)
n/(2−)
Now reset f (n) := n/(2−) , and reset g(n) := n1/2+α with α ∈ (0, 1/2). Applying the
3-SUM Self-Reduction (Theorem 13) with T (n) = n2 /f (n), S(n) = n, and g(n) as above,
we find an algorithm for 3-SUM in
TISP n2−2α + n2−(1/2−α)/(2−) , n1/2+α .
Note that for all > 0 and α ∈ (0, 1/2), the running time bound is truly subquadratic.
Further note that for any α ≥ 1/2, we only have more space to work with, so we clearly
obtain O(n2−δ ) time and O(n1/2+α ) space (for some δ > 0) in that case as well.
J
This lemma can be applied to show that the 3-SUM Conjecture is equivalent to seemingly
much weaker statement:
Reminder of The Small-Space 3-SUM Conjecture (Conjecture 5) On a word-RAM with
O(log n)-bit words, there exists an > 0 such that every algorithm that solves 3-SUM in
O(n1/2+ ) space must take at least n2−o(1) time.
I Theorem 19. The Small-Space 3-SUM Conjecture is equivalent to the 3-SUM Conjecture.
ICALP 2016
XXX:12 Deterministic Time-Space Tradeoffs for k-SUM
Proof. It suffices to show that the 3-SUM Conjecture if true implies the Small-Space 3-SUM
Conjecture and that the refutation of the 3-SUM Conjecture implies the Small-Space 3-SUM
Conjecture. First, we observe that the 3-SUM Conjecture trivially implies the Small-Space
3-SUM Conjecture.
Suppose the 3-SUM Conjecture is false. Then a O(n2− ) time algorithm for 3-SUM
exists, and Lemma 18 implies that for every α > 0, there is a δ > 0 such that 3-SUM is
solvable in O(n2−δ ) time and O(n1/2+α ) space, simultaneously. But this means that for any
choice of 0 > 0 for the Small-Space 3-SUM Conjecture, we can find a truly-subquadratic
0
3-SUM algorithm that uses only O(n1/2+ /2 ) space. This would falsify the Small-Space
3-SUM Conjecture.
J
We conclude that, in order to prove the 3-SUM conjecture, it is sufficient to prove that
no algorithm can solve 3-SUM in TISP(n2− , n0.51 ) for some > 0.
5
5.1
k-SUM
k-SUM Self-Reduction
We now generalize from 3-SUM to k-SUM. Again, we plan to split the lists into g groups
of size O(n/g). By Corollary 11, we will have only O(g k−1 ) subproblems of size O(n/g).
Unlike 3-SUM, where we just used the naive algorithm to solve subproblems, in this section
we use a general algorithm; we reduce from k-SUM to itself (albeit on smaller instances).
I Theorem 20. Suppose real k-SUM can be solved in TISP(T (n), S(n)). Then for any g, it
can also be solved in TISP(g k−1 (n + T (n/g)), n/g + S(n/g)).
Proof. This follows from a generalized analysis of the proof of Theorem 13. We brute force
over which groups the first k − 1 elements are in. We then extract groups where the negative
sum of elements from these first k − 1 groups could land. By Corollary 11 and similar
reasoning as before, there are only O(g k−1 ) tuples of blocks. For each tuple, we make a call
to NextGroup and to our input k-SUM algorithm on a subproblem of size O(n/g). This
gives the desired time and space, completing the proof.
J
5.2
Applying our k-SUM Self-Reduction
We want to apply the self-reduction on efficient deterministic algorithms. One of the best
starting points is the Schroeppel-Shamir 4-SUM algorithm, which we note is actually deterministic and works on reals because it simply uses priority queues and reduces to the
classic 2-SUM algorithm, both of which only use comparisons.
I Lemma 21 (From [23]). Real 4-SUM is solvable in TISP(n2 , n).
Another useful fact observed by Wang is that an algorithm for k-SUM can be transformed
into an algorithm for (k + 1)-SUM by brute-forcing one element:
I Lemma 22 (From [25]). If Real k-SUM is solvable in TISP(T (n), S(n)) then real k+1-SUM
is solvable in TISP(nT (n), S(n) + 1).
Suppose we want to use our results to derive a linear-space algorithm for k-SUM. We
will assume k is a multiple of 4, although Lemma 22 allows us to fill in for the other values
of k. By writing down sums of k/4 elements, we can transform k-SUM to 4-SUM, yielding
a TISP(nk/2 , nk/4 ) algorithm. We can then apply Theorem 20 with g = n(k−4)/k to get a
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
XXX:13
TISP(nk−3+4/k , n) algorithm. Notice that this algorithm runs significantly faster than O(nk )
time; we get O(n11/2 ) for 8-SUM and O(n28/3 ) for 12-SUM. As a coarse upper bound, we
can apply Lemma 22 and round down our savings (to make things cleaner), compensating
for k which are not a multiple of 4, we get:
I Corollary 23. For k ≥ 4, k-SUM is solvable in TISP(nk−3+4/(k−3) , n).
√
Suppose we wanted to use O( n) space instead. We get smaller subproblems by making
√
more groups; choosing g = n(k−2)/k instead yields a TISP(nk−2+2/k , n). Similarly applying
Lemma 22 and round down our savings to compensate for k which are not a multiple of 4,
we get another coarse upper bound:
√
I Corollary 24. For k ≥ 4, k-SUM is solvable in TISP(nk−2+2/k , n).
6
Future Work
We would like to extend these results to derandomize other known randomized algorithms
for k-SUM. To do that, it seems we require a “deterministic simulation” of the hash functions used in those results. Baran, Demaine, and Patrascu use hashing to get subquadratic
algorithms for 3-SUM [7]; Patrascu uses it to reduce 3-SUM to Convolution 3-SUM [20];
Wang uses it to produce a family of linear-space algorithms for k-SUM [25]. Which of these
results, if any, can be derandomized?
The hash families involved have three crucial properties: load-balancing (the hash buckets are not “too large”), few subproblems (the number of k-tuples of hash buckets examined
is “small”), and few false positives (there are few non-k-SUM solutions mapped to k-tuples
of hash buckets examined). Our s-Select algorithm (Lemma 12) and Domination Lemma
(Lemma 10) are used to achieve the first two properties, without using randomization. Can
the last property also be simulated deterministically? (Note that it’s not entirely clear what
it would mean to simulate “few false positives” deterministically.) If so, it is likely that all
these results can be derandomized efficiently.
References
1
2
3
4
5
6
A. Abboud, K. Lewi, and R. Williams. Losing weight by gaining edges. In Algorithms ESA 2014 - 22th Annual European Symposium, Wroclaw, Poland, September 8-10, 2014.
Proceedings, pages 1–12, 2014.
A. Abboud, V. Vassilevska Williams, and O. Weimann. Consequences of faster alignment
of sequences. In Automata, Languages, and Programming - 41st International Colloquium,
ICALP 2014, Copenhagen, Denmark, July 8-11, 2014, Proceedings, Part I, pages 39–51,
2014.
A. Abboud and V. V. Williams. Popular conjectures imply strong lower bounds for dynamic
problems. In 55th IEEE Annual Symposium on Foundations of Computer Science, FOCS
2014, Philadelphia, PA, USA, October 18-21, 2014, pages 434–443, 2014.
N. Ailon and B. Chazelle. Lower bounds for linear degeneracy testing. J. ACM, 52(2):157–
171, 2005.
A. Amir, T. M. Chan, M. Lewenstein, and N. Lewenstein. On hardness of jumbled indexing.
In Automata, Languages, and Programming - 41st International Colloquium, ICALP 2014,
Copenhagen, Denmark, July 8-11, 2014, Proceedings, Part I, pages 114–125, 2014.
P. Austrin, P. Kaski, M. Koivisto, and J. Määttä. Space-time tradeoffs for subset sum:
An improved worst case algorithm. In Automata, Languages, and Programming - 40th
International Colloquium, ICALP 2013, Riga, Latvia, July 8-12, 2013, Proceedings, Part
I, pages 45–56, 2013.
ICALP 2016
XXX:14 Deterministic Time-Space Tradeoffs for k-SUM
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
A
I. Baran, E. D. Demaine, and M. Patraşcu. Subquadratic algorithms for 3sum. In Algorithms and Data Structures, pages 409–421. Springer, 2005.
P. Beame, R. Clifford, and W. Machmouchi. Element distinctness, frequency moments,
and sliding windows. In FOCS, pages 290–299, 2013.
P. Beame, M. E. Saks, X. Sun, and E. Vee. Time-space trade-off lower bounds for randomized computation of decision problems. J. ACM, 50(2):154–195, 2003.
M. Blum, R. W. Floyd, V. Pratt, R. L. Rivest, and R. E. Tarjan. Time bounds for selection.
Journal of computer and system sciences, 7(4):448–461, 1973.
A. Czumaj and A. Lingas. Finding a heaviest triangle is not harder than matrix multiplication. In Proc. SODA, pages 986–994, 2007.
S. Diehl, D. van Melkebeek, and R. Williams. An improved time-space lower bound for
tautologies. J. Comb. Optim., 22(3):325–338, 2011.
J. Erickson. Lower bounds for linear satisfiability problems. In Proceedings of the Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, 22-24 January 1995. San Francisco,
California., pages 388–395, 1995.
L. Fortnow, R. J. Lipton, D. van Melkebeek, and A. Viglas. Time-space lower bounds for
satisfiability. J. ACM, 52(6):835–865, 2005.
A. Gajentaan and M. H. Overmars. On a class of O(n2 ) problems in computational geometry. Computational geometry, 5(3):165–185, 1995.
O. Gold and M. Sharir. Improved bounds for 3sum, k-sum, and linear degeneracy. arXiv
preprint arXiv:1512.05279, 2015.
A. Gronlund and S. Pettie. Threesomes, degenerates, and love triangles. In Foundations of
Computer Science (FOCS), 2014 IEEE 55th Annual Symposium on, pages 621–630. IEEE,
2014.
R. Impagliazzo and R. Paturi. On the complexity of k-SAT. J. Comput. Syst. Sci.,
62(2):367–375, 2001.
R. M. Karp. Reducibility among combinatorial problems. Springer, 1972.
M. Patrascu. Towards polynomial lower bounds for dynamic problems. In Proceedings of
the forty-second ACM symposium on Theory of computing, pages 603–610. ACM, 2010.
S. Pettie and V. Ramachandran. A shortest path algorithm for real-weighted undirected
graphs. SIAM J. Comput., 34(6):1398–1431, 2005.
M. Pătraşcu and R. Williams. On the possibility of faster sat algorithms. In Proceedings of
the Twenty-First Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’10, pages
1065–1075, Philadelphia, PA, USA, 2010. Society for Industrial and Applied Mathematics.
R. Schroeppel and A. Shamir. A T = O(2n/2 ), S = O(2n/4 ) algorithm for certain npcomplete problems. SIAM journal on Computing, 10(3):456–464, 1981.
V. Vassilevska and R. Williams. Finding, minimizing, and counting weighted subgraphs.
In Proceedings of the forty-first annual ACM symposium on Theory of computing, pages
455–464. ACM, 2009.
J. R. Wang. Space-efficient randomized algorithms for k-sum. In Algorithms-ESA 2014,
pages 810–829. Springer, 2014.
R. R. Williams. Time-space tradeoffs for counting NP solutions modulo integers. Computational Complexity, 17(2):179–219, 2008.
s-Select
In addition to NextGroup we have two other s-Select algorithms. We present two
algorithms to solve this subtask. The first requires the values to be integers in the range
[−R, R] and runs in word-TISP(n log R, 1) (recall we are in the word-RAM model and we are
measuring space in terms of the number of words). The other needs no assumptions and
A. Lincoln, V. Vassilevska Williams, J. R. Wang and R. R. Williams
XXX:15
returns the answers for g choices of k in TISP(n2 , g). The NextGroupalgorithm discussed
in subsection 3.2 runs in TISP(n, s).
A.1
Bounded Range s-Select
This first algorithm runs a binary search over the bounded range to locate the sth smallest
element.
Algorithm 2: Bounded Range s-Select Algorithm
Set ` = −R, r = R;
while ` < r do
Set m = b `+r
2 c;
Set c = 0;
for a ∈ L do
if m ≥ a then
Increment c;
if c ≥ s then
Set r = m;
else
Set ` = m + 1;
return `;
I Theorem 25. Algorithm 2 solves s-Select in word-TISP(n log R, 1).
Proof. Algorithm 2 returns the smallest integer v such that there are s values less than or
equal to v. Since all values are integers, by assumption, this is the sth smallest value. The
algorithm runs for O(log R) iterations, but each iteration does a scan of L that takes O(n)
time. The algorithm keeps a constant number of values, so it uses O(1) space.
J
A.2
Batch real k-Select
When we lose the range and integrality assumptions, we can still gain when we have several
s-Select instances with the same list L. In particular, suppose there are g indices we want
to know: s1 , . . . , sg , where g ≤ n, we can go through the list in order in n2 time noting and
ICALP 2016
XXX:16 Deterministic Time-Space Tradeoffs for k-SUM
saving the value of all of those indices. Furthermore, we can use this method over the reals.
Algorithm 3: Batch s-Select Algorithm
Set prev = −∞;
Create a return vector V of length g;
Set i = 1;
while i ≤ n do
Set curr = ∞;
for a ∈ L do
If a > prev, set curr = min(curr, a);
Set dup = 0;
for a ∈ L do
If a = curr, increment dup;
for j = [1, g] do
If kj ∈ [i, i + dup), set V [j] = curr;
Set prev = curr;
Increment i by dup;
return V ;
I Theorem 26. Algorithm 3 solves batch real s-Select in TISP(n2 , g).
Proof. Algorithm 3 repeatedly scans L, each time finding the next largest element. After
it finds the sth smallest element, it checks to see if s was one of the requested indices, and
if so, fills it into its answer. The algorithm performs O(n) scans of L and the kj , but since
g ≤ n, this runs in O(n2 ) time. Keeping g elements around takes O(g) space.
J
| 8cs.DS
|
1
Flow: Architecture and Benchmarking for
Reinforcement Learning in Traffic Control
arXiv:1710.05465v1 [cs.AI] 16 Oct 2017
Cathy Wu∗ , Aboudy Kreidieh† , Kanaad Parvate∗ , Eugene Vinitsky‡ , Alexandre M Bayen∗†§
∗ UC Berkeley, Electrical Engineering and Computer Science
† UC Berkeley, Department of Civil and Environmental Engineering
‡ UC Berkeley, Department of Mechanical Engineering
§ UC Berkeley, Institute for Transportation Studies
Abstract—Flow is a new computational framework, built to
support a key need triggered by the rapid growth of autonomy
in ground traffic: controllers for autonomous vehicles in the
presence of complex nonlinear dynamics in traffic. Leveraging
recent advances in deep Reinforcement Learning (RL), Flow
enables the use of RL methods such as policy gradient for traffic
control and enables benchmarking the performance of classical
(including hand-designed) controllers with learned policies (control laws). Flow integrates traffic microsimulator SUMO with
deep reinforcement learning library rllab and enables the easy
design of traffic tasks, including different networks configurations
and vehicle dynamics. We use Flow to develop reliable controllers
for complex problems, such as controlling mixed-autonomy traffic
(involving both autonomous and human-driven vehicles) in a ring
road. For this, we first show that state-of-the-art hand-designed
controllers excel when in-distribution, but fail to generalize; then,
we show that even simple neural network policies can solve the
stabilization task across density settings and generalize to outof-distribution settings.
Index Terms—Learning and Adaptive Systems; Deep Reinforcement Learning; Traffic Microsimulation
I. I NTRODUCTION
Transportation accounts for 28% of energy consumption in
the US. Workers spent on aggregate over three million driveryears commuting to their jobs [1], with significant impact
on nation-wide congestion. Based on 2012 estimates, U.S.
commuters experienced an average of 52 hours of delay per
year, causing $121 billion of delay and fuel costs annually
[2]. Depending on its use in traffic, automation has the
potential to achieve many benefits or to exacerbate problems
at the system level, with potential amelioration or worsening
of various system metrics including greenhouse gas (GHG)
emissions, vehicle miles traveled (VMT), total travel time
(TTT). Estimates project that 2% of fuel consumption today is
wasted due to congestion, a figure that rises to 4.2% in 2050
[3]. As such, the potential efficiency improvement provided
by autonomous vehicles is two to four percent of total fuel
consumption due to the alleviation of congestion alone.
In recent breakthrough experiments, Stern et al. [4] demonstrated a reduction in fuel consumption over 40% by the
insertion of an autonomous vehicle in ring traffic to dampen
the famous ring instabilities displayed by Sugiyama et al. in
his seminal 2008 experiment [5]. This very disruptive field
Corresponding author: Cathy Wu ([email protected])
Email addresses: {aboudy, kanaad, evinitsky, bayen}@berkeley.edu
operational test is one of the motivations for the present work:
it demonstrates the power of automation and its potential
impact on complex traffic phenomena such as stop-and-go
waves [6].
The breakthrough results [4], [7] are part of a broader
core set of robotics challenges concerning the deployment
of multi-agent automation systems, such as fleets of selfdriving cars [8], [9], coordinated traffic lights [10], [11],
or other coordinated infrastructure. Robotics has already
demonstrated tremendous potential in improving transportation systems through autonomous vehicles research; highly
related problems include localization [12], [13], [14], path
planning [15], [16], collision avoidance [17], and perception
[18] problems. Considerable progress has also been made
in recent decades in vehicle automation, including anti-lock
braking systems (ABS), adaptive cruise control, lane keeping,
automated parking, etc. [19], [20], [21], [22], which also have
great potential to improve energy efficiency and safety in
traffic. Down the road, the emergence of automated districts,
i.e. districts where all vehicles are automated and operate
efficiently with collaborative path-planning, might push this
paradigm to next generation mobility [23]. Fleets of autonomous vehicles have recently been explored in the context
of shared-mobility systems, such as autonomous mobility-ondemand systems, which abstracts out the low-level vehicle
dynamics and considers a queuing theoretic model. Lowlevel vehicle dynamics, however, are of crucial importance, as
exhibited by [24] and because many traffic phenomena, which
affect energy consumption, safety, and travel time are exhibited
at the level of low-level dynamics [5], [25], [26], [27]. In some
settings, model-based controllers enable analytical solutions,
or tractable algorithmic solutions. However, often, due to the
nonlinearity of the models, numerous guarantees are lost in
the process of developing controllers (i.e. optimality, run-time,
complexity, approximation ratio, etc.). For example, while the
ring setting enables elegant controllers to work in practice, the
extension of these results (both theoretical and experimental)
to arbitrary settings (network topologies, number of lanes,
heterogeneity of the fleet, etc.) is challenging.
Deep reinforcement learning (RL), which is the main enabler in our framework, is a powerful tool for control and has
already had demonstrated success in complex but data-rich
problem settings such as Atari games [28], 3D locomotion
and manipulation [29], [30], [31], chess [32], among others.
2
RL testbeds exist for different problem domains, such as the
Arcade Learning Environment (ALE) for Atari games [33],
DeepMind Lab for a first-person 3D game [34], OpenAI gym
for a variety of control problems [35], FAIR TorchCraft for
Starcraft: Brood War [36], MuJoCo for multi-joint dynamics
with Contact [37], TORCS for a car racing game [38], among
others. DeepMind and Blizzard will collaborate to release
the Starcraft II AI research environment [39]. Each of these
RL testbeds enables the study of control through RL of a
specific problem domain by leveraging of the data-rich setting
of simulation. One of the primary goals of this article is to
present a similarly suitable RL testbed for traffic dynamics by
making use of an existing traffic simulator.
These recent advances in deep RL provide a promising
alternative to model-based controller design, which the present
article explores. One key step in the development of such
paradigms is the ability to provide high fidelity microsimulations of traffic that can encompass accurate vehicle dynamics
to simulate the action of these new RL-generated control
policies, a pre-requisite to field experimental tests. This is
precisely one of the aims of the present article. RL promises
an approach to design controllers using black box machine
learning systems. It still requires physical vehicle response
to be incorporated in the simulation to learn controllers that
match physical vehicle dynamics. This problem extends beyond the vehicle for which the controller is to be designed.
For example, although vehicle velocity is intuitive as a control
variable, it is important to keep in mind that other variables,
such as actuator torques, are those actually controlled; another
example is that the input may consist of data from cameras,
LIDAR, or radar. The conversion between the variables might
or might not be direct and may require the design of additional
controllers, the performance of which would also have to be
considered.
In the present article, we propose the first (to our knowledge) computational framework and architecture to systematically integrate deep RL and traffic microsimulation, thereby
enabling the systematic study of autonomous vehicles in
complex traffic settings, including mixed-autonomy and fullyautonomous settings. Our framework permits both RL and
classical control techniques to be applied to microsimulations.
As classical control is a primary approach for studying traffic
dynamics, supporting benchmarking with such methods is
crucial for measuring progress of learned controllers. As an
illustration, this article provides a benchmark of the relative performance of learned and explicit controllers [4] for
the mixed-autonomy ring road setting. The computational
framework encompasses model-free reinforcement learning
approaches, which complement model-based methods such as
model-based reinforcement learning, dynamic programming,
optimal control, and hand-designed controllers; these methods
dramatically range in complexity, sometimes exhibiting prohibitive computational costs. Our initial case study investigates
microscopic longitudinal dynamics (forwards-backwards) [40]
and lateral dynamics (left-right) [41] of vehicles. We study
a variety of network configurations, and our proposed framework largely extends to other reinforcement learning methods
and other dynamics and settings, such as coordinated behaviors
[42], other sophisticated behavior models, and more complex
network configurations.
The contribution of this article includes three components,
(1) a computational framework and architecture, which provides a rich design space for traffic control problems and
exposes model-free RL methods, (2) the implementation of
several instantiations of RL algorithms that can solve complex
control tasks, and (3) a set of use cases that illustrates
the power of the building block and benchmark scenarios.
Specifically, our contributions are:
•
•
•
•
•
•
Flow, a computational framework for deep RL and control
experiments for traffic microsimulation. Flow integrates
the traffic microsimulator SUMO [43] with a standard
deep reinforcement learning library rllab [44], thereby
permitting the training of large-scale reinforcement learning experiments at scale on Amazon Web Services (AWS)
Elastic Compute Cloud (EC2) for traffic control tasks.
Our computational framework is open-source and available at https://github.com/cathywu/flow.
An interface, provided by Flow for the design of traffic
control tasks, including customized configurations of different road networks, vehicle types and vehicle dynamics,
noise models, as well as other attributes provided by a
standard Markov Decision Process (MDP) interface.
Extensions of SUMO to support high frequency simulation and greater flexibility in controllers.
Exposing model-free reinforcement algorithms for discounted finite (partially observed) MDPs such as policy
gradient, with specific examples including Trust Region
Policy Optimization (TRPO) [30] and Generalized Advantage Estimation (GAE) [29], to the domain of traffic
control problems.
Benchmarking of relative performance of learned and
explicit controllers in rich traffic control settings. We
present a benchmark on the mixed-autonomy single-lane
ring road network and find that a reinforcement learning
agent is capable of learning policies exceeding the performance of state-of-the-art controllers. The particular case
of Sugiyama instabilities [5] is used to demonstrate the
power of our tool.
Case studies for building block networks. We demonstrate deep RL results on traffic at the level of multiagent vehicle control. We demonstrate all-human, mixedautonomy, and fully-autonomous experiments on more
complex traffic control tasks, such as a multi-lane ring
road and a figure 8 network. We provide additional
networks, including a merge network and an intersection
networks.
The rest of the article is organized as follows. Section II
provides background on the RL framework used in the rest
of the article. Section III describes the architecture of Flow
and the processes it can handle in the three computational
environments they are run (incl. SUMO and rllab). Section IV presents the various building blocks used by SUMO
for building general networks (underlying maps). Section V
presents the various settings for the optimization, incl. action
/ observation space, reward functions and policies. This is
3
followed by two experimental sections: in Section VI, in which
we benchmark the performance of the RL-based algorithm to
the seminal FollowerStopper controller [4], and Section VII
which presents a series of various other experiments on the
building block network modules of Flow. Finally, Section VIII
presents related work to place this in the broader context of
traffic flow modeling, deep RL and microsimulations.
II. P RELIMINARIES
In this section, we define the notation used in subsequent
sections.
The system described in this article solves tasks which
conform to the standard interface of a finite-horizon discounted
Markov decision process (MDP) [45], [46], defined by the
tuple (S, A, P, r, ρ0 , γ, T ), where S is a (possibly infinite) set
of states, A is a set of actions, P : S × A × S → R≥0
is the transition probability distribution, r : S × A → R
is the reward function, ρ0 : S → R≥0 is the initial state
distribution, γ ∈ (0, 1] is the discount factor, and T is the
horizon. For partially observable tasks, which conform to the
interface of a partially observable Markov decision process
(POMDP), two more components are required, namely Ω, a
set of observations, and O : S × Ω → R≥0 , the observation
probability distribution.
RL studies the problem of how agents can learn to take
actions in its environment to maximize its cumulative reward.
The Flow framework uses policy gradient methods [47], a
class of reinforcement learning algorithms which optimize
a stochastic policy πθ : S × A → R≥0 . These algorithms
iteratively update the parameters of the policy through optimizing the expected cumulative reward using sampled data
from SUMO. The policy usually consists of neural networks,
and may be of several forms. Two policies used in this article
are the Multilayer Perceptron (MLP) and Gated Recurrent Unit
(GRU). MLP is a classical artificial neural network with multiple hidden layers and utilizes backpropagation to optimize its
parameters [48]. GRUs are recurrent neural network capable
of storing memory on the previous states of the system through
the use of parametrized update and reset gates, which are also
optimized by the policy gradient method [49]. This enables
GRUs to make decisions based on both current input and past
inputs.
The autonomous vehicles in our system execute controllers
which are parameterized policies, trained using policy gradient
methods. For all experiments in this article, we use the
Trust Region Policy Optimization (TRPO) [30] policy gradient
method for learning the policy, linear feature baselines as
described in [44], discount factor γ = 0.999, and step size
0.01. For most experiments, a diagonal Gaussian MLP policy
is used with hidden layers (100, 50, 25) and tanh nonlinearity. The experiment stabilizing the ring, described later,
uses a hidden layer of shape (3,3). For experiments requiring
memory, a GRU policy with hidden layers (5,) and tanh nonlinearity is used.
III. OVERVIEW OF F LOW
Flow is created to fill the gap between modern machine
learning and complex control problems in traffic. Flow is
Fig. 1: Flow Process Diagram. Flow interfaces SUMO via TraCI with rllab to
permit the construction and simulation of traffic MDPs and for the training and
evaluation of policies (control laws). After initializing the simulation in some
initial configuration, rllab collects samples by advancing and observing the
simulation. In each step, vehicles are provided actions through a pre-specified
controller or through a policy (via rllab). These actions are then applied via
TraCI and the simulation progresses. At the end of an episode, rllab issues
a reset command to the environment, which returns vehicles to their initial
(possibly random) position.
a computational framework for traffic microsimulation with
RL methods. Although the architecture is agnostic to specific
machine learning and traffic software packages, we chose to
integrate widely used open-source tools to promote access and
extension.
The first of those open-source tools is SUMO (Simulation
of Urban MObility) [43]. SUMO is a continuous-time and
continuous-space microscopic traffic simulator. It is capable of
handling large road networks and of modeling the dynamics of
each vehicle in the simulation. SUMO was chosen particularly
for its extensibility, as it includes an API called TraCI (Traffic
Control Interface). TraCI allows users to extend existing
SUMO functionality through querying and modifying the state
of the simulation, at the single time-step resolution. This
allows the user to easily provide intricate, custom commands
that modify the simulation directly.
Secondly, we use rllab, an open source framework that
enables running and evaluating RL algorithms on a variety
of different scenarios, from classic tasks such as cartpole
balancing to more complicated tasks such as 3D humanoid
locomotion [44]. Flow uses rllab to facilitate the training, optimization, and application of control policies that manipulate
the simulation. By modeling traffic scenarios as reinforcement learning problems, we use rllab to issue longitudinal
and lateral controls to vehicles. Rllab further interfaces with
OpenAI Gym, another framework for the development and
4
evaluation of reinforcement learning algorithms. The SUMO
environments built in Flow are also compatible with OpenAI
Gym.
Flow encapsulates SUMO via TraCI to permit the definition and simulation of traffic MDPs for rllab to train and
evaluate policies. After initializing the simulation in some
initial configuration, rllab collects samples by advancing and
observing the simulation. In each step, vehicles are provided
actions through a pre-specified controller or through a policy.
These actions are then applied via TraCI and the simulation
progresses. After a specified number of timesteps (i.e. the
end of a rollout) or after the simulation has terminated early
(i.e. a vehicle has crashed), rllab issues a reset command
to the environment, which returns vehicles to their initial
(possibly random) position. The interactions between Flow,
SUMO/TraCI, and rllab are illustrated in Figure 1.
In addition to learned policies, Flow supports classical
control (including hand-designed controllers and calibrated
models of human dynamics) for longitudinal and lateral control. Flow also supports the car following models and lanechanging models that are provided in SUMO. These models
work analogously to the policies generated by rllab, providing
longitudinal and lateral controls to vehicles through ordinary
differential equations. Together, these controllers comprise the
overall dynamics of mixed-autonomy, fully-human, or fullautonomy settings.
Additionally, Flow provides various failsafes presented in
Appendix C, including the ones that are built into SUMO, to
prevent the vehicles from crashing and the simulation from
terminating early.
Flow can be used to perform both pure model-based control
experiments by using only pre-specified controllers for issuing
actions, as well as experiments with a mixture of pre-specified
and learned controllers. Together, this permits the study of
heterogeneous or mixed-autonomy settings.
A. Architecture of Flow
supporting components as well as their interactions are summarized in Figure 2.
The scenario for an experiment specifies network configuration in the form of network shape and attributes, for
example two-lane loop road with circumference 200m, or by
importing OpenStreetMap data (see Figure 3). Based on the
specifications provided, the net and configuration files needed
by SUMO are generated. The user also specifies the number
and types of vehicles (car following model and a lane-change
controller), which will be placed in the scenario.
The generator is a predefined class, which allows for
rapid generation of scenarios with user-defined sizes, shapes,
and configurations. The experiments presented in this article
include large loop roads generated by specifying the number
of lanes and ring circumference, figure eight networks with
a crossing intersection, closed loop “merged” networks, and
standard intersections.
The environment encodes the MDP, including functions
to step through the simulation, retrieve the state, sample and
apply actions, compute the reward, and reset the simulation.
The environment is updated at each timestep of the simulation
and, importantly, stores each vehicle’s state (e.g. position and
velocity). Information from the environment is provided to
a controller or passed to rllab to determine an action for a
vehicle to apply, e.g. an acceleration. Note that the amount
of information provided to either RL or to a controller can be
restricted as desired, thus allowing fully observable or partially
observable MDPs. This article studies both fully and partially
observed settings.
When provided with actions to apply, Flow calls the action applicator which uses TraCI to enact the action on
the vehicles. Actions specified as accelerations are converted
into velocities, using numerical integration and based on the
timestep and current state of the experiment. These velocities
are then applied to vehicles using TraCI.
IV. N ETWORKS
Flow currently supports learning policies on a variety of networks with a fixed number of vehicles. These include closed
networks such as single and multi-lane ring roads, figure eight
networks, and loops with merge as well as open networks, such
as intersections. See Figure 3 for various example networks
supported by Flow. In each of these networks, Flow can be
used to study the design or learning of controllers which
optimize the system-level velocity or fuel consumption, in the
presence of different types of vehicles, model noise, etc.
Fig. 2: Flow Architecture. A Flow experiment involves a scenario and
environment, interfaced with rllab and controllers. The experiment scenario
runs a generator to create road networks for use in SUMO, which is started
by the environment. Controllers and rllab take experiment states and return
actions, which are applied through SUMO’s TraCI API. (See Section III-A).
An experiment using Flow requires defining two components: a scenario and an environment. These and several
Single-lane Ring Roads: The ring road network consists
of a circular lane with a specified length, inspired by the
230m track studied by Sugiyama et al. [5]. This network has
been extensively studied and serves as an experimental and
numerical baseline for benchmarking.
Multi-lane Ring Roads: Multi-lane ring roads are a natural extension to problems involving a single lane ring. The
inclusion of lane-changing behavior in this setting makes
studying such problems exceedingly difficult from an analytical perspective, thereby constraining most classical control
techniques to the single-lane case. Many multi-lane models
5
Fig. 3: Various network building blocks supported by the Flow framework. Top left: Single-lane Ring Road Network. Top middle: Multi-Lane Ring Road
Network. Top right: Figure-Eight Road Network. Bottom left: Intersection Network. Bottom Middle: Closed Loop Merge Network. Bottom right: Imported
San Francisco Network (currently operational for forward simulation). Maps in Flow can be generated from OSM data and visualized using SUMO.
forgo longitudinal dynamics in order to encourage tractable
analysis [50], [51], [52], [53]. Recent strides have been made
in developing simple stochastic models that retain longitudinal
dynamics while capturing lane-changing dynamics in a single
lane setting [54]. Modern machine learning methods, however,
do not require a simplification of the dynamics for the problem
to become tractable, as explored in Section VII.
the control zone, the system speeds or slows down vehicles
to either maximize average velocity or minimize experienced
delay. The building block can be used to build a general
schema for arbitrary maps such as one the one shown in
Figure 3 (bottom right).
Figure Eight: The figure eight network is a simple closed
network with an intersection. Two ring roads, placed at opposite ends of the network, are connected by two perpendicular
intersections. Vehicles that try to cross this intersection from
opposite ends are constrained by a right-of-way model provided by SUMO to prevent crashes.
Flow provides an interface for fine-grained traffic control
task design. This section describes the options in the task
design space, beyond the selection of a network configuration,
as described in Section IV.
Loops with Merges: This network permits the study of merging behavior in closed loop networks. This network consists
of two ring roads which are connected together. Vehicles in
the smaller ring stay within this ring, while vehicles in the
larger ring try to merge into the smaller ring and then back
out to the larger ring. This typically results in congestion at
the merge points.
Intersections: This network permits the study of intersection
management in an open network. Vehicles arrive in the control
zone of the intersection according to a Poisson distribution. At
V. TASK SPACE
Action Space: When following a pre-defined route, a vehicle performs longitudinal (acceleration) and lateral (lanechanging) actions. Accordingly, for tasks with k autonomous
vehicles, the action space is a set of accelerations c ∈ Rk
and lane-changing decisions d ∈ [−1, 1]k . The lane-changing
values are rounded to the nearest integer (-1, 0, 1) denoting
lane-change left, do not lane-change, and lane-change right,
respectively; this keeps the action space representation continuous. In cases where the network only has one lane, the action
space may be reduced to solely a set of accelerations.
Observation Space: The observation space may be any set
of state information the user wishes to provide to the agent.
6
This information may fully or partially describe the state of
the environment. For instance, the autonomous vehicles may
observe only the preceding vehicle, only nearby vehicles, or
all vehicles and their corresponding position, relative position,
velocity, lane information, etc.
Custom Reward Functions: The reward function can be
any function of vehicle speed, position, fuel consumption,
acceleration, distance elapsed, etc. Note that existing OpenAI
Gym environments (atari and mujoco) come with a prespecified reward function [35]. However, depending on the
context, a researcher, control engineer, or planner may desire
a different reward function or may even want to study a range
of reward functions.
For all experiments presented in this article, we evaluate the
reward on the average velocity of vehicles in the network. At
times, this reward is also augmented with an added penalty
to discourage accelerations or excessive lane-changes by the
autonomous vehicles.
Heterogeneous Settings: Flow supports traffic settings with
heterogeneous vehicle types, such as those with different controllers or parameters. Additionally, simulations can contain
both learning agents (autonomous vehicles) and vehicles with
pre-specified controllers or dynamics. This permits the use of
Flow for mixed autonomy experiments.
Noise and Perturbations: Arbitrary vehicle-level perturbations can be specified in an experiment, choosing and randomly perturbing a vehicle by overriding its control inputs
and commanding a deceleration for some duration. Gaussian
noise may also be introduced to the accelerations of all human
car-following models in Flow.
Vehicle Placement: Flow supports several vehicle placement
methods that may be used to generate randomized starting
positions. Vehicles may be distributed uniformly across the
length of the network or perturbed from uniformity by some
noise. In addition, vehicles may be bunched together to reduce
the space they take up on the network initially, and spread
out across one or multiple lanes (if the network permits it);
these create configurations resembling traffic jams. Finally, the
sequence in which vehicles are placed in the system may also
be randomly shuffled, and thus their ordering in the state space
may be randomized.
VI. C ASE STUDY: MIXED - AUTONOMY RING
This section uses Flow to benchmark the relative performance of an explicit controller and the reinforcement learning
approach to a given set of scenarios. The next section will
show similar outcomes of our RL approaches, including examples for which there are no known explicit controllers.
The goal of this section is to demonstrate the performance
of the reinforcement learning approach on the problem introduced by the seminal work of Stern et al. [4] on the mixedautonomy single-lane ring, following the canonical singlelane ring setup of Sugiyama et al. [5], consisting of 22
human-driven vehicles on a 230m ring track. The seminal
work of Sugiyama et al. [5] shows that such a dynamical
system produces backward propagating waves, causing part
of the traffic to come to a complete stop, even in the absence of typical traffic perturbations, such as lane changes
and intersections. The breakthrough study of Stern et al. [4]
studies the case of 21 human-driven vehicles and one vehicle
following one of two proposed controllers, which we detail
in Sections VI-1 and VI-2. This setting invokes a cascade of
nonlinear dynamics from n (homogeneous) agents. In this and
following sections, we study the potential for machine learning
techniques (RL in particular) to produce well-performing controllers, even in the presence of highly nonlinear and complex
settings.
We begin by defining the experimental setup and the stateof-the-art controllers that had been designed for the mixedautonomy ring setting. We then benchmark the performance
of the controller learned by Flow under the same experimental
setup against the hand-designed controllers under a partially
observed setting.
Experimental Scenario: In our numerical experiments, we
similarly study 22 vehicles, one of which is autonomous,
with ring lengths ranging between 180m and 380m, resulting
in varying traffic densities. The vehicles are each 5m long
and follow Intelligent Driver Model (IDM) dynamics with
parameters specified by [55]. The IDM dynamics are additionally perturbed by Gaussian acceleration noise of N (0, 0.2),
calibrated to match measures of stochasticity to the IDM
model presented by [56]. We focus on the partially observed
setting of observing only the velocity of the autonomous
vehicle, the velocity of its preceding vehicle, and its relative
position to the preceding vehicle. Each experiment runs for a
finite time horizon, ranging from 150 to 300 seconds.
Definitions: We briefly present the important terms used in
this case study. Uniform flow is an equilibrium state of the
dynamical system (and a corresponding solution to the dynamics) where vehicles are traveling at a constant velocity. In this
article, because the dynamical system has multiple equilibria,
we use uniform flow to describe the unstable equilibrium
in which the velocity is constant. We call this velocity the
equilibrium velocity of the system. Uniform flow differentiates
it from the stable equilibrium in which stop-and-go waves are
formed, which does not exhibit a constant velocity.
Settings: We compare the following controllers and observation settings for the single autonomous vehicle:
•
•
•
•
•
Learned agent with GRU policy with partial observation.
Learned agent with MLP policy with partial observation.
Proportional Integral (PI) controller with saturation with
partial observation. This controller is given in [4] and is
provided in Section VI-2.
FollowerStopper with partial observation and desired
velocity fixed at 4.15 m/s. The FollowerStopper controller
is introduced in [4] and is provided in Section VI-1.
FollowerStopper requires an external desired velocity, so
we selected the largest fixed velocity which successfully
stabilizes the ring at 260m; this is further discussed in
the results.
Fully-human setting using Intelligent Driver Model
(IDM) controller, which is presented in more details in
7
Appendix A. This setting reduces to Sugiyamas result of
traffic jams [5] and serves as a baseline comparison.
Explicit Controllers: In this section, we describe the two
state-of-the-art controllers for the mixed-autonomy ring,
against which we benchmark our learned policies generated
using Flow.
1) FollowerStopper: Recent work by [4] presented two
control models that may be used by autonomous vehicles
to attenuate the emergence of stop-and-go waves in a traffic
network. The first of these models is the FollowerStopper.
This model commands the autonomous vehicles to maintain a
desired velocity U , while ensuring that the vehicle does not
crash into the vehicle behind it. Following this model, the
command velocity v cmd of the autonomous vehicle is:
0
if ∆x ≤ ∆x1
v ∆x−∆x1
if
∆x1 < ∆x ≤ ∆x2
∆x2 −∆x1
(1)
v cmd =
∆x−∆x
2
v + (U − v) ∆x3 −∆x2 if ∆x2 < ∆x ≤ ∆x3
U
if ∆x3 < ∆x
where v = min(max(v lead , 0), U ), v lead is the speed of
the leading vehicles, ∆x is the headway of the autonomous
vehicle, subject to boundaries defined as:
∆xk = ∆x0k +
1
(∆v− )2 , k = 1, 2, 3
2dk
(2)
The parameters of this model can be found in [4].
2) PI with Saturation: In addition to the FollowerStopper
model, [4] presents a model titled the “PI with Saturation
Controller” that attempts to estimate the average equilibrium
velocity U for vehicles on the network, and then drives at that
speed. This average is computed
Pm AVas a temporal average from
1
its own history: U = m
j=1 vj . The target velocity at any
given time is then defined as:
∆x − gl
v target = U + v catch × min max
,0 ,1
(3)
gu − gl
Finally, the command velocity for the vehicle at time j + 1,
which also ensures that the vehicle does not crash, is:
cmd
vj+1
= βj (αj vjtarget + (1 − αj )vjlead ) + (1 − βj )vjcmd
(4)
The values for all parameters in the model can be found in
[4].
Results: Through this detailed case study of the mixedautonomy single-lane ring, we demonstrate that Flow enables the fine-grained benchmarking of classical and learned
controllers. Videos and additional results are available at
https://sites.google.com/view/ieee-tro-flow.
1) Performance: Figure 6 shows several key findings. This
traffic density vs. velocity plot shows the performance of
the different learned and hand-designed controllers. First, we
observe that GRU and MLP controllers (in partially observed
settings) are capable of matching the uniform flow speed very
closely for all trained densities, thereby effectively stabilizing
traffic in all densities in the training range. The PI with
Saturation controller, on the other hand, is only capable of
properly performing at densities less than or equal to the
density at which it was calibrated (less congested settings).
Figure 4 shows the velocity profiles for the different learned
and hand-designed controllers for the 260m ring and additionally includes the FollowerStopper controller. We observe
that although all controllers are able to stabilize the system,
the GRU controller allows the system to reach the uniform
flow equilibrium velocity most quickly. The GRU and MLP
policies stabilize the system with less oscillatory behavior than
the FollowerStopper and PI with Saturation controllers, as
observed in the velocity profiles. In addition, the FollowerStopper controller is the least performant; the controller can
only stabilize a 260m ring road to a speed of 4.15 m/s, well
below the 4.82 m/s uniform flow velocity.
Finally, Figure 5 shows the space-time curves for all vehicles in the system, using a variety of controllers. We observe
that the PI with Saturation and FollowerStopper controllers
leave much smaller openings in the network (smaller headways) than the MLP and GRU policies. The MLP policy
exhibits the largest openings, as can be seen by the large
white portion of the MLP plot within Figure 5. If this were
instead applied in a multi-lane ring study, then the smaller
openings would have the benefit of preventing opportunistic
lane changes, so this observation can lead to better reward
design for more complex mixed-autonomy traffic studies.
2) Robustness: One of the strengths of our GRU and MLP
policies is that it does not rely on external calibration of
parameters that is specific to a particular traffic setting, such
as density.
Although the PI with Saturation controller can conceptually
adjust to different densities, with its moving average filter, we
experimentally found that its performance is sensitive to its
parameters. Using parameters calibrated for 260m ring roads
(as described in [4]), the PI with Saturation controller indeed
performs the best at 260m among the density range considered
in this study. However, this controller’s performance quickly
degrades at higher density (more congested settings), dipping
close to the stop-and-go equilibrium velocity.
Similarly, the FollowerStopper Controller suffers from the
same calibration deficiencies as the PI with Saturation Controller. Additionally, the desired velocity must be provided
beforehand. Interestingly and moreover, we found that this
controller often fails to stabilize the system if provided too
high of a desired velocity, even if it is well below the
equilibrium velocity. Instead, if a lower desired velocity is first
provided as an intermediate control target, then the desired
velocity may then subsequently be achieved. This suggests
that a simple control law such as the FollowerStopper cannot
optimally stabilize a mixed-autonomy ring, and additionally,
that there is additional tuning and augmentation necessary to
use the FollowerStopper controller.
3) Generalization of the Learned Control Policy: Training
on different vehicle densities encourages the learning of a
more robust policy. We found the policy to generalize even
to densities outside of the training regime. Figure 6 shows the
average velocity vehicles in the network achieve for the final
100s of simulation time; the gray regions indicate the testtime densities. Surprisingly, we found that even when training
on different densities but in the absence of acceleration error
in the human driver models, the learned policies successfully
8
Fig. 4: All experiments are run for 300 seconds with the autonomous vehicle acting as a human driver before it is “activated”. As we can see, an autonomous
vehicle trained in a partially observable ring road setting with variable traffic densities is capable of stabilizing the ring in a similar fashion to the FollowerStopper
and PI with Saturation Controller. Among the four controller, the GRU controller allows the system to reach the uniform flow equilibrium velocity most
quickly. In addition, the FollowerStopper controller is the most brittle and can only stabilize a 260m ring road to a speed of 4.15 m/s, well below the 4.82
m/s uniform flow velocity.
9
Fig. 5: Prior to the activation of the autonomous vehicles, all settings exhibit backward propagating waves resulting from stop-and-go behavior. The autonomous
vehicles then perform various controlled policies aimed at attenuating these waves. Top left: Space-time diagram for the ring road with a FollowerStopper
Controller. Top Right: Space-time diagram for the ring road with a PI with Saturation Controller. Bottom left: Space-time diagram for the ring road with an
MLP Controller. Bottom right: Space-time diagram for the ring road with a GRU Controller.
stabilized settings with human model noise during test time.
Fig. 6: The performance of the MLP, GRU, and PI Saturation controllers for
various densities are averaged over ten runs for each of the tested densities.
The GRU and MLP controllers are capable of matching the uniform flow
speed very closely for all trained densities. The PI with Saturation controller,
on the other hand, is only capable of properly performing at densities less than
or equal to the density at which it was calibrated. Remarkably, the GRU and
MLP controllers are also reliable enough to stabilize the system at velocities
close to the uniform flow equilibrium even for densities outside the training
set.
Discussion: This benchmark study demonstrates that deep RL,
policy gradient methods in particular, using the same state
information provided to the hand-designed controllers and
with access to samples from the overall traffic system (via
a black box simulator), can learn a controller which performs
better than state-of-the-art hand-designed controllers for the
given setting, in terms of average system-level velocity.
This study focuses on the partially observed setting, since
it is the more realistic setting for near-term deployments. Furthermore, there are hand-designed controllers in the literature
for this setting, with which we can benchmark. We would
expect that the fully observed setting (with an MLP policy)
would perform as well if not better than our learned policies
in the partially observed setting. Since our policies already
closely track the equilibrium velocity curve, we do not explore
the fully observed setting.
VII. F LOW EXPERIMENTS
This section uses Flow to extend the settings typically
studied for controller design, presented in Section VI, to
include examples more representative of real-world traffic.
As traffic is a complex phenomena, it is crucial to build
tools to explore more complex and realistic settings. In the
following, we demonstrate these for three examples: a setting
with multiple autonomous vehicles, a multi-lane ring road
setting, and a figure eight road setting. A multi-lane ring
10
road is a simple closed network which extends the wellstudied single-lane ring road but adds in lateral dynamics (lane
changes).
Three natural settings for benchmarking are the fully autonomous setting, the mixed-autonomy setting, and the fullyhuman setting. Designed controllers for these settings may be
implemented in Flow for additional benchmarking as well.
In the following experiments, a fully observable setting is
assumed. The experiments are run on Amazon Web Services
(AWS) Elastic Compute Cloud (EC2) instances of model
c4.2xlarge, which have eight CPUs and 15 GB of memory.
the average speed of vehicles in the network. Figure 8 shows
the average velocity of vehicles in the network for different
levels of autonomy. With the inclusion of a single autonomous
vehicle. With one autonomous vehicle, the vehicles begin
moving around 1.5 times as fast as they they had when they
were forced to queue, while the fully autonomous setting
exhibits an improvement of almost three time to the average
velocities of vehicles.
Single-lane ring road with multiple autonomous vehicles:
In this experiment, a total of 22 vehicles are placed in a ring
road with a circumference of 230m. Strings of autonomous
vehicles are placed consecutively, with between three and
eleven autonomous vehicles. A string of consecutive autonomous vehicles learn to drive with a smaller headway than
the human models, resulting in greater roadway utilization,
thereby permitting a higher velocity for the overall system, as
can be seen in Figure 7.
Fig. 8: Velocity profile in the figure eight for different levels of autonomous
penetration. Similar to the platooning setting, the performance of the network
improves as the number of autonomous vehicles improves. In particular, the
inclusion of full autonomy almost triples the average velocity in the network
from around 5 m/s in the absence of autonomy for around 14 m/s.
For each of these benchmarks, more investigation is required
to understand the learned behaviors and policies and thereby
take steps towards a real-world deployment.
VIII. R ELATED W ORK
Fig. 7: Velocity profile for single-lane ring road with multiple autonomous
vehicles. The addition of autonomous vehicles permits in exceeding the
uniform flow equilibrium velocity. This velocity increases as the level of
autonomous penetration increases. At three autonomous vehicles, the average
velocity settles at 3.70 m/s; at 11 autonomous vehicles, the average velocity
settles at 4.44 m/s.
Multi-lane ring road with multiple autonomous vehicles:
The single-lane multiple vehicle experiment is extended to
the multiple lane setting, with 44 vehicles placed in a twolane ring road of circumference 230m. In this setting, a string
of six autonomous vehicles are initialized side-by-side (all in
one lane). The human-driven vehicles follow IDM longitudinal
control and SUMO’s lane changing model for lateral control.
The resulting average velocity is 3.66 m/s, an improvement
over the 3.45 m/s uniform flow equilibrium velocity. This
experiment demonstrates the reinforcement learning policy’s
ability to generalize to settings with discontinuous model
dynamics, such as lane changes.
Figure Eight: For this experiment, 14 vehicles are placed in
a figure eight with a ring radius of 30m and total length of
402m. The intersection in this environment is not controlled by
a traffic light; instead vehicles cross the intersection following
a right-of-way model provided by SUMO to prevent crashes.
In the absence of autonomous vehicles, human drivers begin
queuing at the intersection, leading to a significant reduction in
Traffic Dynamics: Modeling and analysis of traffic dynamics
is notoriously complex and yet is historically considered a prerequisite for traffic control [55], [57]. Researchers classically
trade away the complexity of the model (and thus the realism
of the model) in favor of the tractability of analysis, using high
level abstraction with the goal of designing optimal controllers
or other controllers with desirable properties, such as safety or
comfort [58], [59], [60], [61]. Consequently, results in traffic
control can largely be classified as small-scale simulationbased numerical analysis (for example, [62], [63], [64], [65])
or theoretical analysis on simple settings such as assuming
non-oscillatory responses (e.g. [66]) or focusing on a singlelane ring road (e.g. [67], [68], [69], [70], [71], [72]).
In particular, with the advent of autonomous vehicles, new
frameworks and techniques are urgently needed to establish
a foundation for studying the control and the effects of
autonomous vehicles, thereby preparing the world for their
adoption. Modern reinforcement learning techniques indicate
promise towards the goal of obtaining controllers with desirable (though perhaps not optimal) properties while simultaneously studying complex settings.
Deep RL and Traffic: Several recent studies incorporated
ideas from deep learning in traffic optimization. Deep RL
has been used for traffic prediction [73], [74] and control
[75]. A deep RL architecture was used in [74] to predict
traffic flows, demonstrating success even during special events
with nonlinear features; to learn features to represent states
11
involving both space and time, [73] additionally used hierarchical autoencoding in the traffic flow prediction problem.
A multi-agent deep RL algorithm was introduced in [75]
to learn a policy for ramp metering. For additional uses of
deep learning in traffic, we refer the reader to [76], which
presents an overview comparing non-neural statistical methods
versus neural networks in transportation research. These recent
results demonstrate that deep learning and deep RL are a
promising approach to traffic problems. This article aims to
bridge the gap between deep RL and traffic control problems
by providing a computational framework for learning wellperforming controllers; a preliminary prototype of our architecture is published in [77].
richer control actions allow Flow to support a larger class
of controllers, thus permitting a more realistic and suitable
testbed for reinforcement learning in traffic dynamics. SUMO
also includes a Python API called TRAffic Control Interface
(TraCI), from which the user can retrieve information about
the vehicles’ current states and issue precise commands to
set the vehicles’ velocities, positions, and lanes. Using this
interface, we can interface SUMO with RL libraries, read out
state information, issue actions, define our own car following
models, etc.
Traffic Simulators: Traffic microsimulators include Quadstone Paramics [78], VISSIM [79], [80], AIMSUN [81],
MATSIM [82], POLARIS [83], and SUMO [43]. The first
three are closed-source commercial software, whereas the
latter two are open source software. Each of these tools
are capable of large-scale traffic microsimulation and can
handle a variety of policies and control strategies. Each tool
offers an Application Programming Interface (API) which
permits overriding or extending the default models such as
car following, lane changing, route choice, etc. Each of these
simulators are widely used in the research community. These
tools differ in their precise offerings and features, such as
visualization tools, supported models, and simulation speed.
Because most studies focus their study on a single simulator,
a comprehensive comparison of these tools is unfortunately
lacking.
In the present work, we choose to integrate SUMO, an opensource, extensible, microscopic simulator that can simulate
large road networks. SUMO discretizes time and progresses
the simulation for a user-specified timestep; furthermore, because SUMO is microscopic, individual vehicles are controlled
by car following models—functions of the vehicle’s headway,
velocity and the velocity of the preceding vehicle. The acceleration provided by the car following model is applied as
a change of velocity over the course of the next timestep.
SUMO’s car following models include IDM, IDMM, and
Wiedermann.
SUMO has several current issues which limit its suitability
for RL. First, all SUMO built-in car following models are configured with a minimal time headway, τ , that is used to ensure
safety [84], and do not support time delays. Second, SUMO’s
car following models are calibrated for a simulation timestep
of 1.0 seconds, and their behavior for smaller timesteps is
known to produce unnatural behaviors [85] whereas we would
like to simulate at 10-100ms timesteps. Finally, there does
not yet exist an interface between SUMO and RL libraries.
Because the results of an RL experiment rely on the realism
of the model/simulator, we need the traffic models to capture
more realistic fine-grained dynamics, including operating at a
higher granularity (smaller simulation step), with a different
model of time delays, with acceleration-based control, etc.
Our work aims to address each of these limitations. Flow
extends SUMO to permit rich custom controllers which may
operate at smaller simulation steps and with time delays. These
Flow is a computational framework built on open source
tools; it enables learning policies for autonomous vehicles in
complex traffic settings involving nonlinear vehicle dynamics
and arbitrary network configurations. This article demonstrates
its capabilities and provides several concrete examples and
a case study which effectively benchmarks learned policies
against established control results. The expansion and combination of benchmark networks to additional network types,
including arbitrary grid networks, more complex intersections,
and importing arbitrary map networks, is the subject of
ongoing work, and will be operational soon (it is already
functional for simulation). More advanced RL algorithms
will be developed alongside larger networks because current
algorithms suffer poor sample complexity in the presence of
combinatorial structures such as graphs (road networks) [86]
and multiple agents [87]. Interesting and promising future directions include extending Flow to support additional features,
such as evaluating safety (in addition to efficiency), using
Flow as a tool to design specific controllers (which can be
interpreted or for which properties such as optimality can be
proven), and using it to inform public policy in preparation
for the increased adoption of autonomous vehicles. Finally,
as seen in many traffic management project led by State
agencies, microsimulation tools are the last step before field
implementation, which we hope to see for this work as well.
IX. C ONCLUSION
ACKNOWLEDGMENTS
The authors would like to thank Leah Dickstein and Nathan Mandi for
early prototyping, Nishant Kheterpal, Kathy Jang, Saleh Albeaik and Ananth
Kuchibhotla for helping to build out the features, Rocky Duan and Alex Lee
for rllab support, Jakob Erdmann for SUMO support, and Professor Alexander
Skabardonis for several insightful discussions about vehicle dynamics and failsafes. The team is extremely grateful to Professor Dan Work for technical
conversations about the ring experiment and work, and to the inspirational
work of the Piccoli-Seibold-Sprinkle-Work team.
A PPENDIX A
C LASSICAL CONTROLLERS
This section presents several classical controllers available
in Flow that have been and may be used to model human or
non-human driving behavior during experimentation.
12
A. Longitudinal controllers
Longitudinal Controllers: Longitudinal dynamics are usually
defined by car following models [67]. Standard car following
models (CFMs) are of the form:
ai = v̇i = f (hi , ḣi , vi ),
(5)
where the acceleration ai of vehicle i is some typically nonlinear function of hi , ḣi , vi , which are respectively the headway,
relative velocity, and velocity for vehicle i. A general model
may include time delays from the input signals hi , ḣi , vi to the
resulting output acceleration ai . Example CFMs include the
Intelligent Driver Model (IDM) [88] and the Optimal Velocity
Model (OVM) [89], [90]. Our presented system implements
several known CFMs and provides an easy way to implement
custom CFMs.
Custom longitudinal controllers can be implemented in Flow
using methods similar to the general car following model
equation (5) shown above, in which a vehicle’s acceleration
is some function of its speed, headway, and relative velocity.
Car following models are not limited to those inputs, however;
full access to the state of the environment at each timestep is
provided to controllers.
Out of the box, Flow supports a variety of car following
models, including SUMO default models and custom models
not provided by SUMO. Each model specifies the acceleration
for a vehicle at a given time, which is commanded to that
vehicle for the next time-step using TraCI.
Controllers with arbitrary time delays between perception
and action are supported in Flow. Delays are implemented by
storing control actions in a queue. For delayed controllers, a
new action is computed using the state at each timestep and
enqueued, and an action corresponding to some previous state
is dequeued and commanded. Descriptions of supported carfollowing models follow below.
1) Second-order linear model: The first, and simplest, car
following model implemented is the forward-looking car following model specified in [67]. The model specifies the acceleration of vehicle i as a function of a vehicle’s current position
and velocity, as well as the position and velocity of the vehicle
ahead. Thus: v̇i = kd (di − ddes ) + kv (vi−1 − vi ) + kc (vi − vdes )
where vi , xi are the velocity and position of the i-th vehicle,
di := xi−1 − xi is the headway for the i-th vehicle, kd , kc , kv
are controller gains for the difference between the distance to
the leading car and the desired distance, relative velocity, and
the difference between current velocity and desired velocity,
respectively. In addition, ddes , vdes are the desired headways
and velocities respectively.
2) Intelligent Driver Model: The Intelligent Driver Model
(IDM) is a microscopic car-following model commonly used
to model realistic driver behavior [88]. Using this model, the
acceleration for vehicle α is defined by its bumper-to-bumper
headway sα (distance to preceding vehicle), velocity vα , and
relative velocity ∆vα , via the following equation:
aIDM =
δ ∗
2
vα
s (vα , ∆vα )
dvα
=a 1−
−
dt
v0
sα
(6)
where s∗ is the desired headway of the vehicle, denoted by:
vα ∆vα
∗
(7)
s (vα , ∆vα ) = s0 + max 0, vα T + √
2 ab
where s0 , v0 , T, δ, a, b are given parameters. Typical values for
these parameters can be found in [88].
3) Optimal Velocity Model (OVM): Another car following
model implemented in Flow is the optimal velocity model from
[69]. A variety of optimal velocity functions exist for use in
specifying car following models [91], [55]; [69] uses a cosinebased function to define optimal velocity V (h) as a function
of headway:
h ≤ hst
0
h−hst
vmax
V (h) =
2 (1 − cos(π hgo − hst )) hst < h < hgo
vmax
h ≥ hgo
The values hst , hgo correspond to headway thresholds for
choosing an optimal velocity, so that for headways below hst ,
the optimal velocity is 0, and for headways above hgo , the
optimal velocity is some maximum velocity vmax . The optimal
velocity transitions using a cosine function for headways
between hst and hgo . V (h) is used in the control law for
the acceleration of the i-th vehicle, where v̇i = α[V (hi ) −
vi ] + β[vi−1 − vi ] at each timestep. This controller can also be
implemented with delay to simulate perception and reaction
times for human drivers, in which case v̇i (t) would be a
function of states hi (t − τ ), vi (t − τ ), vi−1 (t − τ ).
4) Bilateral control model (BCM): The bilateral controller
presented by [70], [71] considers not only the relation of a
subject vehicle to the vehicle ahead but also to the vehicle
behind it. In their controller, the subject vehicle’s acceleration
depends on the distance and velocity difference to both the
vehicle ahead and behind, with
v˙i = kd hi + kv ((vi−1 − vi ) − (vi − vi+1 )) + kc (vi − vdes )
where hi := (xi−1 − xi ) − (xi − xi+1 ). In [70], [71], Horn
and Wang argue that bilateral controllers can stabilize traffic.
Lateral Controllers: SUMO has lateral dynamics models
dictating when and how to lane change [92]; however, to
extend lateral control to the RL framework, Flow permits
the easy design of new and higher fidelity lane changing
models. The current implementation of Flow includes a proof
of concept lane-changing model in which vehicles change
lanes stochastically based on speed advantage when adjacent
lanes satisfy a set of constraints. Vehicles in Flow do not
check to change lanes at each timestep, as that might lead
to an excessive number of lane changes. Instead, at some
time interval, the vehicle determines if it should lane change.
SUMO’s existing lane-changing models can also be used in a
Flow experiment in place of custom models.
As with longitudinal controllers, custom lateral controllers
can also be built in Flow. These lane-changing models have
access to the full state of the environment at each time step to
use as potential inputs. This allows, for example, a vehicle to
identify all nearby vehicles in adjacent lanes and their speeds,
and then send a lane-change command if a lane is clear and
offers potentially higher speed. Due to the rich development
13
interface available, Flow supports the integration of complex
lateral controllers.
A PPENDIX B
A DDITIONAL EXPERIMENTS
This section uses Flow to benchmark the controllers described in the previous section. The experiments in this section
contain no learning components.
Single-lane Ring (all human driver models):
1) OVM from uniform initial state: The first experiment
runs the Sugiyama setup from an initial state in which all 22
vehicles were spaced evenly around the ring road and start with
the same velocity. Each of the vehicles was using a Optimal
Vehicle Model (OVM) controller, as described in the section
on controllers above. The experiment begins from a stopped
state, gets up to speed, and proceeds free of traffic shockwaves
for its duration.
A. OVM from a nonuniform motion state (Figure 11)
This experiment simulates the Sugiyama setup but from
a non-uniform initial configuration. Starting with the first
vehicle, the subsequent position of each vehicle is drawn
from a Gaussian distribution with mean equal to the length of
track divided by number of vehicles and a standard deviation
given by one fifth the mean. The unstable starting state also
incorporates a bunching factor, in which no vehicles are placed
on some segment of the track, with the length of that segment
being a user-defined variable. All 22 vehicles use the OVM
controller. Instability is apparent from the beginning, with
traffic rapidly degrading into traffic shockwaves and failing
to recover.
Fig. 11: 22 OVM vehicles from a nonuniform ring road initial state, showing
stop-and-go traffic with an average speed of 7.8 m/s.
Fig. 9: 22 OVM vehicles from uniform state on a ring road, showing an
average speed of 4.87 m/s across all vehicles.
2) OVM with a perturbation (Figure 10): In this experiment, 22 OVM vehicles are run from a uniform, evenly-spaced
starting state. No traffic shockwaves form until the system is
perturbed 9 seconds into the experiment, once the vehicles
have roughly reached their equilibrium velocities from the
unperturbed setting. One vehicle is randomly chosen and an
acceleration of −5 m/s2 is applied for 1.5 seconds. The
braking of that vehicle forces the vehicles behind it to slow
down as well, and the system degrades into stop-and-go traffic.
1) BCM with a perturbation (Figure 12): 22 vehicles implementing the bilateral car following model (BCM), described
in the controllers section, are implemented in this simulation.
The simulation begins from a uniform, evenly-spaced starting
state. As with the experiment above, a random vehicle is
perturbed at an acceleration of −5m/s2 , 9 seconds into the
simulation for 1.5 seconds. Some braking results, but unlike
the OVM case described above, the BCM vehicles recover
from this perturbation and traffic returns to uniform motion
shortly after.
Fig. 12: 22 BCM vehicles on a ring road with a perturbation, showing an
average speed of 7.9 m/s.
Fig. 10: 22 OVM vehicles on a ring road with a perturbation, breaking down
from uniform motion into stop-and-go traffic with an average speed of 7.5
m/s.
2) BCM from an nonuniform state (Figure 13): Again, 22
BCM vehicles are run in this simulation, but from the same
nonuniform starting state as in the nonuniform motion OVM
case, in which vehicles are spaced randomly subject to a
bunching factor. There is some initial instability and small
traffic shockwaves, but again the BCM vehicles recover from
this non-stable state and return to uniform motion.
14
rest at the final position of the rear bumper of the preceding vehicle. If the preceding vehicle is initially at position
xi−1 (0), and decelerates maximally, it will come to rest at
v 2 (0)
position xi−1 (0) + i−1
2a . Because the fail-safe issues the
maximum velocity, if the ego vehicle has delay τ , it will
first travel a distance of vsafe τ and then begins to brake with
maximum deceleration, which brings it to rest at position
safe
.
xi (0) + vsafe · τ + v2a
Fig. 13: 22 BCM vehicles from a nonuniform ring road initial state, showing
an average speed of 7.9 m/s.
3) Mixed BCM/OVM from a nonuniform initial state (Figure 14): Here, 11 BCM vehicles and 11 OVM vehicles
begin from a randomly spaced, and bunched starting state as
described above. The proportion of bilateral control vehicles
proves sufficient to prevent the stop-and-go waves seen in
the unstable OVM setting. Some velocity variation persists,
however, unlike the full-BCM unstable setting which returns
to a completely uniform motion state.
Fig. 14: 11 BCM and 11 OVM vehicles, from a nonuniform ring road initial
state, showing an average speed of 7.1 m/s.
A PPENDIX C
FAIL - SAFES
Flow supplements its car following models with safe driving
rules that prevent the inherently unstable car following models
from crashing. As SUMO experiments terminate when a
collision occurs, Flow provides a fail-safe mechanism, called
the final position rule, which runs constantly alongside other
controllers. Fail-safes are passed in the action commanded by
the vehicle controller, regardless of whether it is an action
specified by RL or a control model. Fail-safes are a standard
feature in any traffic simulator that is required to handle large
perturbations and string unstable traffic. The conservativeness
of the fail-safe affects the braking behavior of the traffic.
In general, fail-safes operate according to the principle of
maintaining a minimum safe distance from the leading vehicle
where the maximum acceleration and deceleration of the
leading vehicle is stochastically generated [93], [94].
Final Position Rule: This fail-safe aims to keep a velocity
such that if the preceding vehicle suddenly starts braking
with max deceleration a, then even if the following vehicle
has a delay τ it can still slow down such that it comes to
SUMO-Imposed Safety Behavior: In addition to incorporating its own safe velocity models, Flow leverages various safety
features from SUMO, which may also be used to prevent
longitudinal and lateral collisions. These fail-safes serve as
bounds on the accelerations and lane-changes human and
autonomous vehicles may perform, and may be relaxed on
any set of vehicles in the network to allow for the prospect of
more aggressive actions to take place.
R EFERENCES
[1] U. DOT, “National transportation statistics,” Bureau of Transportation
Statistics, Washington, DC, 2016.
[2] D. Schrank, B. Eisele, and T. Lomax, “Ttis 2012 urban mobility report,”
Texas A&M Transportation Inst.. The Texas A&M Univ. System, 2012.
[3] Z. Wadud, D. MacKenzie, and P. Leiby, “Help or hindrance? the travel,
energy and carbon impacts of highly automated vehicles,” Transportation Research Part A: Policy and Practice, vol. 86, pp. 1–18, 2016.
[4] R. E. Stern, S. Cui, M. L. D. Monache, R. Bhadani, M. Bunting,
M. Churchill, N. Hamilton, R. Haulcy, H. Pohlmann, F. Wu,
B. Piccoli, B. Seibold, J. Sprinkle, and D. B. Work, “Dissipation
of stop-and-go waves via control of autonomous vehicles: Field
experiments,” CoRR, vol. abs/1705.01693, 2017. [Online]. Available:
http://arxiv.org/abs/1705.01693
[5] Y. Sugiyama, M. Fukui, M. Kikuchi, K. Hasebe, A. Nakayama, K. Nishinari, S.-i. Tadaki, and S. Yukawa, “Traffic jams without bottlenecks–
experimental evidence for the physical mechanism of the formation of
a jam,” New Journal of Physics, vol. 10, no. 3, p. 033001, 2008.
[6] M. Garavello and B. Piccolli, Traffic flow on networks: Conservation
Laws Models. Springfield, MO: American Institute of Mathematical
Sciences, 2006.
[7] F. Wu, R. Stern, S. Cui, M. L. D. Monache, R. Bhadanid, M. Bunting,
M. Churchill, N. Hamilton, R. Haulcy, B. Piccoli, B. Seibold, J. Sprinkle,
and D. Work, “Tracking vehicle trajectories and fuel rates in oscillatory
traffic,” Transportation Research Part C: Emerging Technologies, 2017.
[8] M. Pavone, S. L. Smith, E. Frazzoli, and D. Rus, “Robotic load balancing
for mobility-on-demand systems,” The International Journal of Robotics
Research, vol. 31, no. 7, pp. 839–854, 2012.
[9] e. a. Cathy Wu, “Emergent behaviors in mixed-autonomy traffic,”
Conference on Robot Learning, 2017.
[10] F. Belletti, D. Haziza, G. Gomes, and A. M. Bayen, “Expert level
control of ramp metering based on multi-task deep reinforcement
learning,” CoRR, vol. abs/1701.08832, 2017. [Online]. Available:
http://arxiv.org/abs/1701.08832
[11] X.-F. Xie, S. F. Smith, L. Lu, and G. J. Barlow, “Schedule-driven
intersection control,” Transportation Research Part C: Emerging Technologies, vol. 24, pp. 168–189, 2012.
[12] S. Sukkarieh, E. M. Nebot, and H. F. Durrant-Whyte, “A high integrity
imu/gps navigation loop for autonomous land vehicle applications,”
IEEE Transactions on Robotics and Automation, vol. 15, no. 3, pp.
572–578, Jun 1999.
[13] M. W. M. G. Dissanayake, P. Newman, S. Clark, H. F. DurrantWhyte, and M. Csorba, “A solution to the simultaneous localization
and map building (slam) problem,” IEEE Transactions on Robotics and
Automation, vol. 17, no. 3, pp. 229–241, Jun 2001.
[14] Y. Cui and S. S. Ge, “Autonomous vehicle positioning with gps in urban
canyon environments,” IEEE Transactions on Robotics and Automation,
vol. 19, no. 1, pp. 15–25, Feb 2003.
[15] Z. Shiller and Y. R. Gwo, “Dynamic motion planning of autonomous
vehicles,” IEEE Transactions on Robotics and Automation, vol. 7, no. 2,
pp. 241–249, Apr 1991.
15
[16] S. D. Bopardikar, B. Englot, and A. Speranzon, “Multiobjective path
planning: Localization constraints and collision probability,” IEEE
Transactions on Robotics, vol. 31, no. 3, pp. 562–577, June 2015.
[17] J. Minguez and L. Montano, “Extending collision avoidance methods to
consider the vehicle shape, kinematics, and dynamics of a mobile robot,”
IEEE Transactions on Robotics, vol. 25, no. 2, pp. 367–381, April 2009.
[18] K. Kanatani and K. Watanabe, “Reconstruction of 3-d road geometry
from images for autonomous land vehicles,” IEEE Transactions on
Robotics and Automation, vol. 6, no. 1, pp. 127–132, Feb 1990.
[19] B. Van Arem, C. J. Van Driel, and R. Visser, “The impact of cooperative
adaptive cruise control on traffic-flow characteristics,” IEEE Transactions on Intelligent Transportation Systems, vol. 7, no. 4, pp. 429–436,
2006.
[20] J. Lee, J. Choi, K. Yi, M. Shin, and B. Ko, “Lane-keeping assistance
control algorithm using differential braking to prevent unintended lane
departures,” Control Engineering Practice, vol. 23, pp. 1–13, 2014.
[21] Y. S. Son, W. Kim, S.-H. Lee, and C. C. Chung, “Robust multirate
control scheme with predictive virtual lanes for lane-keeping system
of autonomous highway driving,” IEEE Transactions on Vehicular
Technology, vol. 64, no. 8, pp. 3378–3391, 2015.
[22] S. Lefevre, Y. Gao, D. Vasquez, H. E. Tseng, R. Bajcsy, and F. Borrelli,
“Lane keeping assistance with learning-based driver model and model
predictive control,” in 12th International Symposium on Advanced
Vehicle Control, 2014.
[23] G. Meyer and S. Shaheen, Eds., Disrupting Mobility: Impacts of Sharing
Economy and Innovative Transportation on Cities. Springer, 2017.
[24] D. Sadigh, S. Sastry, S. A. Seshia, and A. D. Dragan, “Planning for
autonomous cars that leverage effects on human actions.” in Robotics:
Science and Systems, 2016.
[25] J. Lee, M. Park, and H. Yeo, “A probability model for discretionary lane
changes in highways,” KSCE Journal of Civil Engineering, vol. 20, no. 7,
pp. 2938–2946, 2016.
[26] J. Rios-Torres and A. A. Malikopoulos, “A survey on the coordination
of connected and automated vehicles at intersections and merging at
highway on-ramps,” IEEE Transactions on Intelligent Transportation
Systems, vol. 18, no. 5, pp. 1066–1077, 2017.
[27] ——, “Automated and cooperative vehicle merging at highway onramps,” IEEE Transactions on Intelligent Transportation Systems,
vol. 18, no. 4, pp. 780–789, 2017.
[28] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller, “Playing atari with deep reinforcement learning,” arXiv preprint arXiv:1312.5602, 2013.
[29] J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “Highdimensional continuous control using generalized advantage estimation,”
arXiv preprint arXiv:1506.02438, 2015.
[30] J. Schulman, S. Levine, P. Abbeel, M. I. Jordan, and P. Moritz, “Trust
region policy optimization,” in ICML, 2015, pp. 1889–1897.
[31] N. Heess, G. Wayne, D. Silver, T. Lillicrap, T. Erez, and Y. Tassa,
“Learning continuous control policies by stochastic value gradients,” in
Advances in Neural Information Processing Systems, 2015, pp. 2944–
2952.
[32] M. Lai, “Giraffe: Using deep reinforcement learning to play chess,”
arXiv preprint arXiv:1509.01549, 2015.
[33] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling, “The arcade
learning environment: An evaluation platform for general agents,” J.
Artif. Intell. Res.(JAIR), vol. 47, pp. 253–279, 2013.
[34] C. Beattie, J. Z. Leibo, D. Teplyashin, T. Ward, M. Wainwright,
H. Küttler, A. Lefrancq, S. Green, V. Valdés, A. Sadik et al., “Deepmind
lab,” arXiv preprint arXiv:1612.03801, 2016.
[35] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, “Openai gym,” arXiv preprint
arXiv:1606.01540, 2016.
[36] G. Synnaeve, N. Nardelli, A. Auvolat, S. Chintala, T. Lacroix, Z. Lin,
F. Richoux, and N. Usunier, “Torchcraft: a library for machine learning
research on real-time strategy games,” arXiv preprint arXiv:1611.00625,
2016.
[37] E. Todorov, T. Erez, and Y. Tassa, “Mujoco: A physics engine for modelbased control,” in Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ
International Conference on. IEEE, 2012, pp. 5026–5033.
[38] B. Wymann, E. Espié, C. Guionneau, C. Dimitrakakis, R. Coulom, and
A. Sumner, “Torcs, the open racing car simulator,” Software available
at http://torcs. sourceforge. net, 2000.
[39] O. Vinyals, “Deepmind and blizzard to release starcraft ii
as an ai research environment,” https://deepmind.com/blog/
deepmind-and-blizzard-release-starcraft-ii-ai-research-environment/,
2016.
[40] M. Brackstone and M. McDonald, “Car-following: a historical review,”
Transportation Research Part F: Traffic Psychology and Behaviour,
vol. 2, no. 4, pp. 181–196, 1999.
[41] Z. Zheng, “Recent developments and research needs in modeling lane
changing,” Transportation research part B: methodological, vol. 60, pp.
16–32, 2014.
[42] A. Kotsialos, M. Papageorgiou, and A. Messmer, “Optimal coordinated
and integrated motorway network traffic control,” in 14th International
Symposium on Transportation and Traffic Theory, 1999.
[43] D. Krajzewicz, J. Erdmann, M. Behrisch, and L. Bieker, “Recent
development and applications of sumo-simulation of urban mobility,”
International Journal On Advances in Systems and Measurements,
vol. 5, no. 3&4, 2012.
[44] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel,
“Benchmarking deep reinforcement learning for continuous control,”
CoRR, vol. abs/1604.06778, 2016. [Online]. Available: http://arxiv.org/
abs/1604.06778
[45] R. Bellman, “A markovian decision process,” DTIC Document, Tech.
Rep., 1957.
[46] R. A. Howard, “Dynamic programming and markov processes,” 1960.
[47] R. S. Sutton, D. A. McAllester, S. P. Singh, Y. Mansour et al., “Policy
gradient methods for reinforcement learning with function approximation.” in NIPS, vol. 99, 1999, pp. 1057–1063.
[48] S. Haykin, Neural networks: a comprehensive foundation. Prentice
Hall PTR, 1994.
[49] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Gated feedback
recurrent neural networks,” in International Conference on Machine
Learning, 2015, pp. 2067–2075.
[50] P. G. Michalopoulos, D. E. Beskos, and Y. Yamauchi, “Multilane
traffic flow dynamics: some macroscopic considerations,” Transportation
Research Part B: Methodological, vol. 18, no. 4, pp. 377–395, 1984.
[51] A. Klar and R. Wegener, “A hierarchy of models for multilane vehicular
traffic i: Modeling,” SIAM Journal on Applied Mathematics, vol. 59,
no. 3, pp. 983–1001, 1998.
[52] A. Sasoh and T. Ohara, “Shock wave relation containing lane change
source term for two-lane traffic flow,” Journal of the Physical Society
of Japan, vol. 71, no. 9, pp. 2339–2347, 2002.
[53] C. F. Daganzo, “A behavioral theory of multi-lane traffic flow. part i:
Long homogeneous freeway sections,” Transportation Research Part B:
Methodological, vol. 36, no. 2, pp. 131–158, 2002.
[54] C. Wu, A. Kreidieh, E. Vinitsky, and A. Bayen, “Multi-lane reduction:
A stochastic single-lane model for lane changing,” in Submission, 2017.
[55] M. Treiber and A. Kesting, “Traffic flow dynamics,” Traffic Flow
Dynamics: Data, Models and Simulation, Springer-Verlag Berlin Heidelberg, 2013.
[56] ——, “The intelligent driver model with stochasticity-new insights into
traffic flow oscillations,” Transportation Research Procedia, vol. 23, pp.
174–187, 2017.
[57] M. Papageorgiou, C. Diakaki, V. Dinopoulou, A. Kotsialos, and Y. Wang,
“Review of road traffic control strategies,” Proceedings of the IEEE,
vol. 91, no. 12, pp. 2043–2067, 2003.
[58] P. A. Ioannou and C.-C. Chien, “Autonomous intelligent cruise control,”
IEEE Trans. on Vehicular technology, vol. 42, no. 4, pp. 657–672, 1993.
[59] A. Vahidi and A. Eskandarian, “Research advances in intelligent collision avoidance and adaptive cruise control,” IEEE transactions on
intelligent transportation systems, vol. 4, no. 3, pp. 143–153, 2003.
[60] Technical Committee ISO/TC 204, Intelligent transport systems, Intelligent transport systems – Adaptive Cruise Control systems – Performance
requirements and test procedures, ISO ISO 15 622:2010, 2010.
[61] E. W. Martin, K. Boriboonsomsin, N. D. Chan, N. Williams, S. A.
Shaheen, and M. Barth, “Dynamic ecodriving in northern california: A
study of survey and vehicle operations data from an ecodriving feedback
device,” in 92nd Annual Meeting of the Transportation Research Board,
Washington, DC, January, 2013.
[62] C.-Y. Liang and P. Huei, “String stability analysis of adaptive cruise
controlled vehicles,” JSME International Journal Series C Mechanical
Systems, Machine Elements and Manufacturing, vol. 43, no. 3, pp. 671–
677, 2000.
[63] A. Bose and P. A. Ioannou, “Analysis of traffic flow with mixed manual
and semiautomated vehicles,” IEEE Trans. on Intelligent Transportation
Systems, vol. 4, no. 4, pp. 173–188, 2003.
[64] P. A. Ioannou and M. Stefanovic, “Evaluation of acc vehicles in mixed
traffic: Lane change effects and sensitivity analysis,” IEEE Transactions
on Intelligent Transportation Systems, vol. 6, no. 1, pp. 79–89, 2005.
[65] M. A. S. Kamal, J.-i. Imura, T. Hayakawa, A. Ohata, and K. Aihara,
“Smart driving of a vehicle using model predictive control for improving
16
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
traffic flow,” IEEE Transactions on Intelligent Transportation Systems,
vol. 15, no. 2, pp. 878–888, 2014.
D. Swaroop, “String stability of interconnected systems: An application
to platooning in automated highway systems,” California Partners for
Advanced Transit and Highways (PATH), 1997.
G. Orosz, R. E. Wilson, and G. Stépán, “Traffic jams: dynamics
and control,” Philosophical Trans. of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, vol. 368, no. 1928,
pp. 4455–4479, 2010.
G. Orosz, J. Moehlis, and F. Bullo, “Delayed car-following dynamics
for human and robotic drivers,” in ASME 2011 International Design
Engineering Technical Conferences and Computers and Information in
Engineering Conference. American Society of Mechanical Engineers,
2011, pp. 529–538.
I. G. Jin and G. Orosz, “Dynamics of connected vehicle systems
with delayed acceleration feedback,” Transportation Research Part C:
Emerging Technologies, vol. 46, pp. 46–64, 2014.
B. K. Horn, “Suppressing traffic flow instabilities,” in Intelligent Transportation Systems-(ITSC), 2013 16th International IEEE Conference on.
IEEE, 2013, pp. 13–20.
L. Wang, B. K. Horn, and G. Strang, “Eigenvalue and eigenvector
analysis of stability for a line of traffic,” Studies in Applied Mathematics,
2016.
C. Wu, A. Bayen, and A. Mehta, “Stabilizing traffic with autonomous
vehicles,” in Submission, 2017.
Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Traffic flow
prediction with big data: a deep learning approach,” IEEE Transactions
on Intelligent Transportation Systems, vol. 16, no. 2, pp. 865–873, 2015.
N. G. Polson and V. O. Sokolov, “Deep learning for short-term traffic
flow prediction,” Transportation Research Part C: Emerging Technologies, vol. 79, pp. 1–17, 2017.
F. Belletti, D. Haziza, G. Gomes, and A. M. Bayen, “Expert level
control of ramp metering based on multi-task deep reinforcement
learning,” CoRR, vol. abs/1701.08832, 2017. [Online]. Available:
http://arxiv.org/abs/1701.08832
M. G. Karlaftis and E. I. Vlahogianni, “Statistical methods versus
neural networks in transportation research: Differences, similarities and
some insights,” Transportation Research Part C: Emerging Technologies,
vol. 19, no. 3, pp. 387–399, 2011.
C. Wu, K. Parvate, N. Kheterpal, L. Dickstein, A. Mehta, E. Vinitsky,
and A. Bayen, “Framework for control and deep reinforcement learning
in traffic,” in Submission, 2017.
G. D. Cameron and G. I. Duncan, “Paramicsparallel microscopic simulation of road traffic,” The Journal of Supercomputing, vol. 10, no. 1,
pp. 25–53, 1996.
M. Fellendorf, “Vissim: A microscopic simulation tool to evaluate
actuated signal control including bus priority,” in 64th Institute of
Transportation Engineers Annual Meeting. Springer, 1994, pp. 1–9.
M. Fellendorf and P. Vortisch, “Microscopic traffic flow simulator
vissim,” in Fundamentals of traffic simulation. Springer, 2010, pp.
63–93.
J. Casas, J. L. Ferrer, D. Garcia, J. Perarnau, and A. Torday, “Traffic simulation with aimsun,” in Fundamentals of traffic simulation. Springer,
2010, pp. 173–232.
K. N. Horni, A. and K. A. (eds.), The Multi-Agent Transport Simulation
MATSim. Ubiquity, London, 2016.
J. Auld, M. Hope, H. Ley, V. Sokolov, B. Xu, and K. Zhang, “Polaris:
Agent-based modeling framework development and implementation for
integrated travel demand and network and operations simulations,”
Transportation Research Part C: Emerging Technologies, vol. 64, pp.
101–116, 2016.
J. Erdmann. (2016) Simulation of urban mobility - wiki:
Car-following-models. [Online]. Available: http://sumo.dlr.de/wiki/
Car-Following-Models#tau
(2016) Simulation/basic definition. [Online]. Available: http://sumo.dlr.
de/wiki/Simulation/Basic Definition#Defining the Time Step Length
H. Dai, E. B. Khalil, Y. Zhang, B. Dilkina, and L. Song, “Learning
combinatorial optimization algorithms over graphs,” arXiv preprint
arXiv:1704.01665, 2017.
R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch, “Multiagent actor-critic for mixed cooperative-competitive environments,”
CoRR, vol. abs/1706.02275, 2017. [Online]. Available: http://arxiv.org/
abs/1706.02275
M. Treiber, A. Hennecke, and D. Helbing, “Congested traffic states in
empirical observations and microscopic simulations,” Physical review E,
vol. 62, no. 2, p. 1805, 2000.
[89] M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and Y. Sugiyama,
“Structure stability of congestion in traffic dynamics,” Japan Journal of
Industrial and Applied Mathematics, vol. 11, no. 2, pp. 203–223, 1994.
[90] ——, “Dynamical model of traffic congestion and numerical simulation,” Physical review E, vol. 51, no. 2, p. 1035, 1995.
[91] M. Batista and E. Twrdy, “Optimal velocity functions for car-following
models,” Journal of Zhejiang University-SCIENCE A, vol. 11, no. 7, pp.
520–529, 2010.
[92] J. Erdmann, “SUMO’s lane-changing model,” in Modeling Mobility with
Open Data. Springer, 2015, pp. 105–123.
[93] R. Dowling, A. Skabardonis, and V. Alexiadis, “Traffic analysis toolbox
volume iii: guidelines for applying traffic microsimulation modeling
software,” Tech. Rep., 2004.
[94] H. Yeo, A. Skabardonis, J. Halkias, J. Colyar, and V. Alexiadis, “Oversaturated freeway flow algorithm for use in next generation simulation,”
Transportation Research Record: Journal of the Transportation Research
Board, no. 2088, pp. 68–79, 2008.
| 3cs.SY
|
arXiv:1705.04596v1 [math.CO] 12 May 2017
A NOTE ON IDENTITIES IN PLACTIC MONOIDS AND
MONOIDS OF UPPER-TRIANGULAR TROPICAL MATRICES
ALAN J. CAIN, GEORG KLEIN, LUKASZ KUBAT, ANTÓNIO MALHEIRO,
AND JAN OKNIŃSKI
Abstract. This paper uses the combinatorics of Young tableaux to prove
the plactic monoid of infinite rank does not satisfy a non-trivial identity, by
showing that the plactic monoid of rank n cannot satisfy a non-trivial identity
of length less than or equal to n. A new identity is then proven to hold for the
monoid of n × n upper-triangular tropical matrices. Finally, a straightforward
embedding is exhibited of the plactic monoid of rank 3 into the direct product
of two copies of the monoid of 3 × 3 upper-triangular tropical matrices, giving
a new proof that the plactic monoid of rank 3 satisfies a non-trivial identity.
1. Introduction
Whether the plactic monoid (the monoid of Young tableaux, which is famed for
its ubiquity [Lot02, Chapter 5]) and its finite-rank variants satisfy non-trivial identities is an actively-researched open question [KO15, Izh17]. Various monoids that are
related to the plactic monoid satisfy non-trivial identities. For instance, the Chinese
monoid [CEK+ 01], which has the same growth type as the plactic monoid [DK94],
satisfies Adian’s identity xyyxxyxyyx = xyyxyxxyyx [JO11, Corollary 3.3.4]. (This
is the shortest non-trivial identity satisfied by the bicyclic monoid [Adi66, Chapter IV, Theorem 2].) Furthermore, various ‘plactic-like’ monoids of combinatorial
objects satisfy non-trivial identities [CM]: for instance, the hypoplactic monoid
(the monoid of quasi-ribbon tableaux [KT97, Nov00]), the sylvester monoid (binary search trees [HNT05]), and the Baxter monoid (pairs of twin binary search
trees [Gir12]).
Let plac denote the (infinite-rank) plactic monoid and let placn denote the plactic
monoid of rank n, for any n ∈ N. First, plac1 is monogenic and thus commutative
and so trivially satisfies xy = yx. The monoid plac2 is isomorphic to the Chinese
monoid of rank 2 and so satisfies Adian’s identity. The third and fifth authors
proved that plac3 satisfies the identity uvvuvu = uvuvvu, where u(x, y) and v(x, y)
are respectively the left and right side of Adian’s identity (and so the identity
uvvuvu = uvuvvu has sixty variables x or y on each side) [KO15, Theorem 2.6].
Furthermore, plac3 does not satisfy Adian’s identity [KO15, p. 111–2]. Izhakian
developed a theory that allows one to show that plac3 is isomorphic to a subdirect
The
first
author
was
supported
by
an
Investigador
FCT
fellowship
(IF/01622/2013/CP1161/CT0001).
For the first and fourth authors, this work was partially supported by by the Fundação para a
Ciência e a Tecnologia (Portuguese Foundation for Science and Technology) through the project
UID/MAT/00297/2013 (Centro de Matemática e Aplicações) and the project PTDC/MHCFIL/2583/2014. Much of the research leading to this paper was undertaken during visits by
the first and fourth authors to the University of Warsaw and these authors thank the university
for its hospitality.
1
2
A.J. CAIN, G. KLEIN, L. KUBAT, A. MALHEIRO, AND J. OKNIŃSKI
product of two copies of the monoid of 3×3 upper-triangular tropical matrices U3 (T)
[Izh17]. Since the same author had already proven that for any n ∈ N the monoid
of n × n upper-triangular tropical matrices Un (T) satisfies a non-trivial identity
[Izh14], this gives an alternative proof that plac3 satisfies a non-trivial identity. It
remains open whether placn satisfies a non-trivial identity for n ≥ 4.
This note contributes further to this topic. First, Section 3 uses the combinatorics of Young tableaux to prove that the infinite-rank plactic monoid plac does
not satisfy a non-trivial identity, in contrast to the Chinese, hypoplactic, sylvester,
and other monoids discussed above. Since plac is the union of all the finite-rank
plactic monoids placn , this implies that there is no non-trivial identity satisfied by
all the placn . However, it is still possible that there is a hierarchy of identities
satisfied by the various placn . Section 4 gives a new, simple, identity satisfied by
the monoid of n × n upper-triangular tropical matrices Un (T); the proof is based
on an approach developed by the fifth author [Okn15]. Finally, Section 5 gives an
elementary proof that plac3 embeds into the direct product of two copies of U3 (T);
this gives another proof that plac3 satisfies a non-trivial identity.
2. Preliminaries and notation
This section briefly recalls some definitions and sets up notation; see the references for further background.
An identity is a formal equality between two words in the free monoid, and is
non-trivial if the two words are distinct. A monoid M satisfies such an identity
if the equality holds in M under every substitution of letters in the words by
elements of M . For example, any commutative monoid satisfies the non-trivial
identity xy = yx.
For background on the plactic monoid, Young tableaux, and Schensted’s algorithm, see [Lot02, Ch.5]. Let A = {1 < 2 < · · ·} be the set of natural numbers
viewed as an infinite ordered alphabet. For n ∈ N, let An = {1 < 2 < · · · < n}
be the set of the first n natural numbers viewed as a finite ordered alphabet. For
any u ∈ A∗ , let Pplac (u) be the Young tableaux (the ‘plactic P-symbol’) computed
from u using Schensted’s algorithm. The plactic congruence ≡plac relates words in
A∗ that have the same image under u 7→ Pplac (u). The plactic monoid plac is the
factor monoid A∗/≡plac ; the plactic monoid of rank n is the factor monoid A∗n /≡plac
(where ≡plac is naturally restricted to A∗n ).
The tropical semiring is T = R ∪ {−∞}, and its addition and multiplication
operations ⊕ and ⊗ are defined by x ⊕ y = max{x, y} and x ⊗ y = x + y. The set
of n × n upper-triangular tropical matrices is denoted Un (T). (An upper-triangular
tropical matrix has all entries below the main diagonal equal to −∞, since this is
the additive identity of T.) For X, Y ∈ Un (T), write X ∼diag Y if X and Y have
equal corresponding diagonal entries.
Let ψ1 : Un (T) → Un−1 (T) be the map that sends a matrix Z ∈ Un (T) to the
matrix in Un−1 (T) obtained by erasing the rightmost column and bottommost row
of Z. Let ψ2 : Un (T) → Un−1 (T) be the map that sends a matrix Z ∈ Un (T) to the
matrix in Un−1 (T) obtained by erasing the leftmost column and topmost row of Z.
The maps ψ1 and ψ2 are monoid homomorphisms.
A NOTE ON IDENTITIES
3
3. The infinite-rank plactic monoid
Proposition 3.1. The plactic monoid placn does not satisfy any non-trivial identity
of length less than or equal to n.
Proof. Suppose, with the aim of obtaining a contradiction, that placn satisfies a nontrivial identity of length less than or equal to n. Without loss of generality, assume
that it satisfies such an identity over the variable set {x, y}, say u(x, y) = v(x, y) of
length equal to n (that is, with |u| = |v| = n). Interchanging u and v if necessary,
assume u is lexicographically less than v (where the variable set is ordered by
x < y). Suppose u(x, y) = u1 · · · un and v(x, y) = v1 · · · vn , where each ui and vi is
a variable x or y. If j is minimal such that uj and vj are different variables, then
uj is x and vj is y.
Let s = 12 · · · n ∈ A∗n and let t = 12 · · · (n − j)(n − j + 2) · · · n ∈ A∗n . (So
the word t does not contain the generator n − j + 1.) Since
placn satisfies the
identity u(x, y) = v(x, y), the equality u Pplac(s), Pplac (t) = v Pplac (s), Pplac (t)
holds. Thus the Young tableaux Pplac u(s, t) and Pplac v(s, t) are equal. In
particular, the height of the leftmost columns of these Young tableaux are equal.
By Schensted’s theorem [Sch61], the height of the leftmost column of Pplac (w) (for
w ∈ A∗ ) is equal to the length of the longest strictly decreasing subsequence in
w. Thus the lengths of the longest strictly decreasing subsequences of u(s, t) and
v(s, t) are equal.
Since s and t only contain symbols in {1, . . . , n}, the longest strictly decreasing
subsequences of u(s, t) and v(s, t) cannot be longer than n.
In u(s, t), there is a strictly decreasing subsequence of length n, namely n(n −
1) · · · 21, where, for i ∈ {1, . . . , n}, the symbol i is chosen from the instance of s or t
that was substituted for un−i+1 . Note in particular that since n−(n−j +1)+1 = j,
the symbol n − j + 1 is chosen from the instance of s substituted for uj (which is
the variable x). By the previous paragraph, n(n − 1) · · · 21 is a longest strictly
decreasing subsequence of u(s, t).
Thus v(s, t) must also have a strictly decreasing subsequence of length n. Since
each word s or t is strictly increasing, such a subsequence contains at most one
symbol chosen from the instance of s or t substituted for each vi . Since |v| = n,
this implies that exactly one symbol is chosen from each instance of s or t. Hence,
for i ∈ {1, . . . , n}, the symbol i must be chosen from the instance of s or t that was
substituted for vn−i+1 . In particular, since n − (n − j + 1) + 1 = j, the symbol
n − j + 1 is chosen from the instance of t substituted for vj (which is the variable
y). This is a contradiction, because the word t does not contain a symbol n − j + 1.
Therefore placn does not satisfy a non-trivial identity of length less than or equal
to n.
Since every finite-rank plactic monoid is a submonoid of the infinite-rank plactic
monoid, the following result is immediate from Proposition 3.1:
Theorem 3.2. The infinite-rank plactic monoid plac does not satisfy any nontrivial identity.
4
A.J. CAIN, G. KLEIN, L. KUBAT, A. MALHEIRO, AND J. OKNIŃSKI
4. A new identity for the monoid of upper-triangular tropical
matrices
Define
u0 (p, q) = p,
v0 (p, q) = q,
u1 (p, q) = pqppq,
v1 (p, q) = pqqpq,
and inductively define
un (p, q) = u1 un−1 (p, q), vn−1 (p, q) ,
vn (p, q) = v1 un−1 (p, q), vn−1 (p, q) .
for n ≥ 2. Note that the preceding equations hold for n = 1, but we only use them
as a definition for n ≥ 2. Note also that u1 (xy, yx) = v1 (xy, yx) is Adian’s identity.
Proposition 4.1. Let X, Y ∈ Un (T) with X ∼diag Y .
vn−1 (X, Y ).
Then un−1 (X, Y ) =
The proof strategy is similar to that used by the fifth author to show that Un (T)
satisfied a different set of identities [Okn15].
Proof. Proceed by induction on n. For n = 1, since the only entry of a 1 × 1
matrix is on its main diagonal, it follows from X ∼diag Y that X = Y ; hence
u0 (X, Y ) = X = Y = v0 (X, Y ). This is the base of the induction.
For the induction step, let n ≥ 2 and assume the result holds for n − 1. Let A =
un−2 (X, Y ) and B = vn−2 (X, Y ). Recall the homomorphisms ψ1 , ψ2 : Un (T) →
Un−1 (T) defined in Section 2. By the induction hypothesis,
un−2 ψ1 (X), ψ1 (Y ) = vn−2 ψ1 (X), ψ1 (Y ) and
un−2 ψ2 (X), ψ2 (Y ) = vn−2 ψ2 (X), ψ2 (Y ) .
Thus
ψ1 (A) = ψ1 un−2 (X, Y ) = ψ1 vn−2 (X, Y ) = ψ1 (B) and
ψ2 (A) = ψ2 un−2 (X, Y ) = ψ2 vn−2 (X, Y ) = ψ2 (B).
Therefore the matrices A and
can differ only in their (1, n)-th entries.
B
Denote the matrix C = cij by (c11 , c1n , cnn ; κ) for some suitable array κ representing the rest of C. That is,
c11
c1n
.
κ
(c11 , c1n , cnn ; κ) denotes
cnn
Note that if D = (d11 , d1n , dnn ; µ), then
CD = c11 + d11 , max{c11 + d1n , c1n + dnn , γ}, cnn + dnn ; ν ,
for some γ and some array ν. In fact, γ is the maximum of the other sums of pairs
in the scalar product of the first row of C and first column of D and so is only
dependent on κ and µ (and not on c11 , c1n , cnn , d11 , d1n , or dnn ) and ν is only
dependent on κ, µ, c11 and dnn (and not on c1n , cnn , d11 , or d1n ).
So A = (c, a, d; κ1 ) and B = (c, b, d; κ1 ) for some c, a, b, d ∈ T and some array
κ1 . Then
AB = 2c, max{b + c, a + d, γ2 }, 2d, κ2
A NOTE ON IDENTITIES
5
for some γ2 ∈ T and array κ2 . Next,
ABA = (AB)A = 3c, max{a + 2c, b + c + d, a + 2d, d + γ2 , γ3 }, 3d; κ3 ,
ABB = (AB)B = 3c, max{b + 2c, b + c + d, a + 2d, d + γ2 , γ3 }, 3d; κ3 ;
for some γ3 ∈ T and array κ3 ; note that the these are equal in (AB)A and (AB)B
because they depend only on equal corresponding entries of A and B. Finally,
ABAAB = (ABA)AB = 5c, max{b + 4c, a + 3c + d, 3c + γ2 , a + 2c + 2d,
b + c + 3d, a + 4d, 4d + γ2 , 2d + γ3 , γ5 }, 5d; κ5 ,
ABBAB = (ABB)AB = 5c, max{b + 4c, a + 3c + d, 3c + γ2 , b + 2c + 2d,
b + c + 3d, a + 4d, 4d + γ2 , 2d + γ3 , γ5 }, 5d; κ5 ;
for some γ5 ∈ T and array κ5 ; note that these are equal in (ABA)AB and (ABB)AB
because they depend only on equal corresponding entries of ABA and ABB. Note
that the expressions for ABAAB and ABBAB only differ in the terms a + 2c + 2d
and b + 2c + 2d, respectively.
Now consider the four possible relative orders of a and b and of c and d. In each
case, there is some element that appears in the max{. . .} terms of the expressions
for both ABAAB and ABBAB and that is greater than both a + 2c + 2d and
b + 2c + 2d:
a ≥ b ∧ c ≥ d =⇒ a + 3c + d ≥ a + 2c + 2d, b + 2c + 2d;
a ≥ b ∧ c ≤ d =⇒ a + 4d ≥ a + 2c + 2d, b + 2c + 2d;
a ≤ b ∧ c ≥ d =⇒ b + 4c ≥ a + 2c + 2d, b + 2c + 2d;
a ≤ b ∧ c ≤ d =⇒ b + c + 3d ≥ a + 2c + 2d, b + 2c + 2d.
So in each case, a + 2c + 2d and b + 2c + 2d are not the maximum terms in the
expressions for ABAAB and ABBAB. Thus u1 (A, B) = ABAAB = ABBAB =
v1 (A, B). Therefore un−1 (X, Y ) = u1 (A, B) = v1 (A, B) = vn−1 (X, Y ).
Hence, by induction, the result holds for all n.
Theorem 4.2. The monoid of n × n upper-triangular tropical matrices Un (T)
satisfies the identity un−1 (xy, yx) = vn−1 (xy, yx).
Proof. Let X, Y ∈ Un (T). Then XY ∼diag Y X; the result is immediate from
Proposition 4.1.
5. A tropical representation of plac3
For any w ∈ A+ and any p, q ∈ A, define:
the maximal length of a nondecreasing subsequence
wpq = in w with entries in the interval [p, q]
−∞
if p ≤ q,
if p > q.
∗
For the identity ε of A , define εpq = 0 if p = q and εpq = −∞ otherwise.
Define a map φn : A∗ → Un (T) by φn (w) = wpq .
Lemma 5.1. The map φn is a homomorphism.
Proof. Proceed by induction on n. If n = 1 then the claim is clear. Assume
that n ≥ 2 and recall the homomorphisms ψ1 , ψ2 : Un (T) → Un−1 (T) defined in
Section 2. Let ϑ1 = ψ1 ◦ φn and ϑ2 = ψ2 ◦ φn . Clearly, ϑ1 is simply φn−1 and is
6
A.J. CAIN, G. KLEIN, L. KUBAT, A. MALHEIRO, AND J. OKNIŃSKI
thus a homomorphism by the induction hypothesis. Furthermore, if w 7→ w is the
∗
map from {2, . . . , n} to A∗n−1 , extending a 7→ a − 1, then ϑ2 (w) = φn−1 (w) and
so is a homomorphism by the induction hypothesis.
By the definition of matrix multiplication in Un (T), it now remains to prove that
for every u, v ∈ Pn the (1, n)-th entries of φn (uv) and φn (u)φn (v) are equal for all
u, v ∈ A∗n . In other words,
(5.1)
(uv)1n = max{u11 + v1n , u12 + v2n , . . . , u1n + vnn }.
However, it is clear that the longest nondecreasing subsequence in the product
uv is of the form αβ, where for some j the word α is the longest nondecreasing
subsequence in u with entries in the interval [1, j] and β is the longest nondecreasing
subsequence of v with entries in the interval [j, n]. Therefore (5.1) indeed holds and
the result follows.
Thus, by induction on n, the map φn is a homomorphism from A∗n to Un (T).
Lemma 5.2. The map φn factors to give a homomorphism φn : placn → Un (T).
Proof. Notice that the plactic relations (baa, aba), (bab, bba), (bca, bac), (cab, acb)
(for a < b < c), with a, b, c ∈ A, preserve wpq for any w ∈ A∗ and any p, q ∈ A.
This is easy to see by a direct verification of the possible cases: the intersection
{a, b, c} ∩ [p, q] is either a single generator, or is equal to {a, b}, {b, c} or {a, b, c}.
It follows that if u, v ∈ A∗n are such that u ≡plac v, then φn (u) = φn (v).
Define fn : A∗n → A∗n as follows. For any i ∈ An , define
fn (i) = n(n − 1) · · · (i + 1)(i − 1) · · · 1,
and for w = a1 · · · ak ∈ A∗n , where ai ∈ A, define fn (w) = fn (ak ) · · · fn (a1 ).
It is straightforward to prove that the antihomomorphism fn factors to give an
antihomomorphism
fn : placn → placn .
Let F = Fij be the n × n tropical matrix with Fi,n+1−i = 0 for i = 1, 2, . . . , n
and all other entries being −∞. Define an involution πn : Un (T) → Un (T) by
x 7→ (F xF )T (conjugation by F composed with transposition); note that πn is an
antihomomorphism (and in fact an antiautomorphism, since it is bijective). Define
σn = πn ◦ φn ◦ fn ; since fn and πn are antihomomorphisms, σn is a homomorphism.
Finally, define
the homomorphism Ψ3 : plac3 → U3 (T) × U3 (T) by Ψ3 (u) =
φ3 (u), σ3 (u) .
Proposition 5.3. The map Ψ3 is an embedding.
Proof. An arbitrary element of plac3 has a unique representation by a word of the
form 3ζ 2δ 3ǫ 1α 2β 3γ , where ζ ≤ δ ≤ α and δ + ǫ ≤ α + β. (This is the ‘row reading’
of a Young tableau.) Notice that, by the definition of φ3 ,
α
α+β
α+β+γ
(5.2)
φ3 (3ζ 2δ 3ǫ 1α 2β 3γ ) = −∞ β + δ max{β + ǫ} + δ + γ .
−∞ −∞
ζ +ǫ+δ
An arbitrary element of plac3 also has a unique representative of the form
u = (321)k1 (21)k2 (31)k3 1k4 (32)k5 2k6 3k7 ,
A NOTE ON IDENTITIES
7
for some ki ∈ N ∪ {0} and either k4 = 0 or k5 = 0. (This is the column reading of
a Young tableau.) Straightforward computation shows that
(21)k7 (31)k6 (31)k5 (21)k5
(32)k3 (21)k3 (32)k2 (31)k2 (32)k1 (31)k1 (21)k1 if k4 = 0,
f3 (u) =
(21)k7 (31)k6 (32)k4 (32)k3
if k5 = 0;
(21)k3 (32)k2 (31)k2 (32)k1 (31)k1 (21)k1
thus
f3 (u) =
(
(321)2k1 +k2 +k3 +k5 (21)k7 (31)k6 1k5 2k3 3k2
(321)2k1 +k2 +k3 (21)k7 (31)k6 (32)k4 2k3 3k2
if k4 = 0,
if k5 = 0.
Regardless of whether k4 = 0 or k5 = 0, the row reading of u (or more precisely
Pplac (u)) has the form:
3k1 2k1 +k2 3k3 +k5 1k1 +k2 +k3 +k4 2k5 +k6 3k7 .
By (5.2), φ3 (u) determines the following numbers:
k7 , k5 + k6 , k1 + k2 + k3 + k4 ,
(5.3)
k1 + k2 , k1 + k3 + k5 , max{k3 + k5 , k5 + k6 }.
The row reading of f (u) (or more precisely Pplac f (u) ) has the form:
32k1 +k2 +k3 +k5 22k1 +k2 +k3 +k5 +k7 32k1 +k2 +k3 +k4 +k5 +k6 12k1 +k2 +k3 +2k5 +k6 +k7 2k3 +k4 3k2 .
Therefore, since πn is bijective, we know by (5.2) that σ3 (u) determines the following
numbers:
k2 , k3 + k4 , 2k1 + k2 + k3 + 2k5 + k6 + k7 ,
(5.4)
2k1 + k2 + k3 + k5 + k7 ,
4k1 + 2k2 + 2k3 + k4 + 2k5 + k6 ,
max{k3 + k4 , 2k1 + k2 + k3 + k4 + k5 + k6 }.
Consider another elemenent u′ ∈ plac3 , with
′
′
′
′
′
′
′
u′ = (321)k1 (21)k2 (31)k3 1k4 (32)k5 2k6 3k7 ,
and suppose that φ3 (u) = φ3 (u′ ) and σ3 (u) = σ3 (u′ ). By (5.3), k7 = k7′ . By (5.4),
k2 = k2′ . Consequently, (5.3) implies that k1 = k1′ . Thus,
k3 + k4 = k3′ + k4′ ,
k3 + k5 = k3′ + k5′ ,
k5 + k6 = k5′ + k6′ ,
respectively by (5.3), (5.4), and (5.3) together with k1 = k1′ .
If k4 = k4′ = 0, then k3 = k3′ , and so k5 = k5′ , and so k6 = k6′ ; thus u = u′ .
Similarly, if k5 = k5′ = 0, then k6 = k6′ and k3 = k3′ , and so k6 = k6′ ; thus u = u′ .
So suppose that k4 = 0 and k5′ = 0. Then
k3 = k3 + k4 = k3′ + k4′ = k3′ + k5′ + k4′ = k3 + k5 + k4′ ;
so k5 +k4′ = 0 and so k5 = 0. Hence k6 = k6′ , and k3 = k3′ and so k4 = k4′ . Therefore
u = u′ . Parallel reasoning shows that k4′ = 0 and k5 = 0 implies u = u′ .
Therefore φ3 (u) = φ3 (u′ ) and σ3 (u) = σ3 (u′ ) implies u = u′ .
Hence the map Ψ3 : plac3 → U3 (T) × U3 (T) with Ψ3 (u) = φ3 (u), σ3 (u) is an
embedding.
Since Ψ3 embeds plac3 into the direct product of two copies of U3 (T), the following result is immediate from Theorem 4.2:
Corollary 5.4. plac3 satisfies u2 (xy, yx) = v2 (xy, yx).
8
A.J. CAIN, G. KLEIN, L. KUBAT, A. MALHEIRO, AND J. OKNIŃSKI
References
[Adi66]
S. Adian. ‘Defining relations and algorithmic problems for groups and
semigroups’. Trudy Mat. Inst. Steklov,
85 (1966),
pp. 3–123. url:
http://www.mathnet.ru/eng/tm2766.
[CEK+ 01] J. Cassaigne, M. Espie, D. Krob, J.-C. Novelli, & F. Hivert. ‘The Chinese Monoid’.
Int. J. Alg. Comput., 11, no. 3 (2001), pp. 301–334. doi: 10.1142/S0218196701000425.
[CM]
A. J. Cain & A. Malheiro. ‘Identities in plactic, hypoplactic, sylvester, Baxter, and
related monoids’. arXiv: 1611.04151.
[DK94]
G. Duchamp & D. Krob. ‘Plactic-growth-like monoids’. In M. Ito & H. Jürgensen, eds,
Words, Languages and Combinatorics, II, p. 124–142, River Edge, NJ, 1994. World
Scientific.
[Gir12]
S. Giraudo. ‘Algebraic and combinatorial structures on pairs of twin binary trees’. J.
Algebra, 360 (2012), pp. 115–157. doi: 10.1016/j.jalgebra.2012.03.020.
[HNT05] F. Hivert, J.-C. Novelli, & J.-Y. Thibon. ‘The algebra of binary search trees’. Theoret.
Comput. Sci., 339, no. 1 (2005), pp. 129–165. doi: 10.1016/j.tcs.2005.01.012.
[Izh14]
Z. Izhakian. ‘Semigroup identities in the monoid of triangular tropical matrices’. Semigroup Forum, 88, no. 1 (2014), pp. 145–161. doi: 10.1007/s00233-013-9507-6.
[Izh17]
Z. Izhakian. ‘Tropical plactic algebra, the cloaktic monoid, and semigroup representations’. 2017. arXiv: 1701.05156v1.
[JO11]
J. Jaszuńska & J. Okniński. ‘Structure of Chinese algebras’. J. Algebra, 346, no. 1
(2011), pp. 31–81. doi: 10.1016/j.jalgebra.2011.08.020.
[KO15]
L. Kubat & J. Okniński. ‘Identities of the plactic monoid’. Semigroup Forum, 90, no. 1
(2015), pp. 100–112. doi: 10.1007/s00233-014-9609-9.
[KT97]
D. Krob & J.-Y. Thibon. ‘Noncommutative Symmetric Functions IV: Quantum Linear
Groups and Hecke Algebras at q = 0’. J. Algebraic Combin., 6, no. 4 (1997), pp. 339–
376. doi: 10.1023/A:1008673127310.
[Lot02]
M. Lothaire. Algebraic Combinatorics on Words. No. 90 in Encyclopedia of Mathematics and its Applications. Cambridge University Press, 2002.
[Nov00]
J.-C. Novelli. ‘On the hypoplactic monoid’. Discrete Mathematics, 217, no. 1-3 (2000),
pp. 315–336. doi: 10.1016/S0012-365X(99)00270-8.
[Okn15]
J. Okniński. ‘Identities of the semigroup of upper triangular tropical matrices’. Comm.
Algebra, 43, no. 10 (2015), pp. 4422–4426. doi: 10.1080/00927872.2014.946141.
[Sch61]
C. Schensted. ‘Longest increasing and decreasing subsequences’. Canad. J. Math., 13
(1961), pp. 179–191. doi: 10.4153/CJM-1961-015-3.
Centro de Matemática e Aplicações (CMA), Faculdade de Ciências e Tecnologia,
Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
E-mail address: [email protected]
Institute of Mathematics, University of Warsaw, Banacha 2, 02–097 Warsaw, Poland
E-mail address: [email protected]
Institute of Mathematics, University of Warsaw, Banacha 2, 02–097 Warsaw, Poland
E-mail address: [email protected]
Centro de Matemática e Aplicações (CMA), Faculdade de Ciências e Tecnologia,
Universidade Nova de Lisboa, 2829–516 Caparica, Portugal
Departamento de Matemática, Faculdade de Ciências e Tecnologia, Universidade
Nova de Lisboa, 2829–516 Caparica, Portugal
E-mail address: [email protected]
Institute of Mathematics, University of Warsaw, Banacha 2, 02–097 Warsaw, Poland
E-mail address: [email protected]
| 4math.GR
|
Prime factorization using quantum annealing and
computational algebraic geometry
arXiv:1604.05796v2 [quant-ph] 16 Jun 2016
Raouf Dridi∗ and Hedayat Alghassi†
1QB Information Technologies (1QBit)
458–550 Burrard Street
Vancouver, British Columbia, Canada V6C 2B5
June 17, 2016
Abstract
We investigate prime factorization from two perspectives: quantum annealing and
computational algebraic geometry, specifically Gröbner bases. We present a novel
autonomous algorithm which combines the two approaches and leads to the factorization of all bi-primes up to just over 200 000, the largest number factored to date
using a quantum processor. We also explain how Gröbner bases can be used to
reduce the degree of Hamiltonians.
1
Introduction
Prime factorization is at the heart of secure data transmission because it is widely believed
to be NP-complete. In the prime factorization problem, for a large bi-prime M , the task
is to find the two prime factors p and q such that M = pq. In secure data transmission,
the message to be transmitted is encrypted using a public key which is, essentially, a large
bi-prime that can only be decrypted using its prime factors, which are kept in a private
key. Prime factorization also connects to many branches of mathematics; two branches
relevant to us are computational algebraic geometry and quantum annealing.
To leverage the problem of finding primes p and q into the realm of computational algebraic
geometry, it suffices to transform it into an
S. This is done
P algebraic system of equations
P
using the binary representation p = 1 + i=1..sp 2i Pi and q = 1 + i=1..sq 2i Qi , which is
plugged into M = pq and expanded into a system of polynomial equations. The reader is
invited to read the sections Methods 4.1 and 4.2 for the details of this construction. The
system S is given by this initial system of equations in addition to the auxiliary equations
∗
†
[email protected]
[email protected]
1
expressing the binary nature of the variables Pi and Qi , carry-on, and connective variables.
The two primes p and q are then given by the unique zero of S. In theory, we can solve
the system S using Gröbner bases; however, in practice, this alone does not work, since
Gröbner basis computation (Buchberger’s algorithm) is exponential.
The connection to quantum annealing can also be easily described. Indeed, finding p
and q can be formulated into an unconstrained binary optimization problem (P), where
the cost function f is the sum of the squares of polynomials in S. The unique zero of S
now sits on the unique global minimum of (P) (which has minimum energy equal to zero).
There are, however, a few non-trivial requirements we need to deal with before solving
the cost function using quantum annealing. These requirements concern the nature of
cost functions that quantum annealers can handle. In particular, we would like the cost
function of (P) to be a positive quadratic polynomial. We also require that the coefficients
of the cost function be rather uniform and match the hardware-imposed dynamic range.
In the present paper, we suggest looking into the problem through both lenses, and
demonstrate that indeed this approach gives better results. In our scheme, we will be using quantum annealing to solve (P), but at the same time we will be using Gröbner bases
to help us reduce the cost function f into a positive quadratic polynomial f + with desired
values for the coefficients. We will be also using Gröbner bases at the important step of
pre-processing f + before finally passing it to the quantum annealer. This pre-processing
significantly reduces the size of the problem. The result of this combined approach is an
algorithm with which we have been able to factorize all bi-primes up to 2 × 105 using the
D-Wave 2X processor. The algorithm is autonomous in the sense that no a priori knowledge, or manual or ad hoc pre-processing, is involved. We refer the interested reader to
Supplementary materials 5 for a brief description of the D-Wave 2X processor, along with
some statistics for several of the highest numbers that we embedded and solved. More
detail about the processor architecture can be found in [JAG+ 11]. Another important
reference is the work of S. Boixo et al. in [BAS+ 13], which presents experimental evidence that the scalable D-Wave processor implements quantum annealing with surprising
robustness against noise and imperfections. Evidence that, during a critical portion of
quantum annealing, the qubits become entangled and entanglement persists even as the
system reaches equilibrium is presented in [LPS+ 14].
Relevant to us also is the work in [PS01], which uses algebraic geometry to solve optimization problems (though not specifically factorization; see Methods 4.4 for an adaptation
to factorization). Therein, Gröbner bases are used to compute standard monomials and
transform the given optimization problem into an eigenvalue computation. Gröbner basis computation is the main step in this approach, which makes it inefficient, given the
high complexity of Gröbner basis computation. In contrast to that work, we ultimately
solve the optimization problem using a quantum annealing processor and pre-process and
adjust the problem with algebraic tools, that is, we reduce the size of the cost function
and adjust the range of its parameters. However, we share that work’s point of view of
using real algebraic geometry, and our work is the first to introduce algebraic geometry,
and Gröbner bases in particular, to solve quantum annealing-related problems. We think
that this is a fertile direction for both practical and theoretical endeavours.
2
Mapping the factorization problem into a degree-4 unconstrained binary optimization
problem is first discussed in [Bur02]. There, the author proposes solving the problem using
a continuous optimization method he calls curvature inversion descent. Another related
work is the quantum annealing factorization algorithm proposed in [SS10]. We will discuss
it in the next section and improve upon it in two ways. The first involves the addition
of the pre-processing stage using Gröbner bases of the cost function. This dramatically
reduces the number of variables therein. The second way concerns the reduction of the
initial cost function, for which we propose a general Gröbner basis scheme that precisely
answers the various requirements of the cost function. In Results 2.2, we present our
algorithm (the column algorithm) which outperforms this improved algorithm (i.e., the
cell algorithm). Using a reduction proposed in [SS10] and ad-hoc simplifications and
tricks, the paper [XZL+ 12] reports the factorization of bi-prime 143 on a liquid-crystal
NMR quantum processor, which until now was the largest reported bi-prime number
factored in any quantum computation realization.
This review is far from complete without mentioning Shor’s algorithm [Sho97] and Kitaev’s
phase estimation [Kit95], which, respectively, solve the factorization problem and the
abelian hidden subgroup problem in polynomial time, both for the gate model paradigm.
The largest number factored using a physical realization of Shor’s algorithm is 15 [MNM+ 16];
see [SSV13] also for a discussion about oversimplification in the previous realizations. Finally, in [Rau13], it has been proved that contextuality is needed for any speed-up in a
measurement-based quantum computation factorization algorithm.
2
Results
The binary multiplication of the two primes p and q can be expanded in two ways: cellbased and column-based procedures (see Methods 4.1 and 4.2 ). Each procedure leads to
a different unconstrained binary optimization problem. The cell-based procedure creates
the unconstrained binary quadratic programming problem
P
(
minZ2 ij Hij 2 ,
(2.1)
(P1 )
with Hij := Qi Pj + Si,j + Zi,j − Si+1,j−1 − 2 Zi,j+1 ,
and the column-based procedure results in the problem
P
minZ2 1≤i≤(sp +sq +1) Hi 2 ,
sq +1+i−m
sq
(P2 )
i
P
P
P i j−i
with Hi :=
Qj Pi−j +
Zj,i − mi −
2 Zi,i+j .
j=0
j=1
(2.2)
j=1
The two problems (P1 ) and (P2 ) are equivalent. Their cost functions are not in quadratic
form, and thus must be reduced before being solved using a quantum annealer. The
reduction procedure is not a trivial task. In this paper we define, for both scenarios:
1) a reduced quadratic positive cost function and 2) a pre-processing procedure. Thus,
we present two different quantum annealing-based prime factorization algorithms. The
3
first algorithm’s decomposition method (i.e., the cell procedure, Methods 4.1 ) has been
addressed in [SS10], without pre-processing and without the use of Gröbner bases in the
reduction step. Here, we discuss it from the Gröbner bases framework and add the important step of pre-processing. The second algorithm, however, is novel in transformation of
its quartic terms to quadratic, outperforming the first algorithm due to its having fewer
variables.
We write R[x1 , . . . , xn ] for the ring of polynomials in x1 , . . . , xn with real coefficients and
V(f ) for the affine variety defined by the polynomial f ∈ R[x1 , . . . , xn ], that is, the set of
zeros of the equation f = 0. Since we are interested only in the binary zeros (i.e., xi ∈ Z2 ),
we need to add the binarization polynomials xi (xi − 1), where i = 1, ..., n, to f and obtain
the system S = {f, xi (xi − 1), i = 1, ..., n}. The system S generates an ideal I by taking
all linear combinations over R[x1 , . . . , xn ] of all polynomials in S; we have V(S) = V(I).
The ideal I reveals the hidden polynomials which are the consequence of the generating
polynomials in S. To
√ hidden polynomials is given by rthe so√ be precise, the set of all
called radical ideal √I, which is defined by I = {g ∈ R[x1 , . . . , xn ]| ∃r ∈ N : g ∈ I}.
In practice, the ideal I is infinite, so we represent such an √
ideal using a Gröbner basis B
which one might take to be a triangularization of the ideal I. In fact, the computation
of Gröbner√
bases generalizes Gaussian elimination
in linear systems. We also have V(S) =
√
V(I) = V( I) = V(B) and I(V(I)) = I. A brief review of Gröbner bases is given in
Methods 4.3.
2.1
The cell algorithm
Suppose we would like to define the variety V(I) by the set of global minima of an
unconstrained optimization problem minZn2 (f + ), where f + is a quadratic polynomial.
For instance, we would like f + to behave like f 2 . Ideally, we want f + to remain in
R[x1 , . . . , xn ] (i.e., not in a larger ring), which implies that no slack variables will be
added. We also want f + to satisfy the following requirements:
√
(i) f + vanishes on V (I) or, equivalently, f + ∈ I.
(ii) f + > 0 outside V (I), that is, f + > 0 over Zn2 − V (I).
(iii) Coefficients of the polynomial f + are adjusted with respect to the dynamic range
allowed by the quantum processor.
Let B be a Gröbner basis for I. We can then go ahead and define
X
f+ =
at t,
t∈B| deg(t)≤2
where the real
√ coefficients ai are subject to the requirements above; note that we already
+
have f ∈ I and thus the first requirement (i) is satisfied.
Let us apply this procedure to the optimization problem (P1 ) above. There, f = Hij
and the ring of polynomials is R[Pj , Qi , Si,j , Si+1,j−1 , Zi,j , Zi,j+1 ]. We obtain the following
Gröbner basis:
4
t1
:= Qi Pj + Si,j + Zi,j − Si+1,j−1 − 2 Zi,j+1 ,
t2
:= (−Zi,j+1 + Zi,j ) Si+1,j−1 + (Zi,j+1 − 1) Zi,j ,
t3
:= (−Zi,j+1 + Zi,j ) Si,j + Zi,j+1 − Zi,j+1 Zi,j ,
t4
:= (Si+1,j−1 + Zi,j+1 − 1) Si,j − Si+1,j−1 Zi,j+1 ,
t5
:= (−Si+1,j−1 − 2 Zi,j+1 + Zi,j + Si,j ) Qi − Si,j − Zi,j + Si+1,j−1 + 2 Zi,j+1 ,
t6
:= (−Si+1,j−1 − 2 Zi,j+1 + Zi,j + Si,j ) Pj − Si,j − Zi,j + Si+1,j−1 + 2 Zi,j+1 ,
t7
:= (−Zi,j+1 + Zi,j+1 Zi,j ) Qi + Zi,j+1 − Zi,j+1 Zi,j ,
t8
:= −Si+1,j−1 Zi,j+1 + Si+1,j−1 Qi Zi,j+1 ,
t9
:= (−Zi,j+1 + Zi,j+1 Zi,j ) Pj + Zi,j+1 − Zi,j+1 Zi,j ,
(2.3)
t10 := −Si+1,j−1 Zi,j+1 + Si+1,j−1 Pj Zi,j+1 .
We have used the lexicographic order plex(Pj , Qi , Si,j , Si+1,j−1 , Zi,j , Zi,j+1 ); see Methods
4.3 for definitions. Note that t1 = Hij . We define
X
X
Hij+ =
at t, that is, Hij+ =
ai ti ,
1≤i≤6
t∈B| deg(t)≤2
where the real coefficients ai are to be found. We need to constrain the coefficients ai
with the other requirements. The second requirement (ii), which translates into a set
of inequalities on the unknown coefficients ai , can be obtained through a brute force
evaluation of Hij and Hij+ over the 26 points of Z62 . The outcome of this evaluation is a set
of inequalities expressing the second requirement (ii) (see Supplementary materials 5 ).
The last requirement (iii) can be expressed in different ways. We can, for instance, require that the absolute values of the coefficients of Hij+ , with respect to the variables
Pj , Qi , Si,j , Si+1,j−1 , Zi,j , and Zi,j+1 , be within [1 − , 1 + ]. This, together with the set of
inequalities from the second requirement, define a continuous optimization problem and
can be easily solved. Another option is to minimize the distance between the coefficients to
one coefficient. The different choices of the objective function and the solution of the corresponding continuous optimization problem are presented in Supplementary materials 5 .
Having determined the quadratic polynomial Hij+ ∈ R satisfyies the important requirements (i, ii, and iii) above, we can now phrase our problem
(P1 ) as the equivalent quadratic
P
unconstrained binary optimization problem minZ2 ij Hij+ . Notice that this reduction is
performed only once for all cases; it need not to be redone for different bi-primes M .
Before passing the problem to the quantum annealer, we use Gröbner bases P
to reduce the
size of the problem. In fact, what we pass to the quantum annealer is H = NFB Hij+ ,
where NF is the normal form and B is now the Gröbner basis cutoff, which we discuss in
the next section. The largest bi-prime number that we embedded and solved successfully
using the cell algorithm is ∼35 000. The following table presents some bi-prime numbers M that we factored using the cell algorithm, the number of variables using both the
customized reduction CustR and the window-based GB reduction, the overall reduction
5
percentage R%, and the embedding and solving status inside the D-Wave 2X processor
Embed.
Cell algorithm
M
31861
34889
46961
150419
2.2
p×q
CustR GB R% Embed
√
211 × 151
111
95 14
√
251 × 139
111
95 14
311 × 151
125
109 13
×
431 × 349
143
125 12
×
The column algorithm (factoring up to 200 000)
The total number of variables in the cost function of the previous method is 2sp sq , before
any pre-processing. Here we present the column-based algorithm where the number of
variables (before pre-processing) is bounded by 1 + sp sq + log2 (sp )(sp + sq ). Recall that
here we are phrasing the factorization problem M = pq as
X
(P2 ) : minP1 ,...,Psp ,Q1 ,...,Qsq ,Z12 ,Z23 ,Z24 ,...∈Z2
Hi 2 ,
i
where Hi , for 1 ≤ i ≤ sp , is
Hi =
sq
X
j=0
Li
i
X
X
Qj Pi−j +
Zj,i − mi −
2j−i Zi,i+j
j=1
(Q0 = P0 = m0 = 1, Li = sq + 1 + i − mi ).
j=1
The cost function is of degree 4 and, in order to use quantum annealing, it must be
replaced with a positive quadratic polynomial with the same global minimum. The idea
is to replace the quadratic terms Qj Pi−j inside the different Hi with new binary variables
Wi−j,j , and add the penalty (Qj Pi−j − Wi−j,j )+ to the cost function (now written in terms
of the variables Wi−j,j ). To find (Qj Pi−j − Wi−j,j )+ , we run Gröbner bases computation
on the system
Qj Pi−j − W,
Qj 2 − Qj ,
(2.4)
2
P
−
P
,
i−j
i−j
Wi−j,j 2 − Wi−j,j .
Following the same steps as in the previous section, we get
(Qj Pi−j − Wi−j,j )+ = a(Qj Wi−j,j − Wi−j,j ) + b(Pi−j Wi−j,j − Wi−j,j ) + c(Pi−j Qj − Wi−j,j ),
with a, b, c ∈ R such that −a − b − c > 0, −b − c > 0, −a − c > 0, c > 0 (e.g., c = 1, a =
b = −2). The new cost function is now
X
X
H=
Hi (W )2 +
(Qj Pi−j − Wi−j,j )+ .
i
ij
6
We can obtain a better Hamiltonian by pre-processing the problem before applying the
W transformation. Indeed, let us first fix a positive integer cutoff ≤ (sp + sq + 1) and let
B ⊂ R [P1 , . . . , Psp , Q1 , . . . , Qsq , Z12 , Z23 , Z24 . . .] be a Gröbner basis of the set of polynomials
{Hi }i=1...cutoff ∪ {Pi (Pi − 1), Qi (Qi − 1), Zij (Zij − 1)}i,j .
In practice, the cutoff is determined by the size of the maximum subsystem of polynomials Hi on which one can run a Gröbner basis computation; it is defined by the hardware.
We also define a cutoff on the other tail of {Hi }, that is, we consider {Hi }i=2ndcutoff...(sp +sq +1) .
Notice that here we are working on the original Hi rather than the new Hi (W ). This is because we would like to perform the replacement Qj Pi−j → Wi−j,j after the pre-processing
(some of the quadratic terms might be simplified by this pre-processing). Precisely, what
we pass to the quantum annealer is the quadratic positive polynomial
X
2 X
H=
NFWi−j,j −LT(NFBc (Qj Pi−j )) (NFBc (Hi )) +
(Wi−j,j − LT (NFBc (Qj Pi−j )))+ .
ij
(2.5)
Here LT stands for the leading term with respect to the graded reverse lexicographic
order. The second summation is over all i and j such that LT (NFB (Qj Pi−j )) is still
quadratic. The outer normal form in the first summation refers to the replacement
LT (NFB (Qj Pi−j )) → Wi−j,j , which is again performed only if LT (NFB (Qj Pi−j )) is still
quadratic.
The columns of the following table present: a sample of bi-prime numbers and their prime
factors, the number of variables using each of a naı̈ve polynomial-to-quadratic transformation P2Q, our novel polynomial-to-quadratic transformation CustR, and our windowbased reduction GB after applying pre-processing. The overall reduction percentage R%
and the embedding and solving status in the D-Wave 2X processor Embed are also shown.
The adjacency matrix of the corresponding positive quadratic polynomial graph H and
its embedded pattern inside the Chimera graph of the D-Wave 2X processor for one of
the bi-primes are also depicted (see Figure 1). Details pertaining to use of the hardware
can be found in Supplementary materials 5.3.
Column Algorithm
M
150419
151117
174541
200099
223357
3
p×q
P 2Q CustR GB R% Embed
√
431 × 349 116
86
73 37
√
433 × 349 117
88
72 38
√
347 × 503 117
86
72 38
√
499 × 401 115
89
75 35
557 × 401 125
96
80 36
×
Discussion
In this work, factorization is connected to quantum annealing through binarization of the
long multiplication. The algorithm is autonomous in the sense that no a priori knowledge,
or manual or ad hoc pre-processing, is involved. We have attained the largest bi-prime
7
Figure 1: The column algorithm: the adjacency matrix pattern (left) and embedding into
the the D-Wave 2X quantum processor (right) of the quadratic binary polynomial for
M = 200 099.
factored to date using a quantum processor, though more-subtle connections might exist.
A future direction that this research can take is to connect factorization (as an instance of
the abelian hidden subgroup problem), through Galois correspondence, to covering spaces
and thus to covering graphs and potentially to quantum annealing. We believe that morerewarding progress can be made through the investigation of such a connection.
4
4.1
Methods
Column factoring procedure
Here we discuss the two single-bit multiplication methods of the two primes p and q.
The first method generates a Hamiltonian for each of the columns of the long multiplication expansion, while the second method generates a Hamiltonian for each of the
multiplying cells in the long multiplication expansion. The column factoring procedure of
p = 2sp Psp + 2sp−1 Psp−1 + ... + 2P1 + 1 and q = 2sq Qsq + 2sq−1 Qsq−1 ... + 2Q1 + 1 is depicted
in the following table:
8
···
Psp
msp+sq+1
...
···
···
···
···
Qsq−1Psp
Qsq Psp Qsq Psp−1
msp+sq msp+sq−1
Pi
···
Pi
···
Q1Pi−1
···
Q2Pi−2
..
···
· · · Qsq−1Pi−sq+1
···
Qsq Pi−sq
···
mi
Psp
Q1Psp−1
Q2Psp−2
..
Qsq−1Psp−sq+1
Qsq Psp−sq
msp
···
···
···
···
···
···
···
···
Psq
Qsq
Psq
Q1Psq−1
Q2Psq−2
..
Qsq−1P1
Qsq
msq
· · · P2 P1 P0 = 1
· · · Q2 Q1 Q0 = 1
· · · P2 P1
1
· · · Q1 P 1 Q1
· · · Q2
···
...
···
m2
m1 m0 = 1
The equation for an arbitrary column (i) can be written as the sum of the column’s multiplication terms (above) plus all previously generated carry-on terms from lower significant
columns (j < i). This sum is equal to the column’s bi-prime term mi plus the carry-ons
generated from higher significant columns. The polynomial equation for the i-th column
is
sq
P
Qj Pi−j +
j=0
i
P
Zj,i = mi +
j=1
Li
P
2j−i Zi,i+j (Q0 = P0 = m0 = 1) .
j=1
The above equation is used as the main column procedure’s equation Hi . The Hamiltonian
generation and reduction is discussed in detail in Results 2.2.
4.2
Cell bi-prime factoring procedure
In the cell multiplication procedure, the ultimate goal is to break each of the column equations discussed above into multiple smaller equations so that each equation contains only
one quadratic term. This not only simplifies the generation of quadratic Hamiltonians,
but also generates Hamiltonians with more-uniform quadratic coefficients in comparison
to the column procedure. The following table depicts the structure of the cell procedure:
.
...
Psp
· · · Pi+j · · ·
0
Q1 P n
Z1,n
Pn
Q1Pn−1
Z1,n−1
..
Z1,n+1
Q2Pn
Z2,n
S25
Q2Pn−1
Z2,n−1
S24
Q2P4
Z2,n−2
..
..
..
..
Zi,j+1
msq+sp+1
Zsq−1,sp+1
Qsq Psp
Zsq,sp
msq+sp
Zsq−2,sp+1
Qsq−1Psp
Zsq−1,sp
Ssq−1,sp−1
Qsq−1Psp−1
Zsq−1,sp−1
Ssq−1,sp−2
Qsq−1Psp−2
Zsq−1,sp−2
Ssq−1,sp−3
Qsq−1Psp−3
Zsq−1,sp−3
Ssq,sp−1
Qsq Psp−1
Zsq,sp−1
msq+sp−1
Ssq,sp−2
Qsq Psp−2
Zsq,sp−2
msq+sp−2
Ssq,sp−3
Qsq Psp−3
Zsq,sp−3
msq+sp−3
Ssq,sp−4
Qsq Psp−4
Zsq,sp−4
msq+sp−4
Si,j
Qi P j
Zi,j
Si+1,j−1
..
..
· · · mi+j · · ·
9
Psq
Qsq
Psq
Q1Psq
Z1,3
···
···
···
S2,sq−1
Q2Psq−1
Z2,sq−1
···
..
..
Ssq−1,1
Qsp−1P1
Zsq−1,1
Ssq−1,0
Qsq−1
0
msq−1
Ssq0
Qsq
0
msq
P3
Q3
P3
Q1P2
Z1,2
P2
Q2
P2
Q1 P 1
Z1,1
S2,1
Q2P1
Z2,1
S20
Q2
0
m2
...
P1
Q1
P1
Q11
0
m1
P0 = 1
Q0 = 1
m0 = 1
Each cell contains one of the total (sp − 1)(sq − 1) quadratic terms in the form of Qi Pj .
To chain a cell to its upper cell, one extra sum variable Si,j is added. Also, each carry-on
variable Zi,j in a cell is the carry-on of the cell directly to its right, so each cell contains
four variables. The sum of three terms Qi Pj , Si,j , and Zi,j is at most 3; thus, it generates an additional sum variable Si+1,j−1 and one carry-on variable Zi,j+1 . Therefore, the
equation for an arbitrary cell indexed (i, j), shown in the centre of the above table, is
Si,j + Qi Pj + Zi,j = Si+1,j−1 + 2Zi,j+1 .
As we can see, only six binary variables are involved in each cell equation and the equation
contains one quadratic term, so it can be transformed into a positive Hamiltonian without
adding slack variables. The Hamiltonian generation and reduction procedure is discussed
in detail in Results 2.1.
4.3
Gröbner bases
Good references for the following definitions are [Stu96, CLO98].
Normal forms. A normal form is the remainder of Euclidean divisions in the ring of
polynomials R[x1 , . . . , xn ]. Precisely, the normal form of a polynomial f ∈ R[x1 , . . . , xn ],
with respect to the set of polynomials B ⊂ R[x1 , . . . , xn ] (usually a Gröbner basis), is the
polynomial NF(f ) ∈ R[x1 , . . . , xn ], which is the image of f modulo B. It is the remainder
of the Euclidean of f by all g ∈ B.
Term orders. A term order on R[x1 , . . . , xn ] is a total order ≺ on the set of all monomials
xa = xa11 . . . xann , which has the following properties: (1) if xa ≺ xb , then xa+c ≺ xb+c
for all positive integers a, b, and c; (2) 1 ≺ xa for all strictly positive integers a. An
example of this is the pure lexicographic order plex(x1 , . . . , xn ). Monomials are compared
first by their degree in x1 , with ties broken by degree in x2 , etc. This order is usually
used in eliminating variables. Another example, is the graded reverse lexicographic order
tdeg(x1 , . . . , xn ). Monomials are compared first by their total degree, with ties broken by
reverse lexicographic order, that is, by the smallest degree in xn , xn−1 , etc.
Gröbner bases. Given a term order ≺ on R[x1 , . . . , xn ], then by the leading term (initial
term) LT of f we mean the largest monomial in f with respect to ≺. A (reduced) Gröbner
basis to the ideal I with respect to the ordering ≺ is a subset B of I such that: (1) the
initial terms of elements of B generate the ideal LT(I) of all initial terms of I; (2) for each
g ∈ B, the coefficient of the initial term of g is 1; (3) the set LT(g) minimally generates
LT(I); and (4) no trailing term of any g ∈ B lies in LT(I). Currently, Gröbner bases are
computed using sophisticated versions of the original Buchberger algorithm, for example,
the F4 algorithm by J. C. Faugère.
4.4
Factorization as an eigenvalue problem
In this section, for completeness, we describe how the factorization problem can be solved
using eigenvalues and eigenvectors. This is an adaptation of the method presented in
10
[PS01] to factorization, which is itself an adaption to real polynomial optimization of the
method of solving polynomial equations using eigenvalues in [CLO98].
Let H be in R[x1 , . . . , xn ] as in (2.5), where we have used the notation xi instead of the
P s, Qs, Zs, and W s. Define
X
Hα := H +
αi xi (xi − 1),
i
which is in the larger ring R[x1 , . . . , xn , α1 , . . . , αn ]. We also define the set of polynomials
C = {∂Hα /∂x1 , . . . , ∂Hα /∂xn , ∂Hα /∂α1 , . . . , ∂Hα /∂αn } .
The variety V(C) is the set of all binary critical points of H. Its coordinates ring is
the residue algebra A := R[x1 , . . . , xn , α1 , . . . , αn ]/C. We need to compute a basis for A.
This is done by first computing a Gröbner basis for C and then extracting the standard
monomials (i.e., the monomials in R[x1 , . . . , xn , α1 , . . . , αn ] that are not divisible by the
leading term of any element in the Gröbner basis). In the simple example below, we
do not need to compute any Gröbner basis since C is a Gröbner basis with respect to
plex(α, x). We define the linear map
mHα : A → A
g 7→ Hα g
Since the number of critical points is finite, the algebra A is always finite-dimensional by
the Finiteness Theorem ([CLO98]). Now, the key points are:
• The value of Hα (i.e., values of H), on the set of critical points V(C), are given by
the eigenvalues of the matrix mHα .
• Eigenvalues of mxi and mαi give the coordinates of the points of V(C).
• If v is an eigenvector for mHα , then it is also an eigenvector for mxi and mαi for
1 ≤ i ≤ n.
We illustrate this in an example. Consider M = pq = 5 × 3 and let
H = 2 + 7 x4 + 2 x3 + 2 x4 x3 − 2 x3 x 2 − x1 − 4 x4 x1 − 2 x3 x1 + x2 x1
be the corresponding Hamiltonian as in (2.5), where x1 = p2 , x2 = q1 , x3 = w2,1 , and
x4 = z2,3 . A basis for the residue algebra A is given by the set of the 16 monomials
{1, x4 , x3 , x4 x3 , x2 , x4 x2 , x3 x2 , x3 x2 x4 , x1 , x4 x1 , x3 x1 , x1 x3 x4 , x2 x1 , x4 x1 x2 , x1 x3 x2 , x1 x3 x2 x4 }.
11
The matrix mHα is
2 7
0 9
0 0
0 0
0 0
0 0
0 0
0 0
mHα :=
0 0
0 0
0 0
0 0
0 0
0 0
0 0
0 0
2
2
0 0 −2
0
4
0 0
4
9
0 0 −2
0
−1 −4 −2
0
−2
0
−5
0
0
0
1
0
0
−2 0
1
0
−3 −4 0
0
1
−7 0
0
0
0
0
0 13 0 0
0
−2
0
0
0
0
0
2 7
0
2
0
0
0
0
0 −4 −2
0
0
0 9
0
2
0
0
0
0
0 −4
0
0
0 0
2
9
0
0
0
0
0
0
−2
0
0
0 0
0
11
0
0
0
0
0
0
0
0
0
0 0
0
0
1
3
0
2
1
0
−2
0
0
0 0
0
0
0
4
0
2
0
1
0
0
0
0 0
0
0
0
0
1
5
0
0
−1
0
0
0 0
0
0
0
0
0
6
0
0
0
0
0
0 0
0
0
0
0
0
0
2
3
−2
0
0
0 0
0
0
0
0
0
0
0
5
0
0
0
0 0
0
0
0
0
0
0
0
0
0
0
0
0 0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
−2
−4
−6
0
−2
0
−1
2
0
5
5
We expect the matrix’s smallest eigenvalue to be zero and, indeed, we get the following
eigenvalues for mHα :
{0, 1, 2, 4, 5, 6, 9, 11, 13}.
This is also the set of values which Hα takes on V(C). The eigenvector v which corresponds
to the eigenvalue 0 is the column vector
v := (1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0)T .
This eigenvector is used to find the coordinates of x̂ ∈ V(C) that cancel (minimize) Hα .
The coordinates of the global minimum x̂ = (x̂1 , . . . , x̂n ) are defined by mxi v = x̂i v, and
this gives x1 = x2 = x3 = 1, x4 = 0, and α1 = 2α2 = α3 = 2, α4 = 5.
12
5
Supplementary materials
5.1
Continuous optimization problems for the requirements (ii–
iii)
In Results 2.1, we describe how a positive quadratic polynomial Hij+ can be extracted
using Gröbner bases. Here we provide the details of the calculation.
The second requirement (ii) is equivalent to each of the following linear polynomials being
greater than zero:
a1 , −a1 + a3 , −a1 − a4 , −a1 + a5 , −a1 + a6 , 2 a1 + a3 , 2 a1 − a4 , −a2 − a1 , −a2 + 2 a1 ,
−2 a1 + a3 + 2 a5 , −2 a1 + a3 + 2 a6 , −2 a1 − a4 + 2 a5 , −2 a1 − a4 + 2 a6 , −a1 + a5 + a6 ,
a1 + a3 − a5 , a1 + a3 − a6 , a1 − a4 − a5 , a1 − a4 − a6 , −a2 − 2 a1 + 2 a5 , −a2 − 2 a1 + 2
a6 , −a2 + a1 − a5 , −a2 + a1 − a6 , −2 a1 + a3 + 2 a5 + 2 a6 , −2 a1 − a4 + 2 a5 + 2 a6 ,
a1 + a3 − a5 − a6 , a1 − a4 − a5 − a6 , −a2 − 2 a1 + a3 − a4 , −a2 − 2 a1 + 2 a5 + 2 a6 ,
−a2 + a1 − a5 − a6 , −a2 + 3 a1 + a3 − a4 , −a2 − 3 a1 + a3 − a4 + 3 a5 ,
−a2 − 3 a1 + a3 − a4 + 3 a6 , −a2 + 2 a1 + a3 − a4 − 2 a5 , −a2 + 2 a1 + a3 − a4 − 2 a6 ,
−a2 − 3 a1 + a3 − a4 + 3 a5 + 3 a6 , −a2 + 2 a1 + a3 − a4 − 2 a5 − 2 a6
For the third requirement (iii), the first choice for the objective function f : R5 → R is
2
2
f (a1 , . . . , a5 ) = (−a1 + a5 + a6 )2 − 1 + (−2 a1 + a3 + 2 a5 + 2 a6 )2 − 1
2
2
+ (a1 − a2 − a5 − a6 )2 − 1 + (a1 − a4 − a5 − a6 )2 − 1
2
2
2
2
+ 2 a2 2 − 1 + a1 2 − 1 + 2 a3 2 − 1 + 2 a4 2 − 1
2
2
2
2
+ 2 a5 2 − 1 + 2 a6 2 − 1 + 4 a5 2 − 1 + 4 a6 2 − 1
The solution is a1 = 0.214, a2 = −1.082, a3 = 0.514, a4 = −1.082, a5 = 0.314, and
a6 = 0.314.
The second choice for f is
2
+ (−2 a1 + a3 + 2 a5 + 2 a6 )2 − a2
2
2
+ (a1 − a2 − a5 − a6 )2 − a2 + (a1 − a4 − a5 − a6 )2 − a2
2
2
2
2
+ 2 a2 2 − a2 + a1 2 − a2 + 2 a3 2 − a2 + 2 a4 2 − a2
2
2
2
2
+ 2 a5 2 − a2 + 2 a6 2 − a2 + 4 a5 2 − a2 + 4 a6 2 − a2
(5.1)
f (a1 , . . . , a5 ) =
(−a1 + a5 + a6 )2 − a2
2
The solution is a1 = 1.0, a2 = −4.0, a3 = 4.0, a4 = −4.0, a5 = 2.0, and a6 = 2.0 (identical
to the solution given in [SS10]).
5.2
Basic description of the quantum annealing processor
Here we introduce the quantum annealing concept that ultimately solves a general Ising
(quadratic unconstrained binary optimization, or ”QUBO”) problem, then talk about the
13
important topic of embedding a QUBO problem into the specific quantum annealer (the
D-Wave 2X processor).
Quantum annealing (QA), along with the D-Wave processors, have been the focus of
much research. We refer the interested reader to [JAG+ 11, CCD15, BAS+ 13, BRI+ 14,
LPS+ 14]. QA is a paradigm designed to find the ground state of systems of interacting
spins represented by a time-evolving Hamiltonian:
S(s) = E(s)HP −
HP = −
X
1X
∆(s)σix ,
2 i
hi σix +
i
X
Jij σiz σjz .
i<j
The parameters hi and Jij encode the particular QUBO problem P into its Ising formulation. QA is performed by first setting ∆ E, which results in a ground state into which
the spins can be easily initialized. Then ∆ is slowly reduced and E is increased until
E ∆. At this point the system is dominated by HP , which encodes the optimization
problem. Thus, the ground state represents the solution to the optimization problem.
An embedding is the mapping of the nodes of an input graph to the nodes of the destination graph. The graph representing the problem’s QUBO matrix needs to be embedded
into the actual physical qubits on the processor in order for it to solve the QUBO problem. The specific existing connectivity pattern of qubits in the D-Wave chip is called the
Chimera graph. Embedding an input graph (a QUBO problem graph) into the hardware
graph (the Chimera graph) is in general NP-hard ([Cho08]).
Figure 1–right shows an embedding of the (column algorithm) QUBO corresponding to
the bi-prime M = 200 099 into the Chimera graph of the D-Wave 2X chip consisting of
a 12 by 12 lattice of 4 by 4 bipartite blocks. The Chimera graph is structured so that
the vertical and horizontal couplers in its lattice are connected only to either side of each
bipartite block. Each node in this graph represents one qubit and each edge represents
a coupling between two qubits. Adjacent nodes in the Chimera graph can be grouped
together to form new effective (i.e., logical) nodes, creating nodes of a higher degree. Such
a grouping is performed on the processor by setting the coupler between two qubits to a
large negative value, forcing two Ising spins to align such that the two qubits end up with
the same values. These effective qubits are expected to behave identically and remain in
the same binary state at the time of measurement. The act of grouping adjacent qubits
(hence forming new effective qubits) is called chain creation or chain identification.
An embedding strategy consists of two tasks: mapping and identification. Mapping is the
assignment of the nodes of the input graph to the single or effective nodes of the destination
graph. Solving such problems optimally is in general NP-hard, but one can devise various
approximations and enhancement strategies to overcome these difficulties, for example,
using statistical search methods like simulated annealing, structure-based methods, or a
combination of both. For a better understanding of current embedding approaches, we
refer the reader to [Cho08, BCI+ 14, JWA14, TAA15]. In Figure 1–right, the blue lines
14
indicate the identified couplers, the yellow lines indicate the problem couplers (i.e., the
edges of the problem graph), and the grey lines indicate empty couplers.
5.3
Embedding and solving details
We have used one of the D-Wave 2X processors, DW2X SYS4, as our quantum annealing
solver. This processor operates at a temperature range of 26(±5) millikelvin (mK) and
has 1100 qubits with a 95.5-qubit yield. To utilize the processor, we used D-Wave’s
SAPI software development kit (version 2.2.1). To embed the problem graph into the
hardware graph we used the sapiFindEmbedding and sapiEmbedProblem modules, and
to solve the problems we used the sapiSolveIsing and sapiUnembedAnswer modules. For
all problems we opted for the maximum number of reads available (10 000) in order to
increase the fraction of ground state samples. The following table shows some statistics
of the embedding and solving stages for several of the highest numbers that we were able
to successfully embed and solve.
Embedding & Solving Statistics
M
31861
34889
150419
151117
174541
200099
n emT ry idC prC #qubits jRatio rT ime
95
33
848 721
815
10
3.52
95
27
803 740
833
10
3.52
73
1
941 830
902
64
3.52
72
7
1001 846
918
64
3.52
72
3
1004 897
966
64
3.52
75
5
884 824
897
64
3.52
In the above table, M stands for the bi-prime, n is the number of variables in the QUBO
problem, emTry is the number of block trials of the sapiFindEmbedding routine, idC
is the total number of identified couplers, prC is the total number of problem couplers,
max({|J |})
#qubits is the total number of (physical) qubits, jRatio is the ratio min({|Jijij|}) , and rTime
is the chip run time in seconds.
6
Acknowledgements
We would like to thank Pooya Ronagh for constructive discussions and helpful comments
on the paper. We also thank Marko Bucyk for proofreading the manuscript.
References
[BAS+ 13]
Sergio Boixo, Tameem Albash, Federico M. Spedalieri, Nicholas Chancellor,
and Daniel A. Lidar, Experimental signature of programmable quantum annealing, Nat Commun 4 (2013).
[BCI+ 14]
Zhengbing Bian, Fabian Chudak, Robert Israel, Brad Lackey, William G
Macready, and Aidan Roy, Discrete optimization using quantum annealing
on sparse ising models, Frontiers in Physics 2 (2014), no. 56.
15
[BRI+ 14]
Sergio Boixo, Troels F. Ronnow, Sergei V. Isakov, Zhihui Wang, David
Wecker, Daniel A. Lidar, John M. Martinis, and Matthias Troyer, Evidence for
quantum annealing with more than one hundred qubits, Nat Phys 10 (2014),
no. 3, 218–224.
[Bur02]
C.J.C. Burges, Factoring as optimization, Tech. Report MSR-TR-2002-83,
Microsoft Research, January 2002.
[CCD15]
Cristian S. Calude, Elena Calude, and Michael J. Dinneen, Guest column:
Adiabatic quantum computing challenges, SIGACT News 46 (2015), no. 1,
40–61.
[Cho08]
Vicky Choi, Minor-embedding in adiabatic quantum computation: I. the parameter setting problem, Quantum Information Processing 7 (2008), no. 5,
193–209.
[CLO98]
David A. Cox, John B. Little, and Donal O’Shea, Using algebraic geometry,
Graduate texts in mathematics, Springer, New York, 1998.
[JAG+ 11]
M. W. Johnson, M. H. S. Amin, S. Gildert, T. Lanting, F. Hamze, N. Dickson,
R. Harris, A. J. Berkley, J. Johansson, P. Bunyk, E. M. Chapple, C. Enderud,
J. P. Hilton, K. Karimi, E. Ladizinsky, N. Ladizinsky, T. Oh, I. Perminov,
C. Rich, M. C. Thom, E. Tolkacheva, C. J. S. Truncik, S. Uchaikin, J. Wang,
B. Wilson, and G. Rose, Quantum annealing with manufactured spins, Nature
473 (2011), no. 7346, 194–198.
[JWA14]
Cai Jun, G. Macready William, and Roy Aidan, A practical heuristic for
finding graph minors, Preprint arXiv:1406.2741 (2014).
[Kit95]
A.Yu. Kitaev, Quantum measurements and the abelian stabilizer problem,
quant-ph/9511026 (1995).
[LPS+ 14]
T. Lanting, A. J. Przybysz, A. Yu. Smirnov, F. M. Spedalieri, M. H. Amin,
A. J. Berkley, R. Harris, F. Altomare, S. Boixo, P. Bunyk, N. Dickson, C. Enderud, J. P. Hilton, E. Hoskinson, M. W. Johnson, E. Ladizinsky, N. Ladizinsky, R. Neufeld, T. Oh, I. Perminov, C. Rich, M. C. Thom, E. Tolkacheva,
S. Uchaikin, A. B. Wilson, and G. Rose, Entanglement in a quantum annealing processor, Phys. Rev. X 4 (2014), 021041.
[MNM+ 16] Thomas Monz, Daniel Nigg, Esteban A. Martinez, Matthias F. Brandl,
Philipp Schindler, Richard Rines, Shannon X. Wang, Isaac L. Chuang, and
Rainer Blatt, Realization of a scalable shor algorithm, Science 351 (2016),
no. 6277, 1068–1070.
[PS01]
Pablo A. Parrilo and Bernd Sturmfels, Minimizing polynomial functions, DIMACS Series in Discrete Mathematics and Theoretical Computer Science
(2001).
[Rau13]
Robert Raussendorf, Contextuality in measurement-based quantum computation, Phys. Rev. A 88 (2013), 022322.
16
[Sho97]
Peter W. Shor, Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer, SIAM J. Comput. 26 (1997), no. 5,
1484–1509. MR 1471990
[SS10]
Gernot Schaller and Ralf Schutzhold, The role of symmetries in adiabatic quantum algorithms, Quantum Information & Computation 10 (2010),
no. 1&2, 109–140.
[SSV13]
John A. Smolin, Graeme Smith, and Alexander Vargo, Oversimplifying quantum factoring, Nature 499 (2013), no. 7457, 163–165.
[Stu96]
Bernd Sturmfels, Gröbner bases and convex polytopes, University Lecture
Series, vol. 8, American Mathematical Society, Providence, RI, 1996. MR
1363949
[TAA15]
Boothby Tomas, D. King Andrew, and Roy Aidan, Fast clique minor generation in chimera qubit connectivity graphs, Preprint arXiv:1507.04774 (2015).
[XZL+ 12]
Nanyang Xu, Jing Zhu, Dawei Lu, Xianyi Zhou, Xinhua Peng, and Jiangfeng
Du, Quantum factorization of 143 on a dipolar-coupling nuclear magnetic
resonance system, Phys. Rev. Lett. 108 (2012), 130501.
17
| 0math.AC
|
NOETHERIANITY OF SOME DEGREE TWO
TWISTED COMMUTATIVE ALGEBRAS
arXiv:1501.06925v2 [math.AC] 29 Sep 2015
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
Abstract. The resolutions of determinantal ideals exhibit a remarkable stability property:
for fixed rank but growing dimension, the terms of the resolution stabilize (in an appropriate
sense). One may wonder if other sequences of ideals or modules over coordinate rings of
matrices exhibit similar behavior. We show that this is indeed the case. In fact, our main
theorem is more fundamental in nature: it states that certain large algebraic structures
(which are examples of twisted commutative algebras) are noetherian. These are important
new examples of large noetherian algebraic structures, and ones that are in some ways quite
different from previous examples.
Contents
1. Introduction
2. Generalities on tca’s
3. ModK and algebraic representations
4. Proof of the main theorems
5. A Gröbner-theoretic approach to the main theorems
References
1
5
8
15
17
20
1. Introduction
Let An be the coordinate ring of the space of symmetric bilinear forms on Cn , that is,
Sym(Sym2 (Cn )). Inside of Spec(An ) is the closed subset V (In,r ) of forms of rank at most r
defined by the determinantal ideal In,r . The resolution of An /In,r over An is explicitly known
by a classical result of Lascoux [L] (see also [We, Chapter 6]). The explicit description of the
resolution reveals an interesting feature: its terms stabilize as n grows. More precisely, the
decomposition of TorpAn (An /In,r , C) into irreducible representations of GLn is independent
of n for n ≫ p, r, when one appropriately identifies irreducibles of GLn with a subset of
those for GLn+1 .
Given this observation, one may wonder if the same phenomenon holds true more generally.
That is, suppose that for each n ≥ 0 we have a finitely generated GLn -equivariant An -module
Mn such that the M• are “compatible” in an appropriate sense. Do the resolutions of the
Mn stabilize?
Date: September 18, 2015.
2010 Mathematics Subject Classification. 13E05, 13A50.
SS was supported by a Miller research fellowship. AS was supported by NSF grant DMS-1303082.
1
2
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
The main result of this paper (Theorem 1.1) implies that the answer to this question is
“yes.” In fact, Theorem 1.1 establishes a more fundamental result: compatible sequences of
finitely generated equivariant An -modules are “noetherian” in an appropriate sense.
1.1. Statement of results. Instead of working with a compatible sequence of An -modules,
we prefer to pass to the limit in n and work with a single module over the ring Sym(Sym2 (C∞ )).
This ring, with its GL∞ action, is an example of a twisted commutative algebra (tca);
see §2.1 for the general definition. Given a tca A, there is a notion of (finitely generated) Amodule, and A is said to be noetherian if any submodule of a finitely generated A-module
is again finitely generated.
Our main result is the following theorem:
V
Theorem 1.1. The tca’s Sym(Sym2 (C∞ )) and Sym( 2 (C∞ )) are noetherian.
We also prove a variant of the above result. A bivariate tca is like a tca, but where the
group GL∞ × GL∞ acts. We prove:
Theorem 1.2. The bivariate tca Sym(C∞ ⊗ C∞ ) is noetherian.
Remark 1.3. Let FIM be the category whose objects are finite sets and where a morphism
X → Y is a pair (f, Γ) consisting of an injection f : X → Y and a perfect matching Γ on Y \
f (X). Then the category of Sym(Sym2 (C∞ ))-modules is equivalent to the category of FIMmodules over C (see [SS3, §4.3], where FIM is called the upwards Brauer category). Thus
Theorem 1.1 shows that finitely generated FIM-modules are noetherian. This is reminiscent
of the noetherianity result for FI-modules (see [CEF, Theorem 1.3]), but much more difficult.
There are analogous reinterpretations for the other two cases.
1.2. Motivation. We offer a few pieces of motivation for our work.
• Our main theorems generalize and place into the proper context the stability phenomena observed in the resolutions of determinantal ideals and related ideals (such
as those considered in [RW]).
• The algebras appearing in Theorems 1.1 and 1.2 are closely related to the representation theory of orthogonal and symplectic groups; for example, see [SS3] or Example 1.4 below. We believe our theorems will have useful applications in this area.
• The tca’s we consider provide important new additions to the growing list of large
noetherian algebraic structures; see §1.3 for further discussion.
• FIM-modules are formally very similar to the FI-modules studied in [CEF, CEFN].
Numerous examples of FI-modules have been found, and the noetherian property
for FI-modules often translates to interesting new theorems about the examples
(e.g., representation stability in the cohomology of configuration spaces). We do
not currently have analogous examples of FIM-modules, but when examples are
found (which we expect), Theorem 1.1 will yield interesting new results about them.
Example 1.4. For δ ∈ C define the Brauer category B(δ) as follows: objects are finite
sets, and morphisms are Brauer diagrams, where composition of Brauer diagrams uses the
parameter δ. One can regard FIM as a subcategory of B(δ), and from this one can deduce
noetherianity of B(δ)-modules from Theorem 1.1. Suppose that δ = n − m with integers n
and m. Then one obtains an interesting B(δ)-module by S 7→ (Cn|m)⊗S , where Cn|m is the
super vector space of the indicated super dimension. This module is closely connected to
the representation theory of the orthosymplectic Lie algebra osp(n|m). Our theorem shows
NOETHERIANITY OF SOME DEGREE TWO TCA’S
3
that any submodule of this module is finitely generated. The second and third author plan
to study B(δ)-modules more closely in a future paper, and the noetherian property will be
of foundational importance.
1.3. Connection to previous work. Theorem 1.1 fits into a theme that has emerged in
recent years where large algebraic structures have been found to be noetherian. See [AH,
Co, HS] for examples of S∞ -equivariant polynomial rings. Some other examples include ∆modules [Sn], FI-modules [CEF, CEFN] (see also [SS1]), FS-modules [SS4], VIC(R)-modules
[PS], and certain spaces of infinite matrices [DK, DE, Eg].
However, the noetherian results of this paper seem fundamentally more difficult than the
previous ones. We do not know how to make this observation precise, but offer the following
observation. One can almost always use Gröbner bases to reduce a noetherianity problem
in algebra to one in combinatorics (see [SS4]). In the previous noetherianity results, the
combinatorial problems ultimately concern words in a formal language, and can be easily
solved using Higman’s lemma. In contrast, the combinatorial problem that naturally arises
in the present case (Question 5.2) is graph-theoretic, and does not seem approachable by
Higman’s lemma. Alternatively, this division can be seen in terms of the asymptotics of
Hilbert functions: in the previous noetherian results, the Hilbert functions have exponential
growth, while in the present case the growth is super-exponential.
Due to this fundamental new difficulty, we have been forced to introduce new methods to
prove the main theorem. We believe these will be useful more generally.
1.4. Outline of proof. We now outline the proof of noetherianity for A = Sym(Sym2 (C∞ )).
Let K = Frac(A) and let Modtors
A be the category of torsion A-modules, where we say that an
A-module M is torsion if M ⊗A K = 0. If I is a non-zero ideal of A then the quotient tca A/I
is “essentially bounded,” and it is not difficult to conclude from this that A/I is noetherian
(see Proposition 2.4); it follows that finitely generated objects of Modtors
are noetherian.
A
,
which
we
denote by ModK .
We next consider the Serre quotient category ModA / Modtors
A
The intuition for ModK comes from the following picture, which is not rigorous:
The scheme Spec(A) is the space of symmetric bilinear forms on C∞ . A-modules
correspond to GL∞ -equivariant quasi-coherent sheaves on Spec(A). Torsion Amodules correspond to sheaves that restrict to zero on the open set U of nondegenerate forms. Thus objects of ModK correspond to equivariant quasi-coherent
sheaves on U. But such sheaves correspond to representations of O∞ , since GL∞
acts transitively on U with stabilizer O∞ .
The above reasoning is fraught with errors. Nonetheless, it leads to a correct statement: we
prove (Theorem 3.1) that ModK is equivalent to the category of algebraic representations of
O∞ , as defined in [SS3]. The results of [SS3] can therefore be transferred to ModK , and give
an essentially complete understanding of this category.
We would now like to piece together what we know about Modtors
and ModK to deduce
A
the noetherianity of A. However, the noetherianity of A is not a formal consequence of what
we have so far: we need to use more information about how ModA is built out of the two
and ModK . We proceed in three steps.
pieces Modtors
A
(1) We show that if M is a finitely generated torsion A-module then M admits a resolution
by finitely generated projective A-modules (Proposition 4.3). The essential input here
is [RW], which explicitly computes the resolutions of certain torsion modules.
4
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
(2) We next show that the section functor ModK → ModA , defined as the right adjoint of
the localization functor ModA → ModK , takes finite length objects of ModK to finitely
generated objects of ModA . This follows from step (1) and the structural results for
ModK (see Proposition 4.8).
(3) Finally, the noetherianity of A is deduced from (2), and our knowledge of Modtors
A and
ModK , by a short argument (see Theorem 4.9).
Remark 1.5. Let us offer some broader context for this proof. Suppose that X is a scheme
equipped with an action of a group G. We say that X is topologically G-noetherian if
every descending chain of G-stable Zariski closed subsets in X stabilizes. We say that X is
(scheme-theoretically) G-noetherian if the analogous statement holds for subschemes1 .
Suppose that U is a G-stable open subscheme of X, and let Z be the complement of U. One
would then like to relate the noetherianity of X to that of U and Z.
For topological noetherianity, there is no problem: if U and Z are topologically Gnoetherian then so is X (see [DK, §5]). This is a fundamental tool used in various topological
noetherianity results, such as [DE, DK, Eg]. Unfortunately, the analogous statement for
scheme-theoretic G-noetherianity does not hold: this is why we cannot directly conclude the
and ModK .
noetherianity of A from that of Modtors
A
The main technical innovation in this paper is our method for deducing (in our specific
situation) scheme-theoretic noetherianity of X from that of U and Z, together with some
extra information. This approach is likely to be applicable in other situations, and could
be very useful: for instance, if one could upgrade the topological results of [DK] to schemetheoretic results, it is likely that one could also get finiteness results for higher syzygies in
addition to results about equations (and not just set-theoretic equations).
1.5. Twisted graded-commutative algebras. One can define a notion of twisted gradedcommutative algebra, the basic examples being exterior algebras on finite length polynomial
representations of GL∞ . The noetherianity problem for these algebras is interesting, and has
applications similar to V
the commutative case. Transpose duality interchanges the algebras
∞
∞
Sym(C ⊗ C ) and (C∞ ⊗ C∞ ), and so noetherianity of the latter is an immediate
V
VV
consequence of Theorem 1.2. However, the noetherianity of (Sym2 (C∞ )) and ( 2 (C∞ ))
cannot be formally deduced from the results of this paper. We treat these algebras in a followup paper [NSS]. The main ideas are the same, but the details are more complicated: for
example,
while Sym(Sym2 (C∞ )) is closely related to the orthogonal group O∞ , the algebra
V
2
(Sym (C∞ )) is closely related to the periplectic superalgebra pe∞ .
1.6. Open questions. We list a number of open problems related to this paper.
(1) Theorem 1.1 states that the tca Sym(V ) is noetherian when V is an irreducible polynomial representation of degree 2. It would be natural to generalize this result by
allowing V to be a finite length representation of degree ≤ 2. Eggermont [Eg] has
shown that these tca’s are topologically noetherian (i.e., radical ideals satisfy the ascending chain condition). This suggests that they are all noetherian. However, new
ideas are needed to actually prove this.
(2) It is desirable to have results (either positive or negative) when V has degree > 2. One
might begin by trying to prove topological noetherianity for degree 3 representations.
The third author is currently investigating this with H. Derksen and R. Eggermont.
1One
should ask that all G-equivariant coherent sheaves are noetherian, not just the structure sheaf.
NOETHERIANITY OF SOME DEGREE TWO TCA’S
5
(3) Are the characteristic p analogs of the tca’s considered in this paper noetherian? Our
methods do not apply there. We point out that there are two versions of tca’s in
positive characteristic: one defined in terms of polynomial representations, and one
defined in terms of symmetric groups.
(4) Theorem 1.1 shows that A = Sym(Sym2 (C∞ )) is noetherian if we make use of the
GL∞ action. On the other hand, it is known that A is not noetherian if one only
makes use of the S∞ action [Dr, Example 2.4]. What happens for other groups? Is A
noetherian with respect to O∞ or Sp∞ ?
(5) In §4.2, we show that torsion modules over Sym(C∞ ⊗ C∞ ) satisfy the property (FT)
by appealing to [RW], which explicitly computes the resolutions of certain torsion
modules. We also show that torsion modules over Sym(Sym2 (C∞ )) satisfy (FT), but
deduce this by a rather clumsy argument from the previous case since the analog of
[RW] is not known in this case. We therefore believe that carrying out the analog of
[RW] for Sym(Sym2 (C∞ )) would be a worthwhile undertaking.
(6) Question 5.2 is an interesting and purely combinatorial question that is needed for the
Gröbner approach to Theorem 1.1.
1.7. Outline of paper. In §2, we review definitions and prove some general properties of
tca’s. These include generalities on the localization functor ModA → ModK and the section
functor S : ModK → ModA used in the proof of the main result. In §3 we prove that for the
specific algebras under consideration, the Serre quotient category ModK can be described
in terms of representations of infinite rank classical groups. The proofs of Theorem 1.1 and
Theorem 1.2 are in §4. Finally, §5 discusses an incomplete Gröbner theoretic approach to
the main theorems.
Remark 1.6. Transpose duality [SS2, §7.4] interchanges the two algebras in Theorem 1.1,
so it suffices to prove the noetherianity for either
V one. We give arguments for both when
convenient, but sometimes omit details for Sym( 2 (C∞ )).
2. Generalities on tca’s
2.1. Definitions. A representation of GL∞ is polynomial if it appears as a subquotient
of a (possibly infinite) direct sum of representations of the form (C∞ )⊗k . Polynomial representations are semi-simple, and the simple ones are the Sλ (C∞ ), where Sλ denotes the Schur
functor corresponding to the partition λ. A polynomial representation is said to have finite
length if it is a direct sum of finitely many simple representations. See [SS2] for details.
A twisted commutative algebra (tca) is a commutative associative unital C-algebra
A equipped with an action of GL∞ by C-algebra homomorphisms such that A forms a
polynomial representation of GL∞ . Alternatively, A is a polynomial functor from vector
spaces to commutative algebras [SS2, Theorem 5.4.1]; when used in this perspective, we use
A(V ) to denote its value on a vector space V .
We write |A| when we want to think of A simply as a C-algebra, and forget the GL∞ action.
An A-module is an |A|-module M equipped with an action of GL∞ that is compatible with
the one on A (i.e., g(ax) = (ga)(gx) for g ∈ GL∞ , a ∈ A, and x ∈ M) and such that M
forms a polynomial representation of GL∞ . An ideal of A is an A-submodule of A, i.e., a
GL∞ -stable ideal of |A|. We denote the category of A-modules by ModA . We write |M|
when we want to think of M as a module over |A|, forgetting its GL∞ -structure.
6
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
We say that A is finitely generated if |A| is generated as a C-algebra by the GL∞
orbits of finitely many elements. Equivalently, A is finitely generated if it is a quotient of
a tca of the form Sym(V ), where V is a finite length polynomial representation of GL∞ .
An A-module M is finitely generated if it is generated as an |A|-module by the GL∞
orbits of finitely many elements. Equivalently, M is finitely generated if it is a quotient of an
A-module of the form A ⊗ V , where V is a finite length polynomial representation of GL∞ .
We note that the A ⊗ V are exactly the projective A-modules. An A-module if noetherian
if every submodule is finitely generated. We say that A is noetherian (as an algebra) if
every finitely generated A-module is noetherian.
Remark 2.1. We say that A is weakly noetherian if it is noetherian as a module over itself,
i.e., if ideals of A satisfy ACC. Of course, noetherian implies weakly noetherian. However,
it is not clear if weakly noetherian implies noetherian: not every A-module is a quotient of
a direct sum of copies of A, due to the equivariance, and so there is no apparent way to
connect the noetherianity of A as an A-module to that of general modules.
There are “bivariate” versions of the above concepts. A representation of GL∞ × GL∞ is
polynomial if it appears as a subquotient of a (possibly infinite) direct sum of representations of the form (C∞ )⊗a ⊗ (C∞ )⊗b . Polynomial representations are again semi-simple, and
the simple ones are the Sλ (C∞ ) ⊗ Sµ (C∞ ). A bivariate tca is a commutative associative
unital C-algebra A equipped with an action of GL∞ × GL∞ by C-algebra homomorphisms
such that A forms a polynomial representation of GL∞ × GL∞ . The remaining definitions
in the bivariate case should now be clear.
Since GL∞ sits inside of GL∞ × GL∞ (diagonally), any action of GL∞ × GL∞ can be
restricted to one of GL∞ . Thus bivariate tca’s can be regarded as tca’s, and similarly for
modules. This restriction process preserves finite generation (of algebras and modules) since
the tensor product of finite length polynomial representations is again finite length.
2.2. Annihilators. Let A be a tca and M be an A-module. The annihilator of M, denoted
Ann(M), is the set of elements a ∈ A such that am = 0 for all m ∈ M. This is an ideal of
|A| and GL∞ stable, and thus an ideal of A.
Proposition 2.2. Let A be a tca and let M be an A-module. Suppose am = 0 for some a ∈ A
and m ∈ M. Then there exists an integer n, depending only on m, such that an (gm) = 0 for
all g ∈ GL∞ (C).
Proof. First, we claim that ak+1 Xk · · · X1 m = 0 for any X1 , . . . , Xk ∈ gl∞ . We proceed by
induction on k. The k = 0 case is simply the statement am = 0, which is given. Suppose
now that ak Xk−1 · · · X1 m = 0. Applying Xk , we obtain
kak−1 (Xk a)(Xk−1 · · · X1 m) + ak (Xk · · · X1 m) = 0.
Multiplying by a kills the first term and shows ak+1 Xk · · · X1 m = 0. This completes the
proof of the claim.
Let M ′ ⊂ M be the GL∞ representation generated by m. Suppose that a belongs to A(V )
with V ⊂ C∞ . Pick m′ ∈ M ′ that also generates M ′ and belongs to M ′ (U) with U ∩ V = 0.
We can write m′ = Xm for some X ∈ U(gl∞ ), and so the claim shows that an m′ = 0 for
some n. We claim that this n works for all elements in M ′ . Indeed, given any g ∈ GL∞ , we
can find f : C∞ → C∞ such that f agrees with g on U and is the identity on V . We then
have f∗ (a) = a and f∗ (m′ ) = gm′ , and so 0 = f∗ (an m′ ) = an (gm′ ).
NOETHERIANITY OF SOME DEGREE TWO TCA’S
7
Corollary 2.3. Suppose |A| is a domain. Let M be a finitely generated A-module such that
M ⊗A Frac(A) = 0. Then Ann M 6= 0.
Proof. Let m1 , . . . , mr be generators for M. Since M ⊗A Frac(A) = 0, we can find a 6= 0 in
A such that ami = 0 for all 1 ≤ i ≤ r. By the proposition, there exists n > 0 such that
an (gmi ) = 0 for all 1 ≤ i ≤ r and all g ∈ GL∞ . Thus 0 6= an ∈ Ann(M).
2.3. Essentially bounded tca’s. We say that a polynomial representation V of GL∞
is essentially bounded if there exist integers r and s such that for any simple Sλ (C∞ )
appearing in V we have λr ≤ s. Similarly, we say that a polynomial representation V of
GL∞ ×GL∞ is essentially bounded if there exist integers r and s such that for any simple
Sλ (C∞ ) ⊗ Sµ (C∞ ) appearing in V we have λr ≤ s and µr ≤ s. The Littlewood–Richardson
rule [SS2, (2.14)] implies that the tensor product of essentially bounded representations is
again essentially bounded. In particular, if V is an essentially bounded representation of
GL∞ × GL∞ then its restriction to the diagonal GL∞ is still essentially bounded. Note also
that any finite length representation is essentially bounded.
Proposition 2.4. Let A be a finitely generated and essentially bounded (bivariate) tca. Then
A is noetherian.
Proof. We treat only the univariate case, the bivariate case is similar. Let P be a finitely
generated projective A-module. Note that P is essentially bounded. We must show that P
is noetherian. Suppose that every partition appearing in P has at most r rows and at most
s columns. Let Cr|s be a super vector space with r-dimensional even part and s-dimensional
odd part.
For any symmetric monoidal category C and choice of object V ∈ C, there is a symmetric
monoidal functor Reppol (GL∞ ) → C that sends Sλ (C∞ ) to Sλ (V ). We apply this with C
the category of super vector spaces, equipped with the usual tensor product and the signed
symmetry (see [SS2, (7.3.3)]), and V = Cr|s . We thus obtain a natural map
{A-submodules of P } → {A(Cr|s )-submodules of P (Cr|s)}.
It follows from [BR, Theorem 3.20] that this map is injective. Since A(Cr|s ) is a finitely
generated superalgebra, the finitely generated module P (Cr|s) is noetherian. Thus the right
side satisfies ACC and so the left side does as well.
Remark 2.5. This argument is modeled on the discussion in [SS2, §9.1].
2.4. Serre quotients. Let A be a tca with |A| a domain, and let K = Frac(|A|). The
field K has an action of GL∞ , and we write |K| when we want to disregard this action. A
K-module is a |K|-vector space V equipped with a compatible action of GL∞ such that
V is spanned over |K| by polynomial elements (i.e., elements generating a polynomial Csubrepresentation). We write ModK for the category of K-modules. If V is a polynomial
representation of GL∞ then K ⊗ V is a K-module. All K-modules are quotients of such
K-modules. Note, however, that K ⊗ V is usually not projective as a K-module.
An A-module M is torsion if M ⊗A K = 0. Write Modtors
for the category of torsion
A
gen
tors
modules. We let ModA be the Serre quotient ModA / ModA , and we let T : ModA →
Modgen
A be the localization functor. The functor ModA → ModK given by M 7→ M ⊗A K
gen
is exact and kills Modtors
A , and thus induces an exact functor F : ModA → ModK . Note
that F (T (M)) = M ⊗A K, by definition. Given a K-module M, let S(M) = M pol be the
8
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
set of polynomial elements in M. This is naturally an A-module, and the resulting functor
S : ModK → ModA is right adjoint to F T . The following diagram summarizes the situation.
ModAdd■
t
T ttt
t
t
t
yy t
t
Modgen
A
F
■■
■■S
■■
■■
// ModK
Proposition 2.6. Let M be a K-module. Then the natural map M pol ⊗A K → M is an
isomorphism of K-modules. In particular, we have a natural isomorphism F T S = id.
Proof.
Injectivity is free. For surjectivity, pick v ∈ M. By definition, we can write v =
Pn
x
i=1 i wi , where xi ∈ K and wi ∈ M is polynomial. Again, by definition, we can write
xi = ai /bi ,Qwhere ai and bi are P
polynomial elements of K. We therefore have v = b−1 w,
n
where b = i=1 bi ∈ K and w = ni=1 (ai b/bi )wi is a polynomial element of M.
Proposition 2.7. The functor F is an equivalence of categories.
Proof. Proposition 2.6 shows that F has a right quasi-inverse, and so is therefore essentially
surjective and full. We show that F is faithful. Let M and N be A-modules, and consider
a morphism fe: M → N in Modgen
mapping to 0 in ModK . Write fe = T (f ) for some
A
′
′
morphism f : M → N/N in ModA , where M ′ and N ′ are submodules of M and N with
M/M ′ and N ′ torsion. Since f : M ′ ⊗ K → N/N ′ ⊗ K is 0, it follows that the image of f
is a torsion submodule of N/N ′ , and therefore of the form N ′′ /N ′ , where N ′′ is a torsionsubmodule of N containing N ′ . But then the image f ′ of f in Hom(M ′ , N/N ′′ ) is 0, and
since fe = T (f ) = T (f ′ ) we have fe = 0.
Proposition 2.8. Let W be a finite length polynomial representation with W GL∞ = 0. Set
A = Sym(W ). Then S(K ⊗ V ) = A ⊗ V for any polynomial representation V .
P
Proof. Suppose that x = si=1 (fi /g) ⊗ vi is a polynomial element of K ⊗ V , written in lowest
terms (that is, gcd(g, f1, . . . , fs ) = 1 and {v1 , . . . , vs } is linearly independent). Let m ≫ 0 be
such that g and each fi belong to A(Cm ), and let n = m + 1. We can think of x as a section
of a vector bundle on Cn having a pole along the divisorP
g = 0. Since x ∈ (K ⊗ V )pol , it
generates a finite dimensional representation of GLn . Let k (fj,k /gj ) ⊗ vj,k for 1 ≤ j ≤ r be
a basis; then every element can be written with common denominator g1 · · · gr . In particular,
the GLn -orbit of the divisor g = 0 is contained in g1 · · · gr = 0 and hence is finite. But GLn
is connected, so the irreducible components of g = 0 are preserved. Thus g is semi-invariant
under GLn . Any one-dimensional polynomial representation of GLn must be of the form
Sd,...,d(Cn ). But g ∈ A(Cm ) and is nonzero, and so it must be the case that g is actually
invariant under GLn (d must be zero because otherwise Sd,...,d(Cn ) = 0), and thus under
GL∞ . Since AGL∞ = C, we conclude that g is constant, and so x ∈ A ⊗ V , as required.
There is also a version of the above discussion for bivariate tca’s. The statements and
proofs are nearly identical.
3. ModK and algebraic representations
S
3.1. The main theorem and its consequences. A representation of O∞ = n≥1 On is
algebraic if it appears as a subquotient of a (possibly infinite) direct sum of tensor powers of
NOETHERIANITY OF SOME DEGREE TWO TCA’S
9
the standard representation C∞ . We write Rep(O∞ ) for the category of such representations.
This category was studied in [SS3].
We let A = Sym(Sym2 (C∞ )) and K = Frac(A) until §3.5. We let e1 , e2 , . . . be a basis
for C∞ , and let xi,j = ei ej , so that A = C[xi,j ]. Define m ⊂ |A| to be the ideal generated
by xi,i − 1 and xi,j for i 6= j. This ideal is not stable by GL∞ , but is stable by O∞ . The
e
quotient A/m is isomorphic to C. For an A-module M, define Φ(M)
= M/mM. This is
naturally a representation of O∞ . The main result of §3 is the following theorem (see §3.5
for analogous results in the other two cases):
e induces an equivalence of categories Φ : ModK → Rep(O∞ ).
Theorem 3.1. The functor Φ
We give the proof in the following subsections. The precise definition of Φ is given in §3.3.
For now, we note the following consequences of this theorem:
Corollary 3.2. We have the following:
(a) Finitely generated objects of ModK have finite length.
(b) If V is a finite length polynomial representation of GL∞ then K ⊗ V is a finite length
injective object of ModK , and all finite length injective objects have this form.
(c) Associating to λ the socle of K ⊗ Sλ (C∞ ) gives a bijection between partitions and
isomorphism classes of simple objects of ModK .
(d) Every finite length object M of ModK has a finite injective resolution M → I• where
each Ik is a finite length injective object.
Proof. These properties are proven for Rep(O∞ ) in [SS3]:
(a) [SS3, Proposition 4.1.5],
(b,c) [SS3, Proposition 4.2.9],
(d) dualize the explicit projective resolutions in [SS3, (4.3.9)].
3.2. Local structure at m of A-modules. The main result of this section is the following:
Proposition 3.3. Let M be an A-module. Then Mm is a free Am -module.
Let M∞ be the set of infinite complex matrices, indexed by Z≥0 . Let B ⊂ M∞ be the
set of upper triangular matrices, and let B ⊂ B be the group of invertible upper triangular
matrices. Let bi,j : M∞ → C be the function taking the (i, j) matrix entry. We let C[B] be
the polynomial ring C[bi,j ]i≤j , and we let C[B] = C[B][b−1
i,i ]. Elements of V ⊗ C[B] can be
thought of as (certain) functions B → V .
If V is a polynomial representation of GL∞ then every v ∈ V spans a finite dimensional
subrepresentation of B. It follows that we can give V the structure of a C[B]-comodule,
that is, we have a map V → V ⊗ C[B]. Explicitly, this map takes v to the function B → V
given by b 7→ bv. In fact, the image of the comultiplication map is contained in V ⊗ C[B].
Let H = B ∩ O∞ . Explicitly, H is the group of diagonal matrices with diagonal entries
±1, almost all of which are 1. If V is a polynomial representation of GL∞ then the map
V → V ⊗ C[B] above actually lands in the H-invariants of the target. Here we let B, and
H, act on C[B] by right translation.
Let M be an A-module. We then obtain a map
M → M ⊗ C[B] → M/mM ⊗ C[B].
10
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
The image lands in the H-invariants (note that m is H-stable, so H still acts on M/mM),
and so we have a map
ϕM : M → (M/mM ⊗ C[B])H .
We now study this map. We first treat the case where M = A. Then A/mA = C, and so
our map takes the form
ϕA : A → C[B]H
The invariant ring C[B]H is easily seen to be the subring of C[B] generated by the bi,j bi,k ,
with i ≤ j, k. Since B acts on A by algebra homomorphisms, the map A → A ⊗ C[B] is
an algebra homomorphism, and so ϕA is an algebra homomorphism as well. Due to this, it
suffices to understand where the generators xi,j go. For m ∈ B, we have
X
mei =
bk,i (m)ek ,
k≤i
and so
m(ei ej ) =
X
k≤i
bk,i (m)ek
!
X
bℓ,j (m)eℓ
ℓ≤j
Thus the map A → A ⊗ C[B] is given by
xi,j 7→
X
!
=
X
bk,i (m)bℓ,j (m)ek eℓ .
k≤i,ℓ≤j
bk,i bℓ,j xk,ℓ .
k≤i, ℓ≤j
To compute ϕA , we now apply the homomorphism A → A/mA = C, which takes xi,j to δi,j .
Set Xi,j = ϕ(xi,j ), we thus find
X
Xi,j = ϕ(xi,j ) =
bk,i bk,j .
k≤i,j
Proposition 3.4. The localization of ϕA at m is an isomorphism.
′
Proof. Let me be the extension of m to C[B]H via ϕA . Let i ≤ j. We have Xi,j = bi,i bi,j +Xi,j
,
′
where Xi,j is the sum of the bk,i bk,j with k < i. Since Xi,j ∈ m for i 6= j and Xi,i − 1 ∈ m,
an easy inductive argument shows that b2i,i − 1 ∈ m and bi,j bi,k ∈ m if i 6= j or i 6= k. In
particular, b2i,i is a unit in the localization. The expression Xi,j Xi,k = b2i,i bi,j bi,k + · · · (where
the missing terms involve only smaller variables) shows, inductively, that bi,j bi,k belongs to
the image of ϕA localized at m. Since these generate C[B]H , the result follows. (It is easy
to see that ϕA , and hence its localization, is injective.)
A monomial character of H is a homomorphism H → C× of the form (z1 , z2 , . . .) 7→
· · · where the ni are integers (it suffices to consider ni ∈ {0, 1}) and ni = 0 for i ≫ 0.
A representation of H is admissible if it is a sum of monomial characters.
z1n1 z2n2
Proposition 3.5. Let V be an admissible representation of H. The localization of (V ⊗
C[B])H at m is a free Am -module, and the fiber at m is canonically isomorphic to V .
Proof. It suffices to treat the case where V is one dimensional. Let χ = zi1 · · · zir be the
corresponding (monomial) character. Then an argument similar to the one in the proof of
Proposition 3.4 shows that for any nonzero v ∈ V , the element v ⊗ (bi1 ,i1 · · · bir ,ir ) is an
H invariant and the localization of (V ⊗ C[B])H at m is a free Am -module generated by
v ⊗ (bi1 ,i1 · · · bir ,ir ). The second statement follows immediately from this.
NOETHERIANITY OF SOME DEGREE TWO TCA’S
11
Now let M be an A-module, and consider the map
ϕM : M → (M/mM ⊗ C[B])H .
The target is naturally a module over the ring C[B]H , which is itself an A-algebra, and one
easily verifies that ϕM is a map of A-modules.
Proposition 3.6. Let M be an A-module. The localization of ϕM at m is an isomorphism.
Proof. Note that for any M, the quotient M/mM is an admissible representation of H.
Since such representations are semi-simple, it follows that the target of ϕM commutes with
direct limits in M. It therefore suffices to treat the case where M is finitely generated as an
A-module. Let N = (M/mM ⊗ C[B])H
m , and let R be the kernel of (ϕM )m . Since M/mM
is an admissible representation of H, Proposition 3.5 shows that N is a free Am -module
whose fiber at m is isomorphic to M/mM. It follows that (ϕM )m is a surjection, since it is a
surjection mod m and N is free. We thus have an isomorphism Mm = R ⊕ N, which shows
that R is finitely generated. Since (ϕM )m induces an isomorphism on the fiber at m, we see
that R/mR = 0. Thus R = 0 by Nakayama’s lemma, which completes the proof.
Proposition 3.3 follows from the above proposition, since as noted in the above proof, the
target of (ϕM )m is a free Am -module.
3.3. Definition of Φ. We begin with some simple observations.
e ⊗ V ) is isomorphic to
Lemma 3.7. If V is a polynomial representation of GL∞ then Φ(A
e
the restriction of V to O∞ . For any A-module M, Φ(M)
is an algebraic representation of
O∞ .
Proof. The first part is clear. For the second part, pick a surjection A ⊗ V → M of Ae is right exact, there is an induced surjection V → Φ(M).
e
modules. Since Φ
As any quotient
of an algebraic representation is algebraic, the result follows.
Lemma 3.8. If I is a non-zero ideal of A then I + m = A.
Proof. Suppose I is a non-zero ideal of A. Let An = Sym(Sym2 (Cn )), regarded as a subring
of |A|. Then A is the union of the An , and so for n sufficiently large, I ′ = I ∩ An is a nonzero GLn -stable ideal of An . Of course, m′ = m ∩ An is a maximal ideal of An . The scheme
Spec(An ) is the space of symmetric bilinear forms on Cn , and m′ ∈ Spec(An ) represents the
sum of squares form, which has maximal rank. Since V (I ′ ) is a proper closed GLn -stable
subset of Spec(An ), it cannot contain any form of maximal rank (as the orbit of any such
form is dense), and so I ′ 6⊂ m′ . It follows that I 6⊂ m, and so I + m = A.
e
Lemma 3.9. If M is a torsion A-module then Φ(M)
= 0.
e commutes with direct limits, it suffices to treat the case where M is finitely
Proof. Since Φ
generated and torsion. By Corollary 2.3, M has non-zero annihilator I, and I + m = A by
the Lemma 3.8. Thus
M/mM = M ⊗A/I (A/(I + m)) = 0.
e is exact.
Lemma 3.10. The functor Φ
Proof. This follows immediately from Proposition 3.3.
12
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
e is an exact functor killing Modtors . It follows that Φ
e factors through the Serre
Thus Φ
A
tors
quotient ModA / ModA , which we identify with ModK . In other words, there exists an exact
e
functor Φ : ModK → Rep(O∞ ), unique up to isomorphism, such that Φ(M)
= Φ(M ⊗A K).
e
Since Φ is compatible with direct limits, so is Φ.
3.4. Proof of Theorem 3.1. We now prove that Φ is an equivalence. We first prove that
it is faithful, then full, and finally essentially surjective.
Lemma 3.11. Φ is faithful.
e ) = 0. The square
Proof. Let f : M → N be a map of A-modules, and suppose Φ(f
M
ϕM
(M/mM ⊗ C[B])H
//
f
f ⊗1
N
ϕN
//
(N/mN ⊗ C[B])H
e ) = f = 0, the right map is 0. Since ϕM and ϕN are isomorphisms
commutes. Since Φ(f
after localizing at m, the induced map f : Mm → Nm is 0. This implies that the induced map
f : M ⊗A K → N ⊗A K is 0, and so f = 0 in ModK . This shows that Φ is faithful.
In what follows, we give GL∞ and B the direct limit topology (thinking of them as the
direct limits of GLn and B ∩ GLn in the Zariski topology).
Lemma 3.12. Let M be an A-module and let x ∈ M ⊗A K. Then there exists a dense
Zariski open subset V of B such that bx ∈ Mm for all b ∈ V .
Proof. Given h ∈ C[B], let Uh = {b ∈ B | h(b) 6= 0} be the corresponding Zariski open subset
of B. Let V = {b ∈ B | bx ∈ Mm }. We can find nonzero a ∈ A such that ax ∈ M. Note that
b ∈ UϕA (a) if and only if ba ∈
/ m; since ϕA (a) is not the zero function by Proposition 3.4, we
can find such a b. Then bx ∈ Mm and so V 6= ∅.
P
We claim that V is open. Suppose b ∈ V and write bx = i mi ⊗ (fi /c) with mi ∈ M,
fi ∈ A, and c ∈ A \ m. Then 1 ∈ UϕA (c) and b′ bx ∈ Mm for each b′ ∈ UϕA (c) . So UϕA (c) b ⊆ V ,
showing that V is open. Finally, since B is a directed union of irreducible spaces, a nonempty
open subset, like V , is dense.
Lemma 3.13. Let M be an A-module, and let x ∈ M ⊗A K. Suppose that there exists a
dense Zariski open subset U of B such that for all b ∈ U we have bx ∈ Mm and bx = 0,
where the overline denotes reduction mod m. Then x = 0.
Proof. Replacing x with ax, for an appropriate a ∈ A, it suffices to treat the case x ∈ M.
Then b 7→ bx defines a function B → M/mM which is continuous for the Zariski topology.
The hypothesis implies that it vanishes on a dense subset of B, and therefore it vanishes on
all of B. So ϕM (x) = 0, and so x = 0 since ϕM is injective after localizing at m.
Lemma 3.14. Let U be a dense Zariski open subset of B. Then for all g ∈ GL∞ the set
O∞ Ug −1 ∩ B contains a dense Zariski open subset of B.
Proof. Pick g ∈ GL∞ ; then g ∈ GLn for n large enough. Since On ∩ Un is a finite set,
the multiplication map On × Un → GLn has dense image (by a dimension count). Since it
is also constructible, it contains a dense open subset which we may assume is closed under
multiplication by On . In particular, we conclude that O∞ Ug −1 contains a Zariski dense open
NOETHERIANITY OF SOME DEGREE TWO TCA’S
13
subset V such that O∞ V = V . By a similar argument O∞ B contains a dense open subset
of GL∞ . This implies that V ∩ O∞ B is nonempty and hence there exists h ∈ O∞ such that
V ∩ hB is nonempty. Multiplying on the left by h−1 shows that V ∩ B is a nonempty open
subset of B. Since B is a directed union of irreducible spaces, we conclude that V ∩ B is a
dense open subset of B.
We now begin the proof of fullness. Let M and N be torsion-free A-modules and let
f : M/mM → N/mN be a map of O∞ representations. The diagram in Lemma 3.11 allows
us to define a map fm : Mm → Nm , and this induces a map f : M ⊗A K → N ⊗A K a |K|linear map. By definition, the map fm is characterized as follows: if x ∈ Mm and y ∈ Nm
then y = fm (x) if and only if f (bx) = by for all b ∈ B, where overlines denote reduction
modulo m. Using Lemma 3.13, we can say more: if f (bx) = by for all b in some dense Zariski
open subset U ⊂ B then y = fm (x). Indeed, putting y ′ = fm (x) we have f (bx) = by ′ for all
b ∈ B, and so by = by ′ for all b ∈ U, and so y = y ′ by the lemma. We now give a similar
characterization for f .
Lemma 3.15. Let x ∈ M ⊗A K and y ∈ N ⊗A K. Then y = f (x) if and only if the following
condition holds:
(∗) There exists a dense Zariski open dense subset U of B such that for all b ∈ U we
have bx ∈ Mm and by ∈ Nm and f (bx) = by.
Proof. Suppose y = f (x). Pick non-zero a ∈ A such that ax ∈ M. Let V be a dense Zariski
open subset of B such that ba ∈ Am and ba−1 ∈ Am and bx ∈ Mm and by ∈ Nm for all b ∈ V
(Lemma 3.12). Put z = f (ax). Since ax ∈ Mm we have z = fm (ax), and so bz = f (bax)
for all b ∈ B. For b ∈ V we have f (bax) = ba · f (bx) and bz = bay = ba · by, and so
ba · by = ba · f (bx). Since ba−1 ∈ Am , it follows that ba 6= 0, and so by = f (bx). So (∗) holds.
Now suppose (∗) holds. Let a be a non-zero element of A such that ax ∈ M. Let z = f (ax).
Since ax ∈ M, we have z = fm (ax), and so bz = f (bax) for all b ∈ B. Let V be a dense
Zariski open subset of B such that ba ∈ Am for all b ∈ V (Lemma 3.12). Then for b ∈ U ∩ V
we have bz = f (bax) = ba · f (bx) = ba · by = bay. It follows from Lemma 3.13 that z = ay,
and so ay = f (ax). Since f is K-linear, we conclude y = f (x).
Lemma 3.16. The map f : M ⊗A K → N ⊗A K is GL∞ -equivariant.
Proof. Let x ∈ M ⊗A K and let y = f (x) and let g ∈ GL∞ . We must show gy = f (gx). Let
U be a dense Zariski open subset of B such that bx ∈ Mm and by ∈ Nm and by = f (bx) for
all b ∈ U (Lemma 3.15). Let V = O∞ Ug −1 ∩ B, and let b ∈ V . We can then write bg = h′ b′
with h′ ∈ O∞ and b′ ∈ U. We have bgx = h′ b′ x ∈ Mm since b′ x ∈ Mm and Mm is stable by
O∞ . Similarly, bgy ∈ Nm . Furthermore,
f (bgx) = f (h′ b′ x) = h′ f (b′ x) = h′ b′ y = bgy.
This is the only place where we use the O∞ -equivariance of f . Since this holds for all
b ∈ V and V contains a dense Zariski open of B (Lemma 3.14), it follows that gy = f (gx)
(Lemma 3.15). This completes the proof.
We have shown that Φ is full. The following lemma completes the proof of the theorem.
Lemma 3.17. Φ is essentially surjective.
14
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
Proof. Since Φ is full and compatible with direct limits, it suffices to show that all finitely
generated objects of Rep(O∞ ) are in the essential image of Φ. Thus let M be such an object.
By the results of [SS3, §4], we can realize M as the kernel of a map f : I → J, where I and
J are injective objects of Rep(O∞ ). Every injective object of Rep(O∞ ) is the restriction to
O∞ of a polynomial representation of GL∞ . Thus I = Φ(M) and J = Φ(N) for some M
and N in ModK , and f = Φ(f ′ ) for some f ′ : M → N in ModK . The exactness of Φ shows
that M ∼
= Φ(ker(f ′ )), and so Φ is essentially surjective.
V
3.5. The other two cases. Everything in this section can be adapted to Sym( 2 (C∞ )).
This is straightforward (and not even logically necessary, per Remark 1.6), so we do not
comment further on it.
Everything can also be adapted to the bivariate tca A = Sym(C∞ ⊗ C∞ ). We will make a
few comments on how this goes. First, we state the analogs of Theorem 3.1 and Corollary 3.2.
A representation of GL∞ is algebraic if it appears as a subquotient of a (possibly infinite)
⊗b
∞
direct sum of representations of the form (C∞ )⊗a ⊗ (C∞
∗ ) . Here C∗ is the restricted dual
∞
∗
of C , defined as the span of the dual basis {ei } in the usual dual space (C∞ )∗ . One easily
checks that C∞
∗ is indeed a representation of GL∞ . We write Rep(GL∞ ) for the category
of algebraic representations. This was also studied in [SS3].
By the “twisted diagonal embedding” GL∞ → GL∞ × GL∞ , we mean the embedding
given by g 7→ (g, t g −1). We note that the algebraic representations of GL∞ are exactly
those appearing as a subquotient of the restriction of a polynomial representation from
GL∞ × GL∞ via the twisted diagonal embedding.
We identify A with C[xi,j ] in the obvious manner, and let m ⊂ |A| be the ideal generated
by xi,i − 1 and xi,j for i 6= j. This ideal is stable under the twisted diagonal GL∞ . For an
e
A-module M, define Φ(M)
= M/mM. This is naturally a representation of GL∞ .
e induces an equivalence Φ : ModK → Rep(GL∞ ).
Theorem 3.18. The functor Φ
Corollary 3.19. We have the following:
(a) Finitely generated objects of ModK have finite length.
(b) If V is a finite length polynomial representation of GL∞ × GL∞ then K ⊗ V is a finite
length injective object of ModK , and all finite length injective objects have this form.
(c) Associating to (λ, µ) the socle of K ⊗Sλ (C∞ ) ⊗Sµ (C∞ ) gives a bijection between pairs
of partitions and isomorphism classes of simple objects of ModK .
(d) Every finite length object M of ModK has finite injective resolution M → I• where
each Ik is a finite length injective object.
Proof. These properties are proven for Rep(GL∞ ) in [SS3]:
(a) [SS3, Proposition 3.1.5],
(b,c) [SS3, Proposition 3.2.14],
(d) dualize the explicit projective resolutions in [SS3, (3.3.7)].
The proof of Theorem 3.18 closely follows that of Theorem 3.1. The main differences occur
in the analog of §3.2. In the present case, one takes B ⊂ M∞ × M∞ to be the set of pairs
of upper-triangular matrices. The group H is replaced with the intersection of B and the
twisted diagonal GL∞ inside of GL∞ × GL∞ , and consists of pairs (h, h−1 ) where h ∈ GL∞
is a diagonal matrix. With these definitions, everything proceeds in a similar way.
NOETHERIANITY OF SOME DEGREE TWO TCA’S
15
4. Proof of the main theorems
4.1. The structure of ideals. We have the following multiplicity-free decompositions:
M
Sym(Sym2 C∞ ) =
S2λ (C∞ )
M
V
Sym( 2 C∞ ) =
S(2λ)† (C∞ )
M
Sym(C∞ ⊗ C∞ ) =
Sλ (C∞ ) ⊗ Sλ (C∞ ).
For a proof, see [M, §I.5, Example 5] for the first two decompositions and [M, §I.4, (4.3)] for
the last one. In all cases, the sum is over partitions λ. For the purposes of stating the next
result we write Eλ for the λ summand. Let Iλ be the ideal generated by Eλ .
Proposition 4.1. Eµ ⊆ Iλ if and only if λ ⊆ µ.
V
Proof. For Sym(Sym2 C∞ ), see [Ab], for Sym( 2 C∞ ), see [AdF, Theorem 3.1], and for
Sym(C∞ ⊗ C∞ ), see [CEP, Theorem 4.1]. Since [Ab] is a difficultVreference to obtain, we
note that the result for Sym(Sym2 C∞ ) follows from that of Sym( 2 C∞ ) because the two
are transpose dual (see [SS2, §7.4]). Proofs of these results will also appear in [NSS].
Corollary 4.2. Let A be one of the three algebras above, and let I be any non-zero ideal of
A. Then A/I is essentially bounded, and, in particular, noetherian.
Proof. Suppose that I is a non-zero ideal of A. Then I contains some Eλ , and thus Iλ . Thus
by the proposition, A/I contains no partition µ satisfying λ ⊂ µ, and is therefore essentially
bounded. Noetherianity of A/I follows from Proposition 2.4.
4.2. The (FT) property. Let B be a (bivariate) tca with B0 = C, so that B+ (the ideal of
B generated by positive degree elements) is maximal. We say that a B-module M is (FT)
over B if TorB
i (M, C) is a finite length representation of GL∞ (or GL∞ × GL∞ ) for all
i ≥ 0. The i = 0 case implies that M is finitely generated as a B-module, by Nakayama’s
lemma [SS2, Proposition 8.4.2]. Conversely, if B is noetherian then any finitely generated
B-module satisfies (FT). We note that if
0 → M1 → M2 → M3 → 0
is a short exact sequence of B-modules and two of the modules are (FT) then so is the third.
The main result we need concerning (FT) is the following proposition:
V
Proposition 4.3. Let A be one of Sym(Sym2 C∞ ), Sym( 2 C∞ ), or Sym(C∞ ⊗ C∞ ), and
let M be a finitely generated torsion A-module. Then M satisfies (FT) over A.
We begin with some lemmas.
Lemma 4.4. Let B → B ′ be a homomorphism of (bivariate) tca’s, with B0 = B0′ = C, and
let M be a B ′ -module. Suppose that B ′ is (FT) over B. Then M is (FT) over B if and only
if M is (FT) over B ′ .
Proof. First suppose that M is (FT) over B ′ . Starting with a free resolution of M over B ′ ,
and a free resolution of B ′ over B, we get an acyclic double complex of B-modules resolving
M. This leads to a spectral sequence
′
B
B
′
E2p,q = TorB
p (Torq (M, C), B ) =⇒ Torp+q (M, C).
16
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
′
Note that TorB
q (M, C) is a trivial B module (meaning B+ acts by 0), and so
′
′
B
B
B
′
′
TorB
p (Torq (M, C), B ) = Torq (M, C) ⊗C Torp (B , C).
Each of the Tor’s on the right has finite length by assumption, and so the left side also has
finite length. It follows that TorB
i+j (M, C) has finite length, and so M is (FT) over B.
Now suppose that M is (FT) over B. In particular, M is a finitely generated B-module,
′
and so also a finitely generated B ′ -module. This shows that TorB
0 (M, C) is finite length. Let
P → M → 0 be a minimal projective cover and let N be the kernel. Since B ′ is (FT) over
B, we conclude that P , and hence N are both (FT) over B. In particular, N is a finitely
′
generated as a module over B, and hence over B ′ . This shows that TorB
1 (M, C) is finite
′
length; to get the statement for TorB
i (M, C), we can iterate this argument i times.
Lemma 4.5. Let A be the bivariate tca Sym(C∞ ⊗ C∞ ). Then A/Iλ satisfies (FT) over A
for all rectangular partitions λ.
Proof. This follows from [RW, Theorem 1.2], taking m = n = ∞ (the results there are
stated for finite m and n, but since the answer is given in terms of Schur functors, it can be
extended to the infinite case): one has to show that the coefficient of w i , as a polynomial
in z, is of bounded degree. To see that, note that fixing w i means that q is bounded from
above, and then the result is clear from the form of the polynomials hr×s (z, w).
V
Lemma 4.6. Let B be Sym(Sym2 C∞ ) or Sym( 2 C∞ ). Then B/Iλ satisfies (FT) over B
for all rectangular partitions λ.
e be the
Proof. Let A = Sym(C∞ ⊗C∞ ). Let Jλ be the ideal in A generated by Sλ ⊗Sλ . Let A
tca obtained from A by restricting to the diagonal GL∞ action. Then there is a surjection of
e → B, induced by the natural map (C∞ )⊗2 → Sym2 (C∞ ), and ϕ(Jλ ) ⊂ Iλ . (Note
tca’s ϕ : A
that ϕ(Jλ ) is nonzero: if λ is a single column, then this is an ideal generated by minors
of a given size and the image of every power of Jλ is nonzero; in general, some power of
a determinantal ideal belongs to Jλ after we specialize to large enough finite-dimensional
vector spaces.)
Since A/Jλ is (FT) over A (Lemma 4.5), each TorA
i (A/Jλ , C) is a finite length GL∞ ×GL∞
module, and hence remains finite length under the restriction to the diagonal copy of GL∞ .
e λ is (FT) over A.
e Also A/J
e λ is essentially bounded (since the bivariate tca A/Jλ is)
So A/J
e λ , thus
and hence noetherian (Proposition 2.4). It follows that B/ϕ(Jλ ) is (FT) over A/J
e as well (Lemma 4.4).
over A
e (the resolution of B over A
e is a Koszul complex) so another
Next, B is (FT) over A
application of Lemma 4.4 gives that B/ϕ(Jλ ) is (FT) over B. Finally, B/Iλ is a finitely
generated module over B/ϕ(Jλ ) and the latter is noetherian (Corollary 4.2), so B/Iλ is (FT)
over B/ϕ(Jλ ). We apply Lemma 4.4 again to deduce that B/Iλ is (FT) over B.
Remark 4.7. It would be interesting to prove directly that B/Iλ satisfies (FT) over B by
∞
∞
computing TorB
i (B/Iλ , C), as is done in [RW] for Sym(C ⊗ C ).
Proof of Proposition 4.3. Let I be the annihilator of M. This is non-zero by Corollary 2.3.
Thus I contains an ideal generated by a rectangular partition; replace I with this ideal.
Since A/I is noetherian (Corollary 4.2), M is (FT) over A/I. By Lemma 4.5 or 4.6, A/I is
(FT) over A. Thus by Lemma 4.4, M is (FT) over A.
NOETHERIANITY OF SOME DEGREE TWO TCA’S
4.3. Completion of the proofs. Let A be one of the tca’s Sym(Sym2 C∞ ) or Sym(
or the bivariate tca Sym(C∞ ⊗ C∞ ), and let K = Frac(A).
17
V2
C∞ ),
Proposition 4.8. If M is a finite length K-module then S(M) satisfies (FT) over A.
Proof. We prove this by induction on the injective dimension of M, which is possible by
Corollary 3.2(d) (and its analogs). If M is injective then S(M) is a finitely generated
projective A-module (Corollary 3.2(b), Proposition 2.8), and thus satisfies (FT). Now let M
be a finite length object of ModK with positive injective dimension. We can then find an
exact sequence
0 → M → I → N → 0,
where I is injective and N has smaller injective dimension than M. Applying S, we obtain
an exact sequence
0 → S(M) → S(I) → S(N) → (R1 S)(M) → 0.
By induction, S(N) is (FT) over A, and so finitely generated. It follows that (R1 S)(M)
is finitely generated. By Proposition 2.6 and the fact that localization is exact, we have
(R1 S)(M) ⊗A K = 0, and so (R1 S)(M) satisfies (FT) over A by Proposition 4.3. Thus S(I),
S(N), and (R1 S)(M) all satisfy (FT) over A, and so S(M) satisfies (FT) over A as well.
The following completes the proof of our main results: Theorem 1.1 and Theorem 1.2.
Theorem 4.9. A is noetherian.
Proof. Let P be a finitely generated projective A-module, and let N1 ⊂ N2 ⊂ · · · be an
ascending chain of A-submodules of P . Since P ⊗A K is finite length (Corollary 3.2(a)),
it follows that the ascending chain Ni ⊗A K stabilizes, and so we may as well assume it is
stationary to begin with. Let M ⊂ P be the common value of S(Ni ⊗A K), which is finitely
generated by Proposition 4.8. Then N• is an ascending chain in M. Let M ′ = M/N1 and
Ni′ = Ni /N1 ⊂ M ′ , so that N•′ is an ascending chain in M ′ . Since M ′ is finitely generated
and M ′ ⊗ K = 0, Corollary 2.3 implies that I = Ann(M ′ ) is non-zero. Thus M is a module
over A/I, which is noetherian (Corollary 4.2), and so N•′ stabilizes. This implies that N•
stabilizes, and so P is noetherian.
Remark 4.10. The above proof has three key ingredients:
(1) Finitely generated objects of ModK are noetherian.
(2) If I is a non-zero ideal of A then A/I is noetherian.
(3) If M is a finite length object of ModK then S(M) is a finitely generated A-module.
Let us make one comment regarding (3). Given a finite length object M in ModK , we can
realize M as the kernel of a map I → J where I and J are finite length injective objects of
ModK . Since S is left-exact, it follows that S(M) is the kernel of the map S(I) → S(J),
and we know that S(I) and S(J) are finitely generated projective A-modules. Thus finite
generation of S(M) would follow immediately if we knew A to be coherent (which exactly
says that the kernel of a map of finitely generated projective modules is finitely generated).
Since coherence is a weaker property than noetherianity, it should be easier to prove; however,
we have not found any way to directly prove coherence.
5. A Gröbner-theoretic approach to the main theorems
In this section we outline a possible approach to proving Theorem 1.1 using Gröbner bases.
This leads to an interesting combinatorial problem that we do not know how to resolve.
18
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
5.1. Admissible weights. A weight of GL∞ is a sequence of non-negative integers w =
(w1 , w2 , . . .) such that
L wi = 0 for i ≫ 0. Every polynomial representation V of GL∞
decomposes as V =
Vw , where Vw is the w weight space. A weight is admissible if wi is 0
or 1 for all i. An admissible weight vector is an element of some Vw with w an admissible
weight. We require the following fact: if V is a polynomial representation of GL∞ then V
is generated, as a representation, by its admissible weight vectors.
5.2. Degree one tca’s. We begin by sketching a Gröbner-theoretic proof that the tca
A = Sym(C∞ ⊕ C∞ ) is noetherian. This proof comes from transferring the proof in [SS4]
that Rep(FI2 ) is noetherian through Schur–Weyl duality, and can easily be adapted to treat
all tca’s generated in degree ≤ 1. Let x1 , x2 , . . . be a basis for the first C∞ , and let y1 , y2 , . . .
be a basis for the second C∞ , so that A is the polynomial ring C[x1 , x2 , . . . , y1 , y2 , . . .].
Let M be the set of pairs Γ = (S, ϕ), where S is a finite subset of N = {1, 2, . . .} and
ϕ : S → {red, blue} is a function. Given Γ, Γ′ ∈ M, we define Γ → Γ′ (a “move”) if one of
the following two conditions hold:
• S ′ is obtained from S by adding a single element and leaving the colors unchanged
(i.e., ϕ′ |S = ϕ).
• There exists some i ∈ S such that i + 1 6∈ S and S ′ is obtained from S by replacing
i with i + 1 (and leaving all colors unchanged).
We define Γ ≤ Γ′ if there is a sequence of moves taking Γ to Γ′ . This partially orders M.
We also define a total order on M, as follows. Given two finite subsets S and S ′ of N,
define S S ′ if max(S) < max(S ′ ), or max(S) = max(S ′ ) = n and S \{n} S ′ \{n}. Given
S ⊂ N and ϕ, ϕ′ : S → {red, blue}, define ϕ ϕ′ by thinking of ϕ and ϕ′ as words in R
and B and using the lexicographic order (with R B, say). Finally, define (S, ϕ) (S ′ , ϕ′ )
using the lexicographic order (i.e., S ≺ S ′ , or S = S ′ and ϕ ϕ′ ).
Given Γ ∈ M, define
(
Y xi if ϕ(i) = red
mΓ =
.
yi if ϕ(i) = blue
i∈S
If f ∈ A is an admissible weight vector of weight w, then f is a linear combination of the
mΓ ’s where Γ has the same support as w. We define the initial variable of f , denoted in(f ),
to be the largest Γ (under ) such that mΓ appears in f with non-zero coefficient.
Now let I be an ideal of A. Let in(I) ⊂ M be the set of in(f )’s where f varies over the
admissible weight vectors in I. One then proves the following two statements:
(1) in(I) is a poset ideal of M, that is, in(I) is closed under moves, and
(2) if I ⊂ J and in(I) = in(J) then I = J.
From this, weak noetherianity of A follows from noetherianity of M, which is an easy exercise.
A slight modification of this argument shows that A is noetherian.
5.3. Degree two tca’s. We now sketch our Gröbner approach to the noetherianity of A =
Sym(Sym2 (C∞ )). Let xi,j , with i ≤ j, be a basis for Sym2 (C∞ ), so that A = C[xi,j ].
Let M be the set of undirected matchings Γ on N. (Recall that a graph is a matching if
each vertex has valence 0 or 1.) Given Γ, Γ′ ∈ M, we define Γ → Γ′ if one of the following
two conditions hold:
• Γ′ is obtained from Γ by adding a single edge.
• There exists an edge (i, j) in Γ such that j + 1 is not in Γ, and Γ′ is obtained from
Γ by replacing (i, j) with (i, j + 1). (Here we allow i < j or j < i.)
NOETHERIANITY OF SOME DEGREE TWO TCA’S
19
We call Γ → Γ′ a “type I move”. We define Γ ≤ Γ′ if there is a sequence of type I moves
transforming Γ to Γ′ . This partially orders M.
We also define a total order on M as follows. First, suppose that i < j and k < ℓ are
elements of N. Define (i, j) (k, ℓ) if j < ℓ, or j = ℓ and i ≤ k. Now, let Γ and Γ′ be two
elements of M, and let e1 · · · en and e′1 · · · e′m be their edges, listed in increasing
order. We define Γ Γ′ if m > n, or if m = n and (e1 , . . . , en ) (e′1 , . . . , e′m ) under the
lexicographic order.
Q
Given Γ ∈ M, define mΓ = (i,j)∈Γ xi,j . Once again, every admissible weight vector is a
sum of mΓ ’s, and we define the initial term in(f ) of an admissible weight vector f to be the
largest Γ (under the order ) for which the coefficient of mΓ is non-zero in f .
Let I be an ideal of A. Define in(I) as before. Once again, in(I) is closed under type I
moves, and therefore forms a poset ideal of (M, ≤). The weak noetherianity of A would
follow from the noetherianity of the poset (M, ≤), but the latter property fails:
Example 5.1. For n ≥ 3, define Γn ∈ M to have edges (2i + 1, 2i + 4) for i = 0, 1, . . . , n − 2
and (2, 2n − 1). Then Γn is supported on {1, . . . , 2n}. It is easy to verify that the Γn are
incomparable, so (M, ≤) is not a noetherian poset.
The above observation is not the end of the road, however: the set in(I) is closed under
more than just type I moves. Suppose Γ ∈ in(I) and that e = (i, j) and e′ = (k, ℓ) are edges
appearing in Γ, with i < j and k < ℓ and j < ℓ. We then have the following observations:
• Suppose k < i < j < ℓ and that every number strictly between k and i that appears
in Γ is connected to a number larger than j. Let Γ′ be the graph obtained by replacing
e and e′ with (k, j) and (i, ℓ). Then Γ′ ∈ in(I).
• Suppose i < k < j < ℓ and that every number strictly between k and j that appears
in Γ is connected to a number larger than j. Let Γ′ be the graph obtained by replacing
e and e′ with (i, k) and (j, ℓ). Then Γ′ ∈ in(I).
Write Γ ⇒ Γ′ to indicate that Γ′ is related to Γ by one of the above two modifications. We
call this a “type II move.” Here is a pictorial representation of these moves (we use labels
a < b < c < d, and the dotted lines indicate that any element there is either not on an edge,
or is connected to a number larger than c):
a
b
c
d
⇒
a
b
c
⇒
d
a
b
c
d
′
We define a new partial order ⊑ on M as follows: Γ ⊑ Γ if there exists a sequence of moves
(of any type) taking Γ to Γ′ . The above observations show that in(I) is a poset ideal of
(M, ⊑). This leads to the important open question:
Question 5.2. Is the poset (M, ⊑) noetherian?
Remark 5.3. The sequence defined in Example 5.1 is comparable in (M, ⊑). Let σi be the
element (i, i + 1) · · · (3, 4)(2, 3) of the symmetric group S2n . For each 2 ≤ i ≤ 2n − 4, we
have type II moves σi Γn → σi+1 Γn , so Γn ⊑ σ2n−3 Γn . Finally, (2n − 1, 2n) is a valid type II
move for σ2n−3 Γn . It is now easy to check that ((2n − 1, 2n)σ2n−3 )Γn embeds into Γm (via
type I moves) for any m > n. This shows Γn ⊑ Γm for any m > n ≥ 3.
A positive answer to Question 5.2 would show that A is weakly noetherian. A slight
modification of this question would give noetherianity. Furthermore, this approach would
even give results in positive characteristic.
20
ROHIT NAGPAL, STEVEN V SAM, AND ANDREW SNOWDEN
References
Silvana Abeasis, The GL(V )-invariant ideals in S(S 2 V ), Rend. Mat. (6) 13 (1980), no. 2, 235–262.
S. Abeasis, A. Del Fra, Young diagrams and ideals of Pfaffians, Adv. in Math. 35 (1980), no. 2,
158–178.
[AH]
Matthias Aschenbrenner, Christopher J. Hillar, Finite generation of symmetric ideals, Trans. Amer.
Math. Soc. 359 (2007), 5171–5192, arXiv:math/0411514v3.
[BR]
A. Berele, A. Regev, Hook Young diagrams with applications to combinatorics and to representations of Lie superalgebras, Adv. in Math. 64 (1987), 118–175.
[CEF] Thomas Church, Jordan Ellenberg, Benson Farb, FI-modules and stability for representations of
symmetric groups, Duke Math. J. 164 (2015), no. 9, 1833–1910, arXiv:1204.4533v4.
[CEFN] Thomas Church, Jordan S. Ellenberg, Benson Farb, Rohit Nagpal, FI-modules over Noetherian
rings, Geom. Top. 18 (2014), 2951–2984, arXiv:1210.1854v2.
[Co]
D. E. Cohen, On the laws of a metabelian variety, J. Algebra 5 (1967), 267–273.
[CEP] C. de Concini, David Eisenbud, C. Procesi, Young diagrams and determinantal varieties, Invent.
Math. 56 (1980), no. 2, 129–165.
[Dr]
Jan Draisma, Noetherianity up to symmetry, Combinatorial algebraic geometry, Lecture Notes in
Math. 2108, Springer, 2014, arXiv:1310.1705v2.
[DE]
Jan Draisma, Rob H. Eggermont, Plücker varieties and higher secants of Sato’s Grassmannian, J.
Reine Angew. Math., to appear, arXiv:1402.1667v3.
[DK]
Jan Draisma, Jochen Kuttler, Bounded-rank tensors are defined in bounded degree, Duke Math. J.
163 (2014), no. 1, 35–63, arXiv:1103.5336v2.
[Eg]
Rob H. Eggermont, Finiteness properties of congruence classes of infinite-by-infinite matrices, Linear
Algebra Appl. 484 (2015), 290–303, arXiv:1411.0526v1.
[HS]
Christopher J. Hillar, Seth Sullivant, Finite Gröbner bases in infinite dimensional polynomial rings
and applications, Adv. Math. 221 (2012), 1–25, arXiv:0908.1777v2.
[L]
Alain Lascoux, Syzygies des variétés déterminantales, Adv. in Math. 30 (1978), no. 3, 202–237.
[M]
I. G. Macdonald, Symmetric Functions and Hall Polynomials, second edition, Oxford Mathematical
Monographs, Oxford, 1995.
[NSS] Rohit Nagpal, Steven V Sam, Andrew Snowden, Noetherianity of some degree two twisted skewcommutative algebras; in preparation.
[RW]
Claudiu Raicu, Jerzy Weyman, The syzygies of some thickenings of determinantal varieties,
arXiv:1411.0151v1.
[PS]
Andrew Putnam, Steven V Sam, Representation stability and finite linear groups,
arXiv:1408.3694v2.
[SS1]
Steven V Sam, Andrew Snowden, GL-equivariant modules over polynomial rings in infinitely many
variables, Trans. Amer. Math. Soc., to appear, arXiv:1206.2233v3.
[SS2]
Steven V Sam, Andrew Snowden, Introduction to twisted commutative algebras,
arXiv:1209.5122v1.
[SS3]
Steven V Sam, Andrew Snowden, Stability patterns in representation theory, Forum Math., Sigma
3 (2015), e11, 108 pp., arXiv:1302.5859v2.
[SS4]
Steven V Sam, Andrew Snowden, Gröbner methods for representations of combinatorial categories,
arXiv:1409.1670v2.
[Sn]
Andrew Snowden, Syzygies of Segre embeddings and ∆-modules, Duke Math. J. 162 (2013), no. 2,
225–277, arXiv:1006.5248v4.
[We]
Jerzy Weyman, Cohomology of Vector Bundles and Syzygies, Cambridge Tracts in Mathematics
149, Cambridge University Press, Cambridge, 2003.
[Ab]
[AdF]
NOETHERIANITY OF SOME DEGREE TWO TCA’S
Department of Mathematics, University of Wisconsin, Madison, WI
Current address: Department of Mathematics, The University of Chicago, Chicago, IL
E-mail address: [email protected]
URL: http://math.uchicago.edu/~nagpal/
Department of Mathematics, University of California, Berkeley, CA
Current address: Department of Mathematics, University of Wisconsin, Madison, WI
E-mail address: [email protected]
URL: http://math.wisc.edu/~svs/
Department of Mathematics, University of Michigan, Ann Arbor, MI
E-mail address: [email protected]
URL: http://www-personal.umich.edu/~asnowden/
21
| 0math.AC
|
EXPANSION FOR MOMENTS OF
REGRESSION QUANTILES WITH
APPLICATIONS TO NONPARAMETRIC
TESTING
arXiv:1306.6179v3 [math.ST] 14 Jul 2017
Enno Mammen§
Heidelberg University, Germany
Ingrid Van Keilegom ∗∗
KU Leuven, Belgium
Kyusang Yu ‡‡
Konkuk University, Seoul, Korea
July 17, 2017
Abstract
We discuss nonparametric tests for parametric specifications of regression quantiles. The test is based on the comparison of parametric and nonparametric fits of
these quantiles. The nonparametric fit is a Nadaraya-Watson quantile smoothing
estimator.
An asymptotic treatment of the test statistic requires the development of new
mathematical arguments. An approach that makes only use of plugging in a Bahadur expansion of the nonparametric estimator is not satisfactory. It requires too
strong conditions on the dimension and the choice of the bandwidth.
Our alternative mathematical approach requires the calculation of moments of
Nadaraya-Watson quantile regression estimators. This calculation is done by application of higher order Edgeworth expansions.
§
Institut für Angewandte Mathematik, Universität Heidelberg, Im Neuenheimer Feld 205, 69120 Heidelberg, Germany. E-mail address: [email protected].
∗∗
ORSTAT, KU Leuven, Naamsestraat 69, 3000 Leuven, Belgium.
E-mail address:
[email protected].
‡‡
Department of Applied Statistics, Konkuk University, Seoul 143-701, Korea, E-mail address:
[email protected] .
1
AMS 1991 subject classifications. primary 62G07, secondary 62G20
Journal of Economic Literature Classification: C14
Keywords and phrases. Nonparametric Regression; Quantiles; Bahadur Expansions; Kernel Smoothing; Nonparametric Testing; Goodness-of-fit tests.
1
Introduction
Consider a data set of n i.i.d. tuples (Xi , Yi ), where Yi is a one-dimensional response
variable and Xi is a d-dimensional covariate. For 0 < α < 1 we denote the conditional
α-quantile of Yi given Xi = x by mα (x). Thus we can write
Yi = mα (Xi ) + εi,α
(i = 1, . . . , n),
(1)
with error variables εi,α that fulfill qα (εi,α |Xi ) = 0. Here, qα (εi,α |Xi ) is the α-quantile of
the conditional distribution of εi,α given Xi . Consider the null hypothesis
H0 : For all α ∈ A there exists a θ(α) ∈ Θ, such that mα = mα,θ(α) ,
(2)
where {mα,θ : θ ∈ Θ} is a parametric class of regression quantiles, Θ is a compact subset
of IRk and A ⊂ (0, 1). The set A can be a singleton A = {α}, but can also be a closed
subset of (0, 1) if a set of quantile functions is checked.
In this paper we aim at studying a test statistic for H0 , and to study its asymptotic
properties under the null and the alternative. We will see that this problem is an example
of a quantile model where the asymptotics cannot be developed by standard tools of
quantile regression. In particular, a direct application of Bahadur expansions requires
assumptions that are too restrictive.
Our test statistic is based on kernel smoothing. Let K(u1 , . . . , ud ) =
Qd
j=1
k(uj ), where
k is a one-dimensional density function defined on [−1, 1], and let h = (h1 , . . . , hd ) be a
d-dimensional bandwidth parameter. We assume that all bandwidths h1 , . . . , hd are of the
same order. For simplicity of notation we further assume that they are identical and by
abuse of notation we write h = h1 = . . . = hd . For any 0 < α < 1 and any x in the support
2
RX of X, let Fεα |X (·|x) be the conditional distribution function of εα = Y − mα (X), given
X = x, and let rα,θ(α) (x) be the α-quantile of Y − mα,θ(α) (X) given that X = x. Define
n
X
x − Xi
τα (Yi − mα,θ(α)
rbα (x) = arg min
K
b (Xi ) − r),
r
h
i=1
b
where τα (u) = αu+ − (1 − α)u− , u+ = uI(u > 0) and u− = uI(u < 0) and where θ(α)
is
an estimator of θ(α).
Note that instead of estimating the conditional quantile rα,θ(α) (x) by the above estimator, we could have considered the alternative estimator
n
X
x − Xi
alt
rbα (x) = arg min
K
τα (Yi − m) − mα,θ(α)
b (x).
m
h
i=1
However, the latter estimator has the important drawback that the consideration of responses Yi in a neighborhood of x induces a smoothing bias, whereas rbα (x) has no smoothing related bias, since it is based on the errors Yi − mα,θ(α) (Xi ), whose conditional quantile
of order α is exactly zero under H0 for all Xi .
We suppose that A is a closed subinterval of (0, 1). We define the following test
statistic :
Z Z
TbA =
A
RX
rbα (x)2 w(x, α)dxdα,
(3)
for some weight function w(x, α). For the case that A contains only one value α we use
Z
rbα (x)2 w(x)dx
(4)
Tbα =
RX
for some weight function w(x). One could also generalize our results to the case that A
is a finite set. To keep notation simple we omit this case in our mathematical analysis.
Our test is an omnibus test that has power against all types of alternatives. It is based
on the comparison of a kernel quantile estimator with the parametric fit. We will show
that the test statistic is asymptotically equivalent to a weighted L2 -distance between the
nonparametric and the parametric estimator. Similar tests have been used in a series of
papers for mean regression. Early references are Härdle and Mammen (1993), GonzálezManteiga and Cao-Abad (1993), Hjellvik, Yao and Tjøstheim (1998), Zheng (1996) and
3
Fan, Zhang and Zhang (2001). Furthermore recent references are Dette and Sprekelsen
(2004), Kreiss, Neumann and Yao (2008), Haag (2008), Leucht (2012), Gao and Hong
(2008) and Ait-Sahalia, Fan and Peng (2009). Most of the more recent work concentrates
on time series data.
The classical way to carry over results from parametric and nonparametric mean regression to quantile regression is the use of Bahadur expansions. The main point is that
asymptotically quantile regression is equivalent to weighted mean regression. This approach has been used in Chaudhuri (1991), Truong (1989), He and Ng (1999), He, Ng
and Portnoy (1998) and more recently in Hoderlein and Mammen (2009), Hong (2003),
Kong, Linton and Xia (2010), Lee and Lee (2008), El Ghouch and Van Keilegom (2009),
Li and Racine (2008), and De Backer, El Ghouch and Van Keilegom (2017). A detailed
review of quantile regression can be found in the book by Koenker (2005). Testing procedures in quantile regression were considered in Zheng (1998), Koenker and Machado
(1999), Bierens and Ginther (2001), Horowitz and Spokoiny (2002), Koenker and Xiao
(2002), and He and Zhu (2003), among others. They all considered tests for the parametric form of the quantile function. More recently, Rothe and Wied (2013) proposed
a test statistic for the hypothesis that the conditional distribution belongs to a certain
parametric class. Tests based on quantiles of the errors have also been considered in Su
and White (2012) in the context of testing conditional independence. Other recent papers are the ones by Volgushev et al. (2013) and Conde-Amboage, Sánchez-Sellero and
González-Manteiga (2015), who considered significance tests in quantile regression and
developed a test statistic based on marked empirical processes.
In this paper we will discuss how results from mean regression carry over to our
case. Whereas elsewhere a first attempt could be based on the application of a Bahadur
expansion, we will see that in our setting the accuracy of a direct application of Bahadur
expansions is too poor. We will shortly explain this here for the testing problem where A
contains only one value α. Suppose for simplicity at this stage that the parametric model
b
contains only one value θ0 = θ0 (α) and that θb = θ(α)
= θ0 . The Bahadur expansion of
4
rbα (x) is given by
Pn
reα (x) = −
x−Xi
h
{I(εi,α ≤ 0) − α}
,
x−Xi
K
f
(0|X
)
i
εα |X
i=1
h
i=1 K
Pn
(5)
where fεα |X is the conditional density of εα given X. This gives the following approximation for Tbα :
Z
Teα =
RX
reα (x)2 w(x)dx.
One can show that up to a logarithmic factor supx |b
rα (x) − reα (x)| and supx |b
rα (x)| are of
order (nhd )−3/4 and (nhd )−1/2 , respectively. This implies that up to a logarithmic factor,
the difference Tbα − Teα is of order (nhd )−5/4 . On the other hand as it is also the case in
mean regression Teα is equal to the sum of a deterministic term and a random term of
order n−1 h−d/2 . Thus the above approximation only helps if (nhd )−5/4 << n−1 h−d/2 or
equivalently if nh3d → ∞ for sample size n going to ∞. E.g. if one applies a bandwidth
h ∼ n−1/(4+d) that leads to rate optimal estimation of twice differentiable functions this
assumption would allow only a one-dimensional setting d = 1. Also in the case of minimax
optimal testing with twice differentiable functions under the alternative (see Ingster (1993)
and Guerre and Lavergne (2002)), the optimal bandwidth h ∼ n−2/(8+d) is only allowed
for dimension d = 1. In this paper we develop an asymptotic theory for L2 -type quantile
tests that works under the assumption that nh3d/2 → ∞. In the above examples this
allows dimensions d ≤ 7 and d ≤ 3. Furthermore, for our asymptotic discussion of
the distribution of the test statistic on the hypothesis we only need the assumption that
nhd → ∞. Thus on the hypothesis, our basic assumptions coincide with conditions needed
for the asymptotics of mean regression. We conjecture that also for the alternative the
assumption nh3d/2 → ∞ could be weakened but that then the asymptotic mean of the
test statistic changes. We will comment on this after the statement of Theorem 2.
In our approach we will make use of the fact that Bahadur expansions of kernel quantile
estimators calculated at two different points are asymptotically independent if they are
calculated at points that are such that the supports of the kernels do not overlap. Thus
the variance of an integral over a Bahadur expansion should be of smaller order than the
5
variance of the Bahadur expansion at a fixed point. The main technical difficulty that will
come up when applying this idea is the need to calculate moments of the kernel regression
quantiles. We will introduce a method for the expansion of such moments that is based on
Edgeworth expansions in a related problem. Our main result gives a bound between the
moments of kernel regression quantiles and the moments of its Bahadur approximation.
The paper is organized as follows. In the next section we will state our result on
moments of kernel regression quantiles. Our main result on the asymptotics of L2 -type
quantile tests is given in Section 3. We will also introduce some kind of wild bootstrap
procedure adapted to quantile regression and give a theoretical result on its consistency.
In Section 4 we present the results of a simulation study, and we analyze data on Engel
curves. The proofs are postponed to the last three sections.
2
Asymptotic moments
In this section we will present an asymptotic result on higher order moments of kernel
regression quantiles. This result will be our most important ingredient for getting our
result on the asymptotic distribution of our test statistic. In our result the moments of
kernel regression quantiles are compared with the moments of their Bahadur approximations. Recall that we are interested in the null hypothesis H0 defined in (2). We suppose
that for all α ∈ A,
mα (·) = mα,θ0 (α) (·) + n−1/2 h−d/4 ∆α (·).
(6)
For the case ∆α ≡ 0 the function mα lies on the hypothesis. In order to develop our
asymptotic theory, we need to work under the following assumptions. In the formulation
of the assumptions and in the proofs we use the convention that C, C1 , C2 , ... are generic
strictly positive constants that are chosen large enough, that c, c1 , c2 , ... are generic strictly
positive constants that are chosen small enough, and that C ∗ , C1∗ , C2∗ , ... are generic strictly
positive constants that are arbitrarily chosen. Using this convention we write Ln =
∗
(log n)C for a sequence with C > 0 large enough and L∗n = (log n)C for a sequence with
6
an arbitrarily chosen constant C ∗ > 0. All these variable names are used for different
constants and sequences, even in the same equation.
We will make use of the following assumptions.
(B1) The support RX of X is a compact convex subset of IRd . The density fX of X
is bounded and bounded away from zero on RX . The function ∆α is uniformly
absolutely bounded for α ∈ A.
(B2) The conditional distribution of εα given X = x allows a density fεα |X (e|x) that is
twice differentiable with respect to e. For this derivative it holds that |fε00α |X (e|x)| ≤
C for |e| ≤ c, x ∈ RX , and α ∈ A. The density fεα |X (e|x) also satisfies fεα |X (e|x) > 0
and |fεα |X (e0 |x0 ) − fεα |X (e|x)| ≤ C(kx0 − xk + |e0 − e|) for x, x0 ∈ RX and e, e0 ∈ R,
where k · k is the Euclidean norm. Moreover, the functions fX (x), mα (x) and ∆α (x)
are continuously differentiable with respect to x.
(B3) The bandwidth h satisfies h = o(1) and nhd /L∗n → ∞. The kernel k is a symmetric, continuously differentiable probability density function with compact support,
[−1, 1], say. It fulfills a Lipschitz condition and it is monotone strictly increasing on
[−1, 0]. It holds that k 0 (k −1 (u)) ≥ min c{uκ , (k(0) − u)κ } for some 0 ≤ κ < 1 where
k −1 : [0, k(0)] → [−1, 0] denotes the inverse of k : [−1, 0] → [0, k(0)].
In our asymptotics, the density fX and the functions ∆α are fixed and do not depend
on n. The cumulative distribution function F (·|x) of Y given X = x may depend on n.
We do not indicate this in our notation.
Assumptions (B1)–(B3) are standard assumptions for the study of smoothing estimators, with the exception of the last assumption in (B3). We now shortly explain why
this assumption is needed here. For fixed u and x = (x1 , . . . , xd )| , define the random
| n
o
Pn x1 −X1,j
xd −Xd,j
∆
h
d −1/2
I(εj,α ≤ ∆α (x) + u(nh )
) − α with
vector Vn = j=1 k( h ), ..., k( h )
−1/2 −d/4
ε∆
h
∆α (Xj ), and where ∆hα (x) is defined in (10) below. In the proof of
j,α = εj,α + n
the following Theorem 1 we will develop Edgeworth expansions for the distribution of Vn .
7
Typically, the summands of Vn do not fullfil non-lattice type assumptions that are needed
for the verification of Edgeworth expansions. But under (B3) a non-lattice assumption
can be verified for the conditional distribution of a finite sum of summands of Vn . For
more details we refer to the proof of Theorem 1. The last assumption in (B3) can be
easily verified. It just puts a simple bound on the derivative of k −1 . E.g., it can be easily
checked for the triangle kernel and for all kernels of the form k (z) = 1 (|z| ≤ 1) cr (1 − z 2 )
r
with r ≥ 1. In case that k 0 is bounded away from zero on bounded intervals of (−1, 0)
the assumption follows if for some l, l∗ ∈ N, it holds that k 0 (x) = (x + 1)l + o((x + 1)l ) for
x ≥ −1 and x + 1 small enough and that k 0 (x) = −x2l
∗ +1
∗
+ o(x2l ) for x in a neighborhood
of 0.
We put
rbα∆ (x)
= arg min
r
n
X
K
i=1
rb∆ (x)
α
r∆
(x)
=
α
0
x − Xi
h
τα (ε∆
i.α − r),
if |b
rα∆ (x)| ≤ Ln (nhd )−1/2
(7)
(8)
otherwise.
In the main result of this section we will consider conditional moments of the truncated
kernel smoothing quantiles r∆
α , conditioned on the number of covariables falling into local
neighborhoods. Note that, with positive probability, kernel smoothing quantiles are not
defined because there is no covariable in the support of the kernel, with positive probability. Thus unconditional moments are not defined. In the following theorem we will
condition on local neighborhoods N − (x) = {u : xj − h ≤ uj ≤ xj + h for all j = 1, . . . , d}
that are designed such that the result can be easily used for the asymptotic analysis
of our test statistic in the next section. Note that N − (x) is the support of the kernel
h−d K(h−1 [x − ·]). The theorem could also easily be stated with other local neighborhoods. The conditional moments of the truncated kernel smoothing quantiles r∆
α will be
compared with the conditional moments of the following modified Bahadur expansion,
denoted by reα∆ (x) :
reα∆ (x) = reα∆,− (x) + ∆hα (x),
8
(9)
where ∆hα (x) is defined such that
h x − X n
oi
j
h
Exj K
I(ε∆
≤
∆
(x))
−
α
= 0.
j,α
α
h
(10)
Here Exj denotes the conditional expectation, given that j ∈ N − (x). Furthermore reα∆,− (x)
is defined as
Pn
reα∆,− (x)
x−Xi
h
h
{I(ε∆
i,α ≤ ∆α (x)) − α}
= − P
n
x−Xi
fεα |X (∆hα (x, Xi )|Xi )
K
i=1
h
Pn
x−Xi
K
{I(εi,α ≤ ∆hα (x, Xi )) − α}
i=1
h
= − P
,
n
x−Xi
h (x, X )|X )
K
f
(∆
i
i
εα |X
α
i=1
h
i=1
K
where
∆hα (x, Xj ) = ∆hα (x) − n−1/2 h−d/4 ∆α (Xj ).
We have the following asymptotic result for the moments of kernel quantile estimators
and their Bahadur approximations.
Theorem 1. Assume (B1)–(B3). Then, for natural numbers l ≥ 1,
o
n
2l
∆,−
2l
−
(x)
−
r
e
(x)
N
(x)
=
m
= O(Ln (nhd )−l−1 ),
E r∆,−
α
α
n
o
2l−1
∆,−
2l−1
−
E r∆,−
(x)
−
r
e
(x)
N
(x)
=
m
= O(Ln (nhd )−l ),
α
α
(11)
(12)
uniformly in x ∈ RX , α ∈ A and C1∗ nhd ≤ m ≤ C2∗ nhd where N − (x) is the random
∆
h
number of Xi ’s that lie in N − (x), and where r∆,−
α (x) = r α (x) − ∆α (x). For the second
moments of the uncentered estimators r∆
eα∆ we have that
α and r
n
o
2
∆
2
−
E r∆
(x)
−
r
e
(x)
N
(x)
=
m
= O(Ln n−3/2 h−5d/4 ).
α
α
(13)
Under the additional assumption that ∆α ≡ 0 we get that
n
o
2
∆
2
−
E r∆
(x)
−
r
e
(x)
N
(x)
=
m
= O(Ln (nhd )−2 ).
α
α
(14)
We can apply the theorem when ∆α ≡ 0, in which case ∆hα ≡ 0 and reα∆ (x) = reα∆,− (x)
∆,−
∆,−
and r∆
eα∆,− (x) replaced by
α (x) = r α (x). Hence, (11) and (12) hold with r α (x) and r
r∆
eα∆ (x). In particular, for l = 1 (14) follows directly from (11).
α (x) and r
9
3
Asymptotic theory
b
We suppose that there exists an estimator θ(α)
that converges to θ0 (α). Hence, on the
hypothesis the true value of θ(α) is equal to θ0 (α). On the alternative, θ0 (α) may depend
b
on the chosen estimator θ(α).
In order to develop the asymptotic distribution of TbA and Tbα , we need the following
additional assumptions.
(B4) We assume that
sup
x∈RX ,α∈A
>
− 1 −c
b
|mα,θ(α)
b (x) − mα,θ0 (α) (x) − (θ(α) − θ0 (α)) γα (x)| = OP (n 2 )
for some function γα (x). The function w(x) is continuous, and the functions w(x, α),
γα (x) and ∆α (x) are continuous with respect to (α, x). For g(x) = w(x, α) and
g(x) = w(x) it holds that |g(x0 )−g(x)| ≤ Ckx0 −xk, and |γα (x0 )−γα (x)| ≤ Ckx0 −xkδ
for some 0 < δ < 1 and for all x, x0 ∈ RX and all α ∈ A.
(B5) For some ρ > 0 it holds that
−1/2 −d/4
∗
−δ − 21 −ρ
∗
− 14 d4
b
sup kθ(α) − θ0 (α)k = OP (n
h
)/Ln ∧ (h n
) ∧ (n h /Ln ) .
α∈A
The first assumption in (B4) can be shown under smoothness conditions on the relation
θ → mα,θ (x). For the case that A contains only one single element, this assumption in (B4)
would directly follow from (B5) and the assumption that θ → mα,θ (x) has a derivative that
b
is continuous in x. Assumption (B5) states that θ(α)
achieves at least a nearly parametric
rate. In the case of linear quantile regression (i.e. mα (X) = θ(α)> X), such an assumption
b
has been shown in Angrist et al (2006). Note that supα∈A kθ(α)
− θ0 (α)k = OP n−1/2
1
1
d
implies (B5) if ρ is chosen such that hδ nρ → 0. Note that n− 2 = o(n− 4 h 4 /L∗n ) because
of nhd /L∗n → ∞.
We now state our main result on the asymptotic distribution of our test statistics.
10
Theorem 2. Assume (B1)-(B5). For the case that ∆α 6≡ 0 make the additional assumption that nh3d/2 /L∗n → ∞. Then,
d
nhd/2 TbA − bh,A → N (DA , VA ),
d
nhd/2 Tbα − bh,α → N (Dα , Vα ),
where
Z Z
∆α (x)2 w(x, α) dx dα,
DA =
A
RX
bh,A = h
VA
Z
Z
w(x, α)
dx dα,
2
A
RX fX (x)fεα |X (0|x)
Z
Z
w(x, α)w(x, β)
2
2
(4)
α (1 − β)
= 4K (0)
dx dα dβ,
2
4
α,β∈A,α<β
RX fX (x)fεα |X (0|x)
−d/2
K
(2)
α(1 − α)
(0)
Z
∆α (x)2 w(x) dx,
Dα =
RX
−d/2
bh,α = h
K
(2)
Z
(0)α(1 − α)
RX
Vα = 4K (4) (0)α2 (1 − α)2
Z
RX
w(x)
dx,
fX (x)fε2α |X (0|x)
w2 (x)
dx,
fX2 (x)fε4α |X (0|x)
and where for any j, K (j) (0) denotes the j-times convolution product of K at 0.
In our theorem for the alternative we make the additional assumption that nh3d/2 /L∗n
converges to ∞. This assumption is used in the proof for the treatment of the deterministic
term Tn,2 , see Lemma 5. The assumption nh3d/2 /L∗n → ∞ can be weakened but with
another limit for Tn,2 . This would result in a limit theorem for the test statistic with
a mean that differs from bh,α . We have added a short discussion of this point after the
statement of Lemma 5.
We expect that Theorem 2 cannot be used for an accurate calculation of critical values.
The asymptotic normality result of Theorem 2 is based on the fact that kernel smoothers
are asymptotically independent if they are calculated at points that differ more than 2h.
Thus the convergence is comparable to the convergence of the sum of h−d independent
summands. This would motivate a rate of convergence of order h−d/2 . As has been
11
suggested for other goodness-of-fit tests in the literature, also here a way out is to use
a bootstrap procedure. We will introduce some kind of wild bootstrap for quantiles in
which the Bahadur expansion reα of rbα is resampled. For the definition of reα see (5) in
Section 2. For the bootstrap, we define
Pn
reα∗ (x)
=−
x−Xi
h
{I(Ui ≤ α) − α}
,
x−Xi b
K
f
(0|X
)
i
ε
|X
α
i=1
h
i=1 K
Pn
where fbεα |X is an estimator of fεα |X and Ui are independent random variables with uniform
distribution on [0, 1] that are independent of the sample. The bootstrap test statistics are
defined as:
TbA∗
Z Z
=
A
RX
reα∗ (x)2 w(x, α)dx dα
and
Tbα∗ =
Z
RX
reα∗ (x)2 w(x)dx.
For proving the consistency of this bootstrap procedure, we do not specify the choice of
the estimator fbεα |X that is used in the construction of the bootstrap procedure. We only
assume that the estimator is consistent:
(B6) It holds that
sup
fbεα |X (0|x) − fεα |X (0|x) → 0,
α∈A, x∈RX
in probability.
The next theorem shows the consistency of the above bootstrap approach.
Theorem 3. Assume (B1)-(B6). Then,
p
dK (L∗ (nhd/2 TbA∗ − bh,A ), N (DA , VA )) → 0,
p
dK (L∗ (nhd/2 Tbα∗ − bh,α ), N (Dα , Vα )) → 0,
where L∗ (...) denotes the conditional distribution, given the sample. Furthermore, dK is
the Kolmogorov distance, i.e. the sup norm of the difference between the corresponding
distribution functions.
12
Theorem 3 remains to hold if we replace (B1)–(B5) by weaker conditions. We do
not pursuit this because we need for consistency of bootstrap that both, Theorem 2 and
Theorem 3, hold.
4
Numerical study
In this section, we present the results of our numerical studies. In our first simulation
we show that a direct application of the Bahadur representation is not accurate enough
for studying the approximation of the distribution of our test statistics Tbα and TbA . For
R
this purpose, we compare the differences D1 = |b
rα (x)2 − reα (x)2 |w(x)dx and D2 =
R
R
| rbα (x)2 w(x)dx − reα (x)2 w(x)dx|. Here D1 is the integrated difference between the
quantile regression and its Bahadur representation and D2 is the difference between the
test statistic and its approximation based on Bahadur representation. It is clear that
D2 ≤ D1 . Our point is not that D2 is smaller than D1 but that the ratio D1 /D2 is large
and that it is decreasing for an increasing bandwidth. This result supports our theory
that a direct use of Bahadur expansions only works under very restrictive assumptions
on the bandwidth. Table 1 shows the results of D1 and D2 for the one dimensional case.
We also simulated a two dimensional model, whose results are shown in Table 2. In
the one dimensional model, we set Yi = Xi + (0.5Xi + 0.5)i , where i has a standard
normal distribution. This results in the αth quantile function mα (x) = x + 0.5zα (x + 1),
where zα stands for the αth quantile of the standard normal distribution. For the two
dimensional model, we set Yi = X1i + X2i + (0.5X1i − 0.5X2i + 1)i , where i has a
standard normal distribution and we get the αth quantile function mα (x1 , x2 ) = x1 +2x2 +
0.5zα (x1 − x2 + 2). For the one dimensional model, we generated Xi from the uniform
distribution supported on the unit interval (0, 1). For the two dimensional model we
generated (X1i , X2i ) from a distribution on the unit square (0, 1)×(0, 1) which has uniform
marginals but where the joint distribution differs from a uniform distribution. This is done
to allow for a dependence between the two regressors. We generated random vectors from
13
a bivariate normal distribution with correlations ρ = 0.2 and 0.8 and then transformed
them with their marginal distribution functions. We generated 400 data sets of size 200
and 400 for each model. For the one dimensional model, we used the bandwidths h =
0.05, 0.08, 0.1, 0.12 and 0.15 and we used the bandwidths h = 0.125, 0.150, 0.175 and 0.200
for the two dimensional model. We used the R package quantreg for fitting quantile
functions. From Table 1 and Table 2, one can see that the ratio of D1 over D2 is large
for small bandwidths and decreases as the bandwidth grows. This observation supports
our approach for the asymptotic theory. This implies that when we approximate the
test statistic it requires less strict assumptions on the bandwidth if we approximate the
integrated function rather than when we approximate the quantile function itself.
The second simulation study is conducted to show the validity of our bootstrap procedure. We considered four scenarios:
I. All quantiles are linear:
Yi = m0 (Xi ) + σ0 (Xi )i .
II. The median is linear and other quantiles are not linear:
Yi = m0 (Xi ) + σ1 (Xi )i .
III. All quantiles are non-linear:
(a)
Yi = m1 (Xi ) + σ0 (Xi )i ;
(b)
Yi = m1 (Xi ) + σ1 (Xi )i .
Here m0 (x) = x, m1 (x) = sin(2π(x − 0.5)), σ0 (x) = 21 (1 + x), and σ1 (x) = 2(1.1 +
sin(2π(x−0.5))). The covariates Xi are generated from a uniform distribution on the unit
interval (0, 1). We generated 200 samples of size 400. We generated 201 bootstrap samples
for each data set. The three scenarios are shown in Figure 1. In the bootstrap procedure,
we used a kernel density estimator for estimating the conditional density fεα |X (0|x).
14
Bandwidth
n = 200 α = 0.25
0.5
0.75
n = 400 α = 0.25
0.5
0.75
0.05
0.08
0.1
0.12
0.15
D1
0.195
0.116
0.091
0.077
0.059
D2
0.067
0.044
0.037
0.034
0.028
Ratio
2.922
2.653
2.473
2.284
2.073
D1
0.231
0.139
0.110
0.091
0.071
D2
0.124
0.078
0.065
0.056
0.046
Ratio
1.865
1.779
1.687
1.635
1.547
D1
0.196
0.117
0.092
0.077
0.059
D2
0.065
0.044
0.037
0.033
0.029
Ratio
3.028
2.696
2.481
2.305
2.070
D1
0.094
0.057
0.045
0.037
0.029
D2
0.033
0.022
0.019
0.017
0.014
Ratio
2.899
2.536
2.338
2.170
2.027
D1
0.109
0.068
0.054
0.045
0.036
D2
0.058
0.039
0.032
0.028
0.023
Ratio
1.867
1.755
1.684
1.614
1.552
D1
0.094
0.056
0.045
0.037
0.029
D2
0.032
0.022
0.019
0.017
0.015
Ratio
2.959
2.604
2.378
2.192
1.982
Table 1: Difference between two approximations: D1 is the integrated squared approximation error of a quantile estimator by its Bahadur representation and D2 is the approximation error of the test statistic Tbα by Teα .
We tried three bandwidths 0.075, 0.100, and 0.125 for the test statistic, and fifteen
choices of bandwidths (0.1, 0.2, 0.3) × (0.050, 0.075, 0.100, 0.125, 0.150) for estimating the
conditional density of the error used in the bootstrap procedure. We applied the proposed bootstrap test for testing the linearity of the lower quartile, the median, and the
upper quartile functions. We also tested the linearity hypothesis over these three different
quantile levels. In our scenarios there are four models under the null hypothesis: all three
quantiles in scenario I and the median in scenario II. In Table 3, we report the summary
statistics of rejection ratios of 45 different choices of bandwidths. Among 45 different
15
16
0.077
3.221
D2
Ratio
2.093
Ratio
0.247
0.142
D2
D1
0.296
3.066
Ratio
D1
0.081
D2
3.221
Ratio
0.249
0.077
D2
D1
0.247
2.093
Ratio
D1
0.142
D2
3.065
Ratio
0.296
0.081
D2
D1
0.249
D1
2.929
0.047
0.138
1.884
0.087
0.166
2.716
0.051
0.139
2.929
0.047
0.138
1.884
0.087
0.164
2.716
0.051
0.139
0.150
2.446
0.035
0.085
1.667
0.061
0.102
2.369
0.036
0.086
2.446
0.035
0.085
1.667
0.061
0.102
2.369
0.036
0.086
0.175
1.990
0.029
0.058
1.500
0.048
0.072
1.983
0.029
0.058
1.990
0.029
0.058
1.500
0.048
0.072
1.983
0.029
0.058
0.200
3.274
0.022
0.073
1.991
0.044
0.087
3.273
0.023
0.074
3.355
0.045
0.153
2.021
0.089
0.180
3.445
0.044
0.153
0.125
Bandwidths
2.666
0.019
0.051
1.785
0.034
0.061
2.716
0.019
0.051
2.831
0.037
0.106
1.832
0.069
0.126
2.828
0.037
0.105
0.150
2.230
0.017
0.037
1.610
0.028
0.046
2.278
0.017
0.038
2.378
0.032
0.077
1.659
0.056
0.094
2.371
0.033
0.077
0.175
1.899
0.015
0.028
1.467
0.024
0.035
1.958
0.015
0.029
2.022
0.029
0.059
1.504
0.048
0.072
2.024
0.029
0.059
0.200
Strong dependence
left panel in the table is the result for ρ = 0.2 and the right panel shows the result for ρ = 0.8.
Table 2: Difference between two approximations under a two dimensional model: D1 is the integrated squared approximation
error of a quantile estimator by its Bahadur representation and D2 is the approximation error of the test statistic Tbα by Teα . The
0.75
0.5
n = 400 α = 0.25
0.75
0.5
n = 200 α = 0.25
0.125
Weak dependence
choices of bandwidths, there was no case where the bootstrap test did not keep the significance level of 5% under scenario I. In slightly more than half of the cases the bootstrap
test did not keep the significance level in testing the linearity of the median under scenario II. These cases appeared when we used large bandwidths. Concerning the power of
the bootstrap test, we observed that almost all choices of the bandwidth showed a power
near one under scenario III-(a). One exception is the case with the smallest bandwidths
where we observe an empirical power around 0.8. One interesting observation is that the
bootstrap test shows a poor power for the lower quartile in scenario III-(b) where the
empirical power ranges from 0.045 to 0.27. This is however natural since the function
is not so far from a linear function as one can see in Figure 1. We also observed that
the median and the upper quartile in scenario III-(b) showed much stronger power. The
result of the test based on the test statistic integrated over levels shows a similar result.
In this case, only scenario I is in the null hypothesis and there was no case where the
empirical size of the bootstrap test is bigger than 5%. The power behavior is also similar.
The test showed the strongest power with the bigger bandwidths.
Figure 2 shows the distributions of estimated p-values by using the proposed bootstrap.
The plots are based on 200 simulated data sets for scenario II. The left panel shows the
distribution of estimated p-values for testing the linearity of the median, which lies in
the null and the right panel shows the distribution of estimated p-values for testing the
linearity of the upper quartile, which lies in the alternative. The distribution in the left
panel is close to the uniform distribution which we expect for the null hypothesis and the
right panel shows that the bootstrap p-values are close to zero, which we also expect.
To summarize our observations from this simulation study, in our setting the bootstrap
test keeps the level well except for cases where we use too large bandwidths. On the other
hand, too small bandwidths lead to relatively poor power. Interesting cases are scenario
II and scenario III-(b). Under scenario II, the median is linear but both quartiles are
not. The simulation result shows that the bootstrap test keeps the level for the median
and has some power for the other quantiles. Under scenario III-(b), the lower quartile
17
Scenario II
0
y
0.5
−2
−0.5
y
2
1.5
Scenario I
0.4
0.8
0.0
0.4
0.8
x
x
Scenario III−(a)
Scenario III−(b)
0
y
0.0
−2
−1.5
y
2
1.5
4
0.0
0.0
0.4
0.8
0.0
x
0.4
0.8
x
Figure 1: Shape of quantile curves in each scenario. The three curves in each panel
represent the 0.25, 0.5, and 0.75−quantile curves in each scenario.
18
Quartile functions
Scenario I
Scenario II
Scenario III (a)
Scenario III (b)
1st Quartile
Median
3rd Quartile
Sum quartiles
Q1
0.005
0.000
0.005
0.000
Med
0.010
0.005
0.010
0.005
Q3
0.010
0.010
0.015
0.010
Q1
0.470
0.030
0.450
0.380
Med
0.575
0.055
0.545
0.550
Q3
0.630
0.105
0.610
0.695
Q1
0.985
1.000
0.995
1.000
Med
1.000
1.000
1.000
1.000
Q3
1.000
1.000
1.000
1.000
Q1
0.065
0.300
0.735
0.545
Med
0.135
0.385
0.865
0.770
Q3
0.155
0.440
0.935
0.895
Table 3: First quartile, median and third quartile (obtained from 45 choices of the bandwidth) of the rejection proportions based on 200 generated samples.
is non-linear but close to the null. We observe that here the bootstrap test has stronger
power for the median and the upper quartile than for the lower quartile.
In the last simulation study, we compared our test with the test proposed by Zheng
(1998). We considered the same four scenarios as in the previous simulation study. In
this simulation, we generated 500 data sets of 400 observations. For the bootstrap we
generated 501 bootstrap samples. The other simulation settings are the same as in the
previous simulation study. To choose the bandwidth for Zheng’s test, we applied the
function npregbw in the R-package np, which is based on cross-validation with AIC. For
the nonparametric quantile estimator in our procedure, we used the bandwidth proposed
in Yu and Jones (1998). The bandwidths for the kernel estimator for the conditional
density in the bootstrap procedure were chosen by a rule of thumb using the function
bw.nrd in R. The level of the tests was set to 0.05. We observe in Table 4 that the level of
our test is close to the nominal level. None of the two tests is always more powerful than
19
0
0
20 40 60 80
Frequency
120
10 15 20 25 30
p−value under H1
5
Frequency
p−value under H0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
p
0.6
0.8
1.0
p
Figure 2: The left panel shows the distribution of estimated p-values under the null and
the right panel shows the distribution of estimated p-values under the alternative.
the other. In the null model both tests keep the level well. In Scenario II, Zheng’s test
shows stronger power than the proposed test, whereas in Scenario III(b), the proposed
test has higher power.
Quartile functions
Scenario I
Scenario II
Scenario III (a)
Scenario III (b)
1st Quartile
Median
3rd Quartile
Sum quartiles
Zheng
0.006
0.000
0.000
·
MVY
0.024
0.052
0.048
0.028
Zheng
0.954
0.008
0.950
·
MVY
0.696
0.002
0.650
0.884
Zheng
0.974
0.994
0.984
·
MVY
0.994
0.992
0.990
1.000
Zheng
0.034
0.334
0.460
·
MVY
0.048
0.404
0.992
0.954
Table 4: Rejection proportions based on 500 generated samples.
Finally, as an illustrating example, we applied the proposed test to a historic data set
20
of Ernst Engel. The data set was used in Koenker (2005), among many other publications.
The data set was first presented by Engel (1857) to support his famous Engel’s law. The
data set has two variables, household income and food expenditure and it contains 235
observations. Figure 3 shows the scatter plot of this dataset and the scatter plot of the
data after a log transform with base 10. As one can see in Figure 3, there is one outlier. We
removed this point from the data. Hence the analysis below is based on 234 observations.
We first analyzed the log transformed income versus the log transformed food expenditure.
We used five different bandwidths for calculating the test statistic. The bandwidth for
the conditional density estimator used in the bootstrap resampling was chosen by a rule
of thumb. To obtain the bootstrap distribution, we resampled the data set 1,001 times.
We test the linearity of quantiles for α = 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9. As
can be seen from Table 5, the test did not reject linearity of quantiles for any of these
values at the significance level 5%. This was the case for all five bandwidth choices. We
also used the log transformed income with the original untransformed food expenditure
as a further example. Figure 4 shows the scatter plot of this dataset together with 0.1,
0.3, 0.5, 0.7, 0.9 linear quantile fits. The figure shows that high level quantiles deviate
from their linear fits. This is also seen by our test, since it rejects the linearity for high
level quantiles. The estimated p-values for the conditional quantiles of level 0.5 or higher
are smaller than 0.05 for every bandwidth we used.
5
Proof of Theorem 1
We need to show equations (11)–(13). Claim (13) follows from (11)–(12) because of
∆hα (x) = O(n−1/2 h−d/4 ). Furthermore, (14) is a direct consequence of (11), see the remark
after the statement of the theorem. It remains to show (11)–(12).
It holds that P (cm0 ≤ N − (x) ≤ Cm0 ) → 1, where we use the shorthand notation
m0 = nhd . At this point and in the following proofs we will make use of our convention
of using the symbols, c, C, ....
21
log10 (income) vs log10 (food expenditure)
Quantile level
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.050
0.289
0.246
0.629
0.999
0.996
1.000
0.999
1.000
1.000
0.075
0.152
0.383
0.997
0.992
0.993
0.979
1.000
1.000
1.000
0.100
0.105
0.745
0.997
0.964
0.971
0.970
0.999
1.000
0.993
0.125
0.100
0.986
0.988
0.908
0.895
0.992
1.000
1.000
0.963
0.150
0.149
0.996
0.976
0.894
0.843
0.997
0.999
1.000
0.935
Bandwidth
log10 (income) vs food expenditure
Quantile level
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.050
0.305
0.675
0.413
0.135
0.031
0.008
0.012
0.004
0.000
0.075
0.569
0.498
0.268
0.140
0.046
0.027
0.003
0.002
0.000
0.100
0.640
0.324
0.181
0.047
0.015
0.004
0.003
0.001
0.000
0.125
0.531
0.235
0.150
0.034
0.009
0.001
0.002
0.001
0.000
0.150
0.664
0.197
0.088
0.012
0.003
0.002
0.002
0.003
0.000
Bandwidth
Table 5: Estimated p-values for testing the linearity of conditional quantiles of Engel’s
data. The upper table shows the estimated p-values for testing the linearity of conditional quantiles of log10 (food expenditure) as a function of log10 (income) and the lower
table shows the estimated p-values for testing the linearity of conditional quantiles of
food expenditure as a function of log10 (income).
22
2000
●
3.2
●
1500
3.0
●
●●
●
1000
2.8
2.6
Y
●
●●
● ●
●
●
● ●
●● ●
●
● ●
●
●
●●
●
●
●
●●
●
●● ●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
2.4
1000
●●●
●
500
engel$foodexp
●
3000
5000
●
●●
●
●●●
●
●●●
● ●
●● ●●
● ●
●
●● ●
●
●
●
●
●
●
● ●●
●● ●
●●
●
●
●●
●
●●
●
●
●
●
●●
● ●● ●
●●● ●
●
●
●
●●
●●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●●●
● ●
●
●●
●●
●
●
●●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●●
●●
●●●
●
●●
●●
●●
●●
●●●
●
●
●
●
●●
●
●●
● ●
●
●●
●●
●●●
●●
●
●●
●●
●
●●
2.6
engel$income
3.0
3.4
X
Figure 3: The left panel shows the scatter plot of the original Engel data. The right panel
shows the scatter plot of log transformed data after removing one influential point. The
lines in the right panel represent linear quantile fits of levels 0.1, 0.3, 0.5, 0.7, and 0.9.
First note that
−1/2
rbα∆,− (x) ≤ um0
if and only if
x − X n
o
X
j
−1/2
∆
h
K
I(εj,α ≤ ∆α (x) + um0 ) − α ≥ 0.
h
−
j∈N (x)
Let
oi
h x − X n
j
−1/2
∆
h
I(εj,α ≤ ∆α (x) + um0 ) − α
gx,α (u) =
K
h
h x − X n
oi
j
−1/2
j
h
= Ex K
I(εj,α ≤ ∆α (x, Xj ) + um0 ) − α
h
h x − X
i
j
−1/2
= um0 Exj K
fεα |X (∆hα (x, Xj )|Xj )
h
h
i
1
x − Xj 0
j
h
+ u2 m−1
E
K
f
(∆
(x,
X
)|X
)
j
j
0
x
εα |X
α
2
h
Exj
−3/2
+O(Ln m0
),
23
(15)
2000
1500
●
●
●
1000
●●
500
food expenditure
●
●
● ●
● ●
●
●
●●
● ●●●
●
●
● ●●
●●●
●●● ●
●
●
●
●
●
●●
●
●
●
● ●●
● ●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
● ●●
●●
●
●
●
●
●●
●●● ●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●● ●
●● ● ●
●●
●●●
● ●
●
●●● ● ●
●
●●●
●
●●● ●
●●
●●●
●● ●●
●
● ●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●●
●
● ●●
●●
●
●●●
●
●● ●
● ●●
●●●●●● ●
● ●
●
●● ● ●
● ●
●●
●
●●● ●●
●● ●
● ●
● ●● ●●
2.6
2.8
3.0
3.2
●
●
●
●
●
●
●
●
●●
●
●
●
●
3.4
log10 income
Figure 4: The figure shows the scatter plot of log transformed income versus food expenditure after removing one influential point. The lines represent linear quantile fits of level
0.1, 0.3, 0.5, 0.7, and 0.9.
uniformly in |u| ≤ C ∗ L∗n , because of Assumption (B2). Then, with
ηj,α,u,x = K
x − X n
o
j
−1/2
h
I(ε∆
≤
∆
(x)
+
um
)
−
α
− gx,α (u),
0
j,α
α
h
we have that
P
rbα∆,− (x)
≤
−1/2
um0
N (x), N (x) = m
−
−
X
= P m−1/2
ηj,α,u,x ≥ −m1/2 gx,α (u) N − (x), N − (x) = m .
j∈N − (x)
We now argue that an Edgeworth expansion holds for the conditional density of Tη =
P
m−1/2 j∈N − (x) ηj,α,u,x , given N − (x), N − (x) = m, that is of the form
σ −1
s−3
X
m−r/2 Pr (−φ : {χν })(σ −1 [· − x]) + O(m−(s−2)/2 [1 + |σ −1 [· − x]|s ]−1 )
r=0
24
(16)
where, the error term holds uniformly in α ∈ A, |u| ≤ L∗n and x ∈ RX for C1∗ m0 ≤ m ≤
C2∗ m0 and constants C1∗ < C2∗ . Here, we use standard notation used e.g. in Bhattacharya
and Rao (1976), p. 53. In particular, σ 2 denotes the conditional variance of ηj,α,u,x , given
that j ∈ N − (x), and Pr (−φ : {χν }) denotes a product of a standard normal density φ
with a polynomial that has coefficients depending only on the conditional cumulants χν of
ηj,α,u,x of order ν ≤ s−1, given that j ∈ N − (x). Note that σ 2 and χν depend on u, α, x and
n and that we do not indicate this in our notation. Furthermore, the cumulants and the
variance converge to constants depending on α, uniformly in |u| ≤ L∗n and x ∈ RX . Note
that for n → ∞ the conditional distribution of ηj,α,u,x , given that j ∈ N − (x), converges
to the distribution of K(U )(Z − α) where U and Z are independent random variables,
U has a uniform distribution on [−1, 1] and Z is {0, 1}-valued with P (Z = 1) = α. This
helps to understand that limit theorems hold uniformly. The function Pr (−φ : {χν }) is
defined as
r
X
1
Pr (−φ : {χν })(u) =
(−1)r
m!
m=1
j
X
1 +...+jm
χj1 +2
χjm +2 (r+2m)
· ... ·
φ
(u),
(j1 + 2)!
(jm + 2)!
=r
see Section 7 in in Bhattacharya and Rao (1976). In our case expansion (16) follows from
Theorem 19.3 in Bhattacharya and Rao (1976). For this claim we have to verify that
their conditions (19.27), (19.29) and (19.30) hold. Our setting is slightly different from
theirs, since we consider triangular arrays of independent identically distributed random
variables instead of a sequence of independent random variables as is the case in Theorem
19.3 in Bhattacharya and Rao (1976). But the same proof applies because in our setting
we can verify the following uniform versions of (19.27), (19.29) and (19.30):
i
h
sup
Exj |ηj,α,u,x |s < ∞,
(17)
α∈A,|u|≤L∗n ,x∈RX ,n≥n0
Z
sup
α∈A,|u|≤L∗n ,x∈RX ,n≥n0
sup
q
gα,u,x
(t)dt < ∞ for some q > 0,
{gα,u,x (t) : |t| ≥ b} < 1 for all b > 0
(18)
(19)
α∈A,|u|≤L∗n ,x∈RX ,n≥n0
with gα,u,x (t) = |Exj [exp(itσ −1 ηj,α,u,x )]| for some n0 > 0. Note that gα,u,x , σ and ηj,α,u,x
depend on n.
25
Claim (17) follows by a direct argument using brute force bounds. For the proof of (18),
n
Pp
xd −Xd,j |
x1 −X1,j
I(ε∆
),
...,
k(
))
we consider the conditional density of Up =
(k(
j,α ≤
j=1
h
h
o
−1/2
−
∆hα (x) + um0 ) − α given the value of ε∆
j,α and given that Xj ∈ N (x) for j = 1, ..., p.
For p = 1 this density evaluated at (u1 , . . . , ud ) can be bounded by a constant times
(u1 · ... · ud )−κ ((k(0) − u1 ) · ... · (k(0) − ud ))−κ by Assumptions (B1) and (B3). This bound
holds uniformly over α, u, x and the value of ε∆
1,α .
We now show that for every κ∗ > 0, there exists p∗ > 0 such that the density of Up∗ can
∗
be bounded by a constant times (u1 ·...·ud )κ . For simplification of notation we assume for
the proof that d = 1. For the proof of the claim, we show first that for p ≥ (1−κ)−1 we get
that the conditional density of Up is uniformly bounded. This follows by an evaluation of
R1
Rv
convolution integrals where one uses that 0 u−κ1 (v − u)−κ2 du = v −(κ1 +κ2 −1) 0 w−κ1 (1 −
w)−κ2 dw ≤ Cv −(κ1 +κ2 −1) . Applied to U2 this gives that the density of U2 is bounded by
−(2κ−1)
Cu1
+ C|k(0) − u1 |−(2κ−1) + C(2k(0) − u1 )−(2κ−1) . Note that the density of U1 is
−κ
−κ
bounded by Cu−κ
≤ Cu−κ
. For the density of U3 we get the
1 (k(0)−u1 )
1 +C(k(0)−u1 )
−(3κ−2)
bound Cu1
+C|k(0)−u1 |−(3κ−2) +C|2k(0)−u1 |−(3κ−2) +C(3k(0)−u1 )−(3κ−2) . Finally,
for p ≥ (1 − κ)−1 it holds that pκ − (p − 1) ≤ 0 and we have that for p ≥ (1 − κ)−1 , the
density of Up is uniformly bounded. We now use that the k-fold convolution of a bounded
density with bounded support [0, z] for some z > 0 is bounded by C|u1 |k−1 . This gives
∗
that the density of Up∗ can be bounded by a constant times uκ1 if p∗ ≤ l0 (κ∗ + 1) with
l0 ≥ (1 − κ)−1 , l ∈ N. This result can be easily extended to d > 1.
P
From this result now we want to conclude that the conditional density of pj=1 ηj,α,u,x =
n
o P
n
Pp
xd −Xd,j
−1/2
x−Xj
x1 −X1,j
p
∆
h
K(
)
I(ε
≤
∆
(x)+um
)−α
=
k(
)·...·k(
)
I(ε∆
0
j,α
α
j,α ≤
j=1
j=1
h
h
h
o
−1/2
∆hα (x) + um0 ) − α is uniformly bounded, given the value of ε∆
j,α and given that Xj ∈
N − (x) for j = 1, ..., p. This follows immediately from the following result. Suppose that
Z = (Z1 , ..., Zd ) has support [0, 1]d and a density f that is bounded by f (z) ≤ Dz1 · ... · zd
26
then Z1 · ... · Zd has a density g that is bounded by D. For a proof of this result, note that
Z
g(u) = ∂u
f (z)dz1 ... dzd
z1 ·...·zd ≤u
v
1
∂u
f
, z2 , ..., zd
dv dz2 ... dzd
z2 · ... · zd
z2 · ... · zd
v≤u
Z
u
1
f
, z2 , ..., zd
dz2 ... dzd
z2 · ... · zd
z2 · ... · zd
Z
u
1
f
, z2 , ..., zd
dz2 ... dzd
z2 · ... · zd
u
Z
u
1
D
· z2 · ... · zd · dz2 ... dzd
z2 · ... · zd
u
Z
=
=
≤
≤
= D.
We now apply the result that for p chosen large enough, the conditional density of
Pp
j=1
ηj,α,u,x is bounded, given that Xj ∈ N − (x) for j = 1, ..., p, uniformly over α, u
and x. This implies that the square of this conditional density is integrable and by the
Fourier Inversion Theorem (see Theorem 4.1 (vi) in Bhattacharya and Rao (1976)) the
same holds for the squared modulus of its Fourier transform. Thus the modulus of the
P
−
Fourier transform of the conditional density of 2p
j=1 ηj,α,u,x , given that Xj ∈ N (x) for
j = 1, ..., 2p, is integrable. This shows (18) for q = 2p.
For the proof of (19) one applies the Riemann-Lebesgue Lemma (see Theorem 4.1
in Bhattacharya and Rao (1976)). Consider for simplicity the case where d = 1. For
Exj [exp(itηj,α,u,x )] one gets that
exp[itgx,α (u)]Exj [exp(itηj,α,u,x )]
Z x+h Z
h
i
x−z
−1/2
h
=
exp itK
{I(e ≤ ∆α (x, z) + um0 ) − α}
h
x−h
e∈R
Z x+h
× fεα |X (e|z)fX (z) de dz
fX (z)dz
x−h
Z
1
Z
=
−1
h
i
−1/2
h
exp itK (v) {I(e ≤ ∆α (x, x + hv) + um0 ) − α}
e∈R
Z
1
× fεα |X (e|x + hv)fX (x + hv) de dv
fX (x + hv)dv.
−1
27
The right hand side of this equation converges to
Z
1
Z
1
exp[it(1 − α)K(v)]dv + (1 − α)
α
exp[−itαK(v)]dv.
−1
−1
−1/2
This convergence holds uniformly in t ∈ R, α ∈ A, |u| ≤ L∗n m0
and x ∈ RX . By using
these facts we get (19) from the Riemann-Lebesgue Lemma.
By applying Theorem 19.3 in Bhattacharya and Rao (1976) with s ≥ 4 we get that
−1/2
P rbα∆,− (x) ≤ um0
N − (x), N − (x) = m
(20)
2 −s
= 1 − Φ µα (u) + m−1/2 ρα (u) 1 − µα (u)2 φ µα (u) + O m−1
(1
+
µ
(u)
)
,
α
0
uniformly in u, α and x for C1∗ m0 ≤ m ≤ C2∗ m0 and constants C1∗ < C2∗ . Here we have
used the fact that terms for r = 2, ..., s − 3 in the expansion (16) can be bounded by
2 −s
O m−1
(1
+
µ
(u)
)
. We used the following notation
α
0
µα (u) = −
3
Exj (ηj,α,u,x
)
m1/2 gx,α (u)
and ρα (u) =
,
3
σα (u)
6σα (u)
2
with σα2 (u) = Exj (ηj,α,u,x
). It is easy to show that, uniformly in |u| ≤ C ∗ L∗n ,
2 i
h x − X
i
−1/2
I(εj,α ≤ ∆hα (x, Xj ) + um0 ) − α
σα2 (u) = Exj K 2
h
+O(Ln m−1
0 )
−1/2
= A1 (α) + um0
3
Exj (ηj,α,u,x
)
=
Exj
h
K
3
A2 (α) + O(Ln m−1
0 ),
x − X
i
h
I(εj,α ≤
−1/2
+O(Ln m0
+
−1/2
um0 )
−α
3 i
)
−1/2
= A3 (α) + O(Ln m0
∆hα (x, Xj )
),
and that
−1/2
µα (u) = −um1/2 m0
1
−1/2
A1 (α)−1/2 A5 (α) − u2 m1/2 m−1
A6 (α)
0 A1 (α)
2
1
−3/2
+ u2 m1/2 m−1
(α)A2 (α)A5 (α) + O(Ln m−1
0 A1
0 )
2
28
with
h x − X
i
i
A1 (α) = Exj K 2
(1 − 2α)P (εj,α ≤ ∆hα (x, Xj )) + α2 ,
h
h x − X
i
i
2
h
j
(1 − 2α)fεα |X (∆α (x, Xj )|Xj ) ,
A2 (α) = Ex K
h
h x − X
i
i
3
3
j
2
h
A3 (α) = Ex K
(1 − 3α + 3α )P (εj,α ≤ ∆α (x, Xj )) − α ,
h
i
h x − X
i
2
2
j
h
A4 (α) = Ex K
(1 − 2α)P (εj,α ≤ ∆α (x, Xj )) + α .,
h
h x − X
i
j
A5 (α) = Exj K
fεα |X (∆hα (x, Xj )|Xj ) ,
h
h x − X
i
j
A6 (α) = Exj K
fε0α |X (∆hα (x, Xj )|Xj ) ,
h
−1/2
Note that µα (−u)2 = µα (u)2 + O(Ln m0
). Thus we get that uniformly in |u| ≤ C ∗ L∗n ,
m−1/2 ρα (u) 1 − µα (u)2 φ µα (u) − m−1/2 ρα (−u) 1 − µα (−u)2 φ µα (−u) (21)
= O(Ln m−1
0 ),
−1/2
m−1/2 ρα (u) 1 − µα (u)2 φ µα (u) = O(Ln m0 ).
−1/2
Note also that with um = um1/2 m0
(22)
, uniformly in |u| ≤ C ∗ L∗n ,
−1/2
1 − Φ µα (u) = 1 − Φ − um A1 (α)A5 (α)
u2
−1/2
−1/2
m
+φ − um A1 (α)A5 (α)
(A1 (α)A6 (α)
1/2
2m
−3/2
−A1
(α)A2 (α)A5 (α)) + O(Ln m−1
0 ).
Hence, uniformly in |u| ≤ C ∗ L∗n ,
h
i
−1/2
1 − Φ(µα (u)) + Φ(−µα (−u)) = 2 1 − Φ − um A1 (α)A5 (α) + O(Ln m−1
0 ).(23)
From (21), (23) and the above calculations it now follows for l ≥ 1 that with Dm (α) =
29
−1/2
m1/2 m0
−1/2
A1
(α)A5 (α) and rbα∆,− (x) = rbα∆ (x) − ∆hα (x)
o
n
−1/2
rα∆,− (x)| ≤ L∗n m0 ) N − (x) = m
E ml0 rbα∆,− (x)2l I(|b
Z
L∗n
= 2l
0
−1/2
v 2l−1 P rbα∆,− (x) > vm0
N − (x) = m dv
Z
0
−2l
−L∗n
Z
L∗n
= 2l
0
−1/2
v 2l−1 P rbα∆,− (x) ≤ vm0
N − (x) = m dv
h
−1/2
v 2l−1 P rbα∆,− (x) > vm0
N − (x) = m
i
−1/2
+P rbα∆,− (x) ≤ −vm0
N − (x) = m dv
Z
L∗n
= 2l
h
v 2l−1 Φ µα (v) − m−1/2 ρα (v) 1 − µα (v)2 φ µα (v)
0
i
+1 − Φ µα (−v) + m−1/2 ρα (−v) 1 − µα (−v)2 φ µα (−v) dv + O(Ln m0−1 )
Z
L∗n
= 4l
v
2l−1
Φ − vDm (α) dv + O(Ln m−1
0 )
0
= 4lDm (α)
−2l
L∗n Dm (α)
Z
w2l−1 Φ(−w) dw + O(Ln m−1
0 )
0
Z
h
∗ 2l
∗
−2l
= 2 (Ln ) Φ − Ln Dm (α) + Dm (α)
L∗n Dm (α)
i
v φ(v) dv + O(Ln m−1
0 )
2l
0
uniformly in C1∗ m0 ≤ m ≤ C2∗ m0 with constants C1∗ < C2∗ . If L∗n = (log n)γ is chosen
with γ > 0 large enough we get that the right hand side of the last equation is equal to
R∞
Dm (α)−2l −∞ v 2l φ(v) dv + O(Ln m−1
0 ). This follows since it can be easily shown that
Z
L∗n Dm (α)
2l
Z
∞
z φ(z) dz −
2
0
∗
∗
z 2l φ(z) dz = o(Ln n−C ) = o(m−C ),
−∞
∗
∗ 2l
∗
(Ln ) Φ − Ln Dm (α) = o(m−C )
with γ chosen depending on C ∗ . Thus we get for l ∈ N that
n
o
−1/2
rα∆,− (x)| ≤ L∗n m0 ) N − (x) = m
E rbα∆,− (x)2l I(|b
Z ∞
l
−l A1 (α)
=m
z 2l φ(z) dz + O(Ln (nhd )−l−1 ).
A2l
(α)
−∞
5
30
(24)
With similar arguments one can show that
n
o
−1/2
E rbα∆,− (x)2l−1 I(|b
rα∆,− (x)| ≤ L∗n m0 ) N − (x) = m
=
(2l−1)/2
A
(α)
m−(2l−1)/2 1 2l−1
A5 (α)
Z
(25)
∞
z 2l−1 φ(z) dz + O(Ln (nhd )−l )
−∞
= O(Ln (nhd )−l ).
In this case one applies (22) instead of (21).
For the proof of (11) and (12) it remains to show that uniformly in C1∗ m0 ≤ m ≤ C2∗ m0 ,
with constants C1∗ < C2∗ , and for l ∈ N
n
o
E reα∆,− (x)2l N − (x) = m
=
κ/2
A (α)
m−κ/2 1κ
A5 (α)
Z
(26)
∞
z 2l φ(z) dz + O(Ln (nhd )−l−1 ),
−∞
n
o
∆,−
2l−1
−
E reα (x)
N (x) = m
(27)
= O(Ln (nhd )−l ).
It remains to show (26) – (27). For the proof of (26) note that for independent random
variables Z1 , ..., Zm with mean zero, variance 1 and bounded 2l-th absolute moment it
holds that
m
n
o Z
X
−1/2
2l
E (m
Zi )
=
∞
z 2l φ(z) dz + O(m−1 ),
−∞
i=1
iid
∗
because for Z1∗ , . . . , Zm
∼ N (0, 1) one has
m
n
o
X
P
−1/2
2l
E (m
Zi )
= m−l ∗ E(Zi1 . . . Zi2l ) + O(m−1 )
i=1
= m−l
P∗
E(Zi∗1 . . . Zi∗2l ) + O(m−1 )
m
o
n
X
−1/2
= E (m
Zi∗ )2l + O(m−1 )
i=1
= E((Z1∗ )2l ) + O(m−1 )
Z ∞
=
z 2l φ(z) dz + O(m−1 ),
−∞
31
P∗
where the sum
runs over all indices i1 , . . . , i2l that are such that each value of an
index appears exactly two times. For the proof of (27) one applies that for independent
random variables Z1 , ..., Zm with mean zero, variance 1 and bounded 2l + 1-th absolute
moment it holds that
m
m
o
n
X
X
2l−1
−l+1 −1/2
−1/2
Zi )
= m
m
E(
Zi )2l−1
E (m
i=1
i=1
P
= m−l+1 m−1/2 ∗∗ E(Zi1 . . . Zi2l−1 ) + O(m−3/2 ) = O(m−1/2 ),
where the sum
P∗∗
runs over all indices that are such that one value of an index appears
three times and for all other 2l − 4 indices each value appears exactly two times. This
concludes the proof of the theorem.
6
Proof of Theorem 2
For the proof of Theorem 2 we will use the following corollary of Theorem 1. For the
statement of the corollary we have to define another construction of local neighborhoods.
For their definition suppose first that X is one-dimensional. Then the support RX is a
compact interval. For arbitrary j and for k ∈ {1, 2, 3}, we can then define
Ijk = [(3j + k − 1)h, (3j + k)h],
and
∗
Ijk
= [(3j + k − 2)h, (3j + k + 1)h].
∗
The set of indices of the Xi (i = 1, . . . , n) that fall inside the interval Ijk
is denoted by
Njk . We write Njk for the number of elements of Njk . An arbitrary x ∈ RX belongs to a
unique Ijk and we define N (x) = Njk and N (x) = Njk . Thus N (x) is an interval of length
3h, such that x lies in the middle subinterval of N (x) of length h. If the dimension of X
is larger than one, this partition of the support into small intervals can be generalized in
an obvious way.
Corollary 1. Assume (B1)–(B3). Then, for natural numbers l ≥ 1,
n
o
∆,−
2l
∆,−
2l
E rα (x) − reα (x) N (x) = m = O(Ln (nhd )−l−1 ),
n
o
2l−1
∆,−
2l−1
E r∆,−
(x)
−
r
e
(x)
N
(x)
=
m
= O(Ln (nhd )−l ),
α
α
32
uniformly in x ∈ RX , α ∈ A and C1∗ nhd ≤ m ≤ C2∗ nhd , where N (x) is the random number
∆
h
of Xi ’s that lie in N (x), and where r∆,−
α (x) = r α (x) − ∆α (x). For the second moments
of the uncentered estimators r∆
eα∆ we have that
α and r
o
n
2
2
∆
∆
E rα (x) − reα (x) N (x) = m = O(Ln n−3/2 h−5d/4 ).
Under the additional assumption that ∆α ≡ 0 we get that
n
o
∆
2
∆
2
E rα (x) − reα (x) N (x) = m = O(Ln (nhd )−2 ).
Proof of Corollary 1.
For m+ ≥ m we have by a simple argument with κ = 2l
o
n
n
κ
+
−
κ
∆,−
κ
(x)
N
(x)
=
m
,
N
(x)
=
m
=
E
or κ = 2l + 1 that E r∆,−
(x)
−
r
e
r∆,−
α
α
α (x) −
o
reα∆,− (x)κ N − (x) = m . Note that N − (x) ≤ N (x) because of N − (x) ⊂ N (x). Using (11)
and
m+
−
+
P N (x) ≤
N (x) = m
≤ C exp(−cnhd ),
4
uniformly in m+ ≥ 12 3d fX (x)nhd we conclude that
o
n
κ
∆,−
κ
+
(x)
−
r
e
(x)
N
(x)
=
m
= O(Ln (nhd )−l−1 ),
E r∆,−
α
α
uniformly in x ∈ RX , α ∈ A and 21 3d fX (x)nhd ≤ m+ ≤ 2 3d fX (x)nhd .
Since
P
1 d
3 fX (x)nhd ≤ N (x) ≤ 2 3d fX (x)nhd for all x ∈ RX
2
→ 1,
we get the statement of the corollary.
We now come to the proof of Theorem 2.
We only prove the statement for TbA . The asymptotic result for Tbα follows similarly.
We need to introduce a few more notations. With δθ,α (x) = −(θ(α) − θ0 (α))> γα (x) +
−1/2 −d/4
n−1/2 h−d/4 ∆α (x) and ε∆
h
∆α (Xi ) we define reα∆ as in (9) and we put
i,α = εi,α + n
n
X
x − Xi
∆
τα (εi.α + δθ,α (Xi ) − r).
rbα,θ (x) = arg min
K
r
h
i=1
∆
∆
Note that rbα∆ (x) = rbα,θ
(x), and that rbα (x) = rbα,
(x) + OP (n−1/2−c ) by Assumption (B4).
0
θb
We also define r∆
α as in (8). Let also
Wni (x, h) = Kh (x − Xi )/[
X
j
33
Kh (x − Xj )],
with Kh (·) = K(·/h)/hd .
The proof of Theorem 2 will make use of the following lemmas.
Lemma 1. Suppose that the assumptions of Theorem 2 are satisfied. Then,
sup sup rbα (x) = OP ((nhd )−1/2 Ln ),
(28)
sup sup rbα∆ (x) = OP ((nhd )−1/2 Ln ).
(29)
α∈A x∈RX
α∈A x∈RX
Proof of Lemma 1.
As is known for the case where there is no parametric part and
where ∆α ≡ 0, one has that
∆
sup sup rbα,θ
(x) − reα (x) = OP ((nhd )−3/4 Ln )
0
α∈A x∈RX
with reα defined as in (5). For a proof see Theorem 2 in Guerre and Sabbah (2012). By
standard smoothing theory we have that (still when ∆α ≡ 0)
sup sup reα (x) = OP ((nhd )−1/2 Ln ).
(30)
α∈A x∈RX
This shows (29) when ∆α ≡ 0. We can move from this case to ∆α 6= 0 by adding to the
observations terms of order OP (n−1/2 h−d/4 ). This changes the local quantiles by at most
this amount, and hence (29) still holds when ∆α 6= 0.
∆
In the case of rbα (x) = rbα,
(x) + OP (n−1/2−c ) we have to add to the observations terms
θb
of the order OP (Ln n−1/2 h−d/4 ) = OP ((nhd )−1/2 Ln ). This shows the first statement of the
lemma.
Lemma 2. Suppose that the assumptions of Theorem 2 are satisfied. Then,
b − θ0 (α))> γα (x) = OP (n− 21 −c ).
sup sup rbα (x) − rbα∆ (x) + (θ(α)
α∈A x∈RX
>
b
Proof of Lemma 2. First note that rbα (x)+(θ(α)−θ
0 (α)) γα (x) is equal to the quantile
estimator we would obtain when we shift all observations Yi in the window around x by
b − θ0 (α))> γα (x), and hence we need to show that the distance between
the amount (θ(α)
1
this latter estimator (say rbα,mod (x)) and rbα∆ (x) is OP (n− 2 −c ) uniformly in α and x.
34
Next, note that if now in addition we perturb all observations in the window around
>
b
x by adding mα,θ(α)
b (Xi ) − mα,θ0 (α) (Xi ) − (θ(α) − θ0 (α)) γα (Xi ), the quantile estimator
rbα,mod (x) will get perturbed by at most the maximal perturbation of the observations,
which is of the order OP (n−1/2−c ) by Assumption (B4).
After these two perturbations, the quantile estimator is now based on Yi −mα,θ0 (α) (Xi )+
b − θ0 (α))> (γα (x) − γα (Xi )) instead of Yi − m b (Xi ). Finally note that if we apply
(θ(α)
α,θ(α)
b
one more perturbation by subtracting (θ(α)
− θ0 (α))> (γα (x) − γα (Xi )) for all Xi in the
window around x, the estimator changes by at most OP (h−δ n−1/2−ρ hδ ) = OP (n−1/2−ρ ) by
Assumption (B5). The so-obtained estimator equals rbα∆ (x), which shows the statement of
the lemma.
Lemma 3. Suppose that the assumptions of Theorem 2 are satisfied. Then,
sup sup rbα∆ (x) − reα∆ (x) = OP ((nhd )−3/4 Ln ).
α∈A x∈RX
Proof of Lemma 3. Write
|b
rα∆ (x) − reα∆ (x)|
≤
=
1
n
X
inf x,α fεα |X (0|x)
i=1
1
n
X
inf x,α fεα |X (0|x)
Wni (x, h)fεα |X (0|Xi )b
rα∆ (x) +
n
X
Wni (x, h) I(ε∆
i,α ≤ 0) − α
i=1
Wni (x, h)fεα |X (0|Xi )b
rα∆ (x) − Fbε∆α |X (b
rα∆ (x)|x) + Fbε∆α |X (0|x)
i=1
d −1
+OP ((nh ) ),
(31)
P
where Fbε∆α |X (y|x) = i Wni (x, h)I(ε∆
i,α ≤ y). The latter equality follows from the fact
that
|Fbε∆α |X (b
rα∆ (x)|x) − α| ≤ |Fbε∆α |X (b
rα∆ (x)|x) − Fbε∆α |X (b
rα∆ (x) − |x)|
= OP ((nhd )−1 ).
The following expansion follows from standard kernel smoothing theory, uniformly for
35
d
x ∈ RX , α ∈ A, |y| ≤ an and for sequences an with a−1
n = O(nh ) :
Fbε∆α |X (y|x) − Fbε∆α |X (0|x)
Z y
X
fεα |X (u − n−1/2 h−d/4 ∆α (Xi )|Xi )du + OP ((nhd )−1/2 Ln a1/2
=
Wni (x, h)
n )
0
i
=
X
Z
Wni (x, h)
fεα |X (u|Xi )du + OP ((nhd )−1/2 Ln an1/2 ) + OP (n−1/2 h−d/4 an )
0
i
=y
y
X
Wni (x, h)fεα |X (0|Xi ) + OP ((nhd )−1/2 Ln an1/2 + a2n ) + OP (n−1/2 h−d/4 an ).
i
We now apply this bound to an = (nhd )−1/2 Ln and y = rbα∆ (x), which is possible thanks
to Lemma 1. This combined with (31) shows the statement of the lemma.
For proving Theorem 2, we will make use of the following decomposition, which follows
from Lemma 2 :
Z Z
TbA =
A
i2
b − θ0 (α))> γα (x) w(x, α) dx dα + oP (n−1 h−d/2 )
rbα∆ (x) − (θ(α)
h
rbα∆ (x)2
RX
Z Z
=
A
h
RX
Z Z
+
A
RX
Z Z
+
A
RX
h
n
oi
∆
2
∆
2
∆
2
∆
2
rα (x) − reα (x) − E rα (x) − reα (x) N (x) w(x, α) dx dα
h
−2
RX
Z Z
−2
A
RX
Z Z
+
A
(b
rα∆ (x)
−
reα∆ (x))
b − θ0 (α))> γα (x)
(θ(α)
i
w(x, α) dx dα
h
i
b − θ0 (α))> γα (x) w(x, α) dx dα
reα∆ (x) (θ(α)
h
i2
b − θ0 (α))> γα (x) w(x, α) dx dα
(θ(α)
RX
Z Z
+
A
i
w(x, α) dx dα
n
o
∆
2
∆
2
E rα (x) − reα (x) N (x) w(x, α) dx dα
Z Z
A
−
2
r∆
α (x)
RX
reα∆ (x)2 w(x, α) dx dα + oP (n−1 h−d/2 )
= Tn1 + ... + Tn7 + oP (n−1 h−d/2 ).
Lemma 4. Suppose that the assumptions of Theorem 2 are satisfied. Then,
Tn1 = oP (an ),
for any sequence {an } of positive constants tending to zero as n → ∞.
36
Proof of Lemma 4. Note that
Tn1 ≤ sup sup
α∈A x∈RX
|b
rα∆ (x)|2
Z Z
A
I |b
rα∆ (x)| > Ln (nhd )−1/2 w(x, α) dx dα.
RX
It is easily seen from Lemma 1 that
Z Z
A
I |b
rα∆ (x)| > Ln (nhd )−1/2 w(x, α) dx dα = oP (an ),
RX
for any an → 0, since the indicator inside the integral will be zero from some point on.
From Corollary 1 we get the following result.
Lemma 5. Suppose that the assumptions of Theorem 2 are satisfied. Then,
n
o
∆
2
∆
2
sup sup E rα (x) − reα (x) N (x) = oP ((nhd/2 )−1 ),
α∈A x∈RX
and hence, Tn2 = oP ((nhd/2 )−1 ).
At this point we needed the additional assumption nh3d/2 /L∗n → ∞ for the case that
∆α 6≡ 0. We now shortly outline what happens if we are on the alternative and if this
assumption does not hold. Note that for |m − m0 | = o(m0 )
o
n
2
∆
2
E r∆
(x)
−
r
e
(x)
N
(x)
=
m
α
α
o
n
h
2
∆,−
h
2
(x)
+
∆
(x))
−
(e
r
(x)
+
∆
(x))
N
(x)
=
m
= E (r∆,−
α
α
α
α
n
o
n
o
∆,−
2
∆,−
2
h
∆,−
= E rα (x) − reα (x) N (x) = m + 2∆α (x)E rα (x) N (x) = m .
For the first term on the right hand side we get from Corollary 1 that it is of order
o((nhd/2 )−1 ). For ∆hα (x) one can show that it is equal to n−1/2 h−d/4 ∆α (x)+O(Ln n−1/2 h−d/4 h2 ).
n
o
For the term E r∆,−
(x)
N
(x)
=
m
one can show that it is equal to (nhd )−1 ρ(x) +
α
O(Ln (nhd )−3/2 ) for some function ρ that does not depend on the function ∆α . This can
be done by using the arguments based on Edgeworth expansions that were central in the
proof of Theorem 1. This gives that
Tn2 = n
−3/2 −5d/4
Z Z
h
A
∆α (x)ρ(x)w(x, α) dx dα + o(n−3/2 h−5d/4 ) + o((nhd/2 )−1 ).
RX
37
Suppose now that nh3d/2 → 0. Then it holds that (nhd/2 )−1 = o(n−3/2 h−5d/4 ) and using
Lemma 10 we get that
−1 −d
Tn = n h K
(2)
Z
Z
α(1 − α)
(0)
A
+n−3/2 h−5d/4
RX
Z Z
A
w(x, α)
dx dα
fX (x)fε2α |X (0|x)
∆α (x)ρ(x)w(x, α) dx dα + oP (n−3/2 h−5d/4 ).
RX
This implies that the test rejects for large values of
R R
A
RX
∆α (x)ρ(x)w(x, α) dx dα. Thus,
in this high-dimensional setting the test behaves like a linear test and not like an omnibus
test.
Lemma 6. Suppose the assumptions of Theorem 2 are satisfied. Then,
Tn3 = OP (Ln n−5/4 h−3d/4 ) = oP ((nhd/2 )−1 ).
Proof of Lemma 6.
For simplicity of exposition of the argument, let us assume that
Xi is one-dimensional. For arbitrary j and for k ∈ {1, 2, 3}, define
Z Z
Ujk =
A
Ijk
h
n
oi
∆
2
∆
2
∆
2
∆
2
rα (x) − reα (x) − E rα (x) − reα (x) N (x) w(x, α) dx dα.
Then we can write Tn3 = Tn31 + Tn32 + Tn33 with Tn3k =
P
j
Ujk (k = 1, 2, 3). The
terms Tn31 , Tn32 and Tn33 are sums of O(h−1 ) conditionally independent summands. The
summands are uniformly bounded by a term of order OP (Ln n−5/4 h−1/4 ). This follows
from Lemma 5, from the fact that supα∈A supx |e
rα∆ (x)| = OP (Ln (nh)−1/2 ), see also (30),
and from the Bahadur representation for r∆
α (x), given in Lemma 3. It now follows that
Tn3k = OP (Ln n−5/4 h−3/4 ), which implies the statement of the lemma for d = 1. For d > 1
one can use the same approach.
Lemma 7. Suppose the assumptions of Theorem 2 are satisfied. Then,
Tn4 = oP ((nhd/2 )−1 ).
Proof of Lemma 7.
1
d
This is obvious, since Tn4 = OP (Ln (nhd )−3/4 n− 4 h 4 /L∗n ) =
oP ((nhd/2 )−1 ), thanks to Assumption (B5) and Lemma 3.
38
Lemma 8. Suppose the assumptions of Theorem 2 are satisfied. Then,
Tn5 = oP ((nhd/2 )−1 ).
Proof of Lemma 8. Write
Z Z Pn K x−Xi {I(ε∆ ≤ 0) − α}
i,α
i=1
h
b − θ0 (α))> γα (x)w(x, α) dx dα
(θ(α)
Tn5 = 2
Pn
x−X
i
A RX
fεα |X (0|Xi )
i=1 K
h
Z Z Pn K x−Xi {I(ε∆ ≤ 0) − α}
i,α
i=1
h
2
=
n A RX
gh,α (x)
b − θ0 (α))> γα (x)w(x, α) dx dα + oP ((nhd/2 )−1 )
×(θ(α)
Z
b − θ0 (α))> 1
= 2 (θ(α)
n
A
with gh,α (x) = E K
x−X
h
n
X
d/2 −1
ρh,α (Xi ){I(ε∆
) ),
i,α ≤ 0) − α} dα + oP ((nh
(32)
i=1
fεα |X (0|X) and
Z
ρh,α (v) =
K
RX
Using the notations Qh,α (Xi ) =
x − v γ (x)w(x, α)
α
dx.
h
gh,α (x)
ρ (X )
Pn h,α i
,
j=1 ρh,α (Xj )
Pn
∆
Fbε∆α (y) =
i=1 Qh,α (Xi )I(εi,α ≤ y) and
Fε∆α (y) = P (ε∆
α ≤ y), we have that
n
1X
ρh,α (Xi ){I(ε∆
i,α ≤ 0) − α}
n i=1
n
h
i 1 X
= Fbε∆α (0) − α
ρh,α (Xi )
n i=1
n
n
h
i 1 X
h
i 1 X
b
= Fε∆α (0) − Fε∆α (0)
ρh,α (Xi ) + Fε∆α (0) − α
ρh,α (Xi )
n i=1
n i=1
= OP (n−1/2 ) + OP (n−1/2 h−d/4 ),
uniformly in α ∈ A, and hence the statement of the lemma holds because of (32) and
(B5).
Lemma 9. Suppose the assumptions of Theorem 2 are satisfied. Then,
Tn6 = oP ((nhd/2 )−1 ).
39
Proof of Lemma 9. The statement of the lemma follows from (B5).
Lemma 10. Suppose the assumptions of Theorem 2 are satisfied. Then,
d
nhd/2 Tn7 − bh,A → N (DA , VA ).
Proof of Lemma 10.
The proof is very similar to the proof of e.g. Proposition 1 in
Härdle and Mammen (1993). Write
x − X x − X
XZ Z
j
i
∆
−2
K
{I(ε∆
Tn7 = n
K
i,α ≤ 0) − α}{I(εj,α ≤ 0) − α}
h
h
A
R
X
i,j
×b
gα (x)−2 w(x, α) dx dα,
∆
P
i
where gbα (x) = n−1 ni=1 K x−X
fεα |X (0|Xi ). By writing I(ε∆
i,α ≤ 0) − α = I(εi,α ≤
h
0) − I(εi,α ≤ 0) + I(εi,α ≤ 0) − α , we can decompose Tn7 into Tn7 = Tn71 + Tn72 + 2Tn73 .
As in Härdle and Mammen (1993), Tn73 is negligible. Straightforward calculations show
that Tn71 = (nhd/2 )−1 (DA + oP (1)). Indeed,
E(Tn71 |X1 , . . . , Xn )
x − X x − X
XZ Z
i
j
−2
K
K
Fεα |X (−n−1/2 h−d/4 ∆α (Xi )|Xi ) − Fεα |X (0|Xi )
=n
h
h
A RX
i,j
× Fεα |X (−n−1/2 h−d/4 ∆α (Xj )|Xj ) − Fεα |X (0|Xj ) gbα (x)−2 w(x, α) dx dα
x − X x − X
XZ Z
j
i
−2
=n
K
K
fεα |X (0|Xi )fεα |X (0|Xj )
h
h
A
R
X
i,j
×n−1 h−d/2 ∆α (Xi )∆α (Xj )b
gα (x)−2 w(x, α) dx dα (1 + oP (1))
Z Z
−1 −d/2
=n h
∆2α (x)w(x, α) dx dα (1 + oP (1))
A
RX
= n−1 h−d/2 DA (1 + oP (1)).
Next, write Tn72 = Tn72a + Tn72b with
Tn72a
n
1 X
=
Unii ,
n2 i=1
Tn72b =
1 X
Unij ,
n2 i6=j
40
where
Z Z
Unij =
K
A
RX
x − X x − X
i
j
K
{I(εi,α ≤ 0) − α}{I(εj,α ≤ 0) − α}
h
h
×b
gα (x)−2 w(x, α) dx dα.
By calculating its mean and variance it can be checked that nhd/2 Tn72a = bh,A + oP (1).
d
Thus for the lemma it remains to check that nhd/2 Tn72b → N (0, VA ). For the proof
of this claim one can proceed as in Härdle and Mammen (1993) and apply the central
limit theorem for U-statistics of de Jong (1987). For this purpose one has to verify that
P
4
n2 hd Var(Tn72b ) → VA , max1≤i≤n nj=1 Var(Unij )/Var(Tn72b ) → 0 and E[Tn72b
]/(Var(Tn72b ))2
→ 3. This can be done by straightforward but tedious calculations.
Proof of Theorem 2. The theorem follows immediately from Lemmas 4–10. Lemmas
4–9 imply the negligibility of the terms Tn1 , .., Tn6 . Lemma 10 shows the asymptotic
normality of nhd/2 Tn7 .
7
Proof of Theorem 3
The theorem can be shown by verification of the conditions of the central limit theorem
for U-statistics of de Jong (1987), in the same way as was done in the proof of Lemma
10. The crucial point in the proof is to note that I(Ui ≤ α) has the same distribution as
I(εi,α ≤ 0), and hence the calculations in the proof of Lemma 10 go through in this proof.
Acknowledgments
Research of the first author was prepared within the framework of a subsidy granted to
the HSE by the Government of the Russian Federation for the implementation of the
Global Competitiveness Program and it was supported by Deutsche Forschungsgemeinschaft through the Research Training Group RTG 1953. The research of the second author
was supported by the European Research Council (2016-2021, Horizon 2020 / ERC grant
41
agreement No. 694409), and by IAP research network grant nr. P7/06 of the Belgian
government (Belgian Science Policy).
References
Ait-Sahalia, Y., Fan, J. and Peng, H. (2009). Nonparametric transition-based tests for
diffusions. Journal of the American Statistical Association 104 1102-1116.
Angrist, J., Chernozhukov, V. and Fernández-Val, I. (2006). Quantile regression under
misspecification, with an application to the U.S. wage structure. Econometrica 74
539-563.
Bhattacharya, R. and Rao, R. (1976). Normal Approximations and Asymptotic Expansions. John Wiley & Sons, New York.
Bierens, H.J. and Ginther, D. (2001). Integrated conditional moment testing of quantile
regression models. Empirical Economics 26 307-324.
Chaudhuri, P. (1991). Nonparametric estimates of regression quantiles and their local
Bahadur representation. Annals of Statistics 19 760-777.
Conde-Amboage, M., Sánchez-Sellero, C. and González-Manteiga, W. (2015). A lack-offit test for quantile regression models with high-dimensional covariates. Computational Statistics and Data Analysis 88 128-138.
De Backer, M., El Ghouch, A. and Van Keilegom, I. (2017). Semiparametric copula
quantile regression for complete or censored data. Electronic Journal of Statistics
11 1660-1698.
de Jong, P. (1987). A central limit theorem for generalized quadratic forms. Probability
Theory and Related Fields 75 261-277.
Dette, H. and Sprekelsen, I. (2004). Some comments on specification tests in nonparametric absolutely regular processes. Journal of Time Series Analysis 25 159-172.
42
El Ghouch, A. and Van Keilegom, I. (2009). Local linear quantile regression with dependent censored data. Statistica Sinica 19 1621-1640.
Fan, J., Zhang, C. and Zhang, J. (2001). Generalized likelihood ratio statistics and
Wilks phenomenon. Annals of Statistics 29 153-193.
Gao, J. and Hong, Y. (2008). Central limit theorems for generalized U-statistics with
applications in nonparametric specification. Journal of Nonparametric Statistics 20
61-76.
González-Manteiga, W. and Cao-Abad, R. (1993). Testing the hypothesis of a general
linear model using nonparametric regression estimation. Test 2 161-188.
Guerre, E. and Lavergne, P. (2002). Optimal minimax rates for nonparametric specification testing in regression models. Econometric Theory 18 1139-1171.
Guerre, E. and Sabbah, C. (2012). Uniform bias study and Bahadur representation
for local polynomial estimators of the conditional quantile function. Econometric
Theory 28 87-129.
Haag, B. (2008). Non-parametric regression tests using dimension reduction techniques.
Scandinavian Journal of Statistics 35 719-738.
Härdle, W. and Mammen, E. (1993). Testing parametric versus nonparametric regression. Annals of Statistics 21 1926-1947.
He, X. and Ng, P. (1999). Quantile splines with several covariates. Journal of Statistical
Planning and Inference 75 343-352.
He, X., Ng, P. and Portnoy, S. (1998). Bivariate quantile smoothing splines. Journal of
the Royal Statistical Society - Series B 60 537-550.
He, X. and Zhu, L.-X. (2003). A lack of fit test for quantile regression. Journal of the
American Statistical Association 98 1013-1022.
43
Hjellvik, V., Yao, Q. and Tjøstheim, D. (1998). Linearity testing using local polynomial
approximation. Journal of Statistical Planning and Inference 68 295-321.
Hoderlein, S. and Mammen, E. (2009). Identification and estimation of local average
derivatives in non-separable models without monotonicity. Econometrics Journal
12 1-25.
Hong, S.Y. (2003). Bahadur representation and its applications for local polynomial
estimates in non-parametric M-regression. Journal of Nonparametric Statistics 15
237-251.
Horowitz, J.L. and Spokoiny, V.G. (2002). An adaptive, rate-optimal test of linearity
for median regression models. Journal of the American Statistical Association 97
822-835.
Ingster, Y.I. (1993). Asymptotically minimax hypothesis testing for nonparametric alternatives I, II, III. Math. Methods of Statistics 2 85-114, 171-189, 249-268.
Koenker, R. (2005). Quantile Regression. Cambridge University Press.
Koenker, R. and Machado, J.A.F. (1999). Goodness of fit and related inference processes
for quantile regression. Journal of the American Statistical Association 94 12961310.
Koenker, R. and Xiao, Z. J. (2002). Inference on the quantile regression process. Econometrica 70 1583-1612.
Kong, E., Linton, O. and Xia, Y. (2010). Uniform Bahadur representation for local
polynomial estimates of M-regression and its application to the additive model.
Econometric Theory 26 1529-1564.
Kreiss, J.P., Neumann, M.H. and Yao, Q. (2008). Bootstrap tests for simple structures
in nonparametric time series regression. Statistics and its Interface 1 367-380.
44
Lee, K.L. and Lee, E.R. (2008). Kernel methods for estimating derivatives of conditional
quantiles. Journal of the Korean Statistical Society 37 365-373.
Leucht, A. (2012). Degenerate U - and V -statistics under weak dependence: Asymptotic
theory and bootstrap consistency. Bernoulli 18 552-585.
Li, Q. and Racine, J.S. (2008). Nonparametric estimation of conditional CDF and
quantile functions with mixed categorical and continuous data. Journal of Business
& Economic Statistics 26 423-434.
Rothe, C. and Wied, D. (2013). Misspecification testing in a class of conditional distributional models. Journal of the American Statistical Association 108 314-324.
Su, L. and White, H.L. (2012). Conditional independence specification testing for dependent processes with local polynomial quantile regression. Advances in Econometrics
29 355-434.
Truong, Y.K. (1989). Asymptotic properties of kernel estimators based on local medians.
Annals of Statistics 17 606-617.
Volgushev, S., Birke, M., Dette, H. and Neumeyer, N. (2013). Significance testing in
quantile regression. Electronic Journal of Statistics 7 105-145.
Yu, K. and Jones, M. C. (1998). Local Linear Quantile Regression. Journal of the
American Statistical Association 93 228-237.
Zheng, J.X. (1996). A consistent test of a functional form via nonparametric estimation
techniques. Journal of Econometrics 75 263-289.
Zheng, J.X. (1998). A consistent nonparametric test of parametric models under conditional quantile regressions. Econometric Theory 14 223-238.
45
| 10math.ST
|
arXiv:1803.04078v2 [stat.ME] 30 Mar 2018
Minimum bias multiple taper spectral
estimation ∗
Kurt S. Riedel and Alexander Sidorenko
Courant Institute of Mathematical Sciences, New York University
New York, New York 10012-1185
EDICS: SP 3.1.1
Abstract
Two families of orthonormal tapers are proposed for multitaper
(k)
spectral analysis:
q minimum bias tapers, and sinusoidal tapers {v },
(k)
where vn = N 2+1 sin Nπkn
+1 , and N is the number of points. The resulting sinusoidal multitaper spectral estimate is Ŝ(f ) =
j
2N +2 ) − y(f
1
2K(N +1)
PK
j=1 |y(f +
j
2
2N +2 )| ,
−
where y(f ) is the Fourier transform of the stationary time series, S(f ) is the spectral density, and K is the number
of tapers. For fixed j, the sinusoidal tapers converge to the minimum
bias tapers like 1/N . Since the sinusoidal tapers have analytic expressions, no numerical eigenvalue decomposition is necessary. Both
the minimum bias and sinusoidal tapers have no additional parameter for the spectral bandwidth. The bandwidth of the jth taper is
simply N1 centered about the frequencies 2N±j+2 . Thus the bandwidth
of the multitaper spectral estimate can be adjusted locally by simply
adding
or deleting tapers. The band limited spectral concentration,
Rw
|V
(f
)|2 df , of both the minimum bias and sinusoidal tapers is very
−w
close to the optimal concentration achieved by the Slepian tapers. In
R 1/2
contrast, the Slepian tapers can have the local bias, −1/2 f 2 |V (f )|2 df ,
much larger than of the minimum bias tapers and the sinusoidal tapers.
∗
The authors thank D. J. Thomson and the referees for useful comments. Research
funded by the U.S. Department of Energy.
1
1
Introduction
We consider a stationary time series, {xn , n = 1 . . . N } with a spectral density, S(f ). A common estimator of the spectral density is to smooth the
square of the discrete Fourier transform (DFT) locally:
L
X
1
j
Ŝ(f ) =
|y(f + )|2 ,
(2L + 1)N j=−L
N
(1)
where y(f ) is the Fourier transform (FT) of the stationary time series:
PN
−i2πnf
y(f ) ≡
. Since (1) is quadratic in the FT, y(f ), it is
n=1 xn e
natural to consider a more general class of quadratic spectral estimators. We
examine quadratic estimators where the underlying self-adjoint matrix has
rank K, where K is prescribed. Using the eigenvector representation, the
resulting quadratic spectral estimator can be recast as a weighted sum of K
orthonormal rank one spectral estimators. This class of spectral estimators
was originally proposed by Thomson [19] under the name of multiple taper
spectral analysis (MTSA). We refer the reader to [3, 8, 11, 12, 16, 18, 19] for
excellent expositions and generalizations of Thomson’s theory.
In MTSA, a rank K quadratic spectral estimate is constructed by choosing an orthonormal family of tapers/spectral windows and then averaging
the K estimates of the spectral density. In practice, only the Slepian tapers
(also known as discrete prolate spheroidal sequences [17]) are routinely used
for MTSA.
In the present paper, we propose and analyze two new orthonormal families of tapers: minimum bias (MB) tapers
and sinusoidal tapers. The MB
R 2
tapers minimize the local frequency bias, f |V (f )|2 df , subject to orthonormality constraints, where V (f ) is the DFT of the taper. For continuous time,
the MB tapers have simple analytic expressions. The first taper in the family is Papoulis’ optimal taper [9]. For discrete time, the MB tapers satisfy a
selfadjoint eigenvalue problem and may be computed numerically.
(k)
In the
q case of discrete time, we define the kth sinusoidal taper, v , as
vn(k) = N2+1 sin Nπkn
, where N is the sequence length. The sinusoidal tapers
+1
are an orthonormal family that converge to the MB tapers with rate 1/N as
N → R∞. These results are given in Section 3. Section 4R compares the local
1/2
w
bias, −1/2 f 2 |V (f )|2 df , and the spectral concentration, −w
|V (f )|2 df , of the
MB tapers, the sinusoidal tapers and the Slepian tapers.
2
In Section 5, we show that the quadratic spectral estimator which minimizes the expected square local error is weighted multitaper estimate using
the MB tapers. A local error analysis is given and the optimal number of
tapers is determined. At frequencies where the spectral density is changing rapidly, fewer tapers should be used. In Section 6, we show that kernel
smoother spectral estimates [4, 10] are multitaper estimates and we show
that smoothing the logarithm of the multitaper estimate significantly reduces the variance in comparison with smoothing athe logarithm of a single
taper estimate. We also describe our data adaptive method for estimating
the spectrum. In Section 7, we apply our spectral estimation techniques to
real data and show that our tapers outperform the Slepian tapers whenever
a variable bandwidth is needed. In the Appendix, we show that the leading
principal components of kernel smoother spectral estimates resemble the MB
tapers.
2
Quadratic Estimators of the Power Spectrum
Let N discrete measurements, x1 , x2 , . . . , xN , be given as a realization of
a stationary stochastic process. We normalize the time interval between
measurements to unity. The Cramer representation of a discrete stationary
stochastic process [4, 12] is
x(t) =
Z 1/2
e2πinf dZ(f ) ,
−1/2
where dZ has independent spectral increments: E[dZ(f )dZ(g)] = S(f )δ(f −
g)df dg. We assume that the spectral density, S(f ), is twice continuously
differentiable.
The spectral inverse problem is to estimate the spectral density, S(f ),
given {xn }. As shown in [3, 8], every quadratic, modulation-invariant power
spectrum estimator has the form:
b )
S(f
=
N
X
qnm e2πi(m−n)f xn xm ,
(2)
n,m=1
where Q = [qnm ] is a symmetric matrix of order N and does not depend on
frequency. Consider the eigenvector decomposition: Q =
3
PK
k=1
µk v(k) v(k)
T
,
where K is the rank of Q , and v(1) , v(2) , . . . , v(K) is an orthogonal system
of eigenvectors. The multitaper representation of the quadratic spectral estimator is
b )
S(f
=
K
X
µk
N
X
2
vn(k) xn e−2πinf
.
(3)
n=1
k=1
In the case K = 1, estimator (3) turns out a tapered periodogram estimator:
Sbv (f ) =
N
X
2
vn xn e
−2πinf
,
(4)
n=1
with a taper v = (v1 , v2 , . . . , vN )T . If the tapering is uniform (i.e. v1 = v2 =
. . . = vN = √1N ), we name (4) the periodogram estimator. The estimator (3)
is a linear combination of K orthogonal tapered periodogram estimators. In
MTSA, K is normally chosen to be much less than N . The multiple taper
spectral estimate can be thought of as a low rank, “principal components”
approximation of a general quadratic estimator. Multiple taper analysis has
also been applied to nonstationary spectral analysis [13, 1].
In practice, one does not begin the analysis with a given quadratic estimator, Q . Instead, one usually specifies a family of orthonormal tapers
{v(1) , . . . , v(K) } with desirable properties. Previously, only the family of
Slepian tapers were used in practice. The goal of this article is to introduce other families of tapers.
We define the kth spectral window, V (k) , to be the FT of the kth taper:
V (k) (f ) =
N
X
vn(k) e−i2πnf .
(5)
n=1
The tapers are normally chosen to have their spectral density localized near
zero frequency. We define two common measures
of frequency localization.
R 1/2 2
The local bias of a spectral window V is −1/2 f |V (f )|2 df . The term
“local bias” is used because it is proportional to the leading order term in
the bias error of a taper estimate as N → ∞.
Rw
The spectral concentration in band [−w, w] is defined as −w
|V (f )|2 df .
The bandwidth, w, is a free parameter. The Slepian tapers are the unique sequences which maximize the spectral concentration subject to the constraint
that they form an orthonormal family. Detailed analysis of the Slepian sequences is given in [17]. We stress that the Slepian tapers depend on the
4
bandwidth parameter, w, and that the first 2N w spectral windows are concentrated in the band [−w, w] while the remaining windows are concentrated
outside.
3
3.1
Minimum Bias Tapers
Continuous Time Case
We consider time-limited signals; the time interval is normalized to [0, 1]. In
theR time domain, the taper ν(t) is a function in L2 [0, 1] which we normalize
to 01 ν 2 (t)dt = 1. The functions {sin(πkt), k = 1, 2, . . .} form a complete
orthogonal basis on [0, 1]. (Completeness can be proven by extending ν(t)
to be an odd function on [−1, 1] and using the completeness of the complex
exponentials on [−1,R 1]. See [5].)
P
to
Setting ak = 2 01 ν(t) sin(πkt)dt, then K
k=1 ak sin(πkt)dt converges
P
2
a
ν(t) in L2 [0, 1] as K → ∞. The taper normalization is equivalent to 21 ∞
k=1 k =
1. The Fourier transform of the taper is the complex-valued, spectral window
function:
Z 1
ν(t)e−i2πf t dt .
(6)
V (f ) =
0
V (f ) is defined
on the frequencyR domain [−∞, ∞], belongs to L2 [−∞, ∞],
R∞
and satisfies −∞ |V (f )|2 df = 2π 01 ν 2 (t)dt = 2π.
The local bias of a taper spectral estimate [4, 9, 10] is
S 00 (f ) Z ∞
|V (h)|2 h2 dh .
2
−∞
−∞
(7)
We consider tapers which minimize the leading order term:
b )]−S(f ) =
E[S(f
Z ∞
−∞
Z ∞
2 2
|V (g−f )|2 (S(g)−S(f ))dg ≈
|V (f )| f df =
Z 1
0
!2
∞
1 d X
ak sin(πkt)
2π dt k=1
∞
1Z 1 X
=
ak k sin(πkt)
4 0 k=1
dt
!2
dt =
∞
1X
a2 k 2 .
8 k=1 k
(8)
The last expression attains the global minimum when a21 = 2, a2 = a3 =
. . . = 0. Hence,
√ the leading order term in the bias expression (7) is minimal
for the taper 2 sin(πt) (This result was obtained by Papoulis [9]). In [9],
Papoulis extends ν(t) to be zero outside of (0, 1), and therefore has a Fourier
5
integral representation of ν(t). We have extended ν(t) to be periodic and
vanish at each integer value. Since ν(t) is optimized for t ∈ [0, 1], both
representations are valid.
Equation (8) implies the more general result:
√
Theorem 3.1 vk (t) = 2 sin(πkt) (k = 1, 2, . . .) is the only system of functions in L2 [0, 1] which satisfy the requirements:
(i)
R1 2
v (t)dt
0
k
= 1, and
∞
(ii)
vk minimizes −∞
|V (k) (f )|2 f 2 df in the subspace of functions orthogonal to v1 , . . . , vk−1 .
R
The kth minimum value is
R∞
−∞
|V (k) (f )|2 f 2 df =
k2
4
.
√
We name vk (t) = 2 sin(πkt) (k = 1, 2, . . .) the continuous time minimum
bias tapers. The Fourier transform of vk (t) is
h
k
k
e−iπ(f − 2 ) sin π f − 2
(k)
√
V (f ) =
i 2 π f − k2
f − k−1
2
= e−iπ(
i
h
− (−1)
sin π f +
k
k
2
π f+
k
2
i
k 2 sin πf − πk
2
)·
2 .
√
k
4 2π f 2 − 2
Thus, |V (k) (f )| decays as f −2 for large frequencies.
3.2
Discrete Time Case
We now consider the discrete time domain {1, 2, . . . , N } with the corresponding normalized frequency domain [− 12 , 12 ]. A taper is a vector, ν =
P
2
(ν1 , . . . , νN ), normalized by N
n=1 (νn ) = 1. By the same argument as in the
previous section, the leading order term of the bias of a taper spectral estiR1
mate is proportional to the local bias, −2 1 |V (f )|2 f 2 df , where the frequency
2
window, V (f ), is defined in Eq. (5).
Lemma 3.2 For the frequency window of a discrete time taper,
Z
1
2
− 12
|V (f )|2 f 2 df = ν A ν
6
∗
,
where A = [anm ] with
anm =
Z
1
2
− 12
( 1
i2π(n−m)f
e
2
f df =
if n = m ;
if n 6= m .
12
(−1)n−m
2π 2 (n−m)2
Corollary 3.3 The tapers ν (1) , ν (2) , . . . , ν (N ) , defined by the requirements
(i)
PN
(ii)
ν (k) minimizes
(k) 2
n=1 (νn )
orthogonal to
= 1,
R
1
2
− 12
|V (k) (f )|2 f 2 df in the subspace of vectors
ν (1) , . . . , ν (k−1) ,
are the eigenvectors of the matrix A sorted in the increasing order of the
R1
eigenvalues. The integral −2 1 |V (k) (f )|2 f 2 df is equal to the kth eigenvalue.
2
We name ν (1) , ν (2) , . . . , ν (N ) the discrete minimum bias tapers. They
can be approximated by the sinusoidal tapers, v(1) , v(2) , . . . , v(N ) , which are
discrete analogs of the continuous time minimum
bias tapers. Namely, we
q
(k)
(k) T
2
(k)
(k)
define v = (v1 , . . . , vN ) with vn = N +1 sin Nπkn
, k = 1, 2, . . . , N .
+1
Lemma 3.4 The sinusoidal tapers, v(1) , v(2), . . . , v(N ) , form an orthonormal
k2
1 + O( N1 ) .
basis in RN and the local bias of v(k) is 4N
2
Corollary 3.5 The multitaper estimate (3) using K sinusoidal tapers has
P
k2
K2
local bias equal to K
k=1 µk 4N 2 + O( N 3 ) and has the following representation:
Ŝ(f ) =
K
X
µj
j
j
|y(f +
) − y(f −
)|2 .
2(N
+
1)
2N
+
2
2N
+
2
j=1
The uniformly weighted estimate, µk =
2
O( K
).
N3
1
,
K
has local bias equal to
(9)
K2
12N 2
+
From (9), the reason for the low bias of the sinusoidal tapers is apparent:
the frequency sidelobe from y(f + 2Nj+2 ) cancels the sidelobe of y(f − 2Nj+2 ).
As a result, the sidelobe of y(f + 2Nj+2 ) minus y(f − 2Nj+2 ) is much smaller
than that of the periodogram.
Our preferred weighting is the parabolic weighting: µj = C(1 − j 2 /K 2 )
because the parabolic weighting minimizes the expected square error in kernel
smoothers as K and N tend to infinity. Since the weights decrease smoothly
to zero, the resulting estimate is smooth in frequency.
7
Corollary 3.6 The uniformly weighted multitaper estimate using K sinusoidal tapers can be computed in O(N ln N ) + O(KN ) operations while the
generic multitaper estimate requires O(KN ln N ) operations plus the cost of
computing the K tapers.
The following
result demonstrates
that the kth
i h
i spectral window is conh
k+1
k−1
k+1
k−1
,
,
−
∪
−
.
centrated on 2(N
+1) 2(N +1)
2(N +1)
2(N +1)
Corollary 3.7 The Fourier transform of v(k) equals
h
V (k) (f ) =
k
e−iπ((N +1)f − 2 ) sin N π f −
k
2(N +1)
i
h
i
sin N π f + 2(Nk+1)
k
h
i − (−1)
h
i
q
sin π f + 2(Nk+1)
i 2(N + 1) sin π f − 2(Nk+1)
(N +1)f − k−1
2
= e−iπ(
sin Nπk
+1
)·q
h
·
2
2(N + 1) sin (πf ) −
i
πk
2
2
πk
sin 2(N +1)
sin (N + 1)πf −
.
q
Thus |V (k) (f )| = N2+1 for |f | = 2(Nk+1) , and |V (k) (f )| ∼ N 13/2 for f =
πk
1
O(1). In particular, |V (k) (f )| ≈ √2N
3/2 when f → 2 . In the intermediate
frequencies, 2(Nk+1) < f 12 , |V (k) (f )| decays as f12 .
Numerical evaluation (see Table 1) shows that kv(k) − ν (k) kL2 < 4(Nk+2)
(k)
− k νν(k) k
L∞
k
. (k·kL∞ is the supremum norm in the time domain.) Figure 1 plots the
2(N +2)
envelopes of |V (k) (f )|2 for both the minimum bias and sinusoidal tapers with
N = 200, k = 1. The spectral energy of both tapers is nearly identical for
|f | < .25. Near the Nyquist frequency, the spectral energy of the sinusoidal
taper is roughly three times larger than that of the minimum bias taper. In
the time domain, the sinusoidal tapers are virtually indistinguishable from
the MB tapers.
for all k. The same rate of convergence is observed in L∞ norm:
4
v(k)
kv(k) kL∞
Comparison of Spectral Localizations
We now compare the local bias and the spectral concentration of three families of orthonormal tapers: minimum bias (MB) tapers, sinusoidal tapers
and Slepian tapers. Both the local bias and the spectral concentration of the
8
<
L∞
Slepian tapers depend on the bandwidth parameter, w. For properties of the
Slepian tapers, we refer the reader to [12, 16, 17, 18, 19].
Since the MB tapers minimize the local bias, clearly the sinusoidal tapers
and the Slepian tapers have larger local bias. The only question
is whether
R 1/2 2 (k)
P
(f )|2 df ,
the difference is large or small. Table 2 gives the local bias, K
k=1 −1/2 f |V
of the three families of tapers for N = 50. The sinusoidal tapers come within
0.2% of achieving the optimal local bias. In contrast, the local bias of the
Slepian tapers can be many times larger.
We compute the local bias for three different values of the bandwidth, w.
The general pattern is that the kth Slepian taper has roughly the same local
bias as the MB taper does when N w < k < 2N w. The ratio of the local bias
of the Slepian tapers to that of the MB tapers is smallest at k ≈ 1.2N w. As
|k − 1.2N w| increases, the local bias rapidly departs from the optimal value.
Table 3 compares the spectral concentration of the tapers for N = 50 and
w = .08 . Both the MB tapers and the sinusoidal tapers are within 1.7% of
the optimal value, except for k = 8,R 9. Notice that 2N w = 8. Although the
w
ratio of the spectral concentration, −w
|V (f )|2 df , for the MB and sinusoidal
tapers to that of the Slepian tapers is usually very close to one, the ratio of
the spectral energy outside of the frequency
band |f | < w can be
quite large.
Rw
Rw
2
Thus our conclusions depend on using −w |V (f )| df and not 1− −w |V (f )|2 df
as the measure of frequency concentration.
Figures 1-3 compare |V (k) (f )|2 of the Slepian and MB tapers for N = 200.
For Figures 1 and 2, we select the Slepian parameter, w, equal to .01 so that
K = 2N w equals four. Figure 2 plots |V1 (f )|2 for the frequencies up to
f = 0.14. The central peak of the MB taper is more concentrated around
f = 0 than the Slepian taper is. The first sidelobe of the MB taper is visible
while the first Slepian sidelobe is much smaller.
Figure 1 plots the logarithm of |V1 (f )|2 over the entire frequency range.
The MB taper has smaller range bias in the frequency range |f | < 0.3w and
in the frequency range |f | > 0.13. In the middle frequency range, the Slepian
taper is clearly better. The Slepian penalty function maximizes the energy
inside the frequency band, [−w, w], and thus it is natural that the Slepian
tapers do better for f ∼ w. By using a discontinuous penalty function, the
Slepian spectral windows experience Gibbs phenomenon and decay only as
1
, (|V (f )|2 ∼ f12 ). The MB spectral windows decay as f12 , and thus, it is
f
natural that the MB tapers have lower bias for f ∼ O(1).
P
Figure 3 plots 3k=1 |V (k) (f )|2 for |f | < 0.14 and on this scale, the MB
9
tapers are clearly preferable to the Slepian tapers. For larger frequencies, the
P
(k)
(f )|2 , is very similar to Fig. 2
energy of the multitaper estimate, K
k=1 |V
on the logarithmic scale provided that K << N .
In summary, the sinusoidal tapers perform nearly as well as the MB tapers
while the Slepian tapers have several times larger local bias (except when
k ≈ 1.2N w). For k 2N w, the Slepian tapers have better broad-band bias
protection than the minimum bias tapers do. For k ∼ 2N w, the minimum
bias tapers provide both smaller local bias and better broad-band protection
due to the Gibbs phenomena which the Slepian tapers experience.
5
Local Error Analysis and Optimal Multitapering
We now give a local error analysis of MTSA and determine the optimal
number of tapers. Our results are the multitaper analog of the local error
analysis of the smoothed periodogram [4, 10]. We assume the time series is
a Gaussian processes and do not consider frequencies near f = 0 and f =
1/2. In this case, the variance of the multitaper estimate is approximately
P
2
Variance[Ŝ(f )] ≈ S(f )2 K
k=1 µk due to the orthonormality of the tapers
Asymptotically, the local bias of the multitaper estimate of Eq. (3) is
Bias[Ŝ] = S(f )
K
X
1
λk µk ,
µk − 1 + S 00 (f )
2
k=1
k=1
K
X
!
R 1/2
where λk = −1/2 f 2 |V (k) (f )|2 df . The second term is the MT generalization
P
of (7). When K
µ 6= 1, the MT estimate has bias even in white noise.
k=1
P k
When we require K
k=1 µk = 1, the local expected loss simplifies:
Theorem 5.1 For a Gaussian process, away from f = 0 and f = 1/2, the
P
expected square error of the multitaper spectral estimate (3) with K
k=1 µk = 1
is asymptotically (to leading order in K/N )
"
2
Bias + Variance ≈
K
X
1 00
S (f )
λk µk
2
k=1
#2
+ S(f )2
K
X
µ2k .
(10)
k=1
Theorem 5.2 The multitaper estimate which minimizes the local loss (10)
(with µk ≥ 0) is constructed with the minimum bias tapers.
10
Proof: We order the µk such that µ1 ≥ µ2 ≥ . . . ≥ µK and define µK+1 =
P
∗
0. Since the weights, µk are fixed, we need to minimize K
k=1 µk uk A uk
over all sets of K orthonormal tapers, u1 , . . . , uK . We split the series in K
subseries and minimize each subseries separately:
min
u1 ,...,uK
K
X
µk uk A u∗k =
min
u1 ,...,uK
k=1
K
X
≥
K
X
(µk − µk−1 )
K
X
uj A u∗j
k
X
(k)
(µk − µk−1 ) min
(k)
(k)
u1 ,...,uk
(µk − µk−1 )
k
X
(k)∗
uj A uj
j=1
K
X
λA,j =
j=1
k=1
j=1
k=1
k=1
=
k
X
µk λA,j , (11)
k=1
where the λA,j are the eigenvalues of A , given in increasing order. The
(k)
(k)
(k)
uj are subject to orthonormality constraints that uj · uj 0 = δj,j 0 , but are
otherwise independent and minimized separately. In the last line of (11), we
P
P
(k)
(k)∗
use Fan’s Theorem [6]: minu(k) ,...,u(k) kj=1 uj A uj = kj=1 λA,j , where the
1
k
(k)
uj
are again subject to orthonormality constraints. The theorem is now
proved because the MB tapers are precisely the eigenvectors of A .
Theorem 5.3 The uniformly weighted multitaper estimate using K sinusoidal tapers has an asymptotic local loss of
"
2
Bias + Variance '
S 00 (f )K 2
24N 2
#2
+
S(f )2
.
K
(12)
Corollary 5.4 The asymptotic local loss of (12) is minimized when the number of tapers is chosen as
12S(f )N 2
∼
S 00 (f )
"
Kopt
#2/5
.
(13)
Thus, the optimal number of tapers is proportional to N 4/5 and varies
with the ratio of S(f ) to S 00 (f ). Intuitively (13) shows that fewer tapers
should be used when the spectrum varies more rapidly. A key advantage
of the MB and sinusoidal tapers is that the tapers need not be recomputed
as K is changed. In contrast, the Slepian tapers are most efficient when
the bandwidth parameter, w, is chosen such that K ∼ 2N w. Thus, when
the number of tapers is changed, as in (13), the Slepian tapers should be
recomputed.
11
6
Smoothed Multitaper Estimates
In our own comparison of kernel smoothing and multitaper estimation [16],
we found that a smoothed multiple taper estimate worked best. We now
evaluate the expected error of the kernel smoothed multitaper estimator and
show that smoothing the logarithm of the multitaper estimate is useful for
estimating the logarithm of the spectrum.
We begin by evaluating that the quadratic estimator (2) which is equivalent to a kernel smoother estimates of the spectrum [4, 10]. Let Ŝ(f ) be
the quadratic spectral estimator (2), and smooth it with a kernel κ(·) of
halfwidth w :
Z w
g b
bb
+ g)dg ,
(14)
S(f ) =
κ( )S(f
w
−w
where w is the bandwidth parameter and κ(·) is a kernel smoother with
domain [−1, 1]. This can be rewritten as
b ) =
S(f
b
N
X
q̃ nm ei2π(m−n)f xn xm ,
(15)
n,m=1
w
where q̃ nm = qnm κ̂m−n with κ̂m = −w
κ(g/w)e2πimg dg. Thus smoothing replaces the original quadratic estimator with matrix [qnm ] by another
quadratic estimator with matrix Q̃ = [q̃ nm ]. By Theorem 5.2, this hybrid
method cannot outperform the pure multitaper method with minimum bias
tapers.
We now show that combining kernel smoothing with multitapering does
improve the estimation of the logarithm of the spectral density, θ(f ) =
log[S(f )]. One standard approach is to kernel smooth the logarithm of the
tapered periodogram. This approach has the disadvantage that |y(f )|2 has a
χ22 distribution and log[χ22 ] has a long lower tail of its distribution. As a result, log[|y(f )|2 ] has an appreciable bias and its variance is inflated by π 2 /6.
A common alternative is to estimate the spectrum either by kernel smoothing
or by multitapering and then to take logarithms. This approach has the disadvantage that the smoothed spectral estimate tends to be more sensitive to
nonlocal bias effects than the corresponding smoothed log-spectral estimate.
To robustify the log-spectral estimate while reducing the variance inflation from the long tail, we propose the following hybrid estimate: 1) compute
the multitaper estimate using the sinusoidal tapers with µk = K1 and then
2) smooth θ̂M T (f ) ≡ ln[ŜM T (f )] − BK /K, where BK is the bias of ln[χ22K ].
R
12
(BK ≡ ψ(K) − ln K where ψ(K) is the digamma function). For white noise,
the variance of θ̂M T (f ) = ψ 0 (K)∼ K1 + 2K1 2 , so the variance enhancement
from the logarithm tends rapidly to zero.
In [15], we show that the asymptotic error for this scheme is
K2
bk w 2 +
24N 2
"
00
θ (f )
2
#2
1
Cκ
1+
+
Nw
2K
2
,
(16)
where bκ and Cκ are constants which depend on the kernel shape. In (16), we
assume uniformly weighted sinusoidal tapers are used and 1 K N w. In
1
) is the variance enhancement from the logarithmic
(16), one factor of (1 + 2K
1
transformation and one factor of (1 + 2K
) arises in the variance calculation
of (15) with sinusoidal tapers. Optimizing (16) with respect to both w and
K yield w ∼ N −1/5 and K ∼ N 8/15 , thus the smoothing halfwidth w is much
larger than K/N . The expected error (16) depends weakly on K provided
that 1 K N w.
For simplicity, we set K = N 8/15 and optimize (16) with respect to the
halfwidth w. The resulting halfwidth depends on θ00 (f ): wopt (θ00 (f )) with
wopt ∼ |θ00 (f )|−2/5 N −1/5 . Thus when the log-spectrum varies rapidly, the
halfwidth should be reduced as |θ00 (f )|−2/5 .
Since θ00 (f ) is unknown, we consider two stage estimators which begin by
making a preliminary estimate of θ00 (f ) prior to estimating θ(f ). We then
insert the estimate θ00d
(f ) into the expression for wopt : w(f ) = wopt (θc00 (f )) and
use a variable halfwidth kernel smoother with halfwidth ŵ(f ) to estimate
θ(f ). Multiple stage kernel estimators are described in [2, 7, 13, 14, 15].
These multiple stage schemes have a convergence rate of N −4/5 and have a
relative convergence rate of at least N −2/9 . A more detailed description is
given in [2, 13, 15].
7
Application
We now compare spectral estimates on an actual data series. We use the
microwave scattering data set which is described in [16]. The data measures
turbulent plasma fluctuations in the Tokamak Fusion Test Reactor at Princeton. The spectrum is dominated by a 1 MHz peak which is quasicoherent.
The spectral density varies by over five orders of magnitude.
13
The bias versus variance trade-off of Sec. 5 shows that fewer tapers should
be used near the peak. To make the spectral estimate smooth, a parabolic
weighting of the tapers is used as described in Sec. 3. To determine how
many tapers to use locally, we use the multiple stage “plug-in” method as
described in the previous section; i.e. we determine the number of tapers
using a pre-estimate on the same data. To reduce the fluctuations from the
estimate of the optimal number, we use a longer data segment to determine
the number of tapers at each frequency. We find the optimal number of
sinusoidal tapers is roughly 24 for frequencies in the 200 to 800 kHz range.
Near the 1 MHz peak, as few as 12 tapers are used to minimize the local bias
error. Between 1300 and 2400 kHz, the spectrum is flatter and we use up to
40 tapers.
The dotted line is the sinusoidal multitaper estimate, and the solid, more
wiggly, curve is the corresponding Slepian estimate using 24 tapers with
w = 60 kHz. The 1 MHz peak is poorly resolved in the Slepian estimate, and
the regions of high curvature are artificially flattened. For f ≥ 1.5 MHz, the
Slepian estimate is artificially bumpy due to statistical noise. The variable
taper number estimate suppresses these bumps by averaging over a larger
frequency halfwidth. We have also used a variable taper number estimate
with the Slepian tapers. Since the Slepian parameter, w was fixed at 100 kHz
to allow for forty tapers, the artificial broadening was even more exreme.
Comparing with a converged estimate of the spectrum based on N =
45, 000 shows that the sinuosidal taper estimate is more accurate. Another
significant difference is that the Slepian multitaper estimate requires much
more CPU time than the sinusoidal multitaper estimate.
8
Conclusion
We have
q proposed and analyzed the minimum bias and the sinusoidal tapers,
(k)
, for multitaper spectral estimation. The resulting sinuvn = N2+1 sin Nπkn
+1
j
soidal multitaper spectral estimate is Ŝ(f ) = 2K(N1 +1) K
j=1 |y(f + 2N +2 ) −
y(f − 2Nj+2 )|2 . The sinusoidal tapers have low bias because the frequency
sidelobe from y(f + 2Nj+2 ) cancels the sidelobe of y(f − 2Nj+2 ).
P
R 1/2
The minimum bias tapers minimize the local bias, −1/2 f 2 |V (k) b(f )|2 df ,
and have good broad-band bias protection as well. Asymptotically, the
quadratic spectral estimate which minimizes the expected local square er14
ror is a multiple taper estimate using the minimal bias tapers.
The sinusoidal tapers have
a simple analytic form and approximate the
minimum bias tapers to O N1 . The kth sinusoidal taper has its spectral
k+1
k−1
≤ |f | ≤ 2(N
.
energy concentrated in the frequency bands 2(N
+1)
+1)
The minimum bias and sinusoidal tapers have no auxiliary bandwidth parameter, and the bandwidth of the spectral estimate is determined solely by
the number of used tapers. By adaptively adding and deleting tapers, a multitaper estimate with the optimal convergence properties of kernel smoothers
can be constructed. In contrast, the Slepian tapers need to be recomputed
with a different bandwidth. Thus the Slepian tapers are only practical for
fixed bandwidth estimation and this is inherently inefficient.
9
Appendix: Multitaper decomposition of kernel estimates
In Sec. 6, we showed that kernel smoother estimators (14) have an equivalent
multitaper representation (3) We now show that the equivalent multitapers
of some popular kernel smoother estimates of the spectrum strongly resemble
the MB/sinusoidal tapers. In one special case, this corresondence is exact;
i.e. the smoothed periodogram can be exactly decomposed into MB tapers.
Theorem 9.1 Let w = 21 and κ(f ) be the parabolic kernel, κ(f ) = 23 − 6f 2 .
The eigenvectors of the kernel smoothed periodogram are exactly the discrete
minimum bias tapers.
Proof: The Q̃ matrix in (15) can be calculated explicitly for this case.
We find Q̃ = [bnm ] = N1 ( 32 I − 6A ) where A is the matrix from Lemma 3.2.
Thus Q̃ and A have the same eigenvectors.
To illustrate that this result is typical even when we apply a taper and
smooth over a small band, we consider a smoothed tapered periodogram
with N = 200. We use Tukey’s split-cosine taper [16] and then smooth the
estimate with a square box kernel with a halfwidth of .01. We then evaluate
the corresponding Q̃ matrix and compute its eigenvectors. Figure
q 5 displays
the first 4 eigenvectors. They are very close to the sinusoids N2+1 sin Nπkn
.
+1
Table 4 shows that this spectral estimate is virtually a K = 4 multiple taper
15
spectral estimate. After k > 4, the eigenvalues decrease sharply, and these
higher eigenvectors contribute very little to the overall estimate.
References
[1] M. Amin, “Optimal estimation of evolutionary spectra,”
I.E.E.E. Trans. on Signal Processing vol. 42, 2?, 1994.
[2] T. Brockman, Th. Gasser and E. Hermann, “Locally adaptive
bandwidth choice for kernel regression estimators,” J. Amer. Stat.
Assoc. 88, 1302-1309 (1994).
[3] T. P. Bronez, Nonparametric Spectral Estimation of Irregularly
Sampled Multidimensional Random Processes, PhD Thesis, Arizona State University, 1985.
[4] U. Grenander and M. Rosenblatt, Statistical Analysis of Stationary Time Series, New York: Wiley, 1957.
[5] A. N. Kolmogorov and S. V. Fomin, Reele Funktionen und Funktionalanalysis, Section 7.3.2, Berlin: VEB Deutscher Verlag der
Wissenschaften, 1975.
[6] A. W. Marshall and I. Olkin, Inequalities: Theory of Majorization
and its Applications, p. 511, New York: Academic Press 1979.
[7] H.-G. Müller and U. Stadtmüller, “Variable bandwidth kernel
estimators of regression curves,” Annals of Statistics, vol. 15, pp.
182-201, 1987.
[8] C. T. Mullis and L. L. Scharf, in Advances in Spectrum Analysis,
S. Haykin, Ed., New York: Prentice-Hall, 1990, Chapter 1, pp.
1-57.
[9] A. Papoulis, “Minimum bias windows for high resolution spectral
estimates,” IEEE Trans. Information Theory, vol. 19, pp. 9-12,
1973.
[10] E. Parzens, “On asymptotically efficient consistent estimates
of the spectral density of a stationary time series,” J. Royal
Stat. Soc., vol. 19, pp. 303-322, 1958.
16
[11] J. Park, C. R. Lindberg, and F. L. Vernon, “Multitaper Spectral
analysis of high frequency seismograms,” J. Geophys. Res., vol.
92B, pp. 12765-12684, 1987.
[12] D. Percival and A. Walden, Spectral Analysis for Physical Applications: Multitaper and Conventional Univariate Techniques.
Cambridge: Cambridge University Press, 1993.
[13] K. S. Riedel, “Optimal kernel estimation of evolutionary spectra,”
I.E.E.E. Trans. on Signal Processing vol. 41, 2439-2447, 1993.
[14] K.S. Riedel and A. Sidorenko, “Function estimation using data
adaptive kernel smoothers- How much smoothing?” Computers
in Physics vol. 8, 402-409, 1994.
[15] K. S. Riedel and A. Sidorenko, “Smoothed log-multitaper spectral estimation and data adaptive implementation,” submitted
for publication.
[16] K. S. Riedel, A. Sidorenko, and D. J. Thomson, “Spectral density
estimation for plasma fluctuations I: Comparison of methods,”
Physics of Plasmas vol. 1, pp. 485-500.
[17] D. Slepian, “Prolate spheroidal wave functions, Fourier analysis,
and uncertainty - V: the discrete case,” Bell System Tech. J., vol.
5, pp. 1371-1429, 1978.
[18] D. J. Thomson, “Spectrum estimation and harmonic analysis,”
Proc. IEEE, vol. 70, pp. 1055-1096, 1982.
[19] D. J. Thomson, “Quadratic inverse spectrum estimates: applications to paleoclimatology,” Phil. Trans. R. Soc. Lond. A, vol.
332, p. 539-597, 1990.
17
Table Captions:
Table 1: Convergence of the sinusoidal tapers toR the minimum bias tapers.
P
1/2
2 PK
2
(k)
(f )|2 df ,
Table 2: Normalized bias term, 4(N + 1)2 K
k=1 −1/2 f
k=1 f |V
for N = 50.
R w PK
(k)
(f )|2 df , for N = 50.
Table 3: Spectral concentration, −w
k=1 |V
Table 4: Eigenvectors of the smooth tapered periodogram estimator.
Figure Captions:
Figure 1: Spectral energy of the minimum bias, sinusoidal and Slepian tapers,
N = 200, k = 1, w = .01.
Figure 2: Spectral energy of theRminimum bias and sinusoidal tapers, N = 200, k = 1.
P
1/2
Figure 3 Spectral energy, 3k=1 −1/2 f 2 |V (k) (f )|2 df , of the minimum bias and
Slepian tapers, N = 200, K = 3, w = .01.
Figure 4: Estimated spectral density of the plasma fluctuations. Dashed
line is sinusoidal multitaper estimate and solid line is estimate using Slepian
tapers with w = 60 kHz. Because the Slepian tapers have a fixed bandwidth,
the corresponding estimate spectral density at 1 MHz is artificially broadened
while being undersmoothed for f ≥ 1.5 MHz.
Figure 6: First eigenvectors of the smooth tapered periodogram estimator.
18
19
20
21
22
23
24
Table 1. Convergence of the sinusoidal tapers to the minimum bias tapers
N
maxk
n
20
50
200
800
N +2
kv(k)
k
− ν (k) kL2
o
1
2
3
4
5
6
7
8
9
10
maxk
0.24750
0.24844
0.24852
0.24844
Table 2. Normalized bias term, 4(N + 1)2
N = 50
K
(
Minimum
bias tapers
1.0095
5.0475
14.1328
30.2846
55.5217
91.8634
141.3284
205.9362
287.7056
388.6562
Sinusoidal
tapers
1.0116
5.0580
14.1622
30.3475
55.6366
92.0528
141.6185
206.3570
288.2899
389.4409
w=0.04
1.3439
5.7724
18.0651
58.9520
154.4818
305.4382
496.7959
721.1743
976.5088
1262.4251
25
R 1/2
k=1 −1/2
PK
N +2
k
(k)
− k νν(k) k
L∞
0.4602
0.4760
0.4829
0.4844
v(k)
kv(k) kL∞
f 2 |V (k) (f )|2 df , for
Slepian tapers
w=0.08
w=0.16
2.6316
5.1039
10.5484
20.4670
23.8086
46.1953
42.5181
82.4018
66.9996
129.2087
99.1800
186.7507
150.0103
255.1797
251.8833
334.6717
437.5993
425.4379
702.1523
527.7433
)
L∞
Table 3. Spectral concentration,
k Minimum bias tapers
1
.9997
2
.9988
3
.9972
4
.9940
5
.9888
6
.9760
7
.9381
8
.6084
9
.1688
10
.0637
Rw
−w
|V (f )|2 df , for N = 50, w = 0.08
Sinusoidal tapers
.9997
.9988
.9972
.9937
.9887
.9753
.9417
.6247
.1780
.0624
Slepian tapers
1.
.9999999
.9999989
.99997
.9995
.9928
.9380
.7002
.2981
.0628
Table 4. Eigenvectors of the smooth tapered periodogram estimator
k
1
2
3
4
5
6
7
Weight of the eigenvector
Normalized local bias
R 1/2
λk (B)/tr(B)
4(N + 1)2 −1/2 f 2 |V (f )|2 df
.2856
.2828
.2519
.1416
.0340
.0037
.0002
1.5138
4.7371
9.6254
19.2095
33.7118
51.3616
72.9747
26
Local bias in comparison with
the minimum bias taper (ratio)
1.509
1.181
1.067
1.198
1.345
1.423
1.486
| 10math.ST
|
Human Detection for Night Surveillance using
Adaptive Background Subtracted Image
Yash Khandhediya
Electronics and Communication Department, L. D.
College of Engineering
[email protected]
Karishma Sav
Electronics and Communication Department, L. D.
College of Engineering
[email protected]
Vandit Gajjar
Electronics and Communication Department, L. D.
College of Engineering
[email protected]
Abstract – Surveillance based on Computer Vision has
become a major necessity in current era. Most of the
surveillance systems operate on visible light imaging, but
performance based on visible light imaging is limited due
to some factors like variation in light intensity during the
daytime. The matter of concern lies in the need for
processing images in low light, such as in the need of
nighttime surveillance. In this paper, we have proposed a
novel approach for human detection using FLIR (Forward
Looking Infrared) camera. As the principle involves
sensing based on thermal radiation in the Near IR Region,
it is possible to detect Humans from an image captured
using a FLIR camera even in low light. The proposed
method for human detection involves processing of
Thermal images by using HOG (Histogram of Oriented
Gradients) feature extraction technique along with some
enhancements. The principle of the proposed technique
lies in an adaptive background subtraction algorithm,
which works in association with the HOG technique. By
means of this method, we are able to reduce execution
time, precision and some other parameters, which result in
improvement of overall accuracy of the human detection
system.
of research being pedestrian detection. According to a
research done in 2015, researchers proposed many
techniques; at times a fusion of two or more; for the
purpose of feature extraction from an image. This
research also showed that the lowest miss rate achieved at
that time was of 22.49 % [1]. However, there are a very
few techniques proposed for nighttime surveillance. In
most of the systems proposed for Nighttime surveillance
for human detection, Thermal Imaging is the most
frequently used method, where the features required for
Human detection are extracted from thermal images.
Thermal imaging uses NIR (Near Infrared) band of
Infrared light, which ranges from (0.75 to 1.4 μm). Some
of the existing methodologies make use of thermal
cameras along with visible light cameras [2], by the
means of image fusion but that requires twice the amount
of hardware. On the other hand, some directly perform
the feature extraction on the thermal images reducing the
hardware but at the same time compromising with
accuracy.
The aim of this paper is to propose a technique, which
provides better accuracy and precision, while minimizing
the hardware and execution time requirements.
Keywords—Adaptive Background Subtraction, Human
Detection, Thermal, FLIR, Nighttime Surveillance,
Dynamic Image Processing
I.
INTRODUCTION
The contemporary human detection systems are used in
applications like Autonomous Vehicles, Headcounters,
Search and Rescue Operations, etc. but these surveillance
systems limit itself in night surveillance due to the use of
RGB cameras. As we know that, there are myriads of
applications like border surveillance, security purposes,
monitoring systems, anomaly or intruder detection, which
most seek a system capable of night surveillance.
Many researchers today have proposed efficient methods
of detecting humans from images, the most common area
II.
NEED OF PROPOSED METHOD
The conventional methods of Human Detection use RGB
imaging as input. However, for the application of Human
Detection in Nighttime Surveillance, we cannot use RGB
imaging as the input to our system because at nighttime
the amount of light intensity available is inadequate to
produce clear images. As a solution to this problem, the
proposed method makes use of an FLIR camera and
processes on thermal images; which produces clear
images irrespective of the amount of light intensity.
The most widely used feature extraction technique is
HOG, Histogram of Oriented Gradients[3]. In spite of
good performance of HOG, there also exist some
limitations. Firstly, using HOG in real time application is
1
not feasible due to large execution time. Secondly, the
output of HOG has high recall but low precision of this
output is a matter of concern. So we always have to use
some extra filtering or feature extracting technique to
enhance the performance of the HOG.
The output of the HOG Feature Extraction block is then fed to
Linearly Trained SVM, which generates the prediction Matrix.
The prediction matrix when multiplied with the input image
provides us with the desired output. The results for this
algorithm are summarised in Fig. 2.
Another technique, which is used in for Human
Detection, is background subtraction from current image.
This technique can be used to enhance the performance of
HOG, as only background subtraction has the capability
to detect anomalies in a scene. But this technique is
currently only applicable to images from a static camera.
On the other hand, the proposed technique will be
applicable to images from dynamic cameras. As proposed
in our technique, a fused image is created using the
original image and the background subtracted image and
the performance of HOG on fused image improves
significantly as compared to when applied on the original
image.
Another improvement proposed by our technique is
adaptive background subtraction for dynamic camera. We
have used FLIR camera in capturing images. FLIR
cameras do not provide a wide-angle view. We can use a
number of cameras to solve this problem but it will not be
cost effective and the system would become more
complex as it would also need synchronization between
all the cameras. Therefore, to increase the span, the
camera is made to rotate to cover a broader area.
III.
Figure 2 Results for Static Camera
PROPOSED METHOD
The proposed method reduces the execution time taken
by HOG technique by means of Background Subtraction.
In the block diagram shown in Fig. 1, round blocks
represent our signals and rectangular blocks represent
functional blocks. Background Image, Input Image,
Background subtracted image are signals and Dynamic
background adaption block, Image Fusion Block, HOG
feature extraction block and linearly trained SVM block
are functional bocks. We will describe the algorithm
proposed by us for two different cases.
A. Static Camera: Constant Background Subtraction
For simplicity, first we have applied our algorithm on
thermal images taken from a static camera. The source
for these images is OTCBVS Benchmark Dataset
Collection [4]. In this case we have a single background
image which is subtracted from every image. After
background subtraction, the original image and the
background subtracted image are fused together; and then
HOG features are extracted from this fused image.
B. Dynamic
Camera:
Subtraction
Adaptive
Background
As we are using FLIR Cameras, we need the camera to be
moving to cover a wider angle. This requires an
altogether modified technique. Therefore, we need to
modify the algorithm, to work with dynamic cameras
with. The part of the block diagram in Fig. 1 shown in
Dotted lines is included to accommodate the input from a
dynamic camera. The purpose of this paper lies in the
design of the algorithm which accommodates a
dynamically changing background. This algorithm
corresponds to the proposed feature extraction technique
which will work effectively with dynamic camera
images. The working of this algorithm is summarised in
the Block Diagram in Fig. 3. However, the detailed
explanation and working of the algorithm is presented in
the Next Section. The images for experimentation are
taken from ETHZ Thermal Infrared Dataset [5].
Figure 3 Dynamic Background Adaptation Block
Figure 1 : Block Diagram for Proposed Algorithm
2
IV.
THE PROPOSED FEATURE : ADAPTIVE
BACKGROUND SUBTRACTION
The working principle of this feature is very simple yet
effective. As the problem arises due to a dynamic camera
whose background changes, the algorithm is designed to
shift a constant background by an amount, which is in
proportion with the rotation of the camera. The precise
working of this algorithm can be subdivided into two
parts for detailed analysis.
A.
Vectored Change Detection
In this part of the algorithm, we are firstly selecting a
particular region of the image, which has a specific object
very distinct in properties from its neighboring pixels;
i.e.; an object which does not blend into the background.
Also the object under consideration should be static with
respect to the background. To demonstrate this logic, let
us consider the image shown in Fig. 4.
Figure 5 Object under Consideration for Shifted Camera Image
|∆|=√∆x 2 + ∆y 2
is the magnitude of the difference
vector;
Arg(∆) = tan-1(∆y/∆x)
is the angle of the difference
vector.
Therefore, we now have a difference vector in terms of
magnitude and angle. The magnitude of the vector
represents the amount of shift while the angle represents
the direction of shift.
B.
Figure 4 Region and Object under Consideration for Initial Background
The outermost rectangle represents our image. The
innermost rectangle shaded with lines represents the
object under consideration, while the rectangle exterior to
the object under consideration represents our selected
region. The region is so selected that initially the object
under consideration is exactly at its center. Let the
coordinates of the object under consideration with respect
to the selected region be (x, y).
Let us now consider a second image, which can be an
input image for any arbitrary angle of rotation. The
Selected Region will remain same. Now the selected
region in this second input image will be scanned for the
Object under consideration. This will be done by taking a
window frame of the size of the object and moving this
frame in the entire Selected region. The error is
calculated for each frame and the window frame which
has the minimum error now becomes the new location of
our Object. To visualize the shift in the new image,
consider Fig. 5. Let the location of the object under
consideration in the input image be (x’, y’). If we
compare the two location coordinates of both the images,
we can get a difference vector which can be utilized to
determine the shift required in the background.
∆x = x’ – x
is the change in x coordinate;
∆y = y’ – y
is the change in y coordinate;
(∆x, ∆y) represent the difference vector coordinates,
where the magnitude and angle of the vector are as
described in the equations as follows:
Shifting Background Image
Now we know the magnitude and direction of shift of a
static object under consideration. Because it is a static
object, i.e. it does not move with respect to the camera,
the difference vector actually indicates a shift or
movement of the camera. As the object is static with
respect to the background, the shift of the background has
to be same as the shift in the object. Therefore, if we shift
the original background by this difference vector we will
have a shifted background. This phenomenon is clearly
depicted in Fig. 6.
Figure 6 Adapting Background to Dynamic Camera
Using the procedure explained above an algorithm is
created which can create Adaptive Backgrounds for
Subtraction from images taken from a Dynamic Camera.
An image taken from a dynamic camera is processed
using both constant backround subtraction and Adaptive
background subtraction and the output and original image
are as shown in Fig. 7.
3
Subtraction, for feature extraction from images taken
from a dynamic camera. The proposed feature has been
true to its necessity as it catered to the system by
increasing precision, allowing the camera some freedom
of movement and reducing the execution time of the
system.
VII.
ACKNOWLEDGEMENT
We would like to express our gratitude to Prof. Usha
Neelkantan, HOD of Electronics and Communication
Department. L. D. College of Engineering for her
constant support and motivation. We would also like to
thank Prof. Kinnar Vaghela for without his guidance we
might not have been able to complete our research
successfully.
Figure 7 Output Comparison between images processed using constant
background subtraction and Adaptive Background Subtraction
It is clearly visible that the proposed technique is more
efficient in extracting features as compared to the
available technique.
Table 1 Comparison of HOG, HOG with Background Subtraction and HOG
with Adaptive Background Subtraction for Static and Dyamic Cameras
Parameter
HOG
HOG+BS for
Static Camera
HOG + BS for
Dynamic Camera
HOG+ABS for
Dynamic Camera
Execution Time ( in
Seconds)
1768.29
23.0047
631.0068
478.0096
Precision
12.36%
83.11%
58.36%
76.09%
100%
100%
100%
100%
Recall
V.
RESULTS & COMPARISON
VIII.
For Comparison purposes, we have first detected humans
using only HOG feature Extraction and the results are
compared when HOG is used in association with
Background Subtraction and also with the results
obtained when HOG is used along with our proposed
technique.
[1] Benenson R., Omran M., Hosang J., Schiele B.
(2015) Ten Years of Pedestrian Detection, What
Have We Learned?. In: Agapito L., Bronstein
M., Rother C. (eds) Computer Vision - ECCV
2014 Workshops. ECCV 2014.
Various significant parameters are compared in Table 1.
The Recall is 100% in all the techniques, but we can see
that there is a drastic improvement in Precision. Also the
execution time is significantly reduced. The execution
time increases if we use a dynamic camera and the
precision also reduces but given the flexibility for
movement and coverage of greater angle of the scene is a
feasible trade-off.
VI.
REFERENCES
[2] Alex Leykin, Yang ran and Riad Hammoud,
“Thermal-Visible Video Fusion for Moving
Target Tracking and Pedestrian Classification”
In: Computer Vision and Pattern Recognition,
2007. CVPR '07.
[3] Shyang-Lih Chang, “Nighttime pedestrian detection
using thermal imaging based on HOG feature” In:
2011 International Conference on System Science
and Engineering.
[4] IEEE OTCBVS WS Series Bench Zhang, M.M,
Choi, J., Daniilidis, K., Wolf, M.T. & Kanan, C.
(2015) VAIS: A Dataset for Recognizing
Maritime Imagery in the Visible and Infrared
Spectrums. In: Proc of the 11th IEEE Workshop
on Perception Beyond the Visible Spectrum
CONCLUSION
In this paper, we have proposed a nighttime surveillance
system based on processing of thermal images taken from
a FLIR camera. In addition to this, we have proposed a
novel technique, named Adaptive Background
4
(PBVS-2015).
[5] ETHZ Thermal Infrared Dataset, ASL Datasets
5
| 1cs.CV
|
Semi-Automated Nasal PAP Mask Sizing using Facial Photographs
arXiv:1709.07166v1 [cs.CV] 21 Sep 2017
Benjamin Johnston, Alistair McEwan and Philip de Chazal
Abstract— We present a semi-automated system for sizing
nasal Positive Airway Pressure (PAP) masks based upon a
neural network model that was trained with facial photographs
of both PAP mask users and non-users. It demonstrated an
accuracy of 72% in correctly sizing a mask and 96% accuracy
sizing to within 1 mask size group. The semi-automated system
performed comparably to sizing from manual measurements
taken from the same images which produced 89% and 100%
accuracy respectively.
Index Terms— OSA, PAP, CPAP, neural networks, machine
learning, facial landmarking, telemedicine, sizing
I. I NTRODUCTION
Positive airway pressure or continuous positive airway
pressure (CPAP) therapy, seminally described by Sullivan et
al. in 1981 [1] is generally accepted as the gold standard
treatment for obstructive sleep apnoea (OSA) [2]. Positive airway pressure effectively treats OSA by providing a
pneumatic splint at the location of soft palate, preventing
collapse onto the pharynx and thus allowing the passage of
air to the lungs. PAP therapy is applied using a specially
manufactured device that consists 3 of main components:
a pressure generator, which is a highly sophisticated air
pump used to compress atmospheric air for delivery into the
pharynx; a mask which the patient wears over the nose and
/ or mouth to ensure the pressurised air is delivered to the
correct location and finally a humidifier unit.
The mask is a critical component of the PAP system as it is
the device that is in intimate contact with the patient and must
be capable of maintaining treatment pressure for the duration
of a night’s sleep. Other than potentially reducing treatment
efficacy, issues with mask performance are often immediately
noticeable for the patient such as uncomfortable air leaks
into the eye or visible red marks on the face caused by mask
rubbing during use. While incidents of side effects due to
mask usage vary, a number of studies have observed that
many patients experience difficulties using their mask [3].
Amfilochiou et al. reported incidences of mask leak and red
marks in 48% and 40% of participants [4]; Gay et al. noted
that mask interface issues are the most commonly reported
PAP related side effects [5].
Research supported by ARC grant FT110101098
Ben Johnston and Philip de Chazal are with the Sleep Research Group, Charles Perkins Centre, School of Electrical and Information Engineering, University of Sydney, Sydney, NSW, 2006,
Australia, (phone: +612 911 41528; e-mails: {ben.johnston,
philip.dechazal}@sydney.edu.au)
Alistair McEwan is with the School
Information
Engineering,
University
of
[email protected])
of Electrical and
Sydney,
(e-mail:
In the context of PAP therapy compliance, incidents of
device related side-effects such as mask leakage are not
inconsequential. Engleman in 1994 observed that patients
who experienced side effects used their PAP devices less
that those who did not [6]. More recently in 2003 Massie
and Hart, while studying differences between nasal and nasal
pillows mask types noted a negative correlation with the
occurrence of air leaks and sore eyes with CPAP usage [7].
These side effects often form a barrier to effective therapy
usage, particularly with new patients [8] being introduced to
PAP for the first time. It has been found that the first week
of therapy is critical in determining long term therapy use
patterns, as it is during this time a patient decides if they
are going to continue or abandon treatment [9], [10]. Given
these characteristics and that patients who swap their mask
due to poor fit or comfort problems are 7 times more likely to
discontinue therapy [11] it is critical that patients are issued
with the correct mask size and type on the first attempt.
A. Current Mask Sizing
While the exact details of sizing PAP masks varies between each of the manufacturers, there are two common
methods which are generally used amongst clinicians. One
process uses a fitting template provided by the manufacturer
such as that illustrated in Figure 1a. When using these
templates the patient places their nose within specific features
of the template and the clinician matches the appropriate size
with indications on the guide. The commonly used alternative
sizing methodology uses the experience and expertise of the
clinician to make the mask choice without the use of a guide.
This paper describes the potential use of a third, automated
method of PAP mask sizing with the aim of improving sizing
accuracy and reducing mask related side effects.
B. Potential Applications of Automated Sizing
Automated PAP mask sizing could provide significant benefit to new patients requiring PAP therapy who live in remote
or rural communities and have limited access to medical
services. Through the use of Home Sleep Tests (HST) an
OSA diagnosis can be issued without the patient leaving
the home. However in order to issue a PAP system with
appropriate mask an in-person consult (possibly requiring
hours of travel) is currently required. An automated sizing
system could enable the use of a telemedicine approach to
PAP consults. A patient would send a facial photograph to a
clinician who would use software to analyse the photograph
and automatically determine the correct mask size. The mask
would then be sent to the patient, thus avoiding an inperson visit. Such a system could also provide additional
benefit by reducing consult time and increasing the capacity
of a clinician to see more patients. In this study we have
developed a semi-automated PAP mask sizing system by
processing facial images.
II. I MAGE DATASET
The dataset used throughout this study comprised 251
facial images of male and female, PAP (49 samples) and nonPAP users that were collected in a variety of locations using
smartphone cameras. Care was taken during photography to
ensure the face was not obscured by glasses, hair or hats and
that the pose of the face was approximately parallel with
the smartphone. Participants had their the ALAR or nose
width (see Figure 1c) measured at the time of image capture
using vernier callipers. These nose width measurements were
applied against a sizing template for an Eson nasal PAP mask
manufactured by Fisher & Paykel (F&P) Healthcare (Figure
1a & Table I) to determine the ‘ground truth’ mask size for
each of the selected participants.
Each image in the set (e.g. Figure 1b) is a frontal photographic shot of the participant with an Australian 20 cent
piece placed on the forehead to provide an indication of
scale.
A. Size Overlap
The PAP mask industry provides a margin of overlap for
mask sizes as the means of determining the correct mask
size is not an exact science (see Section I-A). Those mask
users who fall within the boundary between two sizes will
achieve acceptable mask performance with either of the
adjacent mask sizes. This margin of acceptable error provides
some tolerance to mis-sizing. To reflect this situation in the
field, we applied a 2% tolerance to sizing decisions within
this study. For individuals whose ‘ground truth’ nose width
fell within 2% of a size boundary e.g 36.26 - 37.74mm
for a small / medium mask, the predicted mask size was
considered correct if either a small or medium mask is
predicted.
B. Data Pre-Processing
As nasal mask sizes are solely determined by the width of
a user’s nose, the nasal region of each image was extracted
in grayscale using the opencv implementation of the ViolaJones algorithm[12] and the pre-trained nose haarcascade
classifier provided with opencv (Stages 1 & 2 of Figure 2).
Each 200 × 150 image was unwrapped to form a data vector
(x ∈ R251×30000 ) and the two nasal landmark coordinates
were unwrapped to give a 4 element target vector (x ∈
R251×4 ). The x and y datasets were scaled and centered
between -1 and 1 by dividing by their respective maxima
and subtracting the resulting mean (Stage 3 Figure 2).
C. Manual Mask Sizing from the Images
(a) Eson sizing template
(b) Labelled image
In an attempt to determine the optimum performance that
could be achieved by the system, mask sizes were also
calculated manually using the labelled landmarks of the
test set facial photographs. The physical nose width was
calculated by dividing by the distance between the two
labelled landmarks (pixels) and the scale (pixels / mm) as
provided by the 20 cent piece. This physical measurement
was compared with Table I to determine the sizes referred
to as Manually Measured throughout this study.
D. Neural Network Architecture
(c) Illustration indicating ALAR Width
Fig. 1. Dataset (sample image used with permission from Priyanshu Gupta)
III. M ETHODS
Following data collection, the face from each image was
extracted using the opencv implementation of the ViolaJones algorithm[12] into a sub-image and resized to 512 x
512 pixels. The pixel locations of the left and right lateral
nasal walls were then manually identified and recorded for
each image. The scale of each image was recorded in pixels
/ mm using ImageJ (produced by the US National Institute
of Health) by measuring the number pixels that span the
diameter and dividing by the known diameter of the coin
(28.65mm).
As illustrated in Figure 2 the neural network stage used
during this study was a 3 layer architecture, consisting of an
input stage with 30000 inputs, a hidden layer with 40 hidden
units and an output layer with 4 units. The input and hidden
layers also possessed a bias unit. The hidden layer used
a tanh non-linearity activation function, while the output
Size
Small
Medium
Large
Too Large*
Lower Limit (mm)
0
37
41
45
Upper Limit (mm)
37
41
45
∞
TABLE I
F ISHER & PAYKEL E SON MASK SIZES . *N OT ACTUALLY A SIZE BUT IS
USED TO REPRESENT INDIVIDUALS WHO DO NOT FIT AN E SON MASK
Processing stages of the model
layer possessed a linear activation function. The weights
of both the hidden and output layers were initialised using
random values between -0.05 and 0.05. This architecture was
selected following a hyperparameter sweep that produced
optimal validation set performance using 40 training and 13
validation samples.
Once the network architecture was selected, leave-one-out
cross validation was used over the remaining 198 sample test
data set to evaluate the overall performance of the model.
Predicted Mask Sizes
Actual Mask Size
Fig. 2.
Small
Medium
Large
Too Large
Small
61
26
5
1
Medium
8
55
9
0
Large
1
3
21
1
Too Large
0
1
0
6
TABLE II
C ONFUSION MATRIX OF PREDICTED MASK SIZES ( USING THE NOSE
WIDTH FROM THE MEAN TEST SET LANDMARKS )
E. Forward & Back Propagation
F. Weight Updates
Following backpropagation, the weights at each of the
layers of the neural network were updated to reflect their
respective contributions to the error. For the output layer
using a linear activation function, the optimal weights were
determined using the Moore-Penrose pseudo-inverse [15].
The updates to the hidden layer weights were calculated
using the standard update method with momentum. The
learning rate used to update the weights; α started at 0.002
and was reduced by 5% for each improvement in network
performance. The momentum µ was fixed at 0.95 throughout
training.
G. Determining Predicted Mask Size
The leave-one-out training & test regime was repeated 4
times with network weights re-initialised with random values
prior to each repetition. Thus 4 nasal landmark predictions
were produced for each test image. The predicted landmarks
were then transferred back into image space by re-applying
the mean and maximum values calculated in section III-B.
The mean position of each landmark was then computed for
each of the test set images. The nose width was determined
using the length of the line between the two mean predicted
landmarks and the scale of the image as measured in section
Manually Measured Mask Sizes
Actual Mask Size
During training, the forward propagation process differed
from the standard method [13] used for a neural network
slightly in that the hidden unit layer also incorporated
the use of a 70% dropout stage to improve generalisation
performance [14]. The contribution of each weight within
the network to the overall error was computed by backpropagating the sum of squares error function result from the
output layer through the network [13].
Small
Medium
Large
Too Large
Small
81
12
0
0
Medium
2
66
4
0
Large
0
1
23
2
Too Large
0
0
0
7
TABLE III
C ONFUSION MATRIX OF MASK SIZES MANUALLY MEASURED FROM
IMAGES
II. This nose width estimate was then used with the known
mask sizes in Table I to determine the Predicted mask size.
IV. R ESULTS
Tables II to IV summarise the performance of the neural
network sizing model and the manual measurements taken
from the test images; while Figures 3a & b provide examples
of accurate and inaccurate landmark predictions. Using the
confusion matrix in Table II, the semi-automatic sizing
system achieved an overall accuracy of 72% in correctly
estimating the mask size and achieved a within one size
accuracy of 96%. The within one size accuracy figure states
that the system predicted the correct size or an adjacent
size for 96% of samples. The manually measured accuracy
metrics of 89% and 100% respectively are shown in Table
III.
V. D ISCUSSION
The semi-automated mask sizing system produced an
accuracy of 72%, while the manual measurements taken
from the photograph achieved 89%. It was expected that the
manual annotator would achieve 100% accuracy. A review
of the images revealed that parallax error contributed to the
error of the system. The nasal region in some images was
also obscured thus making landmarking more difficult. Such
an issue could contribute significantly to the measurement /
prediction error through an incorrect identification of nasal
width, particularly affecting those measurements which are
very close to the 2% size tolerance.
Another potential source of error could result from the
location of the 20 cent piece on the forehead. It is possible
that in some images the reference coin is some distance
away from the widest part of the nose in terms of depth.
This difference in depth could introduce error as the physical
distance represented by a single pixel at nose and coin would
differ; thus producing errors in final sizing. The model could
benefit from additional work investigating more reliable
means of defining scale within the images.
Given the modest sample size of the datasets used in this
study the system would almost certainly benefit from additional data. This is particularly evident examining the images
where the system performed least effectively (e.g. Figure 3b);
by increasing the number and variety of nose types within
the data we anticipate that the overall performance of the
system would improve.
(a) Correct prediction
(b) Inaccurate prediction
Fig. 3. Nasal landmark prediction examples (blue dots: predicted landmarks, red dots: labelled landmarks)
VI. C ONCLUSION
We intend to continue to improve the accuracy of the
proposed system: collecting a larger data set, investigating
the use of more advanced techniques such as convolutional
neural networks to improve facial landmarking accuracy as
well as methods to offset errors within the system. In order
Man Pred
Sensitivity (%)
Pos. Predict. (%)
S
M
L
TL
S
M
L
TL
66
76
81
86
87
65
60
75
87
92
88
100
98
84
85
78
TABLE IV
S ENSITIVITY & POSITIVE PREDICTIVITY PERFORMANCE OF PREDICTED
(P RED ) MASK SIZES AND THOSE DETERMINED FROM MANUAL
(M AN ) MEASUREMENTS (S: S MALL , M: M EDIUM , L:
L ARGE , TL: T OO L ARGE )
PHOTOGRAPH
for a future version of the system to be effectively used in
the field, a means of automatically determining scale within
the image must also be developed.
This study presented a semi-automated nasal PAP mask
sizing system based on neural networks. The system achieved
72% accuracy in correctly estimating the Fisher & Paykel
Eson mask size and a 96% accuracy in sizing within one size.
This level of performance was comparable to that achieved
using manual measurements on the same data set. Further
improvements could be made increasing and refining the
data sets used to train and evaluate the model as well as
investigating additional methods of determining scale.
R EFERENCES
[1] C E Sullivan, F G Issa, M Berthon-Jones, and L Eves, “Reversal
of obstructive sleep apnoea by continuous positive airway pressure
applied through the nares,” Lancet, vol. 1, no. 8225, pp. 862–865,
1981.
[2] Anonymous, “Indications and standards for use of nasal continuous
positive airway pressure (CPAP) in sleep apnea syndromes. American
Thoracic Society. Official statement adopted March 1944.,” American
Journal of Respiratory and Critical Care Medicine, vol. 150, no. 6,
pp. 1738–1745, dec 1994.
[3] Richard B Berry, “Improving CPAP compliance man more than
machine,” Sleep Medicine, vol. 1, no. 3, pp. 175–178, jul 2000.
[4] Anastasia Amfilochiou, Venetia Tsara, Likurgos Kolilekas, Evangelia
Gizopoulou, Chrysoula Maniou, Demosthenes Bouros, and Vlasios
Polychronopoulos, “Determinants of continuous positive airway pressure compliance in a group of Greek patients with obstructive sleep
apnea,” European Journal of Internal Medicine, vol. 20, no. 6, pp.
645–650, 2009.
[5] Peter Gay, Terri Weaver, Daniel Loube, and Conrad Iber, “Evaluation
of positive airway pressure treatment for sleep related breathing
disorders in adults.,” Sleep, vol. 29, no. 3, pp. 381–401, 2006.
[6] H M Engleman, S E Martin, and N J Douglas, “Compliance
with CPAP therapy in patients with the sleep apnoea/hypopnoea
syndrome.,” Thorax, vol. 49, no. 3, pp. 263–6, 1994.
[7] Clifford A. Massie and Robert W. Hart, “Clinical outcomes related
to interface type in patients with obstructive sleep apnea/hypopnea
syndrome who are using continuous positive airway pressure,” Chest,
vol. 123, no. 4, pp. 1112–1118, 2003.
[8] S S Dickerson and M C Kennedy, “CPAP devices: encouraging
patients with sleep apnea,” Rehabil Nurs, vol. 31, no. 3, pp. 114–
122, 2006.
[9] T E Weaver, N B Kribbs, A I Pack, L R Kline, D K Chugh, G Maislin,
P L Smith, A R Schwartz, N M Schubert, K A Gillen, and D F Dinges,
“Night-to-night variability in CPAP use over the first three months of
treatment.,” Sleep, vol. 20, no. 4, pp. 278–83, 1997.
[10] Carl J. Stepnowsky, Matthew R. Marler, and Sonia Ancoli-Israel,
“Determinants of nasal CPAP compliance,” Sleep Medicine, vol. 3,
no. 3, pp. 239–247, 2002.
[11] Adel Bachour, Pirjo Vitikainen, and Paula Maasilta, “Rates of initial
acceptance of PAP masks and outcomes of mask switching,” Sleep
and Breathing, vol. 20, no. 2, pp. 733–738, 2016.
[12] Paul Viola and Michael J Jones, “Robust Real-time Object Detection,”
International Journal of Computer Vision, , no. February, pp. 1–30,
2001.
[13] Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner,
“Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2323, 1998.
[14] Pierre Baldi and Peter Sadowski, “The dropout learning algorithm,”
Artif. Intell., vol. 210, no. 1, pp. 78–122, 2014.
[15] Philip De Chazal, Jonathan Tapson, and André Van Schaik, “A comparison of extreme learning machines and back-propagation trained
feed-forward networks processing the mnist database,” in ICASSP,
IEEE International Conference on Acoustics, Speech and Signal
Processing - Proceedings, 2015, vol. 2015-Augus, pp. 2165–2168.
| 1cs.CV
|
arXiv:1608.02801v4 [math.ST] 20 Dec 2017
ON THE ASYMPTOTIC NORMALITY AND THE
CONSTRUCTION OF CONFIDENCE INTERVALS FOR
ESTIMATORS AFTER SAMPLING WITH PROBABILISTIC
AND DETERMINISTIC STOPPING RULES
BEN BERCKMOES AND GEERT MOLENBERGHS
Abstract. A key feature of a sequential study is that the actual sample size is a random variable that typically depends on the outcomes
collected. While hypothesis testing theory for sequential designs is well
established, parameter and precision estimation is less well understood.
Even though earlier work has established a number of ad hoc estimators
to overcome alleged bias in the ordinary sample average, recent work
has shown the sample average to be consistent. Building upon these results, by providing a rate of convergence for the total variation distance,
it is established that the asympotic distribution of the sample average
is normal, in almost all cases, except in a very specific one where the
stopping rule is deterministic and the true population mean coincides
with the cut-off between stopping and continuing. For this pathological
case, the Kolmogorov distance with the normal is found to equal 0.125.
While noticeable in the asymptotic distribution, simulations show that
there fortunately are no consequences for the coverage of normally-based
confidence intervals.
1. Introduction
In surprisingly many settings, sample sizes are random. These include
sequential trials, clusters of random size, incomplete data, etc. [MKA14] and
[MMA16] studied implications of this on estimators in a unified framework;
[MMA15] focused on the specific but important case of a sequential trial,
which is also the setting of interest in this paper.
While formal sequential methodology dates back to World War II ([W45]),
most emphasis has been placed on design and hypothesis testing. Regarding
parameter estimation after sequential trials, it has been reported that commonly used estimators, such as the sample average, exhibit bias at least in
some settings. In response, a variety of alternative estimators have been proposed ([S78, HP88, EF90]). Building upon [LH99] and [LHW06], [MKA14],
[MMA16], and [MMA15] reviewed several of these and actually showed that
the sample average is a consistent estimator in spite of earlier concern, even
though there is a small amount of finite-sample bias. Their approach is
based on considering a class of stochastic stopping rules that lead to the
Key words and phrases. asymptotic normality, confidence interval, Kolmogorov distance, random sample size, rate of convergence, sequential clinical trial, stopping rule,
total variation distance.
Ben Berckmoes is post doctoral fellow at the Fund for Scientific Research of Flanders (FWO); financial support from the IAP research network #P7/06 of the Belgian
Government (Belgian Science Policy) is gratefully acknowledged.
1
2
BEN BERCKMOES AND GEERT MOLENBERGHS
more commonly encountered deterministic stopping rules as limiting cases.
They used incomplete-data ignorable likelihood theory to this end. In addition, they showed that there exists an alternative, conditional likelihood
estimator that conditions on the sample size realized; this one is unbiased
also in small samples but is slightly less efficient than the sample average,
and is implicitly defined through an estimating equation.
While these earlier results are important, the authors did not address the
limiting distribution of the mean estimated from a sequential trial and its
implications for confidence interval estimation. This is the focus of the current paper. To this end, we consider the manageable but generic case where
in a first step n i.i.d. normally distributed N (µ, 1) observations are collected,
after which a stopping rule is applied and, depending on the outcome, a second i.i.d. set of n observations is or is not collected. The probability
of
β
stopping after the first round is assumed to be of the form Φ α + n Kn ,
with Φ(·) the probit function, Kn the sample sum of the first n observations, and α and β a priori fixed parameters. The setting is formalized in
the next section. While there are many cases where other than normal data
are collected, it is a sufficiently common and at the same time tractable case;
extension to the exponential family is of interest but outside of the scope of
this paper. Also for ease of exposition, we consider a study with two possible
sample sizes, n and 2n. Also this can be generalized in a straightforward
fashion. Finally, depending on the situation, Kn /n may or may not be the
core of the test statistic considered, even though the ratio of the sample sum
over a measure of information is very commonly encountered. Calculations
for alternative functions of Kn will follow logic similar to the one displayed
here.
Employing the total variation distance, we establish that for stochastic
stopping rules asymptotic normality applies. Likewise, we show that this is
true too for deterministic stopping rules, provided that µ 6= 0. For these
cases rates of convergence are established. When µ = 0 there is no weak
convergence; we establish the Kolmogorov distance between the true distribution and the normal.
In Section 2, the formal framework is introduced. In Section 3, the main
result is formulated. The behavior in practice is gauged by way of a simulation study, described in Section 4, with some details relegated to the
Appendix. Implications and ramifications are discussed in Section 5.
2. Formal framework
Let X1 , X2 , . . . , Xn , . . . be independent and identically distributed random variables with law N (µ, 1). Also, let N1 , N2 , . . . , Nn , . . . be random
sample sizes such that each Nn takes the values n or 2n, is independent of
Xn+1 , Xn+2 , . . ., and satisfies the conditional law
β
P [Nn = n | X1 , . . . , Xn ] = Φ α + Kn ,
n
(1)
3
where Φ is the standard normal cumulative distribution function,
n
X
Xi ,
Kn =
i=1
R+ .
α ∈ R, and β ∈
Notice that the restriction that β be positive is merely
for convenience, and that the results presented in this paper can be easily
extended for negative β. We also consider the limiting case of (1) where
β → ∞, which corresponds to
P [Nn = n | X1 , . . . , Xn ] = 1{Kn >0} ,
(2)
where 1{Kn >0} stands for the characteristic function of the set {Kn > 0}.
Finally, we define the estimator
1
µ
bNn =
KNn ,
(3)
Nn
which is the classical average of a sample with random size Nn .
In [MKA14], it is shown that µ
bNn , defined by (3), is, for both the stopping rules (1) and (2), a legitimate estimator for µ in the sense that it is
asymptotically unbiased. More precisely, it is established there that, for the
probabilistic stopping rule (1),
!
α + βµ
β
1
p
φ p
,
(4)
E[b
µ Nn ] = µ +
2n 1 + β 2 /n
1 + β 2 /n
and, for the deterministic stopping rule (2),
√
1
E[b
µNn ] = µ + √ φ( nµ),
2 n
(5)
where φ is the standard normal density. Clearly, (4) and (5) both converge to
µ as n tends to ∞. These authors also consider small sample bias corrected
estimators, but this is outside of the scope of this paper.
In this note, we consider a different aspect of the legitimacy of the estimator µ
bNn . More precisely, we examine the asymptotic normality of the
sequence
p
Nn (b
µNn − µ) .
(6)
n
3. Statement of the main result
Recall that the Kolmogorov distance between random variables ξ and η
is given by
K(ξ, η) = sup |P[ξ ≤ x] − P[η ≤ x]| ,
x∈R
and the total variation distance by
dT V (ξ, η) = sup |P[ξ ∈ A] − P[η ∈ A]| ,
A
the supremum running over all Borel sets A ⊂ R. Clearly, the inequality
K ≤ dT V
holds, and it is known to be strict in general. Also, it is well known that
a sequence of random variables (ξn )n converges weakly to a continuously
distributed random variable ξ if and only if K(ξ, ξn ) → 0. Finally, dT V
4
BEN BERCKMOES AND GEERT MOLENBERGHS
metrizes a type of convergence which is in general strictly stronger than
weak convergence. For more information on these distances, and on the
theory of probability distances in general, we refer the reader to [R91] and
[Z83].
In the following theorem, our main result, we show
√ that if the probabilistic
Nn (b
µNn − µ) n converges
stopping rule (1) is followed, then the sequence
in total variation distance to Φ, and we establish a rate of convergence in
this case. Furthermore, we prove that if the√deterministic stopping rule (2)
is followed and µ 6= 0, then the sequence
Nn (b
µNn − µ) n also converges
in total variation distance to Φ, and we again provide a rate of convergence
in this case. Finally, we establish that if the deterministic
stopping rule (2)
√
N
(b
µ
−
µ)) = 1/8. In
is followed and
µ
=
0,
then,
for
each
n,
K(Φ,
n
N
n
√
Nn (b
µNn − µ) n fails to converge weakly to Φ in this case. We
particular,
nevertheless show that in all cases it is plausible to use estimation (3) for
the construction of reliable confidence intervals for µ.
A proof is given in Appendix A.
Theorem 1. Suppose that the probabilistic stopping rule (1) is followed.
Then, for each n,
p
µNn − µ)) ≤ C(α, β, µ, n),
(7)
dT V (Φ, Nn (b
where
C(α, β, µ, n) =
Φ
r
Z
∞
φ(u)
−∞
β
2n
(α + βµ) + p
u
2n + β 2
2n + β 2
!
β
− Φ α + βµ + √ u
n
du,
which, by√the Dominated Convergence Theorem, converges to 0 as n → ∞,
µNn − µ) n converges in total variation distance to Φ. In
Nn (b
whence
particular, considering the Borel set Ax = [−x, x] for x ≥ 0, (7) gives
1
1
2Φ(x) − 1 − P µ
b Nn − √ x ≤ µ ≤ µ
bNn + √ x ≤ C(α, β, µ, n),
Nn
Nn
which makes it plausible to use µ
bNn for the construction of reliable confidence
intervals for µ.
Now suppose that the deterministic stopping rule (2) is followed. Then,
for each n,
p
µNn − µ)) ≤ C(µ, n),
(8)
dT V (Φ, Nn (b
where
Z ∞
√
φ(u) 1{u>−√nµ} − Φ u + 2nµ du,
C(µ, n) =
−∞
which, if µ 6= 0, √
by the Dominated
Convergence Theorem, tends to 0 as
Nn (b
µNn − µ) n converges in total variation distance to
n → ∞, whence
Φ. In particular, considering the Borel set Ax = [−x, x] for x ≥ 0, (8) gives
1
1
2Φ(x) − 1 − P µ
bNn − √ x ≤ µ ≤ µ
bNn + √ x ≤ C(µ, n),
Nn
Nn
which, if µ 6= 0, makes it plausible to use µ
bNn for the construction of reliable
confidence intervals for µ.
5
If µ = 0, then, for each n,
p
µNn − µ)) = 1/8,
(9)
K(Φ, Nn (b
√
and
Nn (b
µNn − µ) n fails to converge weakly to Φ. Nevertheless, for each
x ∈ R+
,
0
1
1
P µ
bNn − √ x ≤ µ ≤ µ
bNn + √ x = 2Φ(x) − 1.
(10)
Nn
Nn
Thus, also in the case where µ = 0, it is plausible to use µ
bNn for the construction of reliable confidence intervals for µ.
4. Simulations
We have conducted a brief simulation study to illustrate Theorem 1, the
tables of which are√ given in Appendix B. We have studied the empirical
distribution En of Nn (b
µNn − µ), based on 1000 simulations, for both the
probabilistic stopping rule (1) (Tables 1 and 2) and the deterministic stopping rule (2) (Table 3), and different values for β, the true parameter µ,
and the number of observations n. In each case, we have compared the theoretical upper bound for the total variation distance between
the standard
√
µNn − µ), as
normal distribution and the theoretical distribution of Nn (b
given in Theorem 1, with the Kolmogorov distance
between the standard
√
µNn − µ). We have also
normal cdf and the empirical distribution of Nn (b
counted the number of times
out of √
1000 where the true
µ is
√ parameter
bNn + 1.96/ Nn , which would
contained in the interval µ
bNn − 1.96/ Nn , µ
√
be a 95%-confidence interval for µ if Nn (b
µNn − µ) were standard normally
distributed.
The predictions by Theorem 1 are confirmed by the simulation study.
More precisely, in the cases where the stopping rule is close to being deterministic andõ = 0, the simulation study indeed points out that the
µNn − µ) deviates from a standard normal distribution
distribution of Nn (b
(red values in the tables). However, it is also confirmed that for the construction
√ of confidence intervals for µ, it is ‘harmless’ to nevertheless assume
µNn − µ) is standard normally distributed.
that Nn (b
5. Discussion
While sequential designs are in common use in medical and other applications, and while the hypothesis testing theory based there upon has
been well established for a long time, there is more confusion about parameter and precision estimation following such a sequential study. [MKA14],
[MMA16], and [MMA15] showed that the sample average is a valid estimator, with both stochastic and deterministic stopping rules, for a wide class
of normal and exponential-family-based models. They established that this
estimator, in spite of small-sample bias and the fact that there is no uniform
minimum-variance unbiased estimator, is consistent and hence asymptotically unbiased.
Building upon this work, in this paper, we have shown that the sample
average in the case of normally distributed outcomes is also asymptotically
normal in a broad range of situations. First, this is true with stochastic
6
BEN BERCKMOES AND GEERT MOLENBERGHS
stopping rule. Second, it applies in almost all deterministic stopping rule
situations within the class considered, except in the very specific case where
the normal population mean µ = 0. Note that the special status of the null
value stems from the fact that the cut-off between stopping and continuing
associated with our deterministic stopping rule is equal to zero. It can easily
be shown, should the cut-off point be shifted to a non-zero value, that then
the problematic value for µ also shifts.
We also showed that the Kolmogorov distance, for µ = 0, equals
p
µNn − µ)) = 1/8,
K(Φ, Nn (b
√
from which it follows that
Nn (b
µNn − µ) n does not converge weakly to
Φ in this case. It is enlightening that the qualitative non-convergence result
is supplemented with a quantitative determination of the deviation from
normality.
To further examine the extent of the result obtained, simulations show
that, indeed, asymptotic normality becomes more problematic when µ approaches zero and the parameter β approaches +∞, with the latter value
corresponding to a deterministic rule. However, asymptotic normality is invoked predominantly to calculate normally based confidence intervals. It is
therefore very reassuring that using such intervals for µ = 0 and a deterministic stopping rule does not lead to any noticeable effect on the coverage
probabilities.
In summary, we can conclude that for relevant classes of stopping rules,
the sample average and corresponding normal confidence interval can be
used without problem. It will be of interest to examine in more detail the
situation of outcomes that follow an exponential family distribution, other
than the normal one.
References
[EF90] Emerson, S. S.;Fleming, T. R. (1990). Parameter estimation following group sequential hypothesis testing. Biometrika 77, 875–892.
[HP88] Hughes, M.D.; Pocock, S.J. (1988). Stopping rules and estimation problems in
clinical trials. Statistics in Medicine 7, 1231–1242.
[LH99] Liu, A.; Hall, W. J. (1999). Unbiased estimation following a group sequential test.
Biometrika 86, 71–78.
[LHW06] Liu, A.; Hall, W. J.; Yu, K. F.; and Wu, C. (2006). Estimation following a group
sequential test for distributions in the one-parameter exponential family. Statistica
Sinica 16, 165–81.
[MMA15] Milanzi, E.; Molenberghs, G.; Alonso, A.; Kenward, M. G.; Tsiatis, A. A.;
Davidian, M.; Verbeke, G. Estimation after a group sequential trial. Stat. Biosci. 7
(2015), 187–205.
[MMA16] Milanzi, E.; Molenberghs, G.; Alonso, A.; Kenward, M. G.; Verbeke, G.; Tsiatis, A. A.; Davidian, M. Properties of estimators in exponential family settings with
observation-based stopping rules. J. of Biometrics Biostatist. 7, 272.
[MKA14] Molenberghs, G.; Kenward, M. G.; Aerts, M.; Verbeke, G.; Tsiatis, A. A.;
Davidian, M.; Rizopoulos, D. On random sample size, ignorability, ancillarity, completeness, separability, and degeneracy: sequential trials, random sample sizes, and
missing data. Stat. Methods Med. Res. 23 (2014), no. 1, 11–41.
[R91] Rachev, S. T. Probability metrics and the stability of stochastic models. Wiley Series
in Probability and Mathematical Statistics: Applied Probability and Statistics. John
Wiley & Sons, Ltd., Chichester, 1991.
[S78] Siegmund, D. (1978). Estimation following sequential tests. Biometrika 64, 191–199.
7
[W45] Wald, A. Sequential tests of statistical hypotheses. Ann. Math. Statist. 16 (1945),
117-186.
[Z83] Zolotarev, V. M. Probability metrics. (Russian) Teor. Veroyatnost. i Primenen. 28
(1983), no. 2, 264–287.
8
BEN BERCKMOES AND GEERT MOLENBERGHS
Appendix A. Proof of Theorem 1
Before writing down the proof of Theorem 1, we give three lemmas. Part
of Lemma 3 can be found in [MKA14], but as it belongs to the heart of our
calculations, we present a complete proof here.
Lemma 1. For A, B ∈ R,
Z ∞
A
φ(x)Φ(A + Bx)dx = Φ √
.
1 + B2
−∞
(11)
Proof. This is standard.
Lemma 2. For k, z ∈ R,
2z − k
k − z − nµ
k − 2nµ
z − nµ
√
√
√
φ √
.
φ
=φ
φ
n
n
2n
2n
(12)
Proof. This follows by a straightforward calculation.
Lemma 3. Let fNn ,KNn be the joint density of Nn and KNn . Then, for the
probabilistic stopping rule (1),
1
k − nµ
βk
√
fNn ,KNn (n, k) = √ φ
(13)
Φ α+
n
n
n
and
1
fNn ,KNn (2n, k) = √ φ
2n
k − 2nµ
√
2n
α+
1 − Φ q
2n+β 2
2n
and, for the deterministic stopping rule (2),
1
k − nµ
√
fNn ,KNn (n, k) = √ φ
1{k>0}
n
n
and
1
fNn ,KNn (2n, k) = √ φ
2n
k − 2nµ
√
2n
βk
2n
,
k
1−Φ √
.
2n
(14)
(15)
(16)
Proof. First suppose that the probabilistic stopping rule (1) is followed.
Notice that
fNn ,KNn (n, k) = fNn ,Kn (n, k) = fKn (k)fNn |Kn (n | k),
(17)
with fKn the density of Kn , and fNn |Kn the conditional density of Nn given
Kn . Now, the Xi being independent and normally distributed with mean µ
and variance 1, we have
1
k − nµ
√
fKn (k) = √ φ
.
(18)
n
n
Furthermore, by (1),
βk
fNn |Kn (n | k) = Φ α +
n
Combining (17), (18), and (19), establishes (13).
.
(19)
9
We now establish (14). Observe that
fNn ,Kn (2n, k) = fNn ,K2n (2n, k)
= fK2n (k) − fNn ,K2n (n, k)
= fK2n (k) − fNn ,Kn (n, ·) ⋆ fP2n
i=n+1
Xi
(k),
(20)
⋆ being the convolution product, and the last equality following by indepence of Nn and Xn+1 , . . . , X2n . Using (13) and the fact that the Xi are
independent and normally distributed with mean µ and variance 1, (20)
equals
Z
1
k − z − nµ
z − nµ
k − 2nµ
1 ∞
βz
√ φ
√
√
√
φ
φ
Φ α+
dz,
−
n −∞
n
n
n
2n
2n
which, by (12),
Z ∞
1
k − 2nµ
k − 2nµ
2z − k
1
βz
√
√
=√ φ
dz.
φ √
− φ
Φ α+
n
n
2n
2n
2n
2n
−∞
(21)
2z−k
√
After performing the change of variables u = 2n , (21) reduces to
Z ∞
k − 2nµ
1
β
βk
√ φ
√
+ √ u du ,
φ(u)Φ α +
1−
2n
2n
2n
2n
−∞
which, by (11), coincides with
α + βk
k − 2nµ
1
√ φ
√
1 − Φ q 2n .
2
2n
2n
1+ β
2n
This proves that (14) holds.
Now suppose that the detereministic stopping rule (2) is followed.
Of course, (17) and (18) continue to hold, and, by (2),
fNn |Kn (n | k) = 1{k>0} ,
from which (15) follows.
To establish (16), notice that (21) also continues to hold, which, by (15),
gives
Z
k − 2nµ
1
z − nµ
k − z − nµ
1 ∞
√
√
√
fNn ,Kn (2n, k) = √ φ
φ
−
φ
dz,
n 0
n
n
2n
2n
which, by (12),
Z ∞
1
k − 2nµ
k − 2nµ
2z − k
1
√
√
=√ φ
φ √
− φ
dz,
n
2n
2n
2n
2n
0
√
,
which, after performing the change of variables u = 2z−k
2n
"
#
Z ∞
1
k − 2nµ
√
= √ φ
1−
√ φ(u)du
2n
2n
−k/ 2n
1
k − 2nµ
k
√
= √ φ
1−Φ √
.
2n
2n
2n
This finishes the proof of (16).
10
BEN BERCKMOES AND GEERT MOLENBERGHS
Proof of Theorem 1. First suppose that the probabilistic stopping rule (1)
is followed.
For n and a Borel set A ⊂ R,
hp
i
Nn (b
µNn − µ) ∈ A
P
KNn − Nn µ
√
∈A
= P
Nn
Kn − nµ
K2n − 2nµ
√
√
= P
∈ A, Nn = n + P
∈ A, Nn = 2n . (22)
n
2n
Plugging in (13) and (14) in (22), gives
i
hp
µNn − µ) ∈ A = I1 + I2 ,
P
Nn (b
with
I1 =
and
I2 =
Z
1A
Z
1A
k − nµ
√
n
k − 2nµ
√
2n
1
√ φ
n
1
√ φ
2n
k − nµ
√
n
k − 2nµ
√
2n
(23)
βk
Φ α+
n
α+
1 − Φ q
dk
βk
2n
2n+β 2
2n
√
shows that
Performing the change of variables u = k−nµ
n
Z
β
φ(u)Φ α + βµ + √ u du,
I1 =
n
A
dk.
(24)
√
gives
and performing the change of variables u = k−2nµ
2n
"
!#
r
Z
2n
β
φ(u) 1 − Φ
I2 =
(α + βµ) + p
u
du.
2n + β 2
2n + β 2
A
(25)
Combining (23), (24), and (25), yields (7).
Now suppose that the deterministic stopping rule (2) is followed. Fix n
and a Borel set A ⊂ R. Of course, (22) continues to hold. Plugging in (15)
and (16) in (22) gives
i
hp
Nn (b
µNn − µ) ∈ A = L1 + L2 ,
(26)
P
with
L1 =
Z
∞
−∞
and
L2 =
Z
∞
−∞
1A
1A
k − nµ
√
n
k − 2nµ
√
2n
1
√ φ
2n
1
√ φ
n
k − nµ
√
n
k − 2nµ
√
2n
1−Φ
√
leads to
Performing the change of variables u = k−nµ
n
Z
φ(u)1{u>−√nµ} du,
L1 =
A
1{k>0} dk
k
√
2n
dk.
(27)
11
√
and performing the change of variables u = k−2nµ
yields
2n
Z
i
h
√
φ(u) 1 − Φ u + 2nµ du.
L2 =
(28)
A
Now (26), (27), and (28) give (8).
We now turn to the case µ = 0. Replacing A by ]−∞, x] in (26), (27),
and (28), shows that, for x ≥ 0,
Z x
Z x
p
φ(u)Φ(u)du −
φ(u)du
Φ(x) − P[ Nn (b
µNn − µ) ≤ x] =
−∞
=
0
2
[Φ(x)] /2 − Φ(x) + 1/2 ,
which assumes the maximal value 1/8 on [0, ∞[, and, for x ≤ 0,
Z x
p
Φ(x) − P[ Nn (b
φ(u)Φ(u)du
µNn − µ) ≤ x] =
−∞
= [Φ(x)]2 /2,
which assumes the maximal value 1/8 on ] − ∞, 0]. This proves (9). Finally,
replacing A by [−x, x] in (26), (27), and (28), shows that, for x ≥ 0,
p
2Φ(x) − 1 − P[−x ≤ Nn (b
µNn − µ) ≤ x]
Z x
Z x
=
φ(u)du
φ(u)Φ(u)du −
−x
=
0
2
[Φ(x)] /2 − [Φ(−x)]2 /2 − Φ(x) + 1/2
= 0,
which proves (10).
12
BEN BERCKMOES AND GEERT MOLENBERGHS
Appendix B. Tables from the simulation study
Table 1. Simulation study for estimation (3) for the probabilistic stopping rule (1); α = 0; β small; µ, true mean for the standard normal from which the sample is taken; n, number of observations; C = C(α, β, µ, n) in Theorem 1; K, the Kolmogorov
distance
between the standard normal cdf and the empirical cdf
√
of Nn (b
µNn − µ) based on 1000 simulations; L, number of times
out of 1000 where
√ the true parameter
√ µ is contained in the interval µ
bNn − 1.96/ Nn , µ
bNn + 1.96/ Nn (which would be a 95%√
µn − µ) were standard normally disconfidence interval if Nn (b
tributed).
β
µ
n
C
K
L
0 -10
10 0.000 0.024 947
0 -10 100 0.000 0.018 948
0 -10 1000 0.000 0.021 941
0 -1
10 0.000 0.026 947
0 -1 100 0.000 0.030 948
0 -1 1000 0.000 0.021 952
0
0
10 0.000 0.021 941
0
0 100 0.000 0.030 953
0
0 1000 0.000 0.011 958
0
1
10 0.000 0.027 954
0
1 100 0.000 0.017 957
0
1 1000 0.000 0.045 957
0 10
10 0.000 0.039 950
0 10 100 0.000 0.026 943
0 10 1000 0.000 0.024 956
β
µ
n
C
K
L
1 -10
10 0.000 0.015 958
1 -10 100 0.000 0.018 947
1 -10 1000 0.000 0.029 949
1 -1
10 0.081 0.024 960
1 -1 100 0.025 0.047 954
1 -1 1000 0.008 0.014 941
1
0
10 0.120 0.042 950
1
0 100 0.039 0.017 952
1
0 1000 0.013 0.012 954
1
1
10 0.071 0.026 957
1
1 100 0.023 0.016 955
1
1 1000 0.008 0.036 941
1 10
10 0.000 0.037 951
1 10 100 0.000 0.026 952
1 10 1000 0.000 0.028 949
Table 2. Same setup as in Table 1. Now β is moderately large.
β
µ
n
C
K
L
10 -10
10 0.000 0.010 949
10 -10 100 0.000 0.020 955
10 -10 1000 0.000 0.024 952
10 -1
10 0.002 0.015 953
10 -1 100 0.000 0.017 955
10 -1 1000 0.000 0.017 947
10
0
10 0.440 0.084 945
10
0 100 0.300 0.021 948
10
0 1000 0.120 0.047 950
10
1
10 0.001 0.028 942
10
1 100 0.000 0.026 973
10
1 1000 0.000 0.019 940
10 10
10 0,000 0.021 967
10 10 100 0.000 0.009 955
10 10 1000 0.000 0.033 948
β
µ
n
C
K
L
100 -10
10 0.000 0.019 948
100 -10 100 0.000 0.024 944
100 -10 1000 0.000 0.023 946
100 -1
10 0.001 0.039 960
100 -1 100 0.000 0.015 953
100 -1 1000 0.000 0.011 941
100
0
10 0.494 0.145 950
100
0 100 0.481 0.080 948
100
0 1000 0.437 0.068 946
100
1
10 0.001 0.035 946
100
1 100 0.000 0.011 938
100
1 1000 0.000 0.021 951
100 10
10 0.000 0.009 954
100 10 100 0,000 0.014 960
100 10 1000 0.000 0.010 954
13
Table 3. Simulation study for estimation (3) for the deterministic stopping rule (2); µ, true mean for the standard normal from which the sample is taken; n, number of observations; C = C(µ, n) in Theorem 1; K, the Kolmogorov distance
between the standard normal cdf and the empirical cdf of
√
Nn (b
µNn − µ) based on 1000 simulations; L, number of times
out of 1000 where
√ µ is contained in the inter√ the true parameter
bNn + 1.96/ Nn (which would be a 95%val µ
bNn − 1.96/ Nn , µ
√
µn − µ) were standard normally disconfidence interval if Nn (b
tributed).
β
µ
n
C
K
L
∞ -10
10 0.000 0.005 943
∞ -10 100 0.000 0.026 949
∞ -10 1000 0.000 0.029 949
∞ -1
10 0.002 0.034 952
∞ -1 100 0.000 0.025 956
∞ -1 1000 0.000 0.018 959
∞
0
10 0.250 0.130 940
∞
0 100 0.250 0.113 941
∞
0 1000 0.250 0.129 958
∞
1
10 0.002 0.022 955
∞
1 100 0.000 0.034 968
∞
1 1000 0,000 0.005 964
∞ 10
10 0.000 0.016 946
∞ 10 100 0,000 0.021 953
∞ 10 1000 0,000 0.023 952
| 10math.ST
|
1
Recognizing Detailed Human Context
In-the-Wild from Smartphones and
Smartwatches
arXiv:1609.06354v4 [cs.AI] 30 Sep 2017
Yonatan Vaizman∗† Katherine Ellis∗ and Gert Lanckriet, senior member, IEEE
This version was accepted to IEEE Pervasive Computing magazine.
Please cite the official (edited and stylized) paper, which appears in IEEE Pervasive Computing, vol. 16,
no. 4, October-December 2017, pp. 62-74.
Abstract—The ability to automatically recognize a person’s behavioral context can contribute to health monitoring, aging care and
many other domains. Validating context recognition in-the-wild is crucial to promote practical applications that work in real-life settings.
We collected over 300k minutes of sensor data with context labels from 60 subjects. Unlike previous studies, our subjects used their
own personal phone, in any way that was convenient to them, and engaged in their routine in their natural environments. Unscripted
behavior and unconstrained phone usage resulted in situations that are harder to recognize. We demonstrate how fusion of
multi-modal sensors is important for resolving such cases. We present a baseline system, and encourage researchers to use our public
dataset to compare methods and improve context recognition in-the-wild.
Index Terms—Context awareness, Mobile sensors, Artificial intelligence, Machine learning, Human activity recognition.
F
1
I NTRODUCTION
The ability to automatically recognize a person’s context
(i.e., where they are, what they are doing, who they are
with, etc.) is greatly beneficial in many domains. Health
monitoring and lifestyle interventions have traditionally
been based on manual, subjective reporting [1], sometimes
by end-of-day recalling [2]. These can improve with automatic (frequent, effortless, and objective) detection of behaviors like exercise, eating, sleeping, or mental states like
stress. Just-in-time interventions (e.g. for addiction) often
prompt the patient at arbitrary times of the day, possibly
missing times when the patient needs support the most [3].
Automatically recognizing context will help detect critical
times and offer immediate support (e.g. an alcoholic patient
may be in high risk of craving or lapse when the context is
“at a bar, with friends”).
The biomedical research community acknowledges the
effects of behavior, lifestyle and environment on health, disease and treatments [4]. Automatic context recognition tools
will be essential to incorporate behavioral and exposure
aspects into large scale studies and to tailor appropriate
treatment for patients. The range of measured exposures
should be broad and cover diverse life style and environmental aspects. Commercial tools that offer superficial
recognition (e.g. of walking, running, and driving) will not
suffice. Personal assistant systems can adjust to context and
better serve the user. Aging care programs can use automated logging of older adults’ behavior to detect early signs
of cognitive impairment, monitor functional independence,
•
•
•
Y. Vaizman, K. Ellis and G. Lanckriet were with the Department of Electrical and Computer Engineering, University of California, San Diego.
∗ Y. Vaizman and K. Ellis are equal contributors.
† Corresponding author. Email: [email protected]
and support aging at home [5].
In order for such applications to succeed in large scale,
the context recognition component has to be unobtrusive
and to work smoothly, without requiring the person to
adjust their behavior. It is important that research emulates
real-world settings, where such applications will eventually
be deployed. In this work we promote context recognition
in-the-wild, meaning capturing people’s authentic behavior
in their natural environments, with the use of every-day
devices — smartphones and smartwatches. We address the
difficulty that in-the-wild conditions add, and show how
multi-modal sensors can help.
Related work
It is common for people to have their phone close to them
most of the time [6]. This growing trend, and the variety
of built-in sensors, make phones popular agents for recognizing human behavior. Smartwatches are a useful sensing
addition. While capturing informative signals about hand
and arm motion, they remain very natural to wear and don’t
add any burden to the user.
Previous works have shown the advantage of fusing
sensors of different modalities, from smartphones and
smartwatches, to improve recognition of basic movement
activities [7] and more complex activities, like smoking
or drinking coffee [8]. However, most past works have
collected data under heavily controlled conditions, with
researchers instructing subjects to perform scripted tasks.
Fitting models to recognize prescribed activities may result
in poor generalization to real life scenarios [9]. To promote
real-life working applications, we argue that research has
to be done in natural and realistic settings, satisfying four
2
in-the-wild conditions:
1)
2)
3)
4)
Naturally used devices. Introducing a foreign device to the user adds a burden and harms natural behavior. Ideally, subjects would use their own
phone, and possibly additional convenient devices,
like watches.
Unconstrained device placement. It has been shown
that the placement and orientation of sensors have
a great influence on the success of recognition [10].
However, this does not mean we should avoid this
difficulty by forcing specific placement — a practical
real-world application cannot restrict users to keep
their phone in pant pocket for the recognition to
work. Instead, research should address the variability in device placement as a challenge, and provide
solutions to overcome it.
Natural environment. The recorded behavior
should be in the subjects’ natural environment and
on their own free time. They should not be instructed where or when to perform their activity.
Natural behavioral content. In many works the
researchers instructed subjects to perform scripted
tasks [7], [8]. The recorded behavior was then simulated, and not natural. Other works let the subjects
behave on their own time, but still prescribed a list
of targeted activities [11], [12], which may cause
the subject to perform actions they are not used to,
like “vacuum cleaning”. In-the-wild studies should
record the behavior that is natural to each individual
subject.
A major challenge is acquiring labels of the behavioral
context. Attaining in-the-wild conditions usually trades off
with other aspects of the data collection effort, resulting in
fewer labeled examples, smaller range of interest labels or
compromised privacy of the subjects.
Previous research addressed some aspects of in-the-wild
data collection in different ways. Han et al. [13] designed
a decision-tree architecture that activates predetermined
sensors to differentiate eight ambulatory and transportation
states. Such a hand-crafted system is hard to scale to more
contexts. They validated their system with an observer that
followed a single user. Ordonez et al. [14] installed a set
of state-change sensors around a home to detect daily home
activities. While such sensors are un-obtrusive and maintain
natural behavior, the complicated device setup limits the
deployment of data collection and practical applications.
It also cannot track the person outside of the monitored
environment. Dong et al. [15] targeted eating periods and
used an unnatural setup of having a smartphone bound
to the wrist. Subjects had to mark start times of eating,
and after data collection they reviewed and corrected their
markings. This resulted in 449 hours of data with 116
eating periods from 43 subjects. Rahman et al. [16] compared
different approaches for subjects to self-report their stress
level (immediately or by recalling later). They suggested
a compromise approach where the subject can report on
their own time but with the help of cues (like location or
surrounding sound level) to remember how they felt at
specific times of the day.
Choudhury et al. [17] designed a system to address
the requirements for a practical context recognition system,
including unobtrusive lightweight devices, long battery life
and multi-modal sensing. However, most of their validation
was done on controlled data, collected in specific locations,
with constrained positioning of device, and with a sequence
of 8 activities that was scripted, observed, and repeated by
12 subjects. Consolvo et al. [18] utilized the same system
(trained on the controlled data) in a field study of an application to promote physical activity. The mobile app used
a combination of the automated recognition (of walking,
cycling, etc.) and manual editing of a daily journal.
Hemminki et al. [19] targeted detecting transportation
modes and specifically designed features that would be less
sensitive to placement of the phone. Ganti et al. [11] gave
eight subjects a Nokia N95 phone for a period of eight
weeks and asked them to go about their regular routine
and use the phone for recording whenever they can, in
any location or time-of-day. The phone was constrained to
be in the pocket or pouch. The interface allowed selecting
an activity from a set of eight activities and marking when
you start and when you finish. They collected a total of 80
hours. Khan et al. [20] targeted recognition of 15 activities.
To collect measurements and annotations, they handed a
NEXUS phone to subjects for a month. Subjects were free
to perform the prescribed activities on their own time and
they used the phone to mark the beginning and end of
the selected activity. They collected about 3000 examples
per activity from 30 subjects, plus a follow-up validation
with eight subjects using the trained real-time recognition
system.
In Yatani et al.’s [21] out-of-lab study, five subject wore
a phone around their neck to take egocentric snapshots
that were later used to label the activity. A similar label
acquisition strategy was taken in large scale by Ellis et al. [22]
with 40 subjects who recorded hip-mounted accelerometer
and GPS data from routine behavior in natural environment
for several days. The subjects wore a SenseCam device
around their neck, which took snapshots periodically, and
the thousands of images were later used by research assistants to annotate the activity. Pirsavash et al. [12] used
a GoPro video camera for both sensor measurements and
ground truth labels. The subjects wore the device around
the chest in a single morning at their own home, and were
prescribed a list of home activities to perform with no
extra specifications. They recorded over 10 hours of video
from 12 people and later annotated household objects and
activities for about 30k frames (every second). Their dataset
is publicly available. While the camera approach may generate more reliable labels in certain cases, the unnatural and
uncomfortable equipment compromises natural behavior.
Furthermore, offline annotation of images is costly, making
it hard to scale, and violates the privacy of the subjects and
people around them. The alternative of self reporting has the
advantage of collecting labels when a camera is not present
(e.g. “shower”), when the context is not clearly visible in the
image (e.g. “singing”), or when the subject knows best what
is happening (e.g. “with family” vs. “with friends”).
3
Our work
In this work we use smartphone and smartwatch sensors
to recognize detailed situations of people in their natural
behavior. We collected labeled data from over 300k minutes
from 60 subjects. Every minute has multi-sensor measurements and is annotated with relevant context labels. To
the best of our knowledge, this dataset, which is publicly
available, is far larger in scale compared to others collected
in the field. Similar to [11], [15], [20], we rely on selfreport. Unlike those works, our data collection app offers an
extensive menu of over 100 context labels and the ability to
select combinations of relevant labels. This facilitates natural
behavior from the subjects, e.g. they are free to “run on a
treadmill”, while “watching TV” if it is natural to them (in
[11], [20], subjects were forced to choose one activity, possibly causing them to act unnaturally). This also provides
rich descriptions of context, as combinations of different aspects, like environment, activities, company, body posture.
Similar combinatorial representations were previously used
to describe objects and actions in images [12] and locations,
objects, humans, and animals in sound clips [23]. In those
cases annotation was done offline, but in our case, attaining
detailed labeling by self-reporting requires attention and
effort from the subjects. To mitigate it, our app’s interface
offers many reporting-mechanisms to minimize interaction
time. Subjects can report the start of activity (as in [11], [15],
[20]). They can manually edit events in a daily calendar that
included automatically recognized contexts (similar to [18]).
We treat only the manual corrections or additions as ground
truth; the automated predictions act as cues, to help the
subjects recall their context (as suggested in [16]).
The main contribution of this work is the emphasis on
in-the-wild conditions, as mentioned in “previous work”:
1)
2)
3)
4)
Naturally used devices. Subjects used their own personal phones, and a smartwatch that we provided.
Unconstrained device placement. Subjects were free
to carry their phone in any way that was convenient
to them.
Natural environment. Subjects collected data in
their own regular environment for about a week.
Natural behavioral content. No script or tasks were
given. We did not target a specific set of activities.
Instead, the context labels that we analyze came from
the data, as the subjects engaged in their routine and
applied any relevant labels (from the large menu)
that fit what they were doing.
Recognizing context in-the-wild is more challenging,
compared to controlled conditions, because of the large
variability in real-life. Diversity in phone devices and sensor
hardware has an effect on the measurements [24]. Our data
represents both iPhone and Android, including many varieties of devices. Variability in behavioral content is clearly
visible in the ground truth labels of out data, including
combinations like {Running, Outside, Exercise, Talking,
With friends}, {Running, Indoors, Exercise, At the gym,
Phone on table}, {Sitting, Indoors, At home, Watching TV,
Eating, Phone on table}, {Sitting, At a restaurant, Drinking
(alcohol), Talking, Eating}, {Sitting, On a bus, Phone in
pocket, Talking, With friends}, {On a bus, Standing}. Such
variability was missed in works that defined behavior with
a small set of mutually exclusive activities. Variability in
manner or style (e.g. different gaits) is less visible, but is still
captured in our sensor measurements. Such variability can
easily be missed in scripted experiments or if restricting how
to use devices. Our analysis demonstrates the difficulty in
resolving context in-the-wild, and the importance of using
complementary sensing modalities. We show that everyday
devices, in their natural usage, can capture information
about a wide range of behavioral attributes.
C ONTEXT RECOGNITION SYSTEM
Figure 1 illustrates the flow of our recognition system.
The system is based on measurements from five sensors
in a smartphone: accelerometer (Acc), gyroscope (Gyro),
location (Loc), audio (Aud), and phone state (PS), as well
as accelerometer measurements from a smartwatch (WAcc).
For a given minute, the system samples measurements from
these six sensors and the task is to detect the combination
of relevant context labels (Figure 1 (A)), i.e. declare for each
label l a binary decision: yl = 1 (the label is relevant to this
minute) or yl = 0 (not relevant).
For this paper we opted for simple computational methods, based on linear classifiers and basic heuristics for sensor
fusion. We model each label separately and treat every
minute as an independent example. We include time-ofday as part of the PS features, but we do not model the
behavioral time-series throughout the day. The goal of this
paper is to show the potential of context recognition in-thewild, and to establish a baseline. Future papers will use nonlinear methods, dynamic-context models, and interaction
among labels.
Single-sensor classifiers use sensor-specific features and
help us understand how informative each sensor can be,
independently of the other sensors, for a given context label
(Figure 1 (B)). We use logistic regression — a linear classifier
that outputs a continuous value (interpreted as probability)
in addition to the binary decision. This is helpful for sensor
fusion. The following procedure was performed for a given
sensor s and a given label l: (a) For each example, compute a
ds -dimensional feature vector xs . Each sensor has a different
set of relevant features. (b) Standardize each feature by
subtracting mean and dividing by standard deviation (these
statistics are estimated on the training set). (c) Learn a ds dimensional logistic regression classifier from the training
set. (d) Apply the logistic regression classifier to a test
example to obtain a binary classification yl and probability
value P (yl = 1|xs ). To overcome the imbalance between the
positive class and the negative class we applied balanced
class weights (inversely proportional to the class frequency
in the training set).
At this point, it is possible introduce some domain
knowledge and assign appropriate sensors to certain labels. For example, the watch accelerometer can be a good
indicator for specific hand-motion activities, like “washing
dishes”, while audio might better predict environmental
contexts like “in class” or “at a party”. These design decision
are not always obvious, so we continue with sensor-fusion
methods that can learn the best predictors from data.
4
Sensor fusion
Our system further combines information from N different
sensors. We propose three alternative ways.
Early fusion (EF) classifiers combine information from
multiple sensors prior to the classification stage (Figure 1 (C)). The following procedure was performed for
a given label l: (a) Start with the sensor-specific feature
vectors {xs }N
s=1 . (b) Concatenate the (standardized) sensorspecific feature vectors into a single vector x of dimension
PN
d =
s=1 ds (c) Learn a d-dimensional logistic regression classifier from the training set. (d) Apply the logistic
regression classifier to a test example to obtain a binary
classification yl and probability value P (yl = 1|x).
Late fusion classifiers. We use ensemble methods to
combine the predictions of the N single-sensor classifiers.
We chose to combine the probability outputs P (yl = 1|xs ),
and not the binary decisions, to take into account the
“confidence” of each of the N classifiers and avoid overinfluence of irrelevant sensors. We explore two methods for
late fusion:
Late fusion using average probability (LFA) (Figure 1 (D))
applies a simple bagging heuristic and averages the probability values from all the single-sensor classifiers to obtain a
final “probability” value, i.e., P (yl = 1|x1 , x2 , . . . , xN ) =
PN
1
s=1 P (yl = 1|xs ). LFA declares “yes” if the average
N
probability is larger than half. No additional training is
performed after the single-sensor classifiers are learned.
This method grants equal weight to each sensor, hoping
that informative sensors will classify with higher confidence
(probability close to 0 or close to 1) and will influence
the final decision more than irrelevant sensors (which will
hopefully predict with probability close to 0.5).
As mentioned earlier, some sensors may be consistently
better suited for certain labels. As a flexible alternative to
deciding apriori how to assign sensors to labels, we can
let sensor-weights be learned from data. Late fusion using
learned weights (LFL) (Figure 1 (E)) is a second type of
late fusion that places varying weight on each sensor. This
method learns a second layer of N -dimensional logistic
regression model. The second layer’s input is the N probability outputs of the single-sensor models, and the output is
a final decision yl .
devices of generations from iPhone4 to iPhone6 and with
operating system (iOS) versions 7, 8 and 9. 26 subjects
were Android users, with various devices (Samsung, Nexus,
Motorola, Sony, HTC, Amazon Fire-Phone, and Plus-One).
The subjects were from diverse ethnic backgrounds (selfdefined), including Chinese, Mexican, Indian, Caucasian,
African-American, and more. The majority of the subjects
(93%) were right handed, and chose to wear the smartwatch
on their left wrist. The dataset is homogeneous with regard
to occupation; almost all the subjects were students or
research assistants. 34 subjects were female and 26 were
male. Table 1 describes additional subject characteristics.
We installed the app on each subject’s personal phone and
provided the watch to the subject (56 out of the 60 agreed
to wear the watch). The subject then engaged in their usual
behaviors for approximately a week, while keeping the app
running in the background on their phone as much as
possible and convenient. The subject was asked to report
as many labels as possible without interfering too much
with their natural behavior. They were free to remove the
watch whenever they wanted and were asked to turn off
the watch-app when they were not wearing it. Basic compensation of $40 was given to each subject, with additional
incentive of up to $35 that depended on the amount of
labeled data they provided.
The resulted ExtraSensory Dataset, contains a total of
308,320 labeled examples (minutes) from sixty users. Table 1
details statistics (over 60 subjects) about the amount of
data collected. Not all the sensors were available at all
times. Table 2 specifies details on the sensors. The dataset is
publicly available and researchers are encouraged to use it
for developing and comparing context recognition methods
(http://extrasensory.ucsd.edu).
Range
Mean (SD)
Age (years)
18–42
24.7 (5.6)
Height (cm)
145–188
171 (9)
Weight (kg)
50–93
66 (11)
18–32
23 (3)
Body mass index (kg/m2 )
Labeled examples 685–9706 5139 (2332)
Additional unlabeled examples
2–6218
1150 (1246)
Average applied labels per example
1.1–9.7
3.8 (1.4)
Participation duration (days)
2.9–28.1
7.6 (3.2)
TABLE 1
Statistics over the 60 users in the dataset (SD: standard deviation).
DATA COLLECTION
For the purpose of large-scale data collection, we developed
a mobile application (app) called ExtraSensory, with versions
for both iPhone and Android smartphones, and a companion application for the Pebble smartwatch that integrates
with both. The app was used to collect both sensor measurements and ground truth context labels. Every minute
the app records a 20sec window of sensor measurements
from the phone and watch. Within that window, the time
samples of different sensors are not guaranteed to be exactly
aligned. The flexible user interface provided the user with
many mechanisms to self-report the relevant context labels
and cover long behavioral time with minimal effort and time
of interaction with the app (Figure 2).
Sixty subjects (users) were recruited using fliers posted
around the UC San Diego campus and campus-based email
lists. 34 of the subjects were iPhone users, with iPhone
E VALUATION AND RESULTS
We evaluated classification performance using five-fold
cross validation: each fold has 48 users in the training set
and the other twelve users in the test set. We also conducted leave-one-user-out experiments (LOO). To measure
performance, classification accuracy is a misleading metric
because of imbalanced data; for a rare label that appears in
1% of the test set, a trivial classifier that always declares
“no” will achieve 99% accuracy but is completely useless. It
is important to consider competing metrics, like sensitivity
and specificity. A common approach is to observe sensitivity
(recall) against precision, or to calculate their harmonic
mean (F1). However, precision and F1 are less fitting, since
they are very sensitive to how rare labels are. Chance level
can be arbitrarily small, and when averaging precision or
5
A
Sensor
Measurements
Context
Labels
Classifier
listening to music
at the beach
walking
outdoors
B
Sensor
Measurements
Features
acceleration
gyroscope
x11
x12
.
.
.
audio
x21
x22
.
.
.
C
Pr(walking|x1)
+
Pr(walking|x2)
Pr(walking|x3)
+
+
x41
x42
.
.
.
phone state
Label
Probability
+
x31
x32
.
.
.
location
Binary
Classifier
Pr(walking|x4)
-
E
D
Concatenated
Features
x11
x12
.
.
.
x21
x22
.
.
.
Binary
Classifier
Label
Probability
Sensorspecific
Probability
Average
Label
Probability
Pr(walking|x)
-
Pr(walking|x2)
Pr(walking|x3)
...
Binary
Classifier
Label
Probability
Pr(walking|x1)
Pr(walking|x1)
+
Sensorspecific
Probability
+
Pr(walking|x)
+
Pr(walking|x2)
Pr(walking|x3)
Pr(walking|x)
-
...
Fig. 1. Context recognition system. (A) While the person is engaged in natural behavior the system uses sensor measurements from the smartphone
and smartwatch to automatically recognize the person’s detailed context. (B) Single-sensor classifiers. Appropriate features are extracted from each
sensor. For a given context label, classification can be done based on each sensor independently. (C) Under Early fusion (EF), features from multiple
sensors are concatenated to form a long feature vector. (D) Late fusion using averaging (LFA) simply averages the output probabilities of the singlesensor classifiers. (E) Late fusion with learned weights (LFL) learns how much to “listen” to each sensor when making the final classification.
6
Fig. 2. Screenshots from the ExtraSensory mobile application (iPhone version). (A) On the history tab the user can see a daily log of activities
and contexts. The server sends real-time body-state predictions (based on preliminary training data from two iPhone users — the researchers).
These predictions appear with question marks and help the user organize the log and recall when their activity may have changed. The user can
update the history records’ labels, add secondary labels like “at home” and “eating”, merge consecutive records into a longer period, and split
records. (B) The label selection menu is indexed by topics and a “frequently use” link to make it easier for the user to select quickly. (C) Using the
“active feedback” button the user can report they will be engaged in a specific context starting immediately and valid for a set period of time. (D)
Periodic notifications remind the user to provide labels. If the user is engaged in the same context as they recently reported they simply need to
reply “correct” to the question. If any element of the context has changed they can press “not exactly” and be directed to a screen where they can
update the labels of the recent time period. These notifications appear on the watch as well, which enables easier responses.
7
Sensor
Accelerometer
Gyroscope
Magnetometer
Watch Accelerometer
Watch Compass
Location
Location precomputed
Audio
Audio power
Phone State
Low frequency sensors
Core
raw measurements
3-axis (40Hz)
3-axis (40Hz)
3-axis (40Hz)
3-axis (25Hz)
heading angle (var)
long-lat-alt (var)
location variability (1pe)
13MFCC (46ms frames)
1pe
1pe
1pe
examples
308,306
291,883
282,527
210,716
126,781
273,737
263,899
302,177
303,877
308,320
308,312
176,941
users
Why does sensor fusion help?
60
57
58
56
53 The performance of single-sensor classifiers on selected
58 labels (Figure 4 (A)) demonstrates the advantage of having
58 sensors of different modalities. As expected, for detecting
60
60 sleeping, the watch is more informative than the phone’s
60 motion sensors (Acc, Gyro) — during sleep the phone may
60 be lying motionless on a nightstand, while the watch records
51 wrist movements. Similarly, contexts such as “shower” or
TABLE 2
The sensors in the dataset. For each sensor, details of the raw
measurements, the number of examples with measurements from that
sensor and the number of users with measurements from that sensor.
“Core” represents examples that have measurements from all six core
sensors that are analyzed in this paper (Acc, Gyro, WAcc, Loc, Aud and
PS). “1pe” means sampled once per example. “var” means variable
sampling rate — gathering updates whenever the value changes.
F1 over many labels, certain labels will unfairly dominate
the score. Additionally, the self-reported data may be noisy,
possibly including cases where a label was actually relevant,
but was not reported by the subject. Precision and F1 will be
too sensitive to such cases. Unlike F1, the balanced accuracy,
BA=0.5*(sensitivity+specificity), does not suffer from these
issues, and can serve as a convenient objective that fairly
balances competing metrics.
First, we assess the potential of single sensors. Figure 3 (A) shows some specific context labels for which
relatively few examples were collected. If we pick our first
(sometimes second) guess of relevant sensor we can achieve
reasonable recognition of these contexts.
Next, to see if we can do better, we evaluate the three
sensor-fusion methods described in “Sensor fusion” and
compare them to single-sensor classifiers. Figure 3 (B) shows
performance for 25 labels from diverse context domains. In
most cases sensor-fusion managed to match the best fitting
single-sensor. The system learned from data how to best
utilize the different sensors, without the need of a human to
guide it, which can be useful for scalable systems, where the
researcher does not necessarily know which sensor to trust
for which label. Furthermore, in many cases sensor-fusion
improved performance, compared to the best single-sensor,
meaning that there is complementary information in different sensors. We see the overall advantage of multi-sensor
systems over single-sensor systems, shown by the average
performance of the different systems in Figure 3 (C). The
three sensor-fusion alternatives seem to perform similarly
well, with LFL slightly ahead. The selection of a sensorfusion method can be guided by the training data available
to the researcher. When having plenty of labeled examples
that have all six sensors available, the simple EF system
can work. Otherwise, late fusion will be more fitting, still
having plenty of data to train each single-sensor classifier
alone. Leave-one-user-out results are consistent with 5-fold
evaluation (figure 3 shows LOO results for the EF system,
marked “EF-LOO”). For some labels, like “running”, the
system benefited from the larger training set in the LOO
evaluation. Full per-label results are provided in supplemental material.
“in a meeting” have unique acoustic signatures (running
water, voices) that allow the audio-based classifier to perform well. When showering, it is reasonable that the phone
will be in a different room than the person, in which
case the watch is an important indicator of the activity.
Figure 4 (A) demonstrates that the LFL method assigns
reasonable weights to the six sensors — sensors that perform
more strongly for a given label are given higher weight.
Investigating where misclassification occurs helps to understand the predictive ability of the system. Figure 4 (B–
G) shows confusion matrices that depict misclassification
rates between related context labels. For example, a classifier
using the phone’s motion sensors (Acc and Gyro) (Figure 4 (B)) to discriminate between body movement/posture
states confuses even dissimilar labels (“running” vs. “lying
down”). Such errors arise in natural, unconstrained behavior; in-the-wild, people do not always carry their phone in
their pocket — subjects were sometimes running on a treadmill with their phone next to them, motionless. The watch
can help in such situations — when the watch accelerometer
features were added to the classifier (Figure 4 (C)), the
confusion between activities was reduced.
The audio signal from the smartphone (Figure 4 (D))
is informative for labels related to the environmental context. We see a hierarchy of misclassification: while there is
some confusion between labels that share similar acoustic
properties (“toilet” vs. “shower”, “class” vs. “meeting”),
there is a sharper distinction between label groups from
different domains (“toilet or shower” vs. “class or meeting”
vs. “restaurant”).
The phone placement itself provides cues about the
user’s activity; when the phone is lying on a table it is
more likely the user is showering than walking to work.
The ability to recognize the phone’s position will improve
overall context recognition. A single modality is not sufficient to fully identify phone position. A classifier based on
motion sensors is sensitive to movement, so when the phone
was in a bag (possibly motionless) it was often mistaken for
being on a table (Figure 4 (E)). On the other hand, a classifier
using audio alone is more sensitive to whether the phone is
enclosed or exposed to environmental sounds, so with this
classifier cases of “phone in bag” were mistaken for “phone
in pocket” and “phone in hand” was often mislabeled as
“phone on table” (Figure 4 (F)). However, by combining motion and audio modalities, the classifier synthesized these
two dimensions of discrimination to better recognize phone
position (Figure 4 (G)). These examples demonstrate the
large variability in behavior in-the-wild and highlight the
utility of fusing multi-modal sensors.
8
A
0.9
Label
Stairs - going up
Stairs - going down
Elevator
Cleaning
Laundry
Washing dishes
Singing
At a party
At the beach
At a bar
examples
399
390
124
1839
473
851
384
404
122
520
Sensor
Gyro
Gyro
Gyro
WAcc
WAcc
WAcc
Aud
Aud
Loc
PS
BA
0.73
0.73
0.76
0.71
0.66
0.70
0.68
0.81
0.72
0.93
Sensor
Acc
Acc
Aud
Gyro
Acc
Aud
Loc
Acc
PS
Gyro
BA
0.70
0.71
0.71
0.64
0.65
0.60
0.61
0.74
0.70
0.66
C
classifier
p99
Acc
Gyro
WAcc
Loc
Aud
PS
EF
LFA
LFL
EF-LOO
accuracy
0.50
0.73
0.70
0.73
0.71
0.75
0.76
0.87
0.84
0.85
0.86
sensitivity
0.50
0.64
0.64
0.67
0.63
0.65
0.74
0.67
0.76
0.76
0.69
specificity
0.50
0.73
0.69
0.72
0.70
0.75
0.76
0.87
0.83
0.85
0.86
BA
0.50
0.68
0.66
0.70
0.67
0.70
0.75
0.77
0.80
0.80
0.78
precision
0.11
0.17
0.16
0.18
0.17
0.18
0.20
0.24
0.23
0.24
0.24
F1
0.13
0.22
0.20
0.22
0.22
0.22
0.24
0.30
0.29
0.30
0.30
B
Balanced accuracy
0.8
0.7
0.6
0.5
Ly
in
(54g do
35 wn
9)
(82 Sitt
90 ing
4
W )
(11 alk
89 ing
2
Ru )
n
(67ning
Bic 5)
(35yclin
23 g
S )
(42leep
92 ing
La 0)
b
(28 wo
98 rk
In )
(28 cla
In 72 ss
am )
At
ma (2 eetin
in 904 g
wo )
(20rkpla
38 ce
2)
(10 Ind
79 oor
44 s
O )
(76utsid
29 e
In )
(36 a c
35 ar
Dr
ive On a )
(
( 1 b
Dr I'm t 185 us
ive
h )
(I'm (e dri
a p 5034ver)
ass )
(16enge
55 r)
At )
At (839hom
a r 77 e
est )
Ph (1 aura
on 32 nt
e i 0)
n
(15poc
30 ket
1
Ex )
(53ercis
84 e
C )
(22ookin
5 g
Sh 7)
op
(89ping
6)
S
t
Dr
ink (4rollin
ing 34 g
( )
Ba alcoh
thi (86 ol
ng 4) )
-s
(11how
86 er
)
0.4
Fig. 3. Overall performance of the single-sensor classifiers (Acc, Gyro, WAcc, Loc, Aud and PS) and the sensor-fusion classifiers (EF, LFA, LFL and
EF-LOO). (A) Specific labels that had few examples with first and second guess of sensors that intuitively seem relevant and the BA score of the
corresponding single-sensor classifiers. (B) BA scores for selected labels from diverse domains (number of examples in parenthesis). Color legend
is in table (C). (C) Average performance metrics over the 25 context labels from B. All average scores were well above the p99 value, which marks
the 99th percentile of random score — scores above the p99 value have less than 1% probability of being achieved randomly (p99 was estimated
from 100 random simulations).
U SER PERSONALIZATION
People move, behave and uses their phone in different
manners. A system that is fine-tuned to its specific user may
outperform a more general model. To explore the potential
of personalization we performed experiments with a single
test user. We compared three models: (1) universal (trained
on data from other users), (2) individual (trained on half of
the data from the same test user) and (3) adapted (merges
both). We tested the three models on the same unseen data.
Figure 5 shows the results of these experiments. The universal model demonstrates good performance. This shows
the basic ability of a trained system to work well for a new
unseen user. As suspected, the individual model performed
better than the universal model for labels that had many
individual examples (“lying down”, “sitting”, “sleeping”,
“at home”, “computer work”, and “at main workplace”).
However, the individual user is missing data for many context labels. For other labels there are only a limited number
of examples a new user can acquire in a few “training” days,
which risks over-fitting to these few examples. In such cases
a universal model is better, having been trained on plenty
of data from many users. The optimal solution is to benefit
from both universal and individual data: the user-adapted
model shows overall improvement in recognition performance, even among the labels that had over 300 examples
for the test user. LFA is a simple heuristic that manages to
demonstrate this advantage. For each label, when there is
not enough data to train an individual model, the adapted
model relies only on the universal model. When there is
enough data to train an individual model, the adapted
model “listens” to the universal and individual models, in
some cases achieving better performance than either model
on its own (e.g. “sleeping”, “at home”).
In practical systems, the logistics of implementing personalization may not be an obvious task. For medical applications, the clinician or patient may decide that their cause
is important and worth dedicating some effort to provide
individual labeled data for a few days, in order to better
adapt the model. However, in commercial applications the
users (clients) may not be motivated to invest the extra
effort in labeling. In such cases semi-supervised methods
can still be used to make the most of unlabeled data from
the individual user and personalize the model.
C ONCLUSIONS
Our novel data collection brings about behavioral variability
in-the-wild that is underrepresented in controlled studies.
This makes context recognition a harder challenge compared to previous scenarios, hence accuracy levels in-thewild are lower than those reported in experiments that had
some restrictions on behavioral conditions. We demonstrate
that everyday devices, in their normal unconstrained usage,
carry information about the person’s natural behavioral
context. We describe a baseline system, suggest three simple
methods for sensor fusion, and reinforce previous findings
that showed the advantage of fusing multi-modal sensors.
We demonstrate how the sensing modalities complement
9
In a meeting
In a car
Acc
Gyro
WAcc
Loc
Aud
PS
Acc
Gyro
WAcc
Loc
Aud
PS
At main workplace
Acc
Gyro
WAcc
Loc
Aud
PS
Sleeping
6
Acc
Gyro
WAcc
Loc
Aud
PS
Bicycling
Acc
Gyro
WAcc
Loc
Aud
PS
A
Bathing - shower
learned
weight
4
2
0
BA
0.9
0.7
C
0.44
Bicycling
F
Phone in pocket
0.4
0.54
0.2
0.0
nb
ag
e
Ph in h
on
e i and
n
Ph poc
on ke
eo t
nt
ab
le
0.74
ei
on
on
ei
nb
Ph e in h
on
e i and
n
Ph poc
on ke
eo t
nt
ab
le
ag
Phone on table
0.6
0.35
on
0.62
Phone on table
on
0.78
Phone in hand
0.43
Phone in pocket
Ph
Ph
on
ei
on
Ph
0.52
0.8
Phone in bag 0.34
0.23
Phone in hand
nb
ag
e
i
Ph n h
on
e i and
n
Ph poc
on ke
eo t
nt
ab
le
Phone on table
0.38
Ph
Phone in pocket
G
Phone in bag 0.38
Ph
Phone in bag 0.16
0.68
At a restaurant
Ph
E
0.52
In a meeting
0.66
do
w
Sit n
t
Wa ing
lkin
Ru g
nn
Bic ing
ycl
ing
Ly
ing
do
w
Sit n
tin
Wa g
lk
Ru ing
nn
Bic ing
ycl
ing
Bicycling
0.44
In class
0.23
Running
0.50
0.39
Bathing - shower
0.56
Walking
0.19
Running
Phone in hand
0.67
Sitting
0.51
Walking
Toilet 0.42
Ly
ing
Sitting
D
Lying down 0.82
Ba
thi T
ng oil
- s et
ho
we
r
I
In n cla
a
s
At me s
a r eti
est ng
au
ran
t
B
Lying down 0.79
Acc
Gyro
WAcc
Loc
Aud
PS
0.5
Fig. 4. Why sensor fusion helps recognition. (A) The bottom row shows the overall performance (BA) of each single-sensor classifier and the top
row shows the weights that the LFL classifier learned to assign to each sensor (taken from the first cross validation fold). (B–G) Confusion matrices
for subsets of mutually exclusive labels. Rows represent ground truth labels and columns represent predicted labels. Rows are normalized so that
a cell in row i and column j displays the proportion of examples of class i that were assigned to class j . The correct classification rates (main
diagonal) are also marked numerically. The sensors used are: (B) Acc and Gyro; (C) Acc, Gyro and WAcc; (D) Aud; (E) Acc, Gyro and WAcc; (F)
Aud; (G) Acc, Gyro, WAcc and Aud.
0.86
0.74
0.87
0.84
0.83
0.87
1.0
0.9
0.8
0.7
0.6
0.5
0.4
1.0
0.8
0.6
0.4
0.2
0.0
0.56
0.50
0.60
0.67
0.68
0.72
15
l
rag abels
)
e(
11
lab
els
)
rag
e(
wo
Av
e
oo
ch
At
s
ma
in
At
Av
e
l (3
rkp 189)
lac
e(
20
76
On
)
ab
us
(95
In
)
ac
ar
(30
2)
29
6)
)
g(
tin
ee
am
In
In
c
las
s(
41
8
)
)
ute
rw
ork
(10
17
(51
26
2)
Co
mp
At
ho
me
t (9
)
92
ran
tau
ho
l) (
At
a
res
co
(al
ing
Dr
ink
Ind
oo
rs
(74
62
)
20
)
(34
7)
65
ing
g(
ep
Sle
lkin
(47
Wa
tin
g
(34
Sit
wn
do
Ly
ing
43
)
Universal
Individual
Adapted
90
)
F1
BA
10
Fig. 5. User adaptation performance. Balanced accuracy and F1 score were assessed for a single user. The x-axis denotes the labels with the total
number of examples for the user. The “universal” model was trained on data from other users, the “individual” model was trained on data from the
same user and the “adapted” model combines the universal and individual models using LFA. The bars on the right hand side of each plot present
the scores averaged over the 15 tested labels, and averaged over 11 labels that had over 300 examples.
each other, and help resolve contexts that arise with uncontrolled behavior (e.g. running on treadmill with phone on
table, motionless).
Combinatorial representation of behavior is very flexible. A well trained system has the potential to correctly
recognize a new specific situation (combination of labels)
that did not appear in the training. To broaden the range of
contexts, researchers can either use supervised methods and
focus on newly added target labels when collecting extra
data, or use unsupervised methods to discover complex
behaviors as common combinations or sequences of simpler
contexts [25]. The labels in our work were interpreted in
a subjective manner. The same location may be considered
as “school” for one subject and “workplace” for another.
We did not tell the subjects how we define “walking” or
“eating” in order to capture the full scope of what people
consider eating. Domain-expert researchers may decide to
define labels clearly to subjects or use more specific labels
like “eating a meal” and “snacking”.
F UTURE DIRECTIONS
New technologies and original solutions for collecting labels
in-the-wild are required to reduce annotation load from
study subjects and increase reliability of labeling. Online
learning can be used to keep improving real-time recognition, which will require less label-correcting effort from new
research subjects. Active learning can be utilized to collect
data in scale, while sparsely probing subjects to provide
annotations. In parallel, semi-supervised methods can be
used to make the best out of plenty unlabeled data (which
is easy to collect) and reduce the dependence on labeled
examples to a minimum.
The public dataset we collected provides a platform to develop and evaluate these methods, as well as explore feature
extraction, inter-label interaction, time series modeling and
other directions that will improve context recognition.
R EFERENCES
K. Servick, “Mind the phone,” Science, vol. 350, no. 6266, pp. 1306–
1309, 2015.
[2] M. Basner, K. M. Fomberstein, F. M. Razavi, S. Banks, J. H. William,
R. R. Rosa, and D. F. Dinges, “American time use survey: sleep
time and its relationship to waking activities,” Sleep, vol. 30, no. 9,
p. 1085, 2007.
[3] I. Nahum-Shani, S. N. Smith, A. Tewari, K. Witkiewitz, L. M.
Collins, B. Spring, and S. Murphy, “Just in time adaptive interventions (jitais): An organizing framework for ongoing health
behavior support,” Methodology Center technical report, no. 14-126,
2014.
[4] S. Intille, “The precision medicine initiative and pervasive health
research,” IEEE Pervasive Computing, vol. 15, no. 1, pp. 88–91, 2016.
[5] M. L. Lee and A. K. Dey, “Sensor-based observations of daily
living for aging in place,” Personal and Ubiquitous Computing,
vol. 19, no. 1, pp. 27–43, 2015.
[6] A. K. Dey, K. Wac, D. Ferreira, K. Tassini, J.-H. Hong, and J. Ramos,
“Getting closer: an empirical investigation of the proximity of
user to their smart phones,” in Proceedings of the 13th international
conference on Ubiquitous computing. ACM, 2011, pp. 163–172.
[7] J. J. Guiry, P. van de Ven, and J. Nelson, “Multi-sensor fusion
for enhanced contextual awareness of everyday activities with
ubiquitous devices,” Sensors, vol. 14, pp. 5687–5701, 2014.
[8] M. Shoaib, S. Bosch, H. Scholten, P. J. Havinga, and O. D. Incel,
“Towards detection of bad habits by fusing smartphone and
smartwatch sensors,” in Pervasive Computing and Communication
Workshops (PerCom Workshops), 2015 IEEE International Conference
on. IEEE, 2015, pp. 591–596.
[9] J. Kerr, R. E. Patterson, K. Ellis, S. Godbole, E. Johnson, G. Lanckriet, and J. Staudenmayer, “Objective assessment of physical activity: Classifiers for public health.” Medicine and science in sports and
exercise, vol. 48, no. 5, pp. 951–957, 2016.
[10] K. Kunze and P. Lukowicz, “Sensor placement variations in wearable activity recognition,” IEEE Pervasive Computing, vol. 13, no. 4,
pp. 32–41, 2014.
[11] R. K. Ganti, S. Srinivasan, and A. Gacic, “Multisensor fusion
in smartphones for lifestyle monitoring,” in 2010 International
Conference on Body Sensor Networks. IEEE, 2010, pp. 36–43.
[12] H. Pirsiavash and D. Ramanan, “Detecting activities of daily
living in first-person camera views,” in Computer Vision and Pattern
Recognition (CVPR), 2012 IEEE Conference on. IEEE, 2012, pp. 2847–
2854.
[1]
11
[13] M. Han, Y.-K. Lee, S. Lee et al., “Comprehensive context recognizer
based on multimodal sensors in a smartphone,” Sensors, vol. 12,
no. 9, pp. 12 588–12 605, 2012.
[14] F. J. Ordóñez, P. de Toledo, and A. Sanchis, “Activity recognition
using hybrid generative/discriminative models on home environments using binary sensors,” Sensors, vol. 13, no. 5, pp. 5460–5477,
2013.
[15] Y. Dong, J. Scisco, M. Wilson, E. Muth, and A. Hoover, “Detecting
periods of eating during free-living by tracking wrist motion,”
IEEE journal of biomedical and health informatics, vol. 18, no. 4, pp.
1253–1260, 2014.
[16] T. Rahman, M. Zhang, S. Voida, and T. Choudhury, “Towards
accurate non-intrusive recollection of stress levels using mobile
sensing and contextual recall,” in International Conference on Pervasive Computing Technologies for Healthcare, 2014.
[17] T. Choudhury, S. Consolvo, B. Harrison, J. Hightower, A. LaMarca,
L. LeGrand, A. Rahimi, A. Rea, G. Bordello, B. Hemingway et al.,
“The mobile sensing platform: An embedded activity recognition
system,” IEEE Pervasive Computing, vol. 7, no. 2, 2008.
[18] S. Consolvo, D. W. McDonald, T. Toscos, M. Y. Chen, J. Froehlich,
B. Harrison, P. Klasnja, A. LaMarca, L. LeGrand, R. Libby et al.,
“Activity sensing in the wild: a field trial of ubifit garden,” in
Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems. ACM, 2008, pp. 1797–1806.
[19] S. Hemminki, P. Nurmi, and S. Tarkoma, “Accelerometer-based
transportation mode detection on smartphones,” in Proceedings of
the 11th ACM Conference on Embedded Networked Sensor Systems.
ACM, 2013, p. 13.
[20] A. M. Khan, A. Tufail, A. M. Khattak, and T. H. Laine, “Activity recognition on smartphones via sensor-fusion and kda-based
svms,” International Journal of Distributed Sensor Networks, vol. 2014,
2014.
[21] K. Yatani and K. N. Truong, “Bodyscope: a wearable acoustic
sensor for activity recognition,” in Proceedings of the 2012 ACM
Conference on Ubiquitous Computing. ACM, 2012, pp. 341–350.
[22] K. Ellis, S. Godbole, J. Kerr, and G. Lanckriet, “Multi-sensor
physical activity recognition in free-living,” in Proceedings of the
2014 ACM International Joint Conference on Pervasive and Ubiquitous
Computing: Adjunct Publication. ACM, 2014, pp. 431–440.
[23] M. Rossi, G. Troster, and O. Amft, “Recognizing daily life context
using web-collected audio data,” in Wearable Computers (ISWC),
2012 16th International Symposium on. IEEE, 2012, pp. 25–28.
[24] A. Stisen, H. Blunck, S. Bhattacharya, T. S. Prentow, M. B. Kjærgaard, A. Dey, T. Sonne, and M. M. Jensen, “Smart devices are
different: Assessing and mitigatingmobile sensing heterogeneities
for activity recognition,” in Proceedings of the 13th ACM Conference
on Embedded Networked Sensor Systems. ACM, 2015, pp. 127–140.
[25] J. Seiter, O. Amft, M. Rossi, and G. Tröster, “Discovery of activity composites using topic models: An analysis of unsupervised
methods,” Pervasive and Mobile Computing, vol. 15, pp. 215–227,
2014.
Yonatan Vaizman Yonatan Vaizman is a Ph.D.
candidate in the department of Electrical and
Computer Engineering at UC San Diego. His
fields of interest are machine learning, signal
processing, and artificial intelligence. In his research, he applies methods from these fields
to music recommendation, mobile sensor processing, and human behavior recognition. He received his B.S. degree in Computer Science and
Computational Biology from the Hebrew University in Jerusalem, Israel, and his M.S. degree in
Electrical Engineering from UC San Diego.
Katherine Ellis Katherine Ellis is a Research
Scientist at Amazon. Her research interests are
in applications of machine learning to physical
activity measurement, mobile health, social network analysis and music recommendation. She
received a B.S. degree in electrical engineering
from the University of Southern California and
M.S. and Ph.D. degrees in electrical engineering
from UC San Diego. She is an IEEE member and
an ACM member.
Gert Lanckriet Gert Lanckriet is a Principal
Applied Scientist at Amazon, and Professor of
Electrical and Computer Engineering at UC San
Diego. His interests are in data science, on
the interplay between machine learning, applied
statistics, and large-scale optimization, with applications to music and video search and recommendation, multimedia, and personalized, mobile health.
He was awarded the SIAM Optimization Prize
in 2008 and is the recipient of a Hellman Fellowship, an IBM Faculty Award, an NSF CAREER Award, and an Alfred
P. Sloan Foundation Research Fellowship. In 2011, MIT Technology
Review named him one of the 35 top young technology innovators in
the world (TR35). In 2014, he received the Best Ten-Year Paper Award
at the International Conference on Machine Learning.
He received a masters degree in electrical engineering from the
Katholieke Universiteit Leuven, Belgium, and M.S. and Ph.D. degrees
in electrical engineering and computer science from UC Berkeley. He is
a senior IEEE member and an ACM member.
12
S UPPLEMENTARY MATERIAL
Supplementary material has technical details about the following components of the work:
•
•
•
•
•
•
•
•
•
Mobile app
Data collection procedure
Sensor measurements
Extracted features
Label processing
Classification methods
Performance evaluation
User personalization assessment
Detailed results tables
•
•
Mobile app
For the purpose of data collection in a large scale we developed a mobile application called ExtraSensory, with versions
for both iPhone and Android smartphones, and a companion application for the Pebble smartwatch that integrates
with both. The app was used for supervised data collection,
meaning it collects both sensor measurements and ground
truth context labels. The app is scheduled every minute
to automatically record measurements for 20 seconds from
the sensors. Sensors are sampled in frequencies appropriate
for their domain, and include motion-responsive sensors,
location services, audio, environment sensors, as well as bits
of information about the phone’s state. When the watch
is available (within Bluetooth range and paired with the
phone) measurements from the watch are also collected by
the app during the 20 second recording session. More details
about the sensors are provided in “Sensor measurements”.
At the end of the 20 second recording session the measurements are bundled in a zip file and sent through the
internet (if a WiFi network is available) to our lab’s server,
which runs a quick calculation and replies with an initial
prediction of the activity (e.g. sitting, walking, running). All
communication between the app and the server is secure
and encrypted, and identified only by a unique universal
identifier (UUID) that was randomly generated for each
user.
In addition to collecting sensor measurements, the app’s
interface provides several mechanisms for the user to report
labels about their context. This was a crucial part of the
research design and we had to overcome a basic trade-off:
on one hand we wanted to collect ground truth labels for as
many examples (minutes) as possible and with much detail
(combination of all the relevant context labels). On the other
hand we did not want the subject to interact with the app
every minute to report labels, both because it would be an
extreme inconvenience for the subject and because it would
impact the natural behavior of the subject and miss the
point of collecting data in-the-wild. To balance this tradeoff, we designed a flexible interface that helps minimize
the user-app interaction time. The following label-reporting
mechanisms were included:
•
A history journal presents the user’s activities
chronologically as a calendar and enables the user
to easily edit the context labels of time ranges in the
past (up to one day in the past). The user can easily
merge a sequence of minutes to a single “event” with
•
•
the same context labels, or split a calendar event to
describe a change in context. See Figure 2 (A). The
real-time predictions from the server assist the user
to recall when their activity changed — consecutive
minutes with the same prediction from the server are
merged to a single item on the history calendar.
The user can also initiate active feedback by reporting labels describing their context in the immediate
future (starting “now” for up to half an hour in the
future). See Figure 2 (B).
Every x minutes (by default, x is 10 minutes, but can
be set by the user) the app presents a notification
to remind the user to provide labels. If the the user
has recently provided labels, the notification asks
whether the user was still engaged in the same
activities — allowing for a quick and easy response
if the answer is “yes”. See Figure 2 (C).
The notifications also appear on the smartwatch,
allowing for an easier response with a click of a
button on the watch, without using the phone itself.
When selecting labels from the menu, a side-bar index allows quick search of the relevant labels, either
by categories (e.g. sports, work, company) or through
a “frequently used labels” menu, which presents
labels that the user has applied in the past. The
category in which a label was presented in the menu
does not matter, and a label can appear under different categories (e.g. “skateboarding” appears under
“sports”, “leisure” and “transportation”) — the only
reason for these categories is to make it easy for the
user to find the relevant label quickly.
Data collection procedure
The study’s research plan and consent form were approved
by the university’s institutional review board (IRB). Human subjects were recruited for the study via fliers across
campus, university mailing lists and word of mouth. Every
subject read and signed the consent form. The researchers
installed the app on each subject’s personal phone (to
maximize authenticity of natural behavior). The subject
then engaged in their usual behaviors for approximately a
week, while keeping the app running in the background
on their phone as much as was possible and convenient.
The subject was asked to report as many labels as possible
without interfering too much with their natural behavior.
Subjects varied in their level of rigorousness with respect
to providing labels: some wanted to be very precise (with
specific detailed combinations of labels, and trying to keep
minute-to-minute precision) and others tended to be less
specific and to dedicate less effort. The subjects who used
the watch, which we supplied them with, were told that it
is fine to get it wet (wash hands, shower) but not submerge
it (swimming). They were also asked to turn off the watch
app whenever they removed the watch from their wrist and
to turn it back on when they wore the watch — so we can
generally assume that whenever watch measurements are
available they were taken from the subject’s wrist.
Using the app consumes the phone’s battery more
quickly than normal. To make up for this, the researchers
provided participants with an external portable battery,
13
which provides one extra charge during the day. The researchers also provided the subject with the Pebble smartwatch (56 of the subjects agreed to use the watch). The
external battery and the smartwatch were returned at the
end of the study. Each subject was compensated for their
participation. The basic compensation was in the amount
of $40, and an additional amount was calculated based on
the amount of labeled data that the subject contributed (as
an incentive to encourage reporting many labels). The total
compensation per subject was between $40 and $75.
Technical difficulties
During the development of the iPhone app, there were
releases of new iOS versions that caused the app to not work
well and required us to change the code.
Since subjects used their personal phones, the app had to
handle different devices, and in the Android case, different
makers. For some of the Android users when we installed
the app we noticed it didn’t work well. In three cases the
workaround was to install a slightly different version of
the app that didn’t use the gyroscope. After installing the
changed app and making sure it works those users began
collecting records (without gyroscope measurements).
On top of the dataset’s 60 subjects, there were four more
subjects that participated and received the basic compensation, but whose data was not included in the dataset. For
two of them the app didn’t work well on their devices. The
other two were too stressed or otherwise occupied during
participation, and produced too little and un-reliable labels,
so we decided to discard their data.
Sensor measurements
Raw sensor measurements are provided in the publicly
available dataset.
High frequency measurements:
Each sensor (and pseudo-sensor) in the following list was
sampled at 40Hz during the ∼20 second recording session
to produce a time series of ∼800 time points. The sampling
relies on the design of the phone’s hardware and operating system and the sampling rate was not guaranteed to
be accurate (especially for the Android devices). For that
reason the time reference of each sample in a time series
was also recorded; the differences between consecutive time
references were approximately 25 milliseconds.
•
•
•
•
Accelerometer. Time series of 3-axis vectors of acceleration according to standard axes of phone devices.
Gyroscope. Time series of 3-axis vectors of rotation
rate around each of the phone’s 3 axes.
Magnetometer. Time series of 3-axis vectors of the
magnetic field.
Processed signals. Both iPhone and Android operating systems provide processed versions of the
signals: The raw acceleration is split to the gravity
acceleration (estimated direction of gravity at every
moment, the magnitude is always 1G) and the usergenerated acceleration (subtraction of the gravity signal from the raw acceleration). For the gyroscope the
OS has a calibrated version that attempts to remove
drift effects. For the magnetometer the OS has an
unbiased version that subtracts the estimated bias of
the magnetic field created by the device itself.
In this paper we used the raw acceleration signal (which
includes the effects of gravity) and the calibrated version of
the gyroscope signal. Acceleration is reported in units of G
(gravitational acceleration on the surface of the Earth) on
iPhone and in units of m/s2 on Android. Before extracting
features we converted the Android acceleration measurements to units of G.
Watch measurements: From the Pebble smartwatch
we collected signals from the two available sensors—
accelerometer and compass. Acceleration was sampled at
25Hz and describes a 3-axis vector of acceleration (in units
of mG) relative to the watch’s axes-system. The compass
does not have a constant sampling rate; it was requested
to provide an update of the heading whenever a change
of more than one degree was detected. The compass takes
some time to calibrate before providing measurements, so
some examples that have watch acceleration measurements
are missing compass measurements.
Location measurements: Both iPhone and Android provide
location services (based on a combination of GPS, WiFi and
cellular communications). The app samples location data in
a non-constant rate, as the location service updates each
time a movement is detected. This creates a time series
of varying length (sometimes just a single time point in a
recording session, sometimes more than 20 points) of location updates. Each update has a relative time reference and
the estimated location measurements: latitude coordinate,
longitude coordinate, altitude, speed, vertical accuracy and
horizontal accuracy (these accuracies describe the range of
reasonable error in location). Some of these values may be
missing at times (e.g. when the phone is in a place with weak
signals). In addition to the time series of location updates,
the app calculates on the phone some basic heuristic location
features: standard deviation of latitude values, standard
deviation of longitude values, total change of latitude (last
value minus first value), total change of longitude, average
absolute latitude derivative and average absolute longitude
derivative (as proxy to the speed of the user).
To protect our study subjects’ privacy (collected examples with label “at home” that also include the exact location
coordinates may reveal the subject’s identity) the app has an
option to select a location (typically their home) they would
like to disguise. For the subjects that opted to use this option, whenever they were within 500 meters of their chosen
location, the app would not send the latitude and longitude
coordinates from the current recording (but it would send
the other estimated location values such as altitude, speed,
as well as the basic heuristic location features).
Low frequency measurements: These measurements were
sampled just once in a recording session (approximately
once per minute). Some of them describe the phone state
(PS): app state (foreground/background), WiFi connectivity
status, battery status (charging, discharging), battery level,
or phone call status. Other low frequency measurements
are taken from sensors built in to the phone, if available:
proximity sensor, ambient light, temperature, humidity, air
pressure.
Audio data: Audio was recorded from the phone’s mi-
14
crophone in 22,050 Hz for the duration of the recording
session (∼20 seconds). Audio was not available for recording when the phone was being used for a phone call. In
order to maintain the privacy of the subjects, the raw audio
recording was not sent to the server. Instead, standard audio
processing features (Mel Frequency Cepstral Coefficients
(MFCC)) were calculated on the phone itself and only the
features were sent to the server. The MFCCs were calculated
for half-overlapping windows of 2048 samples, based on
40 Mel scaled frequency bands and 13 cepstral coefficients
(including the 0th coefficient). As part of the preprocessing
of the recorded audio the raw audio signal was normalized
to have maximal magnitude of 1 (dividing by the maximum
absolute value of the sound wave). This normalizing factor
is also sent as a measurement separately from the calculated
MFCC features.
Extracted features
For the experiments in this work we focused on six sensors:
accelerometer, gyroscope, watch accelerometer, location, audio and phone state. Other sensors’ measurements are available in the public dataset. Every sensor measures different
physical or virtual properties and has a different form of
raw measurements. Therefore we designed specific features
for each sensor. The published dataset includes files with
these features pre-computed for all the users.
Accelerometer and Gyroscope (26 features each): Since in
natural behavior the phone’s position is not controlled we
cannot assume it is oriented in a particular way, and it
also may be changing its axes-system with respect to the
ground (and with respect to the person). For that reason we
had little reason to assume that any of the phone’s axes
will have a particular coherent correspondence to many
behavioral patterns, and we extracted most of the features
based on the overall magnitude of the signal. We calculated
the vector magnitude signal as the euclidean norm of the
3-axis q
acceleration measurement at each point in time, i.e.,
2
2
2
a[t] = ax [t] + ay [t] + az [t] . We extracted the following
features:
•
•
•
Nine statistics of the magnitude signal: mean, standard deviation, third moment, fourth moment, 25th
percentile, 50th percentile, 75th percentile, valueentropy (entropy calculated from a histogram of
quantization of the magnitude values to 20 bins) and
time-entropy (entropy calculated from normalizing
the magnitude signal and treating it as a probability
distribution, which is designed to detect peakiness in
time—sudden bursts of magnitude).
Six spectral features of the magnitude signal: log
energies in 5 sub-bands (0–0.5Hz, 0.5–1Hz, 1–3Hz,
3–5Hz, >5Hz) and spectral entropy.
Two autocorrelation features from the magnitude
signal. The average of the magnitude signal (DC
component) was subtracted and the autocorrelation
function was computed and normalized such that
the autocorrelation value at lag 0 will be 1. The
highest value after the main lobe was located. The
corresponding period (in seconds) was calculated as
the dominant periodicity and its normalized autocorrelation value was also extracted.
•
Nine statistics of the 3-axis time series: the mean and
standard deviation of each axis and the 3 inter-axis
correlation coefficients.
Watch accelerometer (46 features): From the watch acceleration we extracted the same 26 features as from the phone
accelerometer or gyroscope. Since the watch is positioned
in a more controlled way than the phone (it is firmly fixed
to the wrist), its axes have a strong meaning (e.g. motion
along the x-axis of the watch describes a different kind
of movement than motion along the z-axis of the watch).
Hence we added 15 more axis-specific features—log energies in the same 5 sub-bands as before, this time calculated
for each axis’ signal separately. In addition, to account for
the changes in watch orientation during the recording we
calculated 5 relative-direction features in the following way:
we first calculate the cosine-similarity between the acceleration directions of any two time points in the time series
(value of 1 meaning same direction, value of -1 meaning
opposite directions and value of 0 meaning orthogonal
directions). Then we averaged these cosine similarity values
in 5 different ranges of time-lag between the compared time
points (0–0.5sec, 0.5–1sec, 1–5sec, 5–10sec, >10sec).
Location (17 features): In this work we used location
features that were based only on relative locations, and
not on absolute latitude/longitude coordinates. This was
in order to avoid over-fitting to our location-homogeneous
training set that will not generalize well to the outside
world (e.g., mistakenly learning that a specific location in
the UCSD campus always corresponds to “at work”). Six
features were calculated on the phone — this was in order
to have some basic location features in cases where the
subjects opted to hide their absolute location. These quick
features included standard deviation of latitude, standard
deviation of longitude, change in latitude (last value minus
first value), change in longitude, average absolute value
of derivative of latitude and average absolute value of
derivative of longitude.
The transmitted location measurements were further
processed to extract the following 11 features: number of
updates (indicating how much the location changed during
the 20 second recording), log of latitude-range (if latitudes
were transmitted), log of longitude-range (if longitudes
were transmitted), minimum altitude, maximum altitude,
minimum speed, maximum speed, best (lowest) vertical
accuracy, best (lowest) horizontal accuracy and diameter
(maximum distance between two locations in the recording
session, in meters).
Audio (26 features): From the time series of 13dimensional MFCC vectors (typically around 400 time
frames) we calculated the average and standard deviation
of each of the 13 coefficients.
Phone State (34 features): For this work we used only
the discrete phone state measurements. We represented
them with a 26-dimensional one-hot representation—for
each property we created a binary indicator for each of the
possible values the property can take, plus one indicator
denoting missing data. This representation is a redundant
coding of the phone state, but it facilitates the use of simple,
linear classifiers over this long binary vector representation.
The keys were: app state (3 options: active, inactive, back-
15
ground), battery plugged (3 options: AC, USB, wireless),
battery state (6 options: unknown, unplugged, not charging,
discharging, charging, full), in a phone call (2 options: false,
true), ringer mode (3 options: normal, silent no vibrate,
silent with vibrate) and WiFi status (3 options: not reachable,
reachable via WiFi, reachable via WWAN).
In addition, we added a set of features indicating timeof-day information. We used the timestamp of every example and (using San Diego local time) extracted the hour
component (one out of 24 discrete values). In order to get a
flexible, useful representation we defined 8 half-overlapping
time ranges: midnight-6am, 3am-9am, 6am-midday, 9am3pm, midday-6pm, 3pm-9pm, 6pm-midnight and 9pm-3am.
Then we represented each example’s hour with an 8-bit
binary value, where 2 bins will be active for 1 relevant time
range.
•
Label processing
Since the labels are obtained by subjects self-reporting their
own behaviors, the reliability of annotation is not perfect.
In some cases, this was the result of the subject reporting
labels some time after the activity had occurred and misremembering the exact time. More common are cases where
the subject neglected to report labels when relevant activities
occurred (perhaps because the subject was distracted, did
not have time to specify all the relevant labels, or was not
aware of another relevant label in the vocabulary). As part
of cleaning the data, we created adjusted versions for some
labels using two methods: based on location data and based
on other labels.
Location adjusted labels. We collected absolute location
coordinates of the examples that had location measurements
(selecting the location update with best horizontal accuracy
from each example) and visualized them on a map. This
made it easier to correct some labels which were clearly
reported incorrectly. In examples without location data the
original label was maintained.
•
•
“At the beach”. According to the few examples that
reported being at the beach we marked areas that
should be regarded as beach (and manually verified
their validity by viewing them on a map). We then
adjusted the label by applying “At the beach” to any
example with a location within these areas.
“At home”. For each subject we identified the location of their home (by visualizing on a map all
locations of examples where the subject reported
being at home) and marked the coordinates of a
visual centroid. This was only done when it was
clear that we had indeed identified a location of
a home. Three subjects reported being at home in
two different houses, in which case we marked the
two locations as locations of home. Two subjects
never reported being at home but it was clear from
their location to identify their location of residence.
Some subjects had none or very few examples of “at
home” with location data, so no home location was
noted and their original reported labels were used.
To define the adjusted version of the label “at home”,
whenever a subject’s location was within 15 meters
of their marked home location (or either of the two
marked home locations), the adjusted value was set
to “true”; whenever a subject’s location was farther
than 100 meters from all the subject’s marked home
locations the adjusted value was set to “false”. In
other cases (when the location was between 15 and
100 meters from a home location, or when there was
no location data available) we retained the subject’s
originally reported value for “at home”. This adjustment removed some obviously false reports of “at
home” (e.g., when the subject was clearly on a drive
on a freeway). The adjustment manifested mostly
by adding the missing label “at home” to many
examples where the subject was clearly at home but
failed to report it.
“At main workplace”. Similarly to home label we
identified for each subject (if they used the original
label “at work”) the centroid location of their main
workplace and created a new label — “at main workplace” — in a similar way. Some subjects reported
being at work in different locations, so the original
label “at work” is still valid for analysis and may
have a different meaning than “at main workplace”
(which was designed to capture behavioral patterns
typical to the most common place that a person
works in). This adjustment removed some examples
where the label “At work” was probably incorrectly
reported, but more importantly, it added the missing
label in cases where the subject was clearly present
at their most common workplace.
Labels corrected using other labels. We used reported values
of other labels to adjust some labels. In a few cases it was
clear that the reported labels were mistakes (because the
combination of labels was unreasonable). In other cases
a relevant label was simply not reported, even though it
clearly should be relevant according to the other reported
labels.
•
•
“Walking”. We identified a few cases where subjects reported walking together with labels related to
driving. In cases where location data was available,
it was clear on the map that the correct activity
was the drive and not the walk. In the adjusted
version of “walking” we changed the value to “false”
whenever the subject reported “on a bus”, “in a car”,
“drive (I’m the driver)”, “drive (I’m a passenger)”,
“motorbike”, “skateboarding” or “at the pool”.
“Running”. The adjusted version was set to “false”
for the same activities as in the adjusted “walking” label, plus in cases where the subject reported
“playing baseball” or “playing frisbee”. Although
these cases are likely still valid (because the subject
decided to report they were running during these
playing activities), we decided to create the adjusted
“running” version to represent a more coherent running activity (assuming that the playing activities
involve a mixture of running, walking and standing intermittently). While the adjusted versions of
“walking” and “running” may have a few misses
(e.g., some minutes during a baseball game when
the subject was purely running), these misses don’t
harm the integrity of the multi-class experiments,
16
•
•
•
•
which used only examples that had positive labels
of “running”, “walking”, “sitting”, etc..
“Exercise”. The adjusted version was set to “true”
whenever the subject reported “exercising”, “running”, “bicycling”, “lifting weights”, “elliptical machine”, “treadmill”, “stationary bike” or “at the
gym”. This adjustment boosted the representation
of the exercise behavior and also took advantage of
reported specific activities without enough examples
to be analyzed on their own.
“Indoors”. The adjusted version was set to “true”
whenever the subject reported “indoors”, “sleeping”,
“toilet”, “bathing — bath”, “bathing — shower”,
“in class”, “at home”, “at a bar”, “at the gym” or
“elevator”. It is reasonable that many subjects simply
did not bother to report being indoors every time
they did an activity indoors.
“Outside”. The adjusted version was set to “true”
whenever the subject reported “outside”, “skateboarding”, “playing baseball”, “playing frisbee”,
“gardening”, “raking leaves”, “strolling”, “hiking”,
“at the beach”, “at sea” or “motorbike”.
“At a restaurant”. In the adjusted version we
changed the value to “false” whenever the subject
reported “on a bus”, “in a car”, “drive (I’m the
driver)”, “drive (I’m a passenger)” or “motorbike”.
Classification methods
Our system uses binary logistic regression classifiers (with
a fitted intercept). Logistic regression provides a real-valued
output, interpreted as the probability of the relevance of the
label (value larger than 0.5 yielding a decision of “relevant”).
For each context label we used an independent model.
We first randomly partitioned the training examples into
internal training and validation subsets, allocating one third
of the training examples for the validation subset, while
maintaining the same proportion of positive vs. negative
examples in both subsets. We then used grid search to select
the cost parameter C for logistic regression: for each value
(out of {0.001, 0.01, 0.1, 1, 10, 100}) we trained a logistic
regression model on the internal train subset and tested the
model on the validation subset. We selected the value of C
that resulted in highest F1 measure on the validation subset.
We then re-trained a logistic regression model with the
selected value on the entire training set. For the leave-oneuser-out experiment with the EF system we simplified the
procedure and only trained the logistic regression models
with value of C = 1 (instead of performing grid search).
The learned weights from LFL for a set of selected labels
that are presented in Figure 4 (A) are taken from the first (of
five) training set of the cross validation evaluation. To look
at misclassifications and to produce the confusion matrices
in Figure 4 (B–G) we used the multiclass (one-versus-rest)
version of logistic regression, with a fixed cost value of
C = 1. Each multiclass experiment used the set of examples
annotated with exactly one label from the examined label
subset and with data from all of the sensors of interest
(so an experiment with only accelerometer and gyroscope
sensors might have more examples than an experiment with
accelerometer, gyroscope and watch accelerometer). These
experiments were more fitting than binary classification in
cases where missing labels are common. For example, labels
describing the phone’s position were not always consistently annotated. A binary classifier will use all negative
examples to learn a decision boundary, including examples
the subject forgot to label, which may skew the results if
there are many missing labels.
Performance evaluation
In order to make a fair comparison among the different
sensors, evaluation was done on the subset of examples with
data from all six core sensors available (∼177k examples
from 51 subjects). In the training phase, however, a singlesensor classifier was allowed to use all examples available
(e.g., all examples in the dataset had phone state data, so the
PS single-sensor classifier was trained with all examples).
While the early fusion system benefited from the advantage of modeling correlations between features of different
sensors, it was limited to being trained only on examples
with all sensor data available. The late fusion systems, on
the other hand, had the advantage of using single-sensor
classifiers that were trained on many more examples.
Classifier performance was evaluated using 5-fold cross
validation. The subjects were randomly partitioned once
into 5 folds, while equalizing the proportion of iPhone
vs. Android users between folds (To keep a fair evaluation
it was important to partition the subjects, and not randomly
partition the pool of examples, in order to avoid having
examples from the same subject appear in both the training
set and the test set). The cross validation procedure repeats
the following for each fold: (1) hold out the selected fold
to act as the test set (2) train a classifier on the remaining
folds (3) apply the classifier to the held out test set. For
each fold and for each label, we counted the numbers of
true positives (TP. Examples that were correctly classified as
positive), true negatives (TN. Examples that were correctly
classified as negative), false positives (FP. Examples that
were wrongfully classified as positive) and false negatives
(FN. Examples that were wrongfully classified as negative).
At the end of the 5-fold procedure we summed up the total
numbers of TP, TN, FP and FN over the entire evaluation
set and calculated the following performance metrics:
•
•
•
•
•
Accuracy is the proportion of correctly classified examples out of all the examples. This metric is sensitive to imbalanced label proportion in the data.
True positive rate (TPR, also called sensitivity
or recall) is the proportion of positive examples that were correctly classified as positive: recall=TPR=TP/(TP+FN).
True negative rate (TNR, also called specificity) is the
proportion of negative examples that were correctly
classified as negative: TNR=TN/(TN+FP).
Precision (prec) is the proportion of correctly classified examples out of the examples that the classifier
declared as positive: precision=TP/(TP+FP).
Balanced accuracy is a measure that accounts for the
tradeoff between true positive rate and true negative
rate: BA=(TPR+TNR)/2.
17
•
The F1 measure is another such measure, which
takes the harmonic mean of precision and recall:
F1=(2*TPR*prec)/(TPR+prec).
While the balanced accuracy is easy to interpret (chance
level is always 0.5 and perfect performance is 1) the F1
measure is very sensitive to how rare the positive examples
are, so for each label a typical F1 value is different. The
5-fold subject partition is available with the dataset, and
we encourage researchers using the dataset to evaluate
new methods to use the same 5-fold partition, in order to
promote fair comparisons.
Random chance. To assess the statistical significance
of the performance scores we achieved, we evaluated a
distribution of performance scores achieved by a random
classifier. The random classifier declares “relevant” with
probability 0.5 independently for each example and for
each label. To estimate the distribution of scores that such
a classifier obtains, we ran 100 simulations (each time the
classifier randomly assigned binary predictions and the
performance scores were calculated over the entire evaluation dataset). Chance level (expected value of the random
classifier) of balanced accuracy is 0.5 for every label. For
the F1 measure the chance level for each label is dependent
on the proportion of positive and negative examples in the
data. For each performance measure and for each label we
estimated a value which we call p99, the 99th percentile
among the 100 scores achieved in the 100 simulations. Hence
the probability of obtaining a score greater than p99 by
chance is less than 1%. For average (over a set of labels)
scores the p99 value was calculated similarly (computing
the average score for each of the 100 simulations).
User personalization assessment
To assess the advantages of user personalization, we selected a single test subject that had provided relatively
many examples and many labels. We partitioned this user’s
examples into the first half and second half of the examples
(according to their recording timestamps), to simulate an
adaptation training period (the first half) and a deployment
period (the second half). We used early fusion (EF) classifiers
to combine the features from all 6 sensors. The universal
model was the one used in previous experiments, taken
from the fold where the test user was part of the cross
validation test set (so the universal model was trained on 48
other users). The individual model was trained only on data
from the test user, taken from the first half of the subject’s
examples. The adapted model was a combination of both the
universal and individual models using the LFA method (i.e.,
averaging the probability outputs of both models). All three
models were tested on the same set of unseen test examples
(the second half of the subject’s examples). For some labels,
an insufficient number of examples to train an individual
classifier resulted in a trivial classifier (always declaring the
same answer). In those cases the performance was reported
as chance level (BA of 0.5 and F1 of 0).
18
D ETAILED RESULTS TABLES
5-fold cross validation evaluation
19
p99
0.50
0.50
0.51
0.52
0.51
0.50
0.51
0.51
0.51
0.50
0.50
0.51
0.51
0.52
0.51
0.51
0.50
0.52
0.50
0.51
0.51
0.52
0.53
0.52
0.52
0.50
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.69
0.81
0.66
0.79
0.85
0.87
0.86
0.88
0.61
0.68
0.61
0.65
0.69
0.76
0.75
0.75
0.80
0.75
0.66
0.63
0.70
0.80
0.80
0.80
0.66
0.80
0.56
0.48
0.58
0.62
0.69
0.71
0.81
0.84
0.81
0.77
0.83
0.87
0.87
0.87
0.70
0.81
0.62
0.79
0.87
0.88
0.88
0.89
0.62
0.65
0.84
0.71
0.81
0.84
0.87
0.85
0.63
0.57
0.74
0.76
0.67
0.70
0.77
0.80
0.57
0.62
0.63
0.79
0.73
0.80
0.79
0.82
0.49
0.63
0.76
0.65
0.78
0.80
0.80
0.81
0.66
0.67
0.63
0.71
0.72
0.75
0.75
0.76
0.73
0.70
0.66
0.66
0.73
0.74
0.77
0.78
0.65
0.71
0.81
0.77
0.84
0.85
0.86
0.86
0.69
0.67
0.75
0.74
0.82
0.77
0.84
0.83
0.61
0.75
0.82
0.74
0.83
0.84
0.86
0.87
0.71
0.64
0.79
0.76
0.81
0.84
0.84
0.85
0.63
0.66
0.63
0.71
0.70
0.75
0.77
0.78
0.67
0.68
0.58
0.85
0.77
0.76
0.83
0.81
0.75
0.67
0.61
0.64
0.72
0.77
0.77
0.77
0.73
0.77
0.71
0.70
0.77
0.81
0.80
0.81
0.53
0.68
0.57
0.62
0.68
0.71
0.71
0.72
0.70
0.69
0.54
0.59
0.79
0.69
0.76
0.80
0.74
0.72
0.67
0.64
0.75
0.66
0.77
0.74
0.69
0.50
0.56
0.80
0.74
0.70
0.82
0.81
0.55
0.73
0.47
0.63
0.47
0.64
0.67
0.70
0.66
0.70
0.67
0.70
0.75
0.77
0.80
0.80
TABLE S1
5-fold evaluation performance (BA) of the different classifiers on each label. Part 1 of the labels. For each label ne is the number of examples and
ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of random scores
— a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest performing classifier
is marked in bold.
Lying down
Sitting
Walking
Running
Bicycling
Sleeping
Lab work
In class
In a meeting
At main workplace
Indoors
Outside
In a car
On a bus
Drive (I’m the driver)
Drive (I’m a passenger)
At home
At a restaurant
Phone in pocket
Exercise
Cooking
Shopping
Strolling
Drinking (alcohol)
Bathing - shower
average
ne
54359
82904
11892
675
3523
42920
2898
2872
2904
20382
107944
7629
3635
1185
5034
1655
83977
1320
15301
5384
2257
896
434
864
1186
ns
47
50
50
19
22
40
8
13
34
26
51
36
24
24
24
19
50
16
31
36
33
18
8
10
27
p99
0.51
0.52
0.52
0.51
0.50
0.53
0.53
0.55
0.53
0.50
0.50
0.51
0.51
0.51
0.52
0.52
0.53
0.53
0.55
0.50
0.50
0.51
0.51
0.50
0.51
0.50
0.50
Acc
0.72
0.63
0.77
0.69
0.81
0.75
0.71
0.60
0.60
0.57
0.66
0.70
0.79
0.73
0.79
0.76
0.65
0.62
0.69
0.73
0.52
0.70
0.67
0.71
0.53
0.68
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.64
0.71
0.41
0.60
0.51
0.60
0.70
0.68
0.66
0.66
0.38
0.53
0.65
0.58
0.63
0.63
0.52
0.70
0.58
0.60
0.57
0.65
0.70
0.70
0.54
0.56
0.56
0.64
0.67
0.65
0.70
0.68
0.55
0.60
0.54
0.60
0.57
0.59
0.63
0.63
0.71
0.49
0.54
0.81
0.56
0.54
0.76
0.75
0.66
0.53
0.60
0.49
0.93
0.50
0.61
0.66
0.48
0.47
0.72
0.58
0.70
0.50
0.71
0.70
0.64
0.46
0.61
0.68
0.59
0.50
0.65
0.53
0.61
0.60
0.54
0.65
0.65
0.65
0.67
0.67
0.56
0.62
0.63
0.61
0.68
0.68
0.71
0.70
0.58
0.60
0.51
0.61
0.62
0.66
0.65
0.65
0.51
0.58
0.57
0.64
0.59
0.65
0.66
0.66
0.49
0.62
0.59
0.63
0.58
0.60
0.63
0.63
0.52
0.64
0.54
0.65
0.61
0.64
0.67
0.67
0.55
0.58
0.57
0.65
0.70
0.54
0.64
0.61
0.73
0.65
0.55
0.55
0.51
0.58
0.69
0.67
0.73
0.66
0.55
0.55
0.51
0.58
0.71
0.66
0.76
0.44
0.54
0.71
0.51
0.49
0.73
0.73
0.59
0.67
0.54
0.59
0.63
0.68
0.67
0.68
0.59
0.59
0.66
0.64
0.68
0.70
0.70
0.70
0.68
0.56
0.59
0.59
0.61
0.64
0.67
0.66
0.56
0.55
0.59
0.64
0.69
0.67
0.68
0.69
0.61
0.56
0.53
0.55
0.61
0.61
0.62
0.62
0.57
0.61
0.58
0.68
0.67
0.69
0.71
0.72
0.58
0.53
0.54
0.60
0.60
0.55
0.61
0.58
0.60
0.59
0.56
0.62
0.62
0.60
0.67
0.66
TABLE S2
5-fold evaluation performance (BA) of the different classifiers on each label. Part 2 of the labels. For each label ne is the number of examples and
ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of random scores
— a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest performing classifier
is marked in bold.
Cleaning
Laundry
Washing dishes
Watching TV
Surfing the internet
At a party
At a bar
At the beach
Singing
Talking
Computer work
Eating
Toilet
Grooming
Dressing
At the gym
Stairs - going up
Stairs - going down
Elevator
Standing
At school
Phone in hand
Phone in bag
Phone on table
With co-workers
With friends
average
ne
1839
473
851
9412
11641
404
520
122
384
18976
23692
10169
1646
1847
1308
906
399
390
124
22766
25840
8595
5589
70611
4139
12865
ns
22
12
17
28
28
3
4
5
6
44
38
49
33
25
27
6
17
15
8
51
39
37
22
43
17
25
Acc
0.63
0.65
0.40
0.61
0.56
0.74
0.45
0.62
0.57
0.60
0.57
0.59
0.57
0.44
0.51
0.50
0.70
0.71
0.72
0.60
0.59
0.65
0.59
0.60
0.57
0.55
0.59
20
p99
0.38
0.49
0.12
0.01
0.04
0.33
0.03
0.03
0.03
0.19
0.55
0.08
0.04
0.01
0.06
0.02
0.49
0.02
0.15
0.06
0.03
0.01
0.01
0.01
0.01
0.13
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.59
0.71
0.55
0.69
0.78
0.81
0.79
0.82
0.58
0.67
0.59
0.62
0.72
0.75
0.75
0.74
0.38
0.31
0.22
0.19
0.22
0.38
0.39
0.38
0.03
0.04
0.01
0.01
0.01
0.03
0.04
0.04
0.16
0.23
0.18
0.12
0.17
0.31
0.25
0.26
0.53
0.65
0.44
0.63
0.75
0.79
0.77
0.81
0.06
0.06
0.11
0.09
0.11
0.21
0.16
0.19
0.06
0.04
0.07
0.10
0.06
0.13
0.12
0.14
0.04
0.05
0.05
0.11
0.07
0.17
0.10
0.14
0.18
0.28
0.41
0.31
0.42
0.49
0.47
0.50
0.73
0.70
0.68
0.75
0.71
0.78
0.79
0.78
0.20
0.18
0.15
0.15
0.20
0.23
0.26
0.25
0.07
0.10
0.27
0.13
0.16
0.23
0.22
0.23
0.03
0.03
0.07
0.04
0.06
0.07
0.07
0.06
0.09
0.15
0.37
0.16
0.23
0.31
0.31
0.31
0.04
0.04
0.15
0.07
0.08
0.14
0.12
0.12
0.65
0.64
0.63
0.70
0.67
0.74
0.76
0.77
0.03
0.03
0.02
0.08
0.05
0.11
0.07
0.09
0.33
0.26
0.21
0.25
0.28
0.38
0.36
0.37
0.18
0.24
0.14
0.13
0.19
0.27
0.26
0.25
0.03
0.05
0.03
0.04
0.06
0.09
0.07
0.08
0.03
0.02
0.01
0.02
0.04
0.04
0.04
0.04
0.02
0.01
0.01
0.01
0.02
0.03
0.03
0.03
0.02
0.01
0.01
0.04
0.03
0.07
0.07
0.06
0.02
0.04
0.01
0.02
0.01
0.04
0.04
0.05
0.20
0.22
0.22
0.22
0.24
0.30
0.29
0.30
TABLE S3
5-fold evaluation performance (F1) of the different classifiers on each label. Part 1 of the labels. For each label ne is the number of examples and
ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of random scores
— a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest performing classifier
is marked in bold.
Lying down
Sitting
Walking
Running
Bicycling
Sleeping
Lab work
In class
In a meeting
At main workplace
Indoors
Outside
In a car
On a bus
Drive (I’m the driver)
Drive (I’m a passenger)
At home
At a restaurant
Phone in pocket
Exercise
Cooking
Shopping
Strolling
Drinking (alcohol)
Bathing - shower
average
ne
54359
82904
11892
675
3523
42920
2898
2872
2904
20382
107944
7629
3635
1185
5034
1655
83977
1320
15301
5384
2257
896
434
864
1186
ns
47
50
50
19
22
40
8
13
34
26
51
36
24
24
24
19
50
16
31
36
33
18
8
10
27
p99
0.02
0.01
0.01
0.10
0.12
0.01
0.01
0.00
0.00
0.18
0.21
0.11
0.02
0.02
0.02
0.01
0.01
0.00
0.00
0.21
0.23
0.09
0.06
0.45
0.05
0.13
0.08
Acc
0.61
0.58
0.38
0.03
0.19
0.57
0.08
0.05
0.05
0.23
0.74
0.20
0.15
0.04
0.21
0.07
0.66
0.02
0.28
0.21
0.03
0.03
0.02
0.03
0.01
0.22
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.05
0.05
0.01
0.03
0.02
0.05
0.06
0.05
0.01
0.01
0.00
0.01
0.01
0.02
0.01
0.02
0.01
0.03
0.01
0.02
0.01
0.03
0.03
0.04
0.11
0.12
0.12
0.17
0.18
0.21
0.22
0.22
0.14
0.17
0.13
0.17
0.15
0.18
0.19
0.20
0.01
0.00
0.01
0.03
0.01
0.04
0.03
0.04
0.01
0.01
0.01
0.00
0.09
0.00
0.03
0.06
0.00
0.00
0.01
0.00
0.01
0.00
0.02
0.02
0.01
0.00
0.01
0.01
0.01
0.00
0.01
0.01
0.26
0.24
0.21
0.29
0.27
0.29
0.30
0.30
0.25
0.30
0.31
0.30
0.35
0.38
0.39
0.39
0.14
0.15
0.11
0.16
0.15
0.19
0.18
0.17
0.02
0.03
0.02
0.03
0.03
0.04
0.04
0.04
0.02
0.03
0.03
0.04
0.03
0.05
0.04
0.05
0.02
0.03
0.02
0.03
0.03
0.04
0.04
0.04
0.01
0.02
0.01
0.02
0.02
0.03
0.03
0.03
0.02
0.01
0.01
0.01
0.00
0.01
0.02
0.01
0.02
0.01
0.01
0.01
0.00
0.01
0.02
0.01
0.01
0.00
0.00
0.00
0.00
0.00
0.01
0.01
0.27
0.35
0.23
0.27
0.29
0.36
0.34
0.35
0.29
0.30
0.38
0.34
0.39
0.41
0.41
0.41
0.17
0.11
0.13
0.13
0.13
0.16
0.17
0.16
0.08
0.07
0.09
0.11
0.12
0.15
0.14
0.14
0.58
0.51
0.50
0.51
0.56
0.56
0.59
0.58
0.06
0.07
0.06
0.11
0.08
0.13
0.12
0.13
0.17
0.14
0.15
0.18
0.18
0.15
0.19
0.18
0.11
0.11
0.10
0.11
0.12
0.13
0.14
0.14
TABLE S4
5-fold evaluation performance (F1) of the different classifiers on each label. Part 2 of the labels. For each label ne is the number of examples and
ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of random scores
— a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest performing classifier
is marked in bold.
Cleaning
Laundry
Washing dishes
Watching TV
Surfing the internet
At a party
At a bar
At the beach
Singing
Talking
Computer work
Eating
Toilet
Grooming
Dressing
At the gym
Stairs - going up
Stairs - going down
Elevator
Standing
At school
Phone in hand
Phone in bag
Phone on table
With co-workers
With friends
average
ne
1839
473
851
9412
11641
404
520
122
384
18976
23692
10169
1646
1847
1308
906
399
390
124
22766
25840
8595
5589
70611
4139
12865
ns
22
12
17
28
28
3
4
5
6
44
38
49
33
25
27
6
17
15
8
51
39
37
22
43
17
25
Acc
0.05
0.01
0.01
0.14
0.15
0.01
0.00
0.00
0.01
0.25
0.26
0.15
0.03
0.02
0.02
0.01
0.02
0.02
0.01
0.28
0.30
0.17
0.10
0.58
0.06
0.15
0.11
21
Leave-one-user-out evaluation
22
p99
0.50
0.50
0.51
0.52
0.51
0.50
0.51
0.51
0.51
0.50
0.50
0.51
0.51
0.52
0.51
0.51
0.50
0.52
0.50
0.51
0.51
0.52
0.53
0.52
0.52
0.50
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.69
0.81
0.65
0.79
0.84
0.87
0.86
0.88
0.61
0.68
0.61
0.65
0.69
0.76
0.75
0.75
0.80
0.75
0.66
0.63
0.71
0.80
0.80
0.81
0.69
0.80
0.56
0.50
0.57
0.67
0.72
0.76
0.81
0.85
0.80
0.76
0.80
0.87
0.86
0.87
0.70
0.81
0.62
0.80
0.85
0.88
0.88
0.89
0.61
0.65
0.82
0.70
0.84
0.84
0.84
0.84
0.63
0.58
0.74
0.77
0.72
0.74
0.79
0.81
0.59
0.62
0.66
0.78
0.73
0.81
0.80
0.82
0.49
0.64
0.76
0.65
0.80
0.80
0.81
0.82
0.66
0.68
0.63
0.70
0.72
0.76
0.75
0.76
0.74
0.70
0.66
0.67
0.71
0.75
0.78
0.79
0.66
0.71
0.82
0.76
0.83
0.85
0.86
0.87
0.69
0.68
0.72
0.72
0.80
0.78
0.83
0.82
0.62
0.75
0.83
0.75
0.84
0.84
0.86
0.87
0.70
0.64
0.80
0.77
0.82
0.84
0.83
0.84
0.63
0.66
0.62
0.72
0.72
0.76
0.77
0.77
0.68
0.69
0.57
0.84
0.74
0.79
0.84
0.84
0.75
0.67
0.61
0.64
0.71
0.77
0.77
0.77
0.73
0.77
0.71
0.70
0.75
0.81
0.81
0.81
0.55
0.67
0.57
0.62
0.68
0.71
0.72
0.72
0.69
0.68
0.53
0.57
0.79
0.67
0.75
0.78
0.73
0.70
0.63
0.62
0.71
0.67
0.74
0.75
0.70
0.54
0.56
0.79
0.54
0.68
0.79
0.77
0.54
0.74
0.48
0.63
0.48
0.64
0.69
0.72
0.67
0.70
0.66
0.70
0.74
0.78
0.80
0.81
TABLE S5
Leave-one-user-out evaluation performance (BA) of the different classifiers on each label. Part 1 of the labels. For each label ne is the number of
examples and ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of
random scores — a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest
performing classifier is marked in bold.
Lying down
Sitting
Walking
Running
Bicycling
Sleeping
Lab work
In class
In a meeting
At main workplace
Indoors
Outside
In a car
On a bus
Drive (I’m the driver)
Drive (I’m a passenger)
At home
At a restaurant
Phone in pocket
Exercise
Cooking
Shopping
Strolling
Drinking (alcohol)
Bathing - shower
average
ne
54359
82904
11892
675
3523
42920
2898
2872
2904
20382
107944
7629
3635
1185
5034
1655
83977
1320
15301
5384
2257
896
434
864
1186
ns
47
50
50
19
22
40
8
13
34
26
51
36
24
24
24
19
50
16
31
36
33
18
8
10
27
p99
0.51
0.52
0.52
0.51
0.50
0.53
0.53
0.55
0.53
0.50
0.50
0.51
0.51
0.51
0.52
0.52
0.53
0.53
0.55
0.50
0.50
0.51
0.51
0.50
0.51
0.50
0.50
Acc
0.73
0.63
0.77
0.69
0.81
0.75
0.69
0.61
0.62
0.55
0.67
0.72
0.79
0.74
0.80
0.76
0.65
0.62
0.69
0.74
0.52
0.71
0.64
0.72
0.50
0.68
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.63
0.73
0.42
0.62
0.49
0.70
0.71
0.70
0.65
0.66
0.35
0.52
0.78
0.60
0.68
0.70
0.48
0.69
0.54
0.61
0.54
0.66
0.67
0.68
0.54
0.57
0.57
0.67
0.66
0.69
0.72
0.71
0.58
0.59
0.56
0.60
0.57
0.61
0.62
0.62
0.71
0.48
0.70
0.84
0.67
0.52
0.79
0.76
0.69
0.50
0.64
0.62
0.88
0.52
0.71
0.68
0.51
0.52
0.71
0.58
0.69
0.57
0.71
0.71
0.62
0.46
0.70
0.68
0.60
0.48
0.68
0.53
0.61
0.61
0.55
0.66
0.64
0.66
0.68
0.68
0.57
0.62
0.65
0.59
0.67
0.69
0.71
0.69
0.58
0.60
0.53
0.61
0.63
0.66
0.66
0.66
0.52
0.57
0.57
0.63
0.56
0.65
0.65
0.65
0.53
0.62
0.60
0.65
0.53
0.63
0.64
0.66
0.54
0.66
0.53
0.67
0.55
0.66
0.67
0.68
0.56
0.67
0.51
0.67
0.70
0.58
0.67
0.67
0.76
0.65
0.57
0.57
0.48
0.59
0.69
0.66
0.75
0.66
0.54
0.55
0.48
0.57
0.69
0.63
0.70
0.56
0.57
0.70
0.54
0.50
0.62
0.61
0.59
0.67
0.54
0.59
0.62
0.68
0.66
0.67
0.59
0.59
0.68
0.66
0.70
0.72
0.71
0.71
0.68
0.56
0.58
0.58
0.59
0.63
0.67
0.66
0.56
0.56
0.59
0.69
0.69
0.72
0.71
0.73
0.61
0.56
0.52
0.56
0.61
0.61
0.63
0.62
0.57
0.61
0.61
0.68
0.71
0.69
0.73
0.74
0.57
0.54
0.55
0.62
0.59
0.58
0.63
0.61
0.60
0.60
0.57
0.63
0.62
0.62
0.68
0.67
TABLE S6
Leave-one-user-out evaluation performance (BA) of the different classifiers on each label. Part 2 of the labels. For each label ne is the number of
examples and ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of
random scores — a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest
performing classifier is marked in bold.
Cleaning
Laundry
Washing dishes
Watching TV
Surfing the internet
At a party
At a bar
At the beach
Singing
Talking
Computer work
Eating
Toilet
Grooming
Dressing
At the gym
Stairs - going up
Stairs - going down
Elevator
Standing
At school
Phone in hand
Phone in bag
Phone on table
With co-workers
With friends
average
ne
1839
473
851
9412
11641
404
520
122
384
18976
23692
10169
1646
1847
1308
906
399
390
124
22766
25840
8595
5589
70611
4139
12865
ns
22
12
17
28
28
3
4
5
6
44
38
49
33
25
27
6
17
15
8
51
39
37
22
43
17
25
Acc
0.62
0.67
0.36
0.61
0.55
0.73
0.53
0.66
0.56
0.61
0.59
0.59
0.57
0.46
0.51
0.55
0.68
0.70
0.68
0.60
0.60
0.66
0.60
0.60
0.55
0.56
0.59
23
p99
0.38
0.49
0.12
0.01
0.04
0.33
0.03
0.03
0.03
0.19
0.55
0.08
0.04
0.01
0.06
0.02
0.49
0.02
0.15
0.06
0.03
0.01
0.01
0.01
0.01
0.13
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.59
0.71
0.54
0.69
0.76
0.81
0.79
0.82
0.58
0.68
0.58
0.62
0.71
0.75
0.75
0.74
0.37
0.32
0.21
0.19
0.22
0.38
0.39
0.39
0.02
0.04
0.01
0.01
0.01
0.04
0.04
0.04
0.15
0.23
0.16
0.12
0.15
0.30
0.24
0.26
0.53
0.65
0.44
0.64
0.74
0.79
0.76
0.80
0.06
0.06
0.11
0.08
0.11
0.21
0.15
0.18
0.06
0.04
0.08
0.11
0.07
0.13
0.12
0.14
0.04
0.05
0.06
0.11
0.07
0.17
0.11
0.15
0.19
0.29
0.41
0.31
0.43
0.49
0.48
0.52
0.73
0.71
0.68
0.75
0.71
0.79
0.79
0.79
0.20
0.18
0.16
0.16
0.20
0.23
0.26
0.25
0.08
0.10
0.27
0.13
0.16
0.23
0.22
0.23
0.03
0.03
0.05
0.04
0.05
0.07
0.07
0.07
0.09
0.16
0.38
0.15
0.21
0.31
0.31
0.31
0.04
0.04
0.15
0.07
0.08
0.13
0.12
0.11
0.65
0.65
0.63
0.71
0.69
0.75
0.76
0.76
0.03
0.03
0.02
0.07
0.04
0.11
0.07
0.10
0.34
0.26
0.22
0.25
0.27
0.38
0.37
0.37
0.16
0.22
0.14
0.13
0.15
0.26
0.26
0.24
0.03
0.05
0.03
0.04
0.05
0.08
0.07
0.07
0.02
0.02
0.01
0.01
0.04
0.04
0.04
0.04
0.02
0.01
0.01
0.01
0.02
0.02
0.02
0.02
0.02
0.01
0.01
0.04
0.01
0.05
0.06
0.05
0.02
0.04
0.01
0.02
0.01
0.04
0.04
0.05
0.20
0.22
0.22
0.22
0.24
0.30
0.29
0.30
TABLE S7
Leave-one-user-out evaluation performance (F1) of the different classifiers on each label. Part 1 of the labels. For each label ne is the number of
examples and ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of
random scores — a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest
performing classifier is marked in bold.
Lying down
Sitting
Walking
Running
Bicycling
Sleeping
Lab work
In class
In a meeting
At main workplace
Indoors
Outside
In a car
On a bus
Drive (I’m the driver)
Drive (I’m a passenger)
At home
At a restaurant
Phone in pocket
Exercise
Cooking
Shopping
Strolling
Drinking (alcohol)
Bathing - shower
average
ne
54359
82904
11892
675
3523
42920
2898
2872
2904
20382
107944
7629
3635
1185
5034
1655
83977
1320
15301
5384
2257
896
434
864
1186
ns
47
50
50
19
22
40
8
13
34
26
51
36
24
24
24
19
50
16
31
36
33
18
8
10
27
p99
0.02
0.01
0.01
0.10
0.12
0.01
0.01
0.00
0.00
0.18
0.21
0.11
0.02
0.02
0.02
0.01
0.01
0.00
0.00
0.21
0.23
0.09
0.06
0.45
0.05
0.13
0.08
Acc
0.62
0.58
0.38
0.03
0.19
0.56
0.07
0.05
0.06
0.22
0.75
0.21
0.15
0.04
0.21
0.07
0.67
0.03
0.29
0.21
0.03
0.03
0.01
0.03
0.01
0.22
Gyro WAcc
Loc
Aud
PS
EF
LFA LFL
0.04
0.06
0.01
0.03
0.02
0.07
0.06
0.06
0.01
0.01
0.00
0.01
0.02
0.02
0.02
0.02
0.01
0.02
0.01
0.02
0.01
0.03
0.03
0.03
0.11
0.12
0.12
0.19
0.17
0.23
0.22
0.24
0.16
0.16
0.14
0.17
0.15
0.20
0.19
0.19
0.01
0.00
0.01
0.03
0.01
0.02
0.03
0.04
0.01
0.01
0.01
0.02
0.06
0.01
0.04
0.05
0.00
0.00
0.01
0.00
0.01
0.05
0.02
0.02
0.01
0.00
0.01
0.01
0.01
0.00
0.01
0.01
0.25
0.25
0.21
0.30
0.26
0.30
0.30
0.30
0.26
0.30
0.32
0.28
0.34
0.39
0.39
0.39
0.14
0.15
0.11
0.16
0.15
0.18
0.18
0.18
0.02
0.03
0.02
0.03
0.02
0.04
0.04
0.04
0.02
0.04
0.03
0.04
0.02
0.05
0.04
0.05
0.02
0.03
0.02
0.03
0.02
0.04
0.04
0.04
0.01
0.02
0.01
0.02
0.03
0.04
0.03
0.04
0.02
0.01
0.01
0.01
0.00
0.01
0.02
0.01
0.02
0.01
0.01
0.01
0.00
0.01
0.02
0.01
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.27
0.35
0.23
0.27
0.29
0.35
0.34
0.35
0.30
0.29
0.42
0.37
0.40
0.44
0.42
0.42
0.17
0.11
0.13
0.12
0.12
0.16
0.17
0.16
0.08
0.08
0.08
0.13
0.11
0.16
0.15
0.16
0.58
0.51
0.48
0.51
0.56
0.55
0.59
0.58
0.06
0.07
0.07
0.11
0.09
0.13
0.12
0.14
0.17
0.14
0.15
0.20
0.17
0.18
0.20
0.20
0.11
0.11
0.10
0.12
0.12
0.14
0.14
0.14
TABLE S8
Leave-one-user-out evaluation performance (F1) of the different classifiers on each label. Part 2 of the labels. For each label ne is the number of
examples and ns is the number of subjects in the testing (possibly more examples participated in the training). p99 marks the 99th percentile of
random scores — a score above the p99 value has less than 0.01 probability to be achieved randomly. For each label the score of the highest
performing classifier is marked in bold.
Cleaning
Laundry
Washing dishes
Watching TV
Surfing the internet
At a party
At a bar
At the beach
Singing
Talking
Computer work
Eating
Toilet
Grooming
Dressing
At the gym
Stairs - going up
Stairs - going down
Elevator
Standing
At school
Phone in hand
Phone in bag
Phone on table
With co-workers
With friends
average
ne
1839
473
851
9412
11641
404
520
122
384
18976
23692
10169
1646
1847
1308
906
399
390
124
22766
25840
8595
5589
70611
4139
12865
ns
22
12
17
28
28
3
4
5
6
44
38
49
33
25
27
6
17
15
8
51
39
37
22
43
17
25
Acc
0.04
0.01
0.00
0.14
0.14
0.01
0.01
0.00
0.01
0.25
0.28
0.15
0.03
0.02
0.02
0.01
0.01
0.01
0.00
0.28
0.30
0.17
0.11
0.59
0.06
0.16
0.11
| 2cs.AI
|
1
Interference Modeling for Cellular Networks
under Beamforming Transmission
Hussain Elkotby, Student Member IEEE and Mai Vu, Senior Member IEEE
arXiv:1706.00050v1 [cs.IT] 31 May 2017
Abstract
We propose analytical models for the interference power distribution in a cellular system employing
MIMO beamforming in rich and limited scattering environments, which capture non line-of-sight signal
propagation in the microwave and mmWave bands, respectively. Two candidate models are considered:
the Inverse Gaussian and the Inverse Weibull, both are two-parameter heavy tail distributions. We further
propose a mixture of these two distributions as a model with three parameters. To estimate the parameters
of these distributions, three approaches are used: moment matching, individual distribution maximum
likelihood estimation (MLE), and mixture distribution MLE with a designed expectation maximization
algorithm. We then introduce simple fitted functions for the mixture model parameters as polynomials of
the channel path loss exponent and shadowing variance. To measure the goodness of these models, the
information-theoretic metric relative entropy is used to capture the distance from the model distribution
to a reference one. The interference models are tested against data obtained by simulating a cellular
network based on stochastic geometry. The results show that the three-parameter mixture model offers
remarkably good fit to simulated interference power. The mixture model is further used to analyze the
capacity of a cellular network employing joint transmit and receive beamforming and confirms a good
fit with simulation.
Keywords: mmWave cellular; interference model; stochastic geometry; moment matching; maximum
likelihood estimation; mixture distribution.
I. I NTRODUCTION
The current scarcity of wireless spectrum coupled with the predicted exponential increase
in capacity demand has focused attention on denser network deployment and the underutilized
millimeter wave (mmWave) bands (30-300 GHz) [1]–[3]. Fifth generation (5G) wireless systems
aiming to cater for the capacity demand are expected to provide a minimum of 1 Gb/s data rate
anywhere with up to 5 Gb/s for high mobility users and 50 Gb/s data rates for pedestrians [2].
Both dense deployments and mmWave make promising candidates for the 5G systems through
the resources reuse over smaller areas and the huge amount of available spectrum [2], [4].
Modeling and characterization of wireless interference under these scenarios is essential for
cellular system analysis and design. Due to the different characteristics of signal propagation in
The authors are with the Department of Electrical and Computer Engineering, Tufts University, Medford, MA, USA. Emails:
[email protected], [email protected]
2
conventional microwave and mmWave bands, we use the terminology rich and limited scattering
environments to refer to the signal propagation in microwave and mmWave bands.
A. Background and Related Works
We argue for the important role that interference characterization plays in evaluating and predicting the network performance in both microwave and mmWave bands. Traditionally, mmWave
bands are considered for backhaul in cellular systems and for high-volume consumer electronics
such as personal area and local area networks, but not for cellular access due to concerns about
short-range and non-line-of-sight coverage issues [2], [4]. MmWave has, however, recently been
shown to be suitable for cellular communications, provided short cell radius of the order of
100-200 meters and sufficient beamforming gain between communicating nodes [2]. Reducing
the cell radius leads to dense base station (BS) deployments. Even under beamforming, these
high BS and user densities can drive cellular networks to be more interference rather than noise
limited. While large adaptive arrays with narrow beams can boost the received signal power
and hence reduce the impact of out-of-cell interference [3], [5], this interference remains an
important performance-limiting factor in dense mmWave networks [6].
In next generation of wireless networks with a large number of subscribers and dense BSs
deployment, interference modeling is an important step towards network realization. Existing
stochastic geometry based work either uses the interference Laplace transform to evaluate simple transmission schemes [7], [8], or models the interference using moment matching gamma
distribution [9], [10]. Gaussian distribution is another approximation that is considered for the
centered and normalized aggregate wireless interference power [11], [12]. The central limit
theorem which justifies the Gaussian approximation, however, does not apply when some of the
interferers are dominant. Also, the Gaussian distribution does not model the interference very
well at low density of interferers or when the exclusion region, the region with no interferers, is
relatively small, as the cell sizes shrink. The distribution of the interference power at relatively
small exclusion regions has a heavy tail which can not be captured by the Gaussian distribution
[12]. Here we propose analytical distribution models characterized by only a few parameters,
which can be fitted to simple polynomial functions of channel path loss exponent and shadowing
variance, and test their fitness against network simulation based on stochastic geometry. A closed
form distribution of the interference helps in designing a cellular system by allowing the analysis
of system capacity and comparing different transmission techniques, which may not be directly
achievable with a characteristic function of the interference using the Laplace transform. Further,
having the interference distribution can significantly speed up system capacity evaluation by
3
generating the interference directly to use in capacity numerical computation, instead of running
time-consuming simulations.
To evaluate an interference model, we apply it to a cellular network and analyze the system
performance. For cellular network analysis, stochastic geometry is shown to capture the main
performance trends with analytical tractability. Stochastic geometry is used to develop a tractable
model for downlink heterogeneous cellular networks [8], which is applied to analyze coordinated
multipoint beamforming [7]. These stochastic geometry based networks are useful in verifying
that the performance of a cellular network matches experimental trends, and can also be used
to verify the accuracy of an interference model.
B. Approaches to Interference Modeling
The underlying question is "What parameterized distribution can best model out-of-cell interference and how can we estimate the parameters of this distribution?". We focus on parameterized
distributions with as few parameters as possible to make the model the simplest while having
a good fit. We test the goodness of the proposed interference models against simulation of a
cellular network based on stochastic geometry. We note that testing against actual measurement
data is also feasible and is desirable when such data are available.
To estimate the parameters of the proposed interference models, we use both moment matching
and maximum likelihood estimation (MLE) techniques. Specifically, we consider three approaches: analytical moment matching, individual distribution MLE, and mixture distribution
MLE. We exploit the iterative expectation maximization (EM) algorithm to estimate the parameters of the mixture MLE model. To evaluate the goodness of each model, we introduce the use
of the information-theoretic relative entropy or Kullback-Leibler (KL) divergence as a measure
for the relative distance between the modeled distributions and simulated data.
In our interference model development, we consider the interference at a BS in a MIMO
uplink scenario, but the results are also applicable to downlink. The developed interference
models apply to all transmit beamforming techniques, as long as the beamforming vector is
independent of the interference channels. We verify the modeled interference power distribution
against network simulation for a wide range of realistic channel propagation parameters.
C. Main Results and Contributions
We develop out-of-cell interference models that represent the interference power as a random variable with a known, parametrized probability density function. This representation is
important for cellular network performance evaluation, prediction and design. We further apply
the developed models in analyzing the capacity and outage performance of a MIMO cellular
network employing joint transmit and receive beamforming.
4
The main contributions and novelties of this paper are a new analytical model for interference
power distribution and methods for estimating its parameters for both rich scattering and limited
scattering environments, which are summarized as:
1) We propose the use of two distributions to model interference power, each distribution
characterized by two parameters: the inverse Gaussian (IG) as a light-to-heavy tailed
distribution and the inverse Weibull (IW) as a heavy tailed one. Further, we propose a
novel model as a mixture of these two distributions with remarkably good fit to simulation
data while having only three parameters.
2) We apply three approaches to estimate the interference models parameters: moment
matching (MM), individual distribution MLE, and mixture distribution MLE. In the MM
approach, a simple matching of the first two interference power moments is used. In the
other two approaches, a combination of MM and MLE techniques is used in designing
an iterative EM algorithm which maximizes a log likelihood function of the interference
data.
3) We propose the use of the relative entropy or Kullback-Leibler distance from information
theory to measure the goodness of each model. This metric measures the relative distance
between a modeled distribution and reference (simulated) data, which gives a good
indication of how far the proposed interference model is from the referenced interference.
4) We provide simple polynomial functions with fitted coefficients to express the mixture
MLE model parameters in terms of channel characteristics, including the path loss exponent and shadowing standard deviation. These polynomials can be used as a simple
representation of network interference in complex system-level simulations.
5) We apply the interference models to evaluate the performance of a cellular network with
joint transmit and receive beamforming. User’s outage probabilities using the MM and
mixture MLE interference models are compared to stochastic geometry based system
simulation.
6) Our proposed mixture model shows excellent fit in both interference distribution and
network performance evaluation for a wide range of channel propagation parameters,
including path loss exponent from 2 to 5 and shadowing standard deviation from 0 dB
to 9 dB.
This work can be extended to model interference in more complicated propagation environments, including those with a probabilistically parametrized LOS and NLOS channel model, and
to more heterogeneous network deployment such as those involving relaying nodes [13]–[15].
5
The importance of our developed mixture model comes from the fact that we can fit the
data set of out-of-cell interference into a simple mathematical model, a distribution with known
probability density function. This mathematical model can be used to study the performance of
a cellular network where model parameters can be specifically tailored for each network setting.
These parameters can be derived by applying the MLE method to interference data obtained
either via simulation (such as those based on stochastic geometry networks, including point
processes and random shapes), or via actual measurement campaigns. The flexibility of adapting
the parameters of this mixture distribution makes it a versatile tool in modeling interference.
II. M ODELS FOR NETWORK AND CHANNEL PROPAGATION
A. Network model based on stochastic geometry
We consider a cellular system consisting of multiple cells with an average cell radius R0 . Each
cell has a single BS that is equipped with NBS antennas and serves multiple user equipments
(UEs). Each UE is equipped with NUE antennas and uses a distinct resource block in each cell,
hence there is no intra-cell interference. However, each UE suffers from out-of-cell interference
due to frequency reuse in all other cells. In this paper, we denote all transmissions on the same
resource block from all cells other than the cell under study as the out-of-cell interference.
Due to the irregular structure of current and future cellular networks, we employ stochastic
geometry to describe the network. We consider uplink transmissions in this paper, although
results are also applicable to downlink. We model the active UEs in different cells contending
for the same resource block and causing interference to each other as being distributed on a
two-dimensional plane according to a homogeneous and stationary Poisson point process (PPP)
Φ1 with intensity λ̃1 = ηλ1 , where η represents the user density factor and can be varied.
Furthermore, under the assumption that each BS serves a single mobile in a given resource
block, we adopt the model in [16] which places each BS uniformly in the Voronoi cell of its
served UE. According to this model, the distance, rc , between each BS and its served UE is
√
Rayleigh distributed and has an average value of Rc = 1/(2 ηλ1 ). Throughout this paper, we
assume that λ1 is fixed and that Rc changes with the parameter η.
In order to develop a model for the out-of-cell interference power in a stochastic geometry
based network, we consider a cell under study with a typical radius Rc in a field of active UE
interferers with intensity λ̃1 . As such, our network model consists of a typical cell under study
centered at its BS and surrounded by Voronoi cells of other active uplink UEs who interfere
with the considered BS at the origin. Our goal is to model this out-of-cell interference to the
considered BS receiver, taking into account channel propagation features as discussed next.
6
B. Channel propagation model
We consider two channel models for two different environments, the rich and the limited
scattering environments, which are typical for signal propagation in the microwave and mmWave
bands, respectively.
For the rich scattering environment, we consider a complete channel model with shadowing,
path loss and small scale fading. As such, we express a typical MIMO channel with Tx-Rx
distance r in the following form:
H=
p
l (r)H̃
(1)
where H̃ is a random matrix that captures the effects of small scale fading and which can be
modeled as i.i.d. CN (0, 1); and function l (r) captures the large scale fading which includes
both path loss and shadowing. The large scale fading function l (r) is modeled as a log-normal
random variable multiplied with the pathloss as a function of the distance r as follows:
l (r) = Ls βr −α = Ls lp (r),
(2)
where Ls ∼ log(0, σSF ) is a log-normal random variable with standard deviation σSF dB; α is
the pathloss exponent; and β is the intercept of the pathloss formula. The intercept represents
the reference attenuation point that determines the tilt of the path loss model [17].
For the limited scattering environment, as typical in mmWave, the signal propagation charac-
teristics differ considerably. The path loss in mmWave systems is severe with distance-dependent
LOS and NLOS propagation, often modeled as a probabilistic function of mixed LOS and NLOS
[4], [13]. Further, mmWave signals can be severely vulnerable to shadowing and blockage effects
[18]. For limited scattering, however, we only focus on the NLOS interference component in
this paper, leaving the composite interference from both NLOS and LOS components as a future
work. As such, the large scale fading can be modeled as having a single slope as in (2).
The main feature, due to limited scattering, is that the channel becomes highly directional.
Specifically, for the small scale fading, a directional H̃ can be defined for K scatterers as
K
X
H̃ =
ak1 urx θkrx1 u∗tx θktx1
(3)
k1 =1
where K denotes the number of channel path clusters; ak1 is the complex channel coefficient
of the k1th path cluster, assuming (2-D) beamforming; θkrx1 and θktx1 respectively are the horizontal
angles of arrival (AoA) and departure (AoD) of the k1th path cluster; and urx (·) ∈ CNBS and
utx (·) ∈ CNUE are the vector response functions for the BS and UE antenna arrays.
Based on these channel models, we can express the channels of the direct link between the
considered active UE and its BS and the interfering links from other active UEs as follows:
7
H0 =
q
l (kp0 k2 )H̃0 ,
Hk =
q
l (kzk k2 )H̃k ,
(4)
where p0 and zk are vectors representing the 2-D location of the considered active UE and the
k th interfering UE in Φ1 with respect to the origin (i.e. the considered BS). Here H0 is the direct
channel from the considered active UE and Hk is the channel from the k th interfering UE in
Φ1 to the considered BS.
Through a number of recent mmWave channel measurement campaigns, channel parameters
for the model in (2) have been identified [18]–[21]. For channel propagation at 28 GHz, typical
values of the path loss exponent range from 1.68 to 2.55 in LOS and from 2.92 to 5.76 in NLOS
environments. Shadowing standard deviation ranges from 0.2 dB to 8.66 dB in LOS and from
8.7 dB to 9.7 dB in NLOS. In this paper, we develop an interference model that works well for
these ranges of channel parameters and beyond, as shown later in the numerical section.
III. I NTERFERENCE F ORMULATION UNDER MIMO B EAMFORMING
In this section, we present the MIMO beamforming signal model and formulate the out-ofcell interference at the considered BS under transmit beamforming from all active UEs in rich
and limited scattering environments. Beamforming has been adopted as an essential technique
for mmWave communications to overcome the huge propagation loss at high mmWave carrier
frequencies [2], [22]. We then formulate the per-user achievable rate and outage probability under
an example of dominant mode joint transmit and receive beamforming in the rich scattering
environment, and directional analog beamforming in the limited scattering environment.
A. MIMO Beamforming Signal Model
We now describe the signal model for single-stream MIMO beamforming, i.e., no spatial
multiplexing. We model the received signal {y0 ∈ CNBS ×1 } at the considered BS as
y0 = H0 x0 + v0 + z0
(5)
where x0 ∈ CNUE ×1 is the transmitted signal vector from the considered UE; z0 ∈ CNBS ×1 is an
i.i.d. noise vector distributed as CN (0, σ 2 I); H0 ∈ CNBS ×NUE is the considered UE-to-BS channel
matrix; and v0 ∈ CNBS ×1 represents the interference vector received at the considered BS from
interfering UEs in all other cells.
The transmit signal vector under beamforming from each user can be generally described as
p
x = w PU.
(6)
where U is a standard Gaussian signal with zero mean and unit variance; P represents the total
power allocated to the active UE within a single transmission period; and w ∈ CNUE ×1 is the unit
norm beaforming vector at the active UE. We will use this signal model with the appropriate
subscript to denote the transmit vector from an active UE, either intended or interfering one.
8
B. Interference Signal Formulation
For the purpose of modeling network-wide interference, we emphasize that the distribution
of interference is independent of the beamforming scheme employed. The only condition is that
the transmit and receive beamforming vectors, which are designed for the intended channel, are
independent of the interference channels. This condition is realistic in most practical scenarios.
For the interfering active UE in the k th cell, denote the interference channel as Hk and the unitnorm interfering beamforming vectors as wk ∈ CNUE ×1 , which can represent either a baseband
digital precoder or analog RF beamforming or a hybrid of both of these forms.
1) Rich Scattering Environment: Given that wk depends on the direct channel between the
k th active UE and its associated BS, it is independent of the interference channel Hk . We can
denote the effective interference vector from the k th interfering active UE as
q
q
hk = Hk wk = l (kzk k2 )H̃k wk = l (kzk k2 )h̆k ,
(7)
where h̆k is a random vector with i.i.d. elements as CN (0, 1). This follows from the fact that
the inner product of a vector of i.i.d. complex Gaussian variables (the column of H̃k ) with
an arbitrary unit norm vector is a complex Gaussian variable of the same distribution as each
element in the original vector.
The interference vector at the receiving antenna array as presented in (5) can be expressed as
X
Xq
Pk l (kzk k2 )h̆k Uk ,
(8)
v0 =
hk xk =
k6=0
k6=0
where the summation is over all interfering users.
2) Limited Scattering Environment: In this environment, we assume directional analog beamforming only and consider the composite effect of both transmit beamforming vector, wk ,
and receive combining vector, w̄0 ∈ CNBS ×1 , on the received interference power. We further
approximate the BSs and UEs antenna array patterns by a sectored antenna model with M and
m representing the main and back lobes gains, and θM representing the beamwidth of the main
lobe. Then, given the independence between wk , w̄0 , and Hk , the received interference signal
after both transmit and receive beamforming is
Xq
X
∗
∗
w̄0 Hk wk xk =
v̄0 = w̄0 v0 =
Gk Pk l (kzk k2 )ak Uk ,
k6=0
(9)
k6=0
where ak is the effective channel coefficient and Gk is the effective antenna gain from the k th
interfering active UE and can be modeled, assuming channel clusters are well separated, as
9
Gk =
MM,
M 2
with probability p1 = K( θ2π
)
mM,
M
M 2
M
)( θ2π
) + (K 2 − K)( θ2π
)
with probability p2 = 2K( 2π−Kθ
2π
mm,
M 2
with probability p3 = ( 2π−Kθ
)
2π
(10)
where K is the number of clusters (the special case of a single cluster channel model, K = 1,
is discussed in [13], [23]). Note the difference between the formulations in (8) and (9) in that
(8) is an interference vector at the receiving antenna array, before receiver processing, whereas
the interference in (9) applies after receive beamforming.
3) Modeling the Interference: For each spatial and channel fading realization, based on the
large number of interferers, the out-of-cell interference can be modeled as a complex Gaussian
random vector for (8) or a complex Gaussian scalar for (9) with zero mean and covariance Σ0
or variance σ02 . These covariance and variance, however, are random and dependent on user locations and channel fading. The matrix Σ0 is symmetric with diagonal elements representing the
interference power at each receiving antenna element, and the off-diagonal elements representing
the correlation among interference signals at different antenna elements. The variance σ02 on the
other hand represents the total interference power after receiver beamforming. We analyze the
off-diagonal and diagonal elements of the covariance matrix separately in Sec. V-A. We show
later in the numerical analysis section that the correlation among antenna elements is weak and
negligible; hence we focus on modeling the interference power elements in Σ0 and σ02 .
C. Achievable Rate and Outage Probability Formulation
The main goal of this paper is to model the interference as formulated above. In order to
evaluate this model, we apply it in capacity analysis. In this section, we formulate the capacity
and show how it is affected by interference.
To establish the per-user capacity in the rich scattering environment, we assume the transmit
beamforming vector w0 as the right singular vector corresponding to the dominant mode of H0 ,
and consider either an interference-aware (IA) or interference-unaware (IU) receive combining
vectors. The noise plus interference in the received signal in (5) can be treated as a Gaussian
random vector with a random covariance matrix R0 , which is dependent on interfering nodes
locations and their interference channels. This covariance matrix can be expressed as
R0 = Σ0 + σ 2 I,
(11)
where Σ0 is the interference covariance matrix discussed in Sec. V-A. Applying a receive
combining vector sd ∈ CNBS ×1 to the received signal y0 in (5) to get
√
ỹ0 = s∗d y0 = s∗d H0 w0 P U0 + s∗d R0 z̃0 ,
10
where z̃0 is a vector of i.i.d. zero mean and unit variance complex normal entries, we can then
write the per-user capacity for the given combining vector sd as
|s∗d H0 w0 |2
P .
C = log 1 + ∗
sd R 0 sd
(12)
For interference-unaware (IU) combining, we design the receive combining vector without
knowledge of the interference statistics – particularly its covariance. As such, the IU combining
vector is chosen as the left singular vector corresponding to the dominant mode of H0 , sd =
v1 (H0 ), resulting in the following per-user capacity:
|λmax (H0 )|2
CIU = log2 1 + 2
P ,
σ + v1∗ Σ0 v1
(13)
where λmax (H0 ) is the maximum singular value of H0 .
As seen in capacity formula (13), with IU combining, the effective post-combining interference
is σI2 = v1∗ Σ0 v1 . Since Σ0 is a random covariance, this post-combining interference is also
random. Its distribution, however, can be characterized in the same way as the pre-combining
interference power at each antenna element, i.e., the diagonal elements of Σ0 , as stated in Lemma
1. As a result, the IU capacity depends only on the interference power at each antenna element
but not on the correlation between interference at different antenna elements.
Lemma 1 (Interference Power Distribution for IU Combining). The post-combining interference
power v1∗ Σ0 v1 is a random variable that has the same distribution as a diagonal element of Σ0 .
Proof: See Appendix A for detail.
For interference-aware (IA) combining, the receive combining vector can be designed as a
−1/2
1/2
function of the interference statistics. Define s̃d = R0 sd ; then replacing sd = R0
it is straightforward to see that the vector which maximizes (12) is
interference-aware per-user capacity can then be obtained as
CIA = log2 1 + ks̃⋆d k2 P .
s̃⋆d
= (R0 )
s̃d into (12),
−1/2
H0 w0 . The
(14)
This IA capacity depends on the complete interference covariance Σ0 ; it is therefore dependent
on both the interference power at each antenna element and the correlation between interference
at different antenna elements.
In the limited scattering environment, we can express the per-user capacity, assuming perfect
beam alignment between the UE and its serving BS [13], as
M 2 |a0 |2 l (kp0 k2 ) P
,
C = log2 1 +
σ 2 + σ02
(15)
11
where M is the antenna array main lobe gain; a0 is the effective complex channel coefficient;
p0 is the location of the active UE with respect to the BS; and σ02 is the total interference power.
Achieving the capacity in (15) does not require knowledge of the interference covariance matrix.
The rates in Eqs. (13) – (15) are random quantities due to channel fading and interference.
We investigate the outage probability at a given target rate as a performance metric, defined as
po (RT ) = P{C < RT },
(16)
where RT is the target rate at which the active UE under consideration transmits.
IV.
I NTERFERENCE M ODELING A PPROACHES
AND
F ITNESS M ETRIC
In wireless cellular networks at system level simulations, the most difficult part is to simulate
interference from other cells to the cell under study. This is mainly due to the processing and
memory power required. These simulations can be substantially simplified if we can characterize
the interference by a simple parametrized distribution that represents a good fit to the actual
interference. In this section, we introduce the approaches we use to model the interference power
terms on the diagonal of Σ0 and σ02 . Then, we introduce the information-theoretic based metric
used to measure the fitness of each of those models. Finally, we discuss candidate distributions
considered for the interference model, the Inverse Gaussian and Inverse Weibull distributions.
A. Interference Modeling Approaches
We introduce three different approaches for modeling the distribution of the interference power
at the receive antenna. These approaches are Moment Matching (MM), individual distribution
MLE (individual MLE), and mixture distribution MLE (mixture MLE).
1) MM Approach: In this approach, we leverage tools from stochastic geometry to analytically
derive the first two moments of each interference power term. Then, we match these two moments
with those of the two-parameter candidate distributions to estimate these two parameters and
hence specify the distribution. This approach is the simplest and does not require complex
computations to estimate the distributions parameters. Thus, it is often used in the literature to
produce analytically tractable models [10], [24], [25]. We will show, however, that this approach
may not produce a good fit for the model.
2) Individual MLE Approach: In this approach, to simplify parameter estimation but improve
the model fitness, we use a combination of MM and MLE, a powerful estimation technique which
is often used in signal detection and estimation [26]. Here we only match the first moment using
the analysis results derived based on stochastic geometry, which helps in either determining one
12
of the parameters for the candidate distribution, or getting one parameter in terms of the other.
Then, we use MLE to determine the other parameter of the candidate distribution.
3) Mixture Approach: This approach is the most complex but also most accurate, in which we
model the interference probability density function as a mixture of the two candidate distributions.
We again use a combination of MM and MLE techniques to simplify parameter estimation. In
this approach, we do not determine the optimal values of the distributions parameters in closed
form, but design an efficient Expectation Maximization (EM) iterative algorithm to identify these
paramters, as discussed in Sec. VI-C.
B. KL Divergence Fitness Metric
In order to measure the developed models goodness or fitness, we use a concept from information theory as the relative entropy [27]. In this paper, we measure the relative entropy
between a proposed model and the data obtained by simulating the network based on stochastic
geometry as discussed in Sec. II-A, but it should be noted that other data such as those obtained
from measurement campaigns can also be used to test our models via relative entropy. The
relative entropy of a distribution P with respect to another distribution Q, also called KullbackLeibler divergence DKL (P ||Q), reflects the difference, or distance, between these two probability
distributions. The distribution P usually represents the true distribution of data or observations,
while Q is typically used to represent a modeled distribution or an approximation to the true
distribution P . This relative entropy is defined for continuous probability distributions as
Z ∞
p(x)
dx
(17)
DKL (P ||Q) =
p(x) log
q(x)
−∞
where p(x) and q(x) denote the probability density functions of P and Q, respectively. This
definition can also be described as the expectation of the logarithmic difference between the
probabilities P and Q. Kullback-Leibler divergence metric has been used to measure the accuracy
of modeling composite fading and shadowing wireless channels such as the Rayleigh-Inverse
Gaussian and mixture Gamma distributions [28], [29].
C. Candidate Distributions for the Interference Model
We consider two candidate distributions for interference modeling: the inverse Gaussian (IG)
and the inverse Weibull (IW), both are characterized by two parameters. The IG distribution
is also known as the Wald distribution and is usually used to model nonnegative positively
skewed data. For example, a composite Rayleigh-IG distribution is used to approximate the
composite Rayleigh-Lognormal distribution in [28]. The name "Inverse Gaussian" comes from the
13
inverse relationship between the cumulant generating functions of these distributions and those
of Gaussian distributions. It is a two-parameter family of continuous probability distributions
which is specified by a shape parameter λ and a scale parameter µ. Given an IG random
variable γIG (µ, λ), its probability density function is defined as [30]–[32]
r
λ
−λ(t − µ)2
fγIG (t|µ, λ) =
,
t > 0, λ, µ > 0.
exp
2πt3
2µ 2 t
The mean and variance of γIG (µ, λ) can be written in the following form [32]
µ3
E[γIG ] = µ,
var[γIG ] =
.
λ
(18)
(19)
The IW distribution is also known as the complementary Weibull distribution and is usually
used to model nonnegative positively skewed data that exhibits a long right tail. It is also specified
by two parameters, a shape parameter c and a scale parameter b (or λIW in some text). Given
an IW random variable γIW (b, c), its probability
function is defined as [33], [34]
( density
−c−1
−c )
c t
t
fγIW (t|b, c) =
, t ≥ 0, b, c > 0.
(20)
exp −
b b
b
The mean and variance of γIW (b, c) can be written in the following form [33]
2
1
var[γIW ] = b 2 Γ(1 − ), if c > 2,
E[γIW ] = bΓ(1 − ), if c > 1,
c
c
R ∞ t−1 −x
where the Gamma function Γ(t) is defined as Γ(t) = 0 x e dx.
(21)
In a recent and independent work, the Gaussian, Gamma, Inverse Gamma, and Inverse Gaus-
sian distributions are considered as models for the interference distribution in a stochastic
geometry based spatial network [35]. Different point processes are considered in this work,
including PPP, Strauss process, and Poisson cluster process. The inverse Gaussian distribution
is shown to be a reasonable model for the interference created by BSs distributed according to
the Strauss point and Poisson cluster processes. Further, both the Gamma and IG were shown
to be suitable for the PPP network geometry. This prior work, however, considered a bounded
path loss model l(r) = (1 + r α )−1 and Rayleigh fading with no shadowing.
In our work, we find through simulations that the Gaussian, Gamma and Inverse Gamma distributions are all poor models for the interference power when shadowing is present. Shadowing
introduces large variation in the interference power and causes a medium to heavy tail, which
none of these distributions are capable of capturing. Thus we need to find other distributions that
can more closely model heavy tail interference. As such, we examine the Inverse Gaussian and
also introduce the Inverse Weibull as a good candidate for parameterized heavy tail distributions
that can capture shadowing with large standard deviation.
14
V.
A NALYTICAL M OMENTS
OF I NTERFERENCE AND
M OMENT M ATCHING M ODELS
In this section, we analytically derive the first two moments of the diagonal and off-diagonal
elements of the interference covariance matrix. We then use the derived moments of the interference power at each antenna element in rich scattering, and the post receive beamforming
interference power σ02 in the limited scattering environment, to develop the MM interference
models.
A. Analytical Moments of Interference
1) Interference Power: In the rich scattering environment, the diagonal elements of the interference covariance matrix Σ0 represent the interference power at each antenna element and
they are uncorrelated with the off-diagonal elements. A correlation between any two of these
diagonal elements exists and can be determined similarly to the second moment of the offdiagonal elements (Lemma 3). However, since this correlation is not strong as shown in the
numerical analysis in Sec. VII-A, we model the diagonal elements of Σ0 as a random vector of
independent elements. Each of these diagonal elements can be generally expressed as
X
l (kzk k2 )gk Pk .
q0 =
(22)
k>0:zk ∈Φ1
where gk are all i.i.d. exp(1) and represent the power gain of the small-scale Rayleigh channel
fading from the k th interferer.
Similarly, in the limited scattering environment, the interference power can be expressed as
X
(1)
(2)
(3)
σ02 = q̄0 =
Gk l (kzk k2 )gk Pk = q̄0 + q̄0 + q̄0 ,
(23)
k>0:zk ∈Φ2
where we assume that the interferers are distributed according to PPP Φ2 with intensity λ2 . Using
the thinning property of random point processes [23], from Φ2 , we define three independent PPPs
Ni , i ∈ {1, 2, 3} with intensities τi = pi λ2 . Then, we can define the normalized interference power
terms as
(i)
X
q̄0
=
l (kzk k2 )gk Pk ,
G(i) k>0:z ∈N
k
where G
(i)
(24)
i
∈ {MM, Mm, mm}. Comparing Eqs. (22) and (24), we note that the distribution of
the normalized interference power for each thinned process in (24) can be characterized as that
of the interference power in (22) by matching the intensity τi to the intensity of the interferers in
the rich scattering environment, λ̃1 . Hence, we focus on the characterization of the rich scattering
environment interference power in (22), based on which the three separate interference power
terms for the limited scattering environment in (23) can be derived.
15
Lemma 2 characterizes the first two moments of the out-of-cell interference power q0 at each
receiving antenna.
Lemma 2 (Interference Power Moments). For network-wide deployment of MIMO transmit
beamforming, the out-of-cell interference power at each antenna of the interfered receiver has
the following first and second moments:
E [q0 ] = νζ1 Rc2−α ,
where
var [q0 ] = 2ν̃ζ2 Rc2(1−α) .
ζ2 = 2Pk2,
ζ 1 = Pk ,
ν=
(25)
2πλ1 βE[Ls ]
,
α−2
ν̃ =
πλ1 β 2 E[L2s ]
.
2(α − 1)
(26)
Proof: See Appendix B for details.
2) Interference Correlation among Different Antennas: The first moment of the off-diagonal
elements of the covariance matrix Σ0 represents the direct correlation among interference signals
at different receiving antenna elements. The second moment, however, represents the correlation
between received interference powers. These off-diagonal elements are expressed as in (47) for
i 6= j in Appendix A. Assuming antenna elements are spaced sufficiently far apart such that
they experience independent fading, Lemma 3 characterizes their first and second moments.
Lemma 3 (Interference Correlation Moments). For network-wide deployment of MIMO transmit
beamforming, the correlation among interference at different receiving antenna elements of the
BS has the following first and second moments:
1
E [q̃0 ] = 0,
E (q̃0 )2 = var [q0 ] .
2
(27)
where var [q0 ] is defined in Lemma 2.
Proof: See Appendix B for details.
The interference correlation coefficient Cxy can then be written as
Cxy =
q̃0
,
E [q0 ]
(28)
Based on Lemmas 2 and 3, we can derive the variance of this correlation coefficient as
2
2
(α − 2)2 E [L2s ]
(α − 2)2 σζSF
var [Cxy ] = E Cxy = κ
e 2 ,
=κ
(α − 1)
(α − 1) (E [Ls ])2
κ=
ζ2
.
8πλ1 ζ12Rc2
(29)
This variance increases with increasing path loss exponent, α, and shadow fading standard
deviation, σSF . We perform numerical analysis of this correlation coefficient in Sec. VII-A and
show that the correlation among interference at different antenna elements is in general weak.
16
B. Moment Matching Interference Models
Given the first two moments of the interference power in Lemma 2, we can use moment
matching as the simplest and most straightforward method to estimate the parameters of the
candidate distributions. This method, however, may not produce the best fit. Specifically, we use
the two analytically derived moments to estimate the (µ, λ) and (b, c) parameters of the IG and
IW distributions. The fitted distribution can then be used to represent the marginal distribution
of the interference power at each receiving antenna element, as stated in Lemmas 4 and 5 next.
Lemma 4 (Estimation of IG Parameters using MM). Given the derived moments in Lemma 2,
the shape and scale parameters λ and µ of an IG model for the interference power γIG [µ, λ]
can be estimated as
µ = E[q0 ],
(E[q0 ])3
.
λ=
var[q0 ]
(30)
Proof: Follows directly from the definition of λ and µ in (19).
Lemma 5 (Estimation of IW Parameters using MM). Given the derived moments in Lemma 2,
the shape and scale parameters c and b of an IW model for the interference power γIW [b, c]
can be estimated, assuming c > 2, by solving the following equation:
Γ(1 − 2/c)
E[q0 ]
var[q0 ] −
− 1 (E[q0 ])2 = 0,
b=
.
2
Γ (1 − 1/c)
Γ(1 − 1/c)
(31)
If the above equation gives a value of c < 2, then set c = 2.01 and compute b accordingly.
Proof: Follows directly from the equations for b and c in (21).
The equation in (31) can be solved numerically to obtain c, which is then used to obtain b.
Remark 1. For the expressions in (21), we note that the mean is infinite for c < 1 and that the
variance is infinite for c < 2. This is a consequence of the heavy right tail of the IW distribution.
Hence, for the IW moment matching in Lemma 5, we assume that c > 2.
Remark 2. For the path loss exponent value α = 2, we use numerically evaluated mean and
variance to perform moment matching since the expressions in Lemma 2 can not be evaluated
analytically at this value of α.
VI.
I NDIVIDUAL
AND
M IXTURE MLE I NTERFERENCE M ODELS
In this section, we apply maximum likelihood estimation to compute the model parameters. We
discuss both the individual MLE and mixture interference models and develop an EM algorithm
17
to estimate the model parameters. Maximum likelihood is one of the dominant methods of
estimation in statistics. A typical MLE problem is to estimate the parameter set θ of a candidate
distribution p(y|θ) from a given data set y and can be expressed as an optimization as follows:
arg max log p(y|θ)
θ
subject to θ ∈ Θ,
(32)
where log p(y|θ) is the log likelihood of the observed data y, and Θ is a closed convex set.
A. Individual MLE Interference Models
In individual MLE interference model, we use only a single distribution from the candidates
to model the interference power at each antenna. We use the maximum likelihood optimization
problem in (32) to estimate the parameters of each distribution. However, to simplify the
estimation, we use moment matching first. In moment matching first, we match the mean, µY ,
of the observed data set to that of each of the candidate distributions. This matching results in
the scale parameters µ and b for the IG and IW distributions as
µY
µ = µY ,
b=
.
(33)
Γ(1 − 1/c)
Next, we replace these expressions of the scale parameters into the log-likelihood function in
(32) for each candidate distribution. In the case of IG distribution, this substitution results in a
single-parameter log-likelihood function as follows
1
λ (y − µY )2
1
log λ − log 2πy 3 − 2
= log fγIG (y|λ).
(34)
log fγIG (y|µ, λ) =
2
2
2µY
y
Hence, the MLE problem in (32) becomes very simple and has a closed form solution for λ.
Similarly, in the case of IW, the substitution results in a single-parameter log-likelihood function
c
µcY c
µY
log fγIW (y|b, c) = log
−
y −c = log fγIW (y|c). (35)
Γ(1 − 1/c)c y (c+1)
Γ(1 − 1/c)
The derivation and final expressions of the optimal parameters, λ and c, that optimize the fit of
each candidate distribution are presented in Sec. VI-C.
B. Mixture MLE Interference Model
In our mixture model, the probability density function (pdf) of the interference model is now
a mixture of both IG and IW distributions as defined in the next theorem.
Theorem 1. Given a data set Y , the pdf of the mixture interference model can be written as
fY (y|θ) = w1 fγIG (y|λ) + w2 fγIW (y|c),
(36)
P
where {w1 , w2 : 2i=1 wi = 1} are the weight parameters for mixing the IG and IW distributions
and fγIG (y|λ) and fγIW (y|c) are as given in Eqs. (34) and (35) with the scale parameters ob-
tained through moment matching first as in (33). This mixture model is a 3-parameter distribution
with the parameter set as θ = {w1 , λ, c}.
18
Given the pdf of the mixture model in Theorem 1, we note that the individual MLE models
are special cases of this mixture model. By setting w1 = 1 , we have the IG individual MLE
model and by setting w2 = 1, we have the IW individual model. Subsequently we focus on the
mixture model parameters estimation, which can be applied to the individual models as well.
C. MLE Models Parameters Estimation
We have three parameters to be optimized in the mixture model which is significantly more
than the single parameter in the individual MLE model. This mixture leads to a more complex
MLE optimization problem and the EM algorithm [36] becomes an appealing approach. Here, we
identify the complete data X as the observed data Y plus some hidden data Z, i.e. X = (Y, Z),
where Y is the observed set of points that we model by a weighted IG and IW distribution and
Z ∈ {1, 2} is a discrete random variable representing the assignment of each data point to the
two candidate distributions. Then, we can define the pdf of the complete data X ∈ R+ as
fX (x|θ) = fX (Y = y, Z = i|θ) = wi φi (y|θi ),
(37)
where φ1 (y|θ1 ) = fγIG (y|λ) and φ2 (y|θ2 ) = fγIW (y|c).
Next, we present the algorithm for estimating the parameters of the MLE interference models.
We focus on the representative case of the mixture MLE model, which includes the individual
MLE models as special cases. Given n observations, in order to develop the EM algorithm to
estimate the parameters of our mixture model, we use Proposition 1 below [37] which can be
easily verified using the independence assumption, the fact that the observed data Y = f (X) is
a deterministic function of the complete data X for some function f , and Bayes’ rule.
Proposition 1. [37] Let the complete data X consist of n i.i.d. samples: X1 , X2 , ..., Xn , which
Q
satisfies f (X|θ) = ni=1 f (Xi |θ) for all θ ∈ Θ, and let yi = f (xi ), i = 1, 2, ..., n, then
n
X
(m)
Q(θ|θ ) =
Qi (θ|θ(m) )
(38)
i=1
(m)
We also define a responsibility function γij , which represents our guess at the mth iteration
of the probability that the ith sample belongs to the j th distribution component, as
(m)
(m)
γij
= p(Zi = j|Yi = yi , θ
(m)
(m)
wj φj (yi |θj )
) = P2
l=1
(m)
wl
(m)
φl (yi |θl
)
,
i ∈ {1, 2, ..., n}, j ∈ {1, 2}
(39)
where θ1 and θ2 represent the shape parameters λ and c, respectively.
An EM algorithm includes two steps: the E-step to calculate the conditional expectation of
the log likelihood of the complete data, and the M-step to maximize this conditional expectation
function. Lemma 6 provides the final expression for the EM algorithm Q-function in the E-step.
19
Lemma 6. Let the complete data X consist of n i.i.d. samples: X1 , X2 , ..., Xn , which satisfies
Q
f (X|θ) = ni=1 f (Xi |θ) for all θ ∈ Θ, and let yi = T (xi ), i = 1, 2, ..., n, then we have
Q(θ|θ
(m)
) =
2
X
(m)
nj
j=1
(m)
(m)
where nj
=
n
1 (m)
λ X (m) (yi − µY )2
log wj + n1 log λ − 2
γ
2
2µY i=1 i1
yi
(m)
(m)
+ cn2 log µY − cn2 log Γ(1 − 1/c) + n2 log c
c X
n
n
X
µY
(m)
(m)
γi2 yi−c ,
− c
γi2 log yi −
Γ(1
−
1/c)
i=1
i=1
(40)
(m)
Pn
i=1
γij , j ∈ {1, 2}; and θ = {w1 , λ, c}.
Proof: See Appendix C for details.
In the M-step, to update our estimate of the parameter θ, we solve the following optimization
problem for the optimal θ∗ :
arg max Q(θ|θ
θ
(m)
)
subject to
2
X
j=1
The optimal θ∗ is then found as in Theorem 2.
wj = 1, wj ≥ 0 , j ∈ {1, 2}.
(41)
n
o
(m+1)
Theorem 2. The optimal new estimate of θ(m+1) = w1
, λ(m+1) , c(m+1) that maximizes the
Q-function in (40) at the mth iteration are determined as follows:
(m)
(m)
nj
n1 µ2Y
(m+1)
(m+1)
, j ∈ {1, 2},
λ
= Pn (m) (y −µ )2
wj
=
i
Y
n
i=1 γi1
yi
(42)
and c (m+1) is obtained by solving
the following equation numerically:
"
#
0
n
X
µY
1
1
(m)
γi2 log yi
log
=
1−ψ 1−
−
+
c
Γ(1 − c1 ) c
i=1
c
#
"
n
X
yi Γ(1 − c1 ) ψ(1 − c1 )
µ
Y
(m)
−c
+
+
log
γi2 yi
µY
c
Γ 1 − 1c
i=1
(m)
n2
(43)
Proof: See Appendix D for details.
Algorithm 1 summarizes the EM algorithm for the mixture distribution interference model.
In the case of IG individual MLE model, we can set w1 = 1 and the resulting Q-function in
(52) represents the log-likelihood function of the IG distribution. Hence, problem (41) results in
the MLE of a single parameter, λ. The optimal value of λ can be obtained from (42) by setting
(m)
n1
(m)
= n and γi1 = 1. Similarly, the optimal value of c for the IW individual MLE model can
(m)
be obtained from the numerical solution of (43) after setting w2 = 1, n2
(m)
= n and γi2 = 1.
20
1.
(0)
Initialize wj , λ(0) , and c(0) , j ∈ {1, 2}, and compute the initial log-likelihood:
Initialization:
L(0)
(m)
=
n
1X
(0)
(0)
log w1 fγIG yi |λ(0) + w2 fγIW yi |c(0) .
n i=1
(m)
as given in Eq. (39) and nj
2.
E-step:
Compute γij
3.
M-step:
Compute the new estimate for wj
(m+1)
=
Pn
i=1
(44)
(m)
γij .
, λ(m+1) , and c(m+1) , j ∈ {1, 2} as given
in Eqs. (42), and (43), respectively.
4.
Convergence check:
Compute the new log-likelihood function
L(m+1) =
n
1X
(m+1)
(m+1)
fγIW yi |c(m+1) .
log w1
fγIG yi |λ(m+1) + w2
n i=1
(45)
Return to 2 if |L(m+1) − L(m) | ≥ δ for a preset threshold δ; Otherwise end the algorithm.
Algorithm 1. MLE Models EM Algorithm
D. Functional Fitting of MLE Model Parameters to Channel Propagation Characteristics
By performing the MLE parameters estimation, we observe that these parameters can be fitted
directly to rather simple functions of the channel features, including the shadowing variance and
path loss exponent. These functions present excellent fit with simulated data for the range of
mmWave propagation parameters as measured in recent campaigns [18]–[21].
Corollary 1 presents the polynomial function form that we use to fit the MLE parameters to
the channel shadowing variance. Note the same functional form applies to all three parameters
of the mixture MLE model, each parameter with a different set of fitting coefficients.
Corollary 1. For a given path loss exponent, the parameters of the mixture MLE algorithm can
be estimated directly as a function of the shadowing standard deviation σSF as
3
2
log10 (θ) = a3 σSF
+ a2 σSF
+ a1 σSF + a0
(46)
where θ represents any of the parameters {w1 , c, λ}. The function coefficients for each parameter
are obtained via numerical fitting based on the EM estimated parameter values. Parameters b
and µ can then be obtained analytically based on Eq. (33) and Lemma 2.
In Appendix E, we provide representative values of the functional coefficients fitted against
simulation data based on stochastic geometry.
Similarly, the MLE parameters can be fitted to functions of the channel path loss exponent
using a simple least square optimization procedure. Corollary 2 presents the polynomial function
forms that we use to fit the MLE parameters to the path loss exponent.
Corollary 2. At a fixed value of the shadowing variance, the IG weight w1 and the IW shape
parameter c of the mixture MLE model can be estimated directly as a 4th order polynomial
21
function of the path loss exponent α. The logarithm of the IG shape parameter log10 (λ) can be
estimated as a polynomial function of α of the 3rd degree. The coefficients for each function can
be obtained via least square fitting based on the EM estimated parameter values.
Using these functional fits, we can bypass the MLE iterative algorithm in Algorithm 1 and
directly use the table of fitted function coefficients. These functional fits therefore provide a
simple and powerful way of modeling the interference and its direct dependence on the channel
propagation features.
Note that the coefficients examples in Appendix E are verified against simulated data using
stochastic geometry based network settings and need further verification against actual measurement data, for which the coefficient values may change slightly. We expect, however, that the
identified function forms in (46) remain a good fit. To fit the function coefficients, apply the
MLE algorithm to the measured data to identify the interference model parameters. Then fit
these model parameters to the function forms in (46) to identify the coefficient values. Once
the fit is verified, the function with the identified coefficients can be used directly to model
the interference in a new environment with given shadowing variance and path loss exponent,
without the need to measure new data or perform the MLE algorithm again.
VII.
N UMERICAL P ERFORMANCE R ESULTS
In this section, we use simulation data obtained from the stochastic geometry based network
setting in Sec. II-A to numerically verify the validity of the analytical results and proposed
interference models. We first verify the validity of our moment matching interference models.
We then discuss the fit of the MLE individual and mixture interference models. Last, we apply
these interference models to evaluate the outage performance of a cellular system employing
MIMO joint transmit-receive dominant mode beamforming.
Since interference in the limited scattering environment can be derived as the sum of interference components from independent, thinned processes, each of which exhibits similar
characteristics as the interference in a rich scattering environment but with a different interferer
density as in [23], we focus on the rich scattering environment in our simulation. For the
simulations setting, we consider a typical cell of radius Rc = 150 m, which is related to the
active UEs density λ1 as ηλ1 = 0.25Rc−2 , where η = 1 in the typical case. Later we vary η to
examine the effect of user density on performance.
A. Interference Correlation among Different Antenna Elements
This part aims to examine the correlation between interference at different antenna elements
to see if this correlation is negligible. We use a kernel distribution to fit the actual interference
22
0.95
Probability(|C
xy
|≤0.3)
1
0.95
0.9
0.9
0.85
0.8
0.85
0.75
0.7
10
0.8
8
5
6
4
4
σ SF (dB)
3
2
0
α
0.75
2
Fig. 1: The probability of weak interference correlation versus shadowing standard deviation and path loss exponent.
correlation. A kernel distribution is a nonparametric representation of the probability density
function of a random variable. It can be used when a parametric distribution cannot properly
describe the simulation data, or when assumptions about the distribution of the data are to be
avoided, as is the case for testing the fit of our proposed parameterized interference distributions.
We examine the interference correlation via the correlation coefficient defined in Sec. V-A2.
In Fig. 1, we plot the probability that the correlation coefficient is smaller than 0.3, and show
that this probability is significant for a wide range of the shadowing standard deviation and the
path loss exponent. This figure demonstrates that the interference correlation among different
antenna elements is in general weak. As shown later in the system performance evaluation (Sec.
VII-E), the impact of interference correlation on transmission rate differs depending on the
receive combining vector, as discussed in Sec. III-C. For IA combining, interference correlation
can have an impact on capacity; whereas for IU combining, the interference correlation has no
impact and can be ignored. In this paper we do not model the interference correlation, hence
IU combining is more appropriate.
B. Moment Matching Interference Models Evaluation
We now use the relative entropy metric described in Sec. IV-A to measure the goodness of the
MM interference models in Section V-B. In Fig. 2a, we plot the relative entropy versus shadowing
standard deviation σSF at a sample path loss exponent value of α = 3.5. The result shows that
IG approximates the simulated interference distribution well at low values of shadowing standard
deviation σSF and diverges at higher values.
The IW distribution, on the other hand, does not show a good fit at the simulated path loss
value. Our extensive simulation and testing suggests that IW can be a good approximation of
the simulated data distribution only for low values of α (lower than shown in Fig. 2). The
moment matching IW distribution diverges significantly as the path loss exponent increases.
This divergence is caused by the fact that the optimal shape parameter c, as we will see later in
23
1.8
2.5
×10
4
Simulation Data
Moment Matching IW
Moment Matching IG
Moment Matching IG
Moment Matching IW
1.6
1.4
2
1
1.5
PDF
Relative Entropy
1.2
0.8
0.6
1
0.4
0.2
0.5
0
-0.2
0
0
1
2
3
4
5
6
7
8
9
σ SF (dB)
(a) Relative entropy at α = 3.5.
0
0.5
1
1.5
2
2.5
Interference Power (W)
3
×10
-4
(b) Sample PDF at α = 3.5, σSF = 4 dB.
Fig. 2: Relative entropy and sample interference distributions of moment matching interference models. System
parameters: Pmax = 30 dBm, η = 1, δ = 10−6 , σSF = 4 dB, α = 3.5.
this section, obtained at high values of α is close to or less than 2. At this range of c values,
the variance of the IW distribution is infinite and the moment matching model is not applicable.
In Fig. 2b, we show a sample interference power distribution at one antenna element for
α = 3.5 and σSF = 4 dB to visually confirm the match, or rather the mismatch, of the momentmatched IW interference distribution. The moment-matched IG distribution also starts to diverge
from the simulated data at this value of σSF , as indicated in Fig. 2a.
C. Individual and Mixture MLE Interference Models Evaluation
Next, we evaluate the fit of the individual and mixture MLE interference models. Fig. 3 shows
the relative entropy measures for these 3 models against simulated data as well as representative
sample distributions. In Fig. 3a, we see that our proposed mixture model outperforms both the
individual IG and IW models. This result aligns with our expectation since the pdf of the mixture
model is a combination of both the individual IG and IW distributions. Hence, performance of
the mixture model should be at least as good as that of the individual models. The fact that the
mixture model performs better than either individual one is because the relative entropy measure
is a non-linear function of the distribution densities.
Comparing between Figs. 2 and 3 also reveals that all MLE models perform much better than
the MM models: at the worst σSF = 9 dB, the IW distribution under individual MLE has a
maximum relative entropy of about merely 0.23 compared to a maximum of 1.8 under MM.
Further, we see that the proposed mixture model follows the simulated data distribution almost
exactly even at large shadowing: the maximum relative entropy is barely 0.03.
Consistently in all our extensive simulations and testing, the mixture MLE interference model
provides the best fit. This model requires an iterative algorithm to fit the parameters initially,
before fitting the functional coefficients as in Corollary 1. The MM models are attractive from
the simplicity point of view since the model parameters can be determined analytically without
24
7000
0.25
Mixture Model
IG MLE
IW MLE
0.2
Simulation Data
Mixture Model
IW MLE
IG MLE
6000
5000
4000
PDF
Relative Entropy
0.15
0.1
3000
0.05
2000
0
1000
-0.05
0
0
1
2
3
4
5
6
7
8
9
0
σ SF (dB)
1
2
3
4
5
Interference Power (W)
(a) Path loss exponent α = 3.5.
6
7
8
×10
-4
(b) Sample PDF at α = 3.5, σSF = 9 dB.
Fig. 3: Relative entropy and sample interference distributions of the individual and mixture MLE models. System
parameters: Pmax = 30 dBm, η = 1, δ = 10−6 , σSF = 9 dB, α = 3.5.
optimization. However, our simulation results suggest that the IG MM model should only be
used for the range of shadowing standard deviation σSF < 3 dB, at which the relative entropy
of the IG MM model is approximately within 1% of the mixture model.
In Fig. 4, we plot the number of iterations required by Algorithm 1 to obtain the optimal values
of the mixture MLE model parameters versus σSF at two samples of the path loss exponent
α = {2.5, 4}, where the preset threshold value is chosen to be δ = 10−6 , the model weights are
(0)
(0)
initialized as w1 = w2 = 0, and the model shape parameters λ(0) and c(0) are initialized to the
optimal values of the corresponding individual MLE models. The number of iterations required
by the EM algorithm to obtain the optimal mixture MLE model parameters is in general less
than 200. In Fig. 5, we study the sensitivity of Algorithm 1 to the initial values of the parameters.
We consider the sample data obtained at path loss exponent α = 2, shadowing value σSF = 4
(m)
dB, and density factor η = 1. In this figure, we plot the IG weight parameter w1
iteration (m) for different initial values
(0)
w1 ,
at each
which shows that the weight parameter converges
to the same value after at most 170 iterations. We also note from simulations that the likelihood
function is concave in both of the shape parameters c and λ at a fixed value of the weight
parameter, therefore it has a unique maximum. Hence, we can conclude based on numerical
evidence that Algorithm 1 converges to the same optimal parameter values irrespective of their
initializations.
D. Mixture Model Parameters and Function Coefficients Fitting
In this section, we use the EM algorithm discussed in Sec. VI-C. along with Theorem 2
to obtain the mixture model parameters. We then fit the estimated parameters w1 , c, λ, into a
function in terms of σSF as in Corollary 1. After that, Eq. (33) can be used to determine the
parameter b. We can also use (25) to determine the parameter µ directly, but instead we provide
25
0.9
220
w 1(0) = 0.1
α = 2.5
α =4
200
w (0)
= 0.5
1
0.8
w 1(0) = 0.9
180
0.7
160
0.6
w (m)
1
# Iterations
140
120
0.5
0.4
100
0.3
80
0.2
60
0.1
40
20
0
1
2
3
4
5
6
7
8
9
0
10
0
10
1
10
2
Iteration # (m)
σ (dB)
Fig. 4: Number of iteration required to obtain optimal Fig. 5: Sensitivity analysis of the MLE EM algorithm
parameters of the mixture MLE model versus σSF at path
using sample data obtained at path loss exponent and
loss exponent values α = {2.5, 4} using the threshold
shadowing standard deviation values of α = 2, σSF =
value δ = 10−6 .
4 dB and η = 1.
a function fitted to the simulated data for this parameter, since (25) is invalid at α = 2. The
coefficients of the fitted functions are provided for representative values in Appendix E.
In Fig. 6, we plot the mixture model weight factor obtained from both simulation (cross
markers) and functional fit (lines). The estimated parameters and fitted functions show an
excellent fit to the simulated data. The results confirm that larger shadowing results in an increase
in the IW component and a decrease in the IG component in our mixture interference model. The
increase in shadowing standard deviation σSF implies a higher probability of high interference
power, leading to a heavier interference distribution tail. As the path loss exponent α increases,
however, the IW distribution starts to diverge from the simulated data and hence reduces in
weight, thus the IG weight w1 increases as shown. Note also that as α increases, shadowing
has less effect on the mixture model weights, especially at low active UEs density. As such, the
weight distribution at high path loss is less sensitive to shadowing.
In Fig. 7, we plot the IW and IG shape parameters {c, λ} versus σSF at α = 3. These curves
match the simulated data perfectly, verifying the functional fits in Corollary 1. At higher path
loss exponents, however, we noted some irregularities in the values of simulated c specifically
at α ≥ 4 and σSF ≤ 4 (not shown in the figure). This irregularity is caused by the fact that the
IG distribution has a much higher weight than IW in this range, hence the IW distribution can
assume a looser range of values for its parameters without affecting the goodness of the fit. The
functional fit for c in (46) thus can still be applied in this high path loss range.
These plots demonstrate that the fitted functions for parameters of the mixture MLE model are
accurate for the practical ranges of channel path loss and shadowing. These functions therefore
provide an extremely simple yet accurate way to estimate interference distribution parameters.
26
1
Sim Data
α =2
α =3.5
α =5
0.9
0.8
10
0
IW, c
Shape Parameters
0.7
IG weight
0.6
0.5
0.4
0.3
0.2
10
-1
10
-2
Sim Data
η=1
η=3
IG, λ
0.1
0
0
1
2
3
4
5
6
7
8
0
9
1
2
3
4
5
6
7
8
9
σ SF (dB)
σ SF (dB)
Fig. 6: IG weight, w1 , in the mixture MLE model. Fig. 7: IW and IG parameters comparison between
System parameters: Pmax = 30 dBm, η = 3, δ = 10−6 ,
simulation (markers) and fitted functions in Corollary
α = {2, 3.5, 5}.
1 (lines). System settings: Pmax = 30 dBm, α = 3,
η = {1, 3}, δ = 10−6 .
E. Rate CDF and Outage Performance
We now apply the proposed interference models to evaluate the cellular system performance.
In particular, we evaluate the transmission rate cumulative distribution function (CDF), which
is related to the outage probability, of a cell edge user at a distance of 145 m from the base
station with η = 1 (note Rc = 150 m and ηλ1 = 0.25Rc−2 ). In this simulation scenario, we do
not consider uplink scheduling; instead, we consider a random user with a random channel to
the considered BS. We further normalize the simulation results by the channel bandwidth and
show rates in bps/Hz. The uplink maximum transmission power is 30 dBm. We use the dominant
mode transmit beamforming discussed in Sec. III-C and assume a noise variance of σ 2 = −124
dBm/Hz and a path loss intercept β = −72.3 dB.
In Fig. 8, we plot the interference-aware capacity in (14) and compare the user transmission
rate CDFs based on IG MM and mixture interference models to the simulated data. Both models
show good accuracy for the interference power as in the left plot at a low σSF value. At larger
shadowing, however, as in the right plot with σSF = 9 dB, only the mixture model remains
accurate while the IG MM model diverges significantly. This result also suggests that interference
has a strong impact on user transmission rate CDF even at the somewhat high noise variance
of σ 2 = −124 dBm/Hz. We also note that the slight mismatch between the mixture model and
simulation is due to the interference correlation, which we did not model but does have an impact
on capacity in an interference-aware combining scheme.
To further examine the accuracy of the mixture interference power model, we consider the
interference-unaware combining scheme with capacity in (13). IU capacity does not depend on
interference correlation as a result of Lemma 1. Results in Fig. 9 shows that the capacity using
interference power model matches with simulation almost perfectly in all considered antenna
1
1
0.9
0.9
0.8
0.8
0.7
0.7
Outage Probability
Outage Probability
27
0.6
0.5
0.4
0.3
0.2
0.6
0.5
0.4
0.3
0.2
Mixture Model
Moment Matching IG
Sim Data
0.1
Mixture Model
Moment Matching IG
Sim Data
0.1
0
0
0
1
2
3
4
5
6
0
Rate (bps/Hz)
(a) σSF = 1 dB.
1
2
3
4
5
6
7
8
Rate (bps/Hz)
(b) σSF = 9 dB.
Fig. 8: User transmission rate CDF comparison among simulation, IG MM model, and mixture MLE model.
System settings: Pmax = 30 dBm, η = 1, α = 3, σSF = {1, 9} dB.
configurations, even at a high shadowing variance of σSF = 9 dB. This result confirms the
validity of our interference power distribution in capacity evaluation.
Our interference distribution models have been validated against simulation based on stochastic
geometry. The ideal next step would be to evaluate the models against actual measurement data,
but at the point of writing this paper, such data are not available. As stochastic geometry has
been shown to reasonably capture cellular performance compared to actual data [8], we expect
that the mixture model with adjusted functional coefficients will also present a good fit to actual
measurement data. The mixture MLE model can provide a valuable tool for cellular system
performance evaluation and prediction.
VIII. C ONCLUSION
We introduced new interference models based on moment matching and maximum likelihood
estimation techniques for both the rich scattering and limited scattering NLOS propagations. A
novel model as a mixture between the Inverse Gaussian and Inverse Weibull presents remarkably
good fit with simulation, capturing the heavy-tail characteristics of interference especially in
high shadowing environments. We designed an expectation maximization algorithm to estimate
the parameters of this mixture model. We also fitted the mixture model parameters to simple
polynomial functions of the channel propagation features, which provide a simple way to model
the interference without any optimization. The fitted mixture model can then be integrated into
a more complex system level simulation to evaluate and predict cellular performance, or used
to aid system design. The next step would be to test the mixture model with appropriately
adjusted parameters and functional fitting coefficients against data from measurement campaigns
to evaluate the model in an actual system setting.
28
1
0.9
0.8
Outage Probability
0.7
0.6
0.5
0.4
0.3
0.2
Analysis, N UE = N BS = 2
Analysis, N UE = N BS = 4
Analysis, N UE = N BS = 8
0.1
Sim Data
0
0
1
2
3
4
5
6
7
8
Rate (bps/Hz)
Fig. 9: User transmission rate CDF comparison among simulation and mixture MLE model. System settings:
Pmax = 30 dBm, η = 1, α = 3, σSF = 9 dB, NUE = NBS = {2, 4, 8}.
A PPENDIX A: P ROOF
OF
L EMMA 1
Given the interference vector in (8), the element in the ith row and j th column of the
interference covariance matrix Σ0 can be expressed as
X
[Σ0 ]ij =
l (kzk k2 )h̆(k,i) h̆∗(k,j) Pk ,
(47)
k>0:zk ∈Φ1
where h̆(k,i) , h̆(k,j) are i.i.d. CN (0, 1) representing the Rayleigh small-scale fading channels
from the k th UE to the ith and j th antenna element of the considered BS, respectively. The
post-combining interference power term v1∗ Σ0 v1 can then be written as
!
!∗
X
X
X
Pk l (kzk k2 )
v1∗ Σ0 v1 =
[v1 ]i h̆(k,i)
[v1 ]i h̆(k,i) .
i
k>0:zk ∈Φ1
(48)
i
Since the direct channel small-scaling fading component is assumed to be independent of
interference channels small-scale fading, v1 is independent of all h̆k . Thus the distribution of
P
∗
i [v1 ]i h̆(k,i) is CN (0, 1) as v1 is unit-norm, which renders the distribution of v1 Σ0 v1 in (48)
the same as that of the diagonal elements of Σ0 in (47).
A PPENDIX B: P ROOF
OF
L EMMAS 2 AND 3
We first derive the Laplace transform of the interference power term, q0 , which is then used
to characterize its moments. We develop the Laplace transform of the interference power at the
destination as follows:
Y
Lq0 (s) = Eq0 e−sq0 = EΦ1 ,T
zk ∈Φ1 \z0
= EΦ1
Y
zk ∈Φ1 \z0
e−sgk l (kzk k2 )Pk , i 6= j
LJ0 (s, kzk k2 ) = exp −2πλ1
Z
∞
Rc
(1 − LJ0 (s, r)) rdr ,
(49)
29
where the last equality follows from the Laplace functional expression for PPP using polar
coordinates and assuming the field of interferers outside a cell of fixed radius Rc ; LT (s) is the
d
Laplace transform of the composite shadowing and small scale fading channel, where T = Ls G
with G ∼ exp(1) as an exponential random variable, and LJ0 (s, kzk k2 ) is expressed as
LJ0 (s, kzk k2 ) = LT (slp (kzk k2 )Pk )
(50)
Then, we obtain the results in Lemma 2 based on the interference power Laplace transform
and evaluating the following formulas:
E [q0 ] =−
∂Lq0 (s)
∂s
,
var [q0 ] =
s=0
∂ 2 Lq0 (s)
− (E [q0 ])2 .
2
∂s
s=0
(51)
The proof to Lemma 3 follows the same procedure with the only difference in that we replace
q0 in Eqs. (49) and (51) with q̃0 , and the Laplace transform, LT (s), in eq. (50) with the Laplace
d
transform LT ∗ (s) of the composite random variable T ∗ = Ls H̃1 H̃2 where H̃1 , H̃2 ∼ CN (0, 1)
are independent and identically distributed complex normal random variables.
A PPENDIX C: P ROOF
OF
L EMMA 6
We start by writing the expression for the EM algorithm Q-function using Proposition 1 as
Q(θ|θ
(m)
) =
n
X
i=1
=
EXi |Yi ,θ(m) [log fXi (xi |θ)] =
2
n X
X
(m)
γij
log wj +
n
X
i=1
i=1 j=1
(m)
γi1
2
n X
X
i=1 j=1
p(Zi = j|Yi = yi , θ(m) ) log fXi (xi |θ)
log φ1 (yi |θ1 ) +
n
X
i=1
(m)
γi2 log φ2 (yi |θ2 ).
(52)
Using the probability density functions of IG and IW distributions in Eqs. (18) and (20) along
with moment matching first, we can write log φ1 (yi |λ) and log φ2 (yi |c) as in Eqs. (34) and
P
(m)
(m)
(35), respectively. Then, substituting into Eq. (52), replacing ni=1 γij by nj , and dropping
constant terms without affecting the EM algorithm, we get the final expression in Eq. (40).
A PPENDIX D: P ROOF
OF
T HEOREM 2
We first solve for wj using the method of Lagrange multipliers as follows.
!
2
2
X
X
(m)
J(w, ν) =
nj log wj + ν 1 −
wj
j=1
∂J
∂wj
=
(m)
nj
wj
j=1
(m)
− ν = 0 , j ∈ {1, 2}
⇒
nj
ν=
.
wj
(53)
30
Substituting into the constraint
(m)
(m+1)
wj
nj
= Pk
P2
(m)
j=1 nj
j=1 wj
= 1, we get the weight wj update as
(m)
(m)
= Pk
j=1
nj
Pn
i=1
(m)
γij
(m)
nj
= Pn
nj
=
n
i=1 1
, j ∈ {1, 2}.
(54)
Then, we can find the update of λIW by simply taking the partial derivative of the Q-function
w.r.t. λIW as
∂Q θ|θ(m)
∂λ
n
1 (m)
1 X (m) (yi − µY )2
=
n − 2
γ
= 0.
2λ 1
2µY i=1 i1
yi
(55)
Solving for λ (m+1) results in Eq. (42).
Similarly, taking the partial derivative of the Q-function w.r.t. c to get
"
# X
n
∂Q θ|θ(m)
1
1
µ
Y
(m)
(m)
+
1 − ψ(1 − )
log
= n2
−
γi2 log yi
∂c
c
Γ(1 − 1c ) c
i=1
#c n
#c ! n
"
"
X
X (m)
d
µY
µY
(m) −c
γ
y
log
y
−
γi2 yi−c . (56)
+
i
i2
i
1
1
dc
Γ(1 − c ) i=1
Γ(1 − c )
i=1
Equating to zero, we obtain the equation in (43).
A PPENDIX E: F ITTED F UNCTIONAL C OEFFICIENTS
FOR THE
M IXTURE MLE M ODEL
The fitted coefficient values in Table I are obtained for the functional fits in (46) by comparing
our Mixture model against simulation data. These coefficients are estimated for a maximum
transmission power of Ps = 30 dBm. To estimate the mixture distribution parameters at a different
transmission power Pnew dBm, we only need to update the b, λ and µ parameters by scaling
each of them by the corresponding transmit power scaling factor, i.e. bnew = 100.1(Pnew −Ps ) × b.
Parameter c stay unchanged with respect to the transmission power.
R EFERENCES
[1]
S. Rajagopal, S. Abu-Surra, and M. Malmirchegini, “Channel feasibility for outdoor non-line-of-sight mmwave mobile
communication,” in IEEE Vehicular Tech. Conf. (VTC Fall), Sept 2012, pp. 1–6.
[2]
W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and F. Aryanfar, “Millimeter-wave beamforming
as an enabling technology for 5G cellular communications: theoretical feasibility and prototype results,” IEEE Comm.
Magazine, vol. 52, no. 2, pp. 106–113, February 2014.
[3]
F. Boccardi, R. Heath, A. Lozano, T. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,” IEEE
Comm. Magazine, vol. 52, no. 2, pp. 74–80, Feb. 2014.
[4]
T. Bai, V. Desai, and R. W. Heath, “Millimeter wave cellular channel models for system evaluation,” in IEEE ICNC,
2014, pp. 178–182.
[5]
T. Bai, A. Alkhateeb, and R. Heath, “Coverage and capacity of millimeter-wave cellular networks,” IEEE Comm. Magazine,
vol. 52, no. 9, pp. 70–77, September 2014.
31
coeff.
α
2
a3
a2
a1
a0
-111.99
-589.09
-1021.1
-873.55
coeff.
a3
a2
a1
a0
2
7.5381
-23.171
-133.81
685.25
α
3
-41.663
-214.46
-443.27
-433.55
3
5.9903
-11.084
-81.932
314.72
3.5
13.4287
-43.972
-230.91
-243.78
3.5
7.0852
8.8851
-105.21
239.38
4
16.8309
-17.117
-137.71
-149.15
4
15.616
15.137
-142
223.73
5
5.7399
-1.4256
-40.481
-122.15
5
11.126
-6.9276
-121.18
224.88
(a) IG weight parameter w1 (×103 ).
coeff.
(b) IW shape parameter c(×103 ).
a3
a2
a1
a0
2
-0.6594
12.124
17.277
1134.1
3
1.258
47.252
128.5
3.5
1.604
56.81
4
8.84
56.04
5
13.99
49.3
α
coeff.
a3
a2
a1
a0
2
0.05993
19.152
44.478
-351.06
-2502.3
3
0.25864
105.43
313.42
-2929.1
117.1
-4102
3.5
0.82885
105.83
311.6
-4186.3
78.43
-5577
4
0.96339
105.86
311.24
-5398.9
65.3
-8338
5
1.1855
105.94
310.68
-7752.1
α
(c) IG shape parameter λ(×103 ).
(d) IG scale parameter µ(×103 ).
TABLE I: Fitting coefficients for the mixture model parameters w1 , b, λ, and µ at η = 1.
[6]
M. Rebato, M. Mezzavilla, S. Rangan, F. Boccardi, and M. Zorzi, “Understanding noise and interference regimes in 5g
millimeter-wave cellular networks,” in European Wireless 2016; 22th European Wireless Conference, May 2016, pp. 1–5.
[7]
A. Giovanidis and F. Baccelli, “A stochastic geometry framework for analyzing pairwise-cooperative cellular networks,”
CoRR, vol. abs/1305.6254, 2013.
[8]
J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans. on
Comm., vol. 59, no. 11, pp. 3122–3134, November 2011.
[9]
H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for modeling, analysis, and design of multi-tier and
cognitive cellular wireless networks: A survey,” IEEE Comm. Surveys Tutorials, vol. 15, no. 3, pp. 996–1019, 2013.
[10]
R. Heath, M. Kountouris, and T. Bai, “Modeling heterogeneous network interference using Poisson point processes,” IEEE
Trans. on Signal Processing, vol. 61, no. 16, pp. 4114–4126, Aug 2013.
[11]
S. Ak, H. Inaltekin, and H. V. Poor, “Gaussian approximation for the downlink interference in heterogeneous cellular
networks,” in 2016 IEEE International Symposium on Information Theory (ISIT), July 2016, pp. 1611–1615.
[12]
M. Aljuaid and H. Yanikomeroglu, “Investigating the gaussian convergence of the distribution of the aggregate interference
power in large wireless networks,” IEEE Transactions on Vehicular Technology, vol. 59, no. 9, pp. 4418–4424, Nov 2010.
[13]
T. Bai and R. Heath, “Coverage and rate analysis for millimeter-wave cellular networks,” IEEE Trans. on Wireless Comm.,
vol. 14, no. 2, pp. 1100–1114, Feb 2015.
[14]
H. Elkotby and M. Vu, “Uplink user-assisted relaying in cellular networks,” IEEE Trans. on Wireless Comm., vol. 14,
no. 10, pp. 5468–5483, Oct 2015.
[15]
A. Altieri, L. Rey Vega, P. Piantanida, and C. Galarza, “Analysis of a cooperative strategy for a large decentralized wireless
network,” IEEE/ACM Trans. on Networking, vol. 22, no. 4, pp. 1039–1051, Aug 2014.
[16]
T. Novlan, H. Dhillon, and J. Andrews, “Analytical modeling of uplink cellular networks,” IEEE Trans. on Wireless
Comm., vol. 12, no. 6, pp. 2669–2679, June 2013.
[17]
G. R. MacCartney, J. Zhang, S. Nie, and T. S. Rappaport, “Path loss models for 5G millimeter wave propagation channels
in urban microcells,” in 2013 IEEE Global Comm. Conf. (GLOBECOM), pp. 3948–3953.
32
[18]
M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter wave channel
modeling and cellular capacity evaluation,” IEEE Journal on Sel. Areas in Comm., vol. 32, no. 6, pp. 1164–1179, 2014.
[19]
M. Samimi, T. Rappaport, and G. Maccartney, “Probabilistic omnidirectional path loss models for millimeter-wave outdoor
communications,” IEEE Wireless Comm. Letters, vol. 4, no. 4, pp. 357–360, Aug 2015.
[20]
T. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. Wong, J. Schulz, M. Samimi, and F. Gutierrez,
“Millimeter wave mobile communications for 5G cellular: It will work!” IEEE Access, vol. 1, pp. 335–349, 2013.
[21]
T. Rappaport, F. Gutierrez, E. Ben-Dor, J. Murdock, Y. Qiao, and J. Tamir, “Broadband millimeter-wave propagation
measurements and models using adaptive-beam antennas for outdoor urban cellular communications,” IEEE Trans. on
Antennas and Propagation, vol. 61, no. 4, pp. 1850–1859, April 2013.
[22]
Y. Niu, Y. Li, D. Jin, L. Su, and A. V. Vasilakos, “A survey of millimeter wave (mmwave) communications for 5G:
Opportunities and challenges,” CoRR, vol. abs/1502.07228, 2015. [Online]. Available: http://arxiv.org/abs/1502.07228
[23]
E. Turgut and M. C. Gursoy, “Coverage in heterogeneous downlink millimeter wave cellular networks,” arXiv preprint
arXiv:1608.01790, 2016.
[24]
A. Afzal and S. A. Hassan, “A stochastic geometry approach for outage analysis of ad hoc SISO networks in Rayleigh
fading,” in 2013 IEEE Global Communications Conference (GLOBECOM), Dec 2013, pp. 336–341.
[25]
S. Akoum and R. W. Heath, “Interference coordination: Random clustering and adaptive limited feedback,” IEEE
Transactions on Signal Processing, vol. 61, no. 7, pp. 1822–1834, April 2013.
[26]
[27]
H. V. Poor, An introduction to signal detection and estimation.
Springer Science & Business Media, 2013.
T. M. Cover and J. A. Thomas, Differential Entropy. John Wiley & Sons, Inc., 2005, pp. 243–259. [Online]. Available:
http://dx.doi.org/10.1002/047174882X.ch8
[28]
R. Agrawal et al., “On efficacy of Rayleigh-Inverse Gaussian distribution over K-distribution for wireless fading channels,”
Wireless Communications and Mobile Computing, vol. 7, no. 1, pp. 1–7, 2007.
[29]
S. Atapattu, C. Tellambura, and H. Jiang, “A Mixture Gamma Distribution to Model the SNR of Wireless Channels,”
IEEE Transactions on Wireless Communications, vol. 10, no. 12, pp. 4193–4203, December 2011.
[30]
M. C. Tweedie, “Statistical Properties of Inverse Gaussian Distributions. I.” The Annals of Math. Stat., pp. 362–377, 1957.
[31]
M. C. Tweedie et al., “Statistical Properties of Inverse Gaussian Distributions. II.” The Annals of Math. Stat., vol. 28,
no. 3, pp. 696–705, 1957.
[32]
R. S. C. J. L. Folks, “The Inverse Gaussian Distribution and Its Statistical Application – A Review,” Journal of the Royal
Stat. Society. Series B (Methodological), vol. 40, no. 3, pp. 263–289, 1978.
[33]
H. Rinne, The Weibull distribution: a handbook.
CRC Press, 2008.
[34]
K. Sultan, N. Alsadat, and D. Kundu, “Bayesian and Maximum Likelihood Estimations of the Inverse Weibull parameters
under progressive type-II censoring,” Journal of Stat. Computation and Sim., vol. 84, no. 10, pp. 2248–2265, 2014.
[35]
M. Kountouris and N. Pappas, “Approximating the interference distribution in large wireless networks,” in 2014 11th
IEEE Int. Symp. on Wireless Comm. Systems (ISWCS), pp. 80–84.
[36]
A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum Likelihood from Incomplete Data via the EM Algorithm,”
Journal of the royal statistical society. Series B (methodological), pp. 1–38, 1977.
[37]
Y. Chen and M. R. Gupta, “EM demystified: An Expectation-Maximization Tutorial,” University of Washington, 2010.
| 7cs.IT
|
arXiv:1803.09928v1 [cs.LG] 27 Mar 2018
Entropy Controlled Non-Stationarity for Improving Performance of
Independent Learners in Anonymous MARL Settings
Tanvi Verma, Pradeep Varakantham and Hoong Chuin Lau
School of Information Systems, Singapore Management University, Singapore
[email protected], [email protected],[email protected]
Abstract
With the advent of sequential matching (of supply and demand) systems (uber, Lyft, Grab for
taxis; ubereats, deliveroo, etc for food; amazon prime, lazada etc. for groceries) across
many online and offline services, individuals
(taxi drivers, delivery boys, delivery van drivers,
etc.) earn more by being at the ”right” place
at the ”right” time. We focus on learning techniques for providing guidance (on right locations to be at right times) to individuals in the
presence of other ”learning” individuals. Interactions between indivduals are anonymous,
i.e, the outcome of an interaction (competing
for demand) is independent of the identity of
the agents and therefore we refer to these as
Anonymous MARL settings.
Existing research of relevance is on independent learning using Reinforcement Learning
(RL) or on Multi-Agent Reinforcement Learning (MARL). The number of individuals in aggregation systems is extremely large and individuals have their own selfish interest (of
maximising revenue). Therefore, traditional
MARL approaches are either not scalable or
assumptions of common objective or action coordination are not viable. In this paper, we
focus on improving performance of independent reinforcement learners, specifically the
popular Deep Q-Networks (DQN) and Advantage Actor Critic (A2C) approaches by exploiting anonymity. Specifically, we control nonstationarity introduced by other agents using entropy of agent density distribution. We demonstrate a significant improvement in revenue for
individuals and for all agents together with our
learners on a generic experimental set up for aggregation systems and a real world taxi dataset.
1
Introduction
Systems that aggregate supply to ensure a better matching of demand and supply have significantly improved
performance metrics in taxi services, food delivery and
grocery delivery. From the individual perspective (taxi
drivers, delivery boys, delivery van drivers), learning to
be at the ”right” place at the ”right” time is essential to
maximize their revenue. From an individual perspective,
this learning problem is challenging because of the uncertainty about demand (which will determine the revenue
earned) and the locations of other ”learning” individuals
(which will determine whether the demand is assigned to
the current individual).
Given the presence of multiple learning agents, a key
focus area of relevance is Multi-Agent Reinforcement
Learning (MARL) [Littman, 1994; Claus and Boutilier,
1998; Busoniu et al., 2008], where multiple agents are
learning to maximize their individual rewards. There are
two main issues with such an approach in case of multiagent sequential matching problems: (a) The definition
of fairness when learning for multiple agents is unclear
as the agent actions are not synchronised, agents can
have different starting beliefs and demand is dynamic or
non-stationary. (b) The scale of matching problems is
typically very large (with thousands of agents) and hence
computing policies for all agents is computationally very
challenging.
A thread of MARL has also focussed on teams of agents
that maximize the global reward [Perolat et al., 2017;
Nguyen et al., 2017] or agents cooperate to achieve a common goal [Hausknecht and Stone, 2015; Foerster et al.,
2016]. The learning approaches there are also limited in
scalability with respect to number of agents. Furthermore,
the problems of interest in this paper are focussed on individual agents that are maximizing their own revenue
rather than a team revenue. Therefore, existing MARL
approaches are not suitable for the problem and hence we
focus on improving performance of independent learners
(ILs) in this paper.
Our work is motivated by the recent success of use entropy
regularizers in improving performances of reinforcement
learning algorithms [Mnih et al., 2016; Haarnoja et al.,
2017b; Schulman et al., 2017]. We use entropy of agent
population distribution as the regularizer. The intuition
is, as the agents are competing for resources, the policy which encourages even agent population distribution
is more advantageous for individual as well collective
agents.
Existing work [Verma et al., 2017] in multi-agent sequential matching problems has employed Reinforcement
Learning (RL) for each individual to learn a decision making policy. However, this line of work has assumed that
other agents employ stationary and fixed strategies, rather
than adaptive strategies. Due to the non-stationarity introduced when multiple agents learn independently, the performance of traditional table based Q-learning approaches
is extremely poor.
The use of deep neural networks (instead of table based
representations) has become hugely popular in the reinforcement learning community [Mnih et al., 2015; Silver et al., 2016, 2017]. Policy gradient methods and
Q-learning methods have been successfully used in deep
neural networks.
We make the following key contributions:
• Extending on past work in multi-agent planning under uncertainty [Varakantham et al., 2012, 2014], we
describe the notion of interaction anonymity in the context of multi-agent reinforcement learning.
• We then provide mechanisms to exploit interaction
anonymity by extending the well known DQN (Deep QNetwork) and A2C (Advantage Actor-Critic) methods.
Specifically, we control the non-stationarity introduced
by other agents by finding policies that also maximize
or minimize the entropy on agent count distribution.
• To demonstrate the utility of our approaches, we performed extensive experiments on a synthetic data set
(that is representative of multiple multi-agent sequential matching problems) and a real taxi data set. We
observe that our individuals learners based on DQN
and A2C are able to learn policies that are reasonably
fair (minor differences in value) and most importantly,
the individual and joint (social welfare) values of the
learning agents are significantly higher than existing
benchmarks.
2
Motivating Problems
This paper is motivated by the following matching problems, where there are multiple agents and matching of
agents to customer demand has to be done on a continuous basis. We refer to these problems as Multi-Agent
Sequential Matching Problems (MSMPs) and our focus
is on enabling individual agents to learn effectively and
efficiently.
Taxi Aggregation: Companies like Uber, Lyft, Didi,
Grab etc. all provide taxi supply aggregation systems.
The goal is to ensure wait times for customers is minimal
or amount of revenue earned is maximized by matching
taxi drivers to customers on a regular basis. However,
these goals of the aggregation companies may not be correlated to the individual driver objective of maximizing
their own revenue. The methods provided in this paper
can be used to guide individual drivers based on their
past experiences of customer demand and taxi supply (obtained directly/indirectly from the aggregation company),
while also considering that other drivers are learning to
improve their revenue.
Food or Grocery Delivery: Aggregation systems have
also become very popular for food delivery (Deliveroo,
Ubereats, Foodpanda, DoorDarsh etc.) and grocery delivery (AmazonFresh, Deliv, RedMart etc.) services. They
offer access to multiple restaurants/grocery stores to the
customers and use services of delivery boys/delivery vans
to deliver the food/grocery. Similar to taxi aggregation
systems, our work can be used to guide delivery boy/van
to improve their individual revenue based on current location and presence of other delivery boys/vans near that
location (obtained directly/indirectly from the aggregation
company).
Supply Aggregation in Logistics: More and more online buyers now prefer same day delivery services and
tradition logistic companies which maximize usage of
trucks, drivers and other resources are not suitable for it.
Companies like Amazon Flex, Postmates, Hitch etc. connect shippers with travellers/courier personnels to serve
same day/on-demand delivery requests. The courier personnels in this system can employ the proposed method
to learn to be at ”right” place at ”right” time by learning
from past experiences.
Sensor Networks: Another motivating problem is target
tracking by a team of sensors [Nair et al., 2005]. Each
sensor node can scan in one of the four directions. To
track a target accurately and receive the associated reward, a small number of sensors must scan the target
simultaneously. Every sensor can interact with only a
limited number of neighbour sensors. In this example
having more number of sensors scanning same location
is required for high reward as opposed to the previous
examples where more number of agents decrease individual rewards. Hence, by modifying our proposed methods
slightly, it can be used by individual sensors to learn to
align in a direction based on past rewards and alignments
of neighbouring sensor nodes.
3
Background: Reinforcement Learning
(RL)
In this section, we briefly describe the Reinforcement
Learning (RL) problem and leading approaches (of relevance to this paper) for solving large scale RL problems.
The RL problem for an agent is to maximize the long run
reward while operating in an environment represented as
a Markov Decision Process (MDP). Formally, an MDP is
represented by the tuple S, A, T , R , where S is the set
of states encountered by the agent, A is the set of actions,
T (s, a, s0 ) represents the probability of transitioning from
state s to state s0 on taking action a and R(s, a) represents the reward obtained on taking action a in state s.
The RL problem is to learn a policy that maximizes the
long term reward while only obtaining experiences (and
not knowing the full reward, R and transition, T models) of transitions and reinforcements. An experience is
defined as (s, a, s0 , r) and typically there is an episode
of experiences that ends when s0 is a terminal state. In
this paper, we are interested in episodic problems (ex: an
episode ends once the taxi driver gets a customer).
We now describe three approaches that are of relevance
to work described in this paper.
3.1
Q-Learning
One of the most popular approaches for RL is Qlearning Watkins and Dayan [1992], where the Q function
is represented as a table (and initialised to arbitrary fixed
values) with an entry in the table for each state and action combination. Q-values are updated based on each
experience given by (s, a, s0 , r):
0
0
Q(s, a) = R(s, a) + γ ∗ max
Q(s , a ) − Q(s, a) (1)
0
a
3.2
Deep Q-Networks (DQN)
Instead of a tabular representation for Q function employed by Q-Learning, the DQN approach Mnih et al.
[2015] employs a deep network to represent the Q function. Unlike with the tabular representation that is exhaustive (stores Q-value for every state, action pair), a deep
network predicts Q-values based on similarities in state
and action features. This deep network for Q-values is
parameterised by a set of parameters, θ and the goal is to
learn values for θ so that a good Q-function is learnt.
We learn θ using an iterative approach that employs gradient descent on the loss function. Specifically, the loss
function at each iteration is defined as follows:
L(θ) = E(e∼U (J )) [(y DQN − Q(s, a; θ− ))2 ]
where y DQN = r + γmaxa0 Q(s0 , a0 ; θ− ) is the target
value computed by a target network parameterised by
previous set of parameters, θ− . Parameters θ− are frozen
for some time while updating the current parameters θ. To
ensure independence across experiences (a key property
required by Deep Learning methods to ensure effective
learning), this approach maintains a replay memory J
and then samples experiences from that replay memory.
Like with traditional Q-Learning, DQN also typically
employs an -greedy policy.
3.3
Advantage Actor Critic (A2C)
DQN learns the Q-function and then computes policy
from the learnt Q-function. A2C is a policy gradient
method that directly learns the policy while ensuring that
the expected value is maximized. To achieve this, A2C
employs two deep networks, a policy network to learn
policy, π(a|s; θp ) parameterized by θp and a value network to learn value function (and not Q function) of the
computed policy, V π (s; θv ) parameterized by θv .
The policy network parameters are updated based on the
policy loss, which in turn is based on the advantage associated with the value function. Formally, the policy loss
is given by:
To ensure a good explore-exploit tradeoff, an -greedy
policy is typically employed.
π(s) =
arg maxa∈A Q(s, a), with probability 1 −
a
a1 , with probability ∗ P e 1 ak
e
a2 , with probability ∗
...
ai ∈A
ea 2
P
ai ∈A
ea k
Q-learning is guaranteed to converge to the optimal solution for stationary domains Watkins and Dayan [1992].
L(θp ) = ∇θp logπ(a|s; θp )A(s, a; θv )
where
A(s, a; θv ) =
i=k−1
X
γ i rt+i + γ k V (sk ; θv ) − V (s; θv )
i=0
(2)
The first term in equation 2 is the k-step (or till the end
of the episode) discounted reward following the current
policy and using a discount factor γ.
are dependent on other agents but only on counts. We
focus on developing independent learners for Anonymous
MARL settings.
Dynamic Bayes Network (DBN) in Figure 1 provides a graphical representation for MSMPs of interest. As can be observed from the DBN, agents
are independent of each other given the agent
counts vector, D = hD1 , D2 , . . . Dt , . . .i, where Dt =
(dt (s1), dt (s2), . . .).
In this paper, our focus is on learning the policy from past
experiences when the reward and transition functions are
not known. Transition and reward functions are unknown
or uncertain because of two reasons: (i) Matching of
an agent to customer demand is done by aggregation
companies like Uber and hence individual agents may not
be aware of what demand is assigned to them; and (ii)
Due to uncertainty in future demand, individual agents
are not aware of the consequences of moving to a specific
state.
Experience for each agent is given by
hs, d(s), a, r, s0 , d(s0 )i,
Figure 1: A DBN depicting relationship between policy
of agents and agent population density distribution Dt .
Where rti is the immediate reward received agent i by
taking action ait in state Sti at time t.
4
Exploiting Anonymity for Independent
Learning
In the Multi-agent Sequential Matching Problems
(MSMPs) described in Section 2, we have individual
agents learning to improve their expected reward in presence of other learning agents. The model for an individual agent, i is the same as in RL, i.e., S, A, T , R .
However, the key difference is that reward and transition functions are dependent on other agents and more
specifically on counts of agents. The reward function,
R(s, a, d(s)) refers to the reward for taking action a in
state s when there are d(s) other agents in state s 1 . The
transition function, T (s, a, s0 , d(s)) refers to the transition to s’ on taking action a from state s, when there are
d(s) other agents in state s. In the literature Varakantham et al. [2012, 2014], this dependence of reward and
transition functions on agent counts is referred to as interaction anonymity. We refer to these as Anonymous
MARL problems, as the reward and transition functions
1
This can be easily extended to consider count of number
of agents in same or relevant state, action pairs as described in
Kumaret al. Kumar and Varakantham [2017]
where d(s) and d(s’) represent the count of other agents2
in states s and s0 respectively.
The RL agent learning in MSMPs is faced with nonstationarity due to changing policies of other learning
agents (that are dependent on policy of the RL agent).
However, the useful structure in MSMPs is that each
agent is affected in the same way (through rewards and
transitions that are dependent on agent counts) by every
other agent due to interaction anonymity. We exploit the
interaction anonymity for independent learning using four
key steps.
First, to ensure better learning, we normalize each agent
counts vector into a probability distribution. That is to
say, d(s) will
Pnow refer to the the fraction of agents in
state s. and s d(s) = 1.
Second, we make policy, Q-function and V-function dependent on distribution of agent count. This is to ensure
that the learnt policy or value function captures any dependency3 on observed agent counts in states. That is to
say:
π : S × A × [0, 1] → [0, 1],
where π(a|s, d(s)) is the probability of taking action a in
2
These counts can typically be obtained directly (provided
by the aggregation company) or indirectly (through customer
application).
3
If there is no dependency on observed counts, then this
method should return same policies as without observed counts
in state space
state s when there is d(s) fraction of agents in state s.
Q : S × A × [0, 1] → R
where Q(s, a, d(s)) is the Q-value for taking action a in
state s when there are d(s) fraction of agents in state s.
V : S × [0, 1] → R,
where V (s, d(s)) is the value for state s.
Third, we modify deep network architectures employed
by Deep RL approaches such as DQN and A2C to exploit
anonymity captured through agent counts vector. Figure
3 shows how architecture of DQN can be modified into
Density Entropy (explained later in the section) based
DQN (DE-DQN). Similarly, figure 3 depicts transformation of standard A2C network into Density Entropy based
A2C (DE-A2C). Similar to DQN, DE-DQN estimates
action-value function. Unlike DQN, DE-DQN predicts
agent density distribution, D(s0 |s, d(s); θ) based on past
experience.
Similarly, apart from maintaining policy and value estimates, DE-A2C (figure 3) also maintains an estimate of
agent population density distribution. We use a softmax
layer to predict D. In the deep network π, V, Q and D are
parameterised by the network parameter θ as shown in
the figures 2 and 3.
policy output and a shared network for state-action/value
output and density prediction output.
Finally and most importantly, apart from maximizing the
value of the independent learner, we also maximize (or
minimize) entropy of the agent density distribution, D.
Entropy for the agent count distribution, D is given by:
X
HD = −
D(s0 |s, d(s); θ)log(D(s0 |s, d(s); θ)
s0 ∈S
As described in literature Haarnoja et al. [2017a], individual policy entropy is maximized so as to deal with
non-stationarity introduced by environments (which ensures we have different policies that work for different
scenarios of non-stationarity). However, in anonymous
MARL settings of interest in this paper, non-stationarity
for a learning agent is also introduced by other learning
agents. Therefore, depending on domain characteristics4 ,
we maximize/minimize density entropy to control the nonstationarity introduced by other agents. In other words,
we find a policy that will ensure other agents choose policies that will force density distribution to have high/low
entropy depending on the domains.
We now provide specific changes required to the target
value and the loss function to maximize density entropy
in DQN and A2C.
4.1
Figure 2: DQN and DE-DQN networks
DE-DQN
As we wish to maximize the density entropy, we augment
the reward with the density entropy. β is the hyperparameter which controls the effect of density entropy term. The
target value y for a given experience can be computed as
follows
(
r + βHD if s0 is terminal
y=
r + βHD + γmaxa0 Q(s0 , a0 , d(s0 ); θ− ) otherwise
θ− is the network parameter for the target network.
DE-DQN optimizes a single combined loss Lθ with respect to the joint parameter θ. Lθ combines Q-loss LQ
and D-loss LD :
Lθ = LQ + λLD
(3)
λ is the weightage term used for the density prediction
loss. Q-loss is computed using standard mean squared
Figure 3: A2C and DE-A2C networks
All the output layers can either have separate non-output
layers each or they can share the non-output layers. Without loss of generality, we assume separate network for
4
For taxi aggregation, food/grocery delivery and supply aggregation in logistics, revenue is reduced if agents are present
in one state and hence entropy needs to be maximized for those
settings. For sensor network (and other coverage) problems,
revenue is increased if more agents are present in one state and
hence entropy needs to be minimized in those cases.
error loss as given in equation 4. Equation 5 shows how
LD is computed using local observation .
LQ = E(e∼U (J )) [(y − Q(s, a, d(s); θ))2 ]
0
0
LD = −E(e∼U (J )) [d(s )logD(s |s, d(s); θ)]
(4)
(5)
As the agent does not have full observation of the density
distribution LD , we computes partial loss for the current
experience.
Algorithm 1 DE-DQN
1: Initialize replay memory J
2: Initialize counter T = 0
3: Initialize action-value function Q(s, d(s), a; θ) and population density distribution function D(.|s, d(s); θ) with
random weight θ
4: Initialize target network parameters θ− = θ
5: repeat
6:
T ←T +1
7:
Take action at using -greedy policy according to
Q(st , d(st ), a; θ) and observe rt , st+1 and d(st+1 )
8:
Store transition < st , d(st ), at , rt , st+1 , d(st+1 ) > to
the replay memory.
9:
if T mod Iupdate == 0 then
10:
dθ ← 0
Sample
batch
of
transitions
<
11:
s, d(s), a, r, s0 , d(s0 ) >from the replay memory
12:
Accumulate gradients wrt θ : dθ ← dθ + ∇θ Lθ
13:
Update θ based on dθ
14:
if T mod Itarget == 0 then
15:
Update target network parameter, θ− ← θ
16:
if T mod Ireplay == 0 then
17:
Reset replay memory
18: until T > Tmax
Algorithm 1 shows detailed steps of DE-DQN. Network
parameters are updated after every Iupdate steps. Itarget
is the number of fixed steps after which target network
parameter is updated and J is reset after every Ireplay
steps.
4.2
equation 8.
DE-A2C
Similar to A2C algorithm, DE-A2C maintains a policy
output π(a|s, d(s); θp ) and an output for the value function estimate V (s, d(s); θv ). It also maintains a density
prediction output D(s0 |s, d(s); θv ). We want to update
the policy such that the density entropy is maximized,
hence we add the entropy to the advantage term. A small
replay memory is also used which contains episodes generated from π(a|s, d(s); θp ) and replay memory is reset
after every update of θp . The value network loss Lθv
is combination of V-loss LV and D-loss LD as given in
LV = E(e∼U (J ) [(R − V (s, d(s); θv ))2 ]
(6)
LD = −E(e∼U (J ) [d(s0 )log(D(s0 |s, d(s); θv )])
(7)
Lθv = LV + λLD
(8)
(9)
Algorithm 2 shows the steps for DE-A2C.
Algorithm 2 DE-A2C
1: Initialize replay memory
2: Initialize counter T = 0
3: Initialize value function V (s, d(s); θv ) with random weight
θv
4: Initialize policy function π(a|s, d(s); θp ) with random
weight θp
5: repeat
6:
tstart ← T
7:
repeat
Take action at according to policy π(st , d(st ); θp )
8:
and observe rt , st+1 and d(st+1 )
9:
T ←T +1
10:
until (
terminal or T − tstart == k
0
if st is terminal
11:
R=
V (st , d(st ); θv ) otherwise
12:
for i ∈ T − 1, ..., tstart do
13:
R ← ri + γR
14:
add < si , di , ai , R, si+1 , d(si+1 ) > to J
15:
if T mod Iupdate == 0 then
16:
sample batch of experience from J
17:
Accumulate gradients wrt θp : dθp ← dθ +
∇θp logπ(a|s, d(s); θp )(R + βHD − V (s, d(s); θv ))
18:
Accumulate gradients wrt θv : dθv ← dθv +∇θv Lθv
19:
Update θp and θv using dθp and dθv respectively.
20:
Clear replay memory J .
21: until T > Tmax
5
Experiments
To demonstrate the performances of DE-DQN and DEA2C, we simulate a multi-agent taxi domain in grid world.
The grids are associated with a demand arrival rate and
multiple agents roam in the grid world to satisfy these
demands. The revenue of an agent can be affected by
various features of the taxi domain.
• Demand-to-Agent Ratio (DAR): This is the average
number of customers per time step per agent.
• Trip pattern: The average length of trips can be uniform
for all the grids or there can be few grids which always
get longer trips (for ex. airports which are usually
outside the city) whereas few grids get relatively shorter
trips (city center).
• Demand pattern: The demand can be uniformly spread
across all the grids (dispersed demand) or there might
be few grids where high demand is present and few
grids where there is no demand ( concentrated demands)
• Demand arrival rate: The Poisson arrival rate of demand can be either homogeneous (static arrival rate) or
non-homogeneous (dynamic arrival rate)
We perform experiments by varying values of these
features using various dimensions of the grid and
different number of agents. The experiments are done
on two sets of data. The first set is synthetic data and
we control how demand arrival rate, reward and cost is
generated. The second dataset is a one month data from a
real-world taxi fleet from a developed Asian city. From
the real world data set we compute demand between
zones and represent each zone as a grid. Travel time
between zones is also computed from the data set. We
evaluate the performances on two metrics:
(1) Revenue of IL agents present in the domain.
(2) Average revenue of all the ILs present in the domain.
Taxi Simulation Domain: Figure 4 shows an example
grid world with 6 × 6 grids. The numbers displayed in
the grids are the Poisson arrival rate of demand in the
corresponding grid. There are no demand in the grids
with no number in it. Demands are generated with a timeto-live value and the demand expires if it is not served
within time-to-live time periods. Furthermore, to emulate
the real-world taxi fares, the trip fares are generated based
on distance of the trip (distance between the origin and
destination grids) plus a fixed starting fare. At every time
step, it is tried to assign a trip to the agents based on the
agent population density at the grid.
Hyperparameters For DQN algorithms we used Adam
optimizer with a learning rate of 1e-4, whereas for A2C
algorithms we used RMSprop optimizer with learning rate
of 1e-5 for policy network and 1e-4 for value network.
Two hidden layers are used with 128 nodes per layer.
Density entropy regularizer β was set to 1e-2, whereas
weigtage of D-loss λ was set to 1e3. We also used policy
entropy regularizer for A2C algorithms and the regularizer
was set to 1e-3.
6
Experimental results
In this section, we benchmark the performance of our
approaches (DE-DQN and DE-A2C) with respect to tabular Q-learning, DQN/A2C without density distribution
input and finally DQN/A2C with density distribution input. One evaluation period consists of 1000 (1e3) time
steps5 and revenue of all the agents is reset after every
evaluation period. All the graphs plotted in the upcoming
sub-sections provide running average of revenue over 100
evaluation periods (1e5 time steps). Unless specified, the
number of agents employed was 20.
Tabular Q-Learning, Standard DQN and Standard
A2C: Figure 5 shows the comparison between tabular
Q-learning, DQN; with and without agent density input
and A2C;with and without agent density input. We used
medium demand-to-agent ratio with non-uniform trip pattern as explained in section 6. We can see that DQN with
Figure 5: Comparison of tabular Q-learning, standard
DQN and standard A2C .
Figure 4: A taxi grid world with demand arrival rates.
In our experiments we make a realistic assumption that
agents do not take action at every time step and they can
go to any grid from any grid and time taken to travel
between grids is proportional to the shortest distance between the grids. These on-way/on-trip agents are not a
candidate for the trip assignment and they do not contribute to the agent population density distribution computation. Hence D is prediction of agent population density
distribution of active learning agents.
agent density input and A2C with agent density input perform significantly better than the other three algorithms.
Even for other demand we obtained similar results. Therefore, we do not include tabular Q-learning, DQN and
A2C without agent density input in the following results.
Henceforth, DQN refers to DQN with density input and
A2C refers to A2C with density input.
Low DAR (Demand-to-Agent Ratio) uniform trip pattern and dispersed demand: In this case, static demand
5
A time step represents a decision and evaluation point in
the simulator. From a learner perspective, an episode is typically
multiple time steps long.
(a) Models
(b) DQN
(c) DE-DQN
(d) A2C
(e) DE-A2C
Figure 6: Low demand-to-agent ratio.
arrival rate was used. The effective demand-to-agent ratio
was 0.5. Figures 6a(b) and 6 provide the joint and individual normalized rewards obtained by different approaches.
Here are the key observations: (1) DE-DQN and DQN
perform better than A2C and DE-A2C. This can be attributed to the static nature of the domain.
(2) DE-DQN, A2C and DE-A2C all provide a reasonably
fair distribution of revenues for individual revenues, with
DE-DQN providing higher individual revenues.
Medium DAR with concentrated demand:
In this setting, there were few grids with high demand
whereas remaining grids did not have any demand and the
overall demand-to-agent ratio was 0.8. The poisson arrival rate was homogeneous and uniform trip pattern was
used. Here are the key observations: (1) Figure 7a shows
that both DE-A2C performs better than other algorithms,
with DE-DQN providing the second best performance.
(2) DQN, DE-DQN and A2C provide reasonably fair distribution of individual revenues, with DE-A2C providing
slightly uneven distribution of revenues for individual
agents.
Medium DAR with non-uniform trip patterns: In this
setting, we employed a demand-to-agent ratio of 0.8 and
non-uniform trip patterns. Arrival of demand was generated based on homogeneous poisson rates and to generate
a non-uniform trip pattern, we associated each grid cell
with a maximum range for destination grids for trips originating from the grid cell. Here are the key observations:
(1) We observed that as we increase complexity of demand
pattern, the performance of DE-DQN starts improving
with respect to DQN. Graph in figure 8a compares the
performances of all the algorithms.
(2) Figure 8 shows that DE-DQN and A2C provide low
variance across individual revenues.
High DAR with non-uniform trip patterns: We generated high demand-to-agent ratio (1.2) in this setting.
Once again, we observe that in Figues 9a and 9, DE-DQN
provides better joint revenue and less variance across individual agent revenues.
Medium DAR with dynamic demand arrival rate: To
simulate peak and off-peak hours we divided the evaluation period (1e3 time steps) into three parts. The demand for the first and third part is generated based on low
demand-to-agent ratio (0.6) and for the second part it is
generated using high demand-to-agent ratio (1.2).
Figure 10 shows that the performances are similar to what
we have seen for other settings, with DE-DQN providing
higher values while having lower variance across individual agents.
Taxi domain simulated using real-world data set For a
more realistic simulation of the taxi domain, we used realworld data to compute mean demand between two grids.
First, based on the GPS data of taxi-fleet, we divided the
map into 110 zones and represented each zone as a grid
in a 10 × 11 grid world. Since we focus on independent
learners, our learning is very scalable. However, simulating thousands of learning agents at the same time requires
extensive computer resources and with the academic resources available, we could not perform a simulation with
1000 agents. Hence, we computed proportional demand
for 100 agents and simulated for 100 agents. Time steps
taken to travel between two grids is also approximated by
computing average travel time between two zones using
the real-world dataset.
Figures 11a and 11c show that DE-DQN again performs
better than other approaches while having very small
variance across individual revenues. DE-A2C and other
approaches had a significant amount of variance across
individual revenues.
Summary of Results Our two contributions, DE-DQN
and DE-A2C that employ density entropy have significantly improved over the well known DQN and A2C algorithms in all the simulation settings and the real world
taxi data set. In addition, DE-DQN provides revenues
of individual agents which are very similar, thereby representing a fair learning approach from an MARL perspective. In other words, if there is an application that
provides guidance based on DE-DQN then the individual
agents following the advice will all be treated fairly.
(a) Models
(b) DQN
(c) DE-DQN
(d) A2C
(e) DE-A2C
Figure 7: Medium demand-to-agent ratio with concentrated demand pattern.
(a) Models
(b) DQN
(c) DE-DQN
(d) A2C
(e) DE-A2C
Figure 8: Medium demand-to-agent ratio with non-uniform trip pattern.
(a) Models
(b) DQN
(c) DE-DQN
(d) A2C
(e) DE-A2C
(d) A2C
(e) DE-A2C
Figure 9: High demand-to-agent ratio.
(a) Models
(b) DQN
(c) DE-DQN
Figure 10: Dynamic demand arrival rate.
(a) Models
(b) DQN
(c) DE-DQN
(d) A2C
Figure 11: Simulation using real-world data set
(e) DE-A2C
References
Busoniu, L., Babuska, R., and De Schutter, B. (2008).
A comprehensive survey of multiagent reinforcement
learning. IEEE Trans. Systems, Man, and Cybernetics,
Part C, 38(2):156–172.
Claus, C. and Boutilier, C. (1998). The dynamics of reinforcement learning in cooperative multiagent systems.
AAAI/IAAI, 1998:746–752.
Foerster, J., Assael, Y., de Freitas, N., and Whiteson, S.
(2016). Learning to communicate with deep multiagent reinforcement learning. In Advances in Neural
Information Processing Systems, pages 2137–2145.
Haarnoja, T., Tang, H., Abbeel, P., and Levine, S. (2017a).
Reinforcement learning with deep energy-based policies. arXiv preprint arXiv:1702.08165.
Haarnoja, T., Zhou, A., Abbeel, P., and Levine, S. (2017b).
Soft actor-critic: Off-policy maximum entropy deep
reinforcement learning with a stochastic actor.
Hausknecht, M. and Stone, P. (2015). Deep reinforcement
learning in parameterized action space. arXiv preprint
arXiv:1511.04143.
Kumar, R. R. and Varakantham, P. (2017). Exploiting anonymity and homogeneity in factored dec-mdps
through precomputed binomial distributions. In Proceedings of the 16th Conference on Autonomous Agents
and MultiAgent Systems, pages 732–740. International
Foundation for Autonomous Agents and Multiagent
Systems.
Littman, M. L. (1994). Markov games as a framework
for multi-agent reinforcement learning. In Machine
Learning Proceedings 1994, pages 157–163. Elsevier.
Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap,
T., Harley, T., Silver, D., and Kavukcuoglu, K. (2016).
Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning,
pages 1928–1937.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M.,
Fidjeland, A. K., Ostrovski, G., et al. (2015). Human-
level control through deep reinforcement learning. Nature, 518(7540):529.
Nair, R., Varakantham, P., Tambe, M., and Yokoo, M.
(2005). Networked distributed pomdps: A synthesis
of distributed constraint optimization and pomdps. In
AAAI, volume 5, pages 133–139.
Nguyen, D. T., Kumar, A., and Lau, H. C. (2017). Policy
gradient with value function approximation for collective multiagent planning. In Advances in Neural
Information Processing Systems, pages 4322–4332.
Perolat, J., Leibo, J. Z., Zambaldi, V., Beattie, C., Tuyls,
K., and Graepel, T. (2017). A multi-agent reinforcement learning model of common-pool resource appropriation. In Advances in Neural Information Processing
Systems, pages 3646–3655.
Schulman, J., Abbeel, P., and Chen, X. (2017). Equivalence between policy gradients and soft q-learning.
arXiv preprint arXiv:1704.06440.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L.,
Van Den Driessche, G., Schrittwieser, J., Antonoglou,
I., Panneershelvam, V., Lanctot, M., et al. (2016). Mastering the game of go with deep neural networks and
tree search. nature, 529(7587):484–489.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,
I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M.,
Bolton, A., et al. (2017). Mastering the game of go
without human knowledge. Nature, 550(7676):354.
Varakantham, P., Adulyasak, Y., and Jaillet, P. (2014).
Decentralized stochastic planning with anonymity in
interactions. In AAAI, pages 2505–2512.
Varakantham, P., Cheng, S.-F., Gordon, G., and Ahmed,
A. (2012). Decision support for agent populations in
uncertain and congested environments. In AAAI.
Verma, T., Varakantham, P., Kraus, S., and Lau, H. C.
(2017). Augmenting decisions of taxi drivers through
reinforcement learning for improving revenues. In
International Conference on Automated Planning and
Scheduling, pages 409–417.
Watkins, C. J. and Dayan, P. (1992). Q-learning. Machine
learning, 8(3-4):279–292.
| 2cs.AI
|
Performance Analysis on Molecular Dynamics Simulation of Protein Using
GROMACS
A.D.Astuti and A.B. Mutiara
Department of Informatics Engineering, Faculty of Industrial Technology
Gunadarma University
E-mail: [email protected]
Abstracts
Development of computer technology in chemistry, bring many application of chemistry. Not only the
application to visualize the structure of molecule but also to molecular dynamics simulation. One of them is
Gromacs. Gromacs is an example of molecular dynamics application developed by Groningen University. This
application is a non-commercial and able to work in the operating system Linux. The main ability of Gromacs is to
perform molecular dynamics simulation and minimization energy.
In this paper, the author discusses about how to work Gromacs in molecular dynamics simulation of some
protein. In the molecular dynamics simulation, Gromacs does not work alone. Gromacs interact with pymol and
Grace. Pymol is an application to visualize molecule structure and Grace is an application in Linux to display
graphs. Both applications will support analysis of molecular dynamics simulation.
Keywords: molecular dynamics, Gromacs, Linux, pymol, grace, chemical
1. Introduction
1.1. Backgrounds
Computer is necessary for life of
society, especially in chemistry. Now, many
non-commercial application of chemistry is
available in Windows version and also
Linux. The applications are very useful not
only in visualization molecule structure but
also to molecular dynamics simulation.
Molecular dynamics is a simulation
method with computer which allowed
representing interaction molecules of atom
in certain time period. Molecular dynamics
technique is based on Newton law and
classic mechanics law. Gromacs is one of
application which able to do molecular
dynamics simulation based on equation of
Newton law. Gromacs was first introduced
by Groningen University as molecular
dynamics simulation machine.
1.2. Definition Of Problem
This writing is focused at usage of
Gromacs application. In this writing, writer
tell about how to install Gromacs, Gromacs
concepts, file format in Gromacs, Program
in Gromacs, and analysis result of
simulation.
2. Theory
2.1. Protein
Protein is complex organic compound
that has a high molecular weight. Protein is
also a polymer of amino acid that has been
linked to one another with a peptide bond.
Structure of protein divided into three,
namely the structure of primary, secondary,
tertiary and quaternary. Primary structure is
amino acid sequence of a protein linked to it
through a peptide bond. Secondary structure
is a three-dimensional structure of local
range of amino acids in a protein stabilized
by hydrogen bond. Tertiary structure is a
combination
of
different
secondary
structures that produce three-dimensional
form. Tertiary structure is usually a lump.
Some of the protein molecule can interact
physically without covalent bonds to form a
stable oligomer (e.g. dimer, trimer, or
kuartomer) and form a Quaternary structure
(e.g. rubisco and insulin).
2.2. Molecular Dynamics
perception, from starting position, it is
possible to calculate the next position and
velocities of atoms at a small time interval
and force in the new position. This can be
repeated many times, even up to hundreds of
times.
Molecular dynamics procedure can be
described with the flowchart as follows:
Molecular dynamics is a method to
investigate exploring structure of solid,
liquid, and gas. Generally, molecular
dynamics use equation of Newton law and
classical mechanics.
Molecular
dynamics
was
first
introduced by Alder and Wainwright in the
late 1950s, this method is used to study the
interaction hard spheres. From these studies,
they learn about behavior of simple liquids.
In 1964, Rahman did the first simulations
using realistic potential for liquid argon. And
in 1974, Rahman and Stillinger performed
the first molecular dynamics simulations
using a realistic system that is simulation of
liquid water. The first protein simulations
appeared in 1977 with the simulation of the
bovine pancreatic trypsin inhibitor (BPTI) [8].
Figure 2.1 Flowchart of molecular
dynamics [13]
The main purposes of the molecular
dynamics simulation is:
•
Generate trajectory molecules in the
limited time period.
•
Become the bridge between theory and
experiments.
•
Allow the chemist to make simulation
that can’t bo done in the laboratory
From The figure above can be seen the
process of molecular dynamics simulation.
The arrow indicates a path sequence the
process will be done. The main process is
calculating forces, computing motion of
atoms, and showing statistical analysis the
configuration for each atom.
2.3. The Concepts of Molecular Dynamics
3. Discussion
In molecular dynamics, force between
molecules is calculated explicitly and the
motion of is computed with integration
method. This method is used to solve
equation of Newton in the constituents
atomic. The starting condition is the position
and velocities of atoms. Based on Newton’s
3.1. Gromacs Concepts
Gromacs is an application that was first
developed by department of chemistry in
Groningen University. This application is
used to perform molecular dynamics
simulations and energy minimization.
The concept used in Gromacs ia a
periodic boundary condition and group.
Periodic boundary condition is classical way
used in Gromacs to reduce edge effect in
system. The atom will be placed in a box,
surrounded by a copy of the atom.
Gromacs applications can be downloaded in
http://www.gromacs.org. How to install
Gromacs is as follows:
1. Download FFTW in http://www.fftw.org
2. Extract file FFTW
% tar xzf fftw3-3.0.1.tar.gz
% cd fftw3-3.0.1
3. Configuration
% ./configure -prefix=/home/anas/fftw3 --enable-float
4. Compile fftw
% make
5. Installing fftw
% make install
6. After fftw installed then install Gromacs.
Extract Gromacs.
% Tar xzf gromacs-3.3.1.tar.gz
% cd gromacs-3.3.1
Figure 2.2: Periodic boundary condition in
Two Dimensions [7]
In Gromacs there are some model boxes.
That is triclinic, cubic, and octahedron.
The second concept is group. This
concept is used in Gromacs to show an
action. Each group can only have a
maximum number of 256 atoms, where each
atom can only have six different groups.
7. Configuration
% Export CPPFLAGS =I/home/anas/fftw3/include
% export LDFLAGS=L/home/anas/fftw3/lib % Export
LDFLAGS =- L/home/anas/fftw3/lib
%. /configure –
prefix=/home/anas/gromacs %. /
Configure-prefix = / home / Anas /
gromacs
8. Compile and install gromacs
3.2. Install Gromacs
Gromacs applications can run on the
operating system Linux and windows. To
run Gromacs on multiple computer, then the
required MPI (Message Passing Interface)
library for parallel communication.
%make & make install
3.3. Flowchart Of Gromacs
Gromacs need several steps to set up a
file input in the simulation. The steps can be
seen in flowchart below:
4. Energy minimization
The process of adding hydrogen bond or
termination may cause atoms in protein
too close, so that the collision occurred
between the atoms. The collision
between atoms can be removed by
energy minimization. Gromacs use mdp
file for setup parameters. Mdp file
specified number of step and cut-off
distance. Use grompp to generate input
file and mdrun to run energi
minimization. The energy minimization
may take some time, depending on the
CPU [21].
5. Molecular dynamics simulation
Figure 3.1 Flowchart Gromacs [16].
Flowchart above illustrates how to do
molecular dynamics simulation of a protein.
The steps are divided into:
1. Conversion of the pdb file
At this step pdb is converted to gromos
file (gro) with pdb2gmx. Pdbgmx also
created topology file (.top)
2. Generate box
At this step, the editconf will determine
the type of box and the box size that will
be used in the simulation. on Gromacs
there are three types of box, namely
triclinic, cubic, and octahedron.
The process of molecular dynamics
simulation is the same as energy
minimization. Grompp prepare the input
file to run mdrun. Molecular dynamics
simulations also need mdp file for setup
parameters. Most option of mdrun on
molecular dynamics is used in energy
minimization except –x to generate
trajectory file.
6. Analysis
After the simulation has finished, the last
step is to analyze the simulation result
with the following program:
•
Ngmx to perform trajectory
•
G_energy to monitor energy
•
G_rms to calculated RMSD (root
mean square deviation)
3. Solvate protein
The next step is solvate the protein in
box. The program genbox will do it.
Genbox will generate a box defined by
editconf based on the type. Genbox also
determined the type of water model that
will be used and add number of water
molecule for solvate protein the water
model commonly used is SPC (Simple Point
Charge).
3.4. File Format
In Gromacs, there are several types of
file format:
1. Trr: a file format that contains data
trajectory for simulation. It stores
information about the coordinates,
velocities, force, and energy.
2. Edr: a file format that stores information
about energies during the simulation and
energy minimization.
•
Triclinic, a box-shaped triclinic
•
Cubic, a square-shaped box with all
four side equal
3. Pdb: a form of file format used by
Brookhaven protein data bank. This file
contains information about position of
atoms in structure of molecules and
coordinates based on ATOM and
HETATM records.
•
Octahedron, a combination
octahedron and dodecahedron.
4. Xvg: a form of file format that can be
run by Grace. This file is used to perform
data in graphs.
5. Xtc: portable format for trajectory. This
file shows the trajectory data in
Cartesian coordinates.
6. Gro: a file format that provides
information about the molecular
structure in format gromos87. The
information displayed in columns, from
left to right.
7. Tpr: a binary file that is used as input file
in the simulation. This file can not be
read through the normal editor.
8. Mdp: a file format that allows the user to
setup the parameters in simulation or
energy minimization.
of
3.5.3. Grompp
Grompp is a pre-processor program.
Grompp have some ability that is:
•
Reading a molecular topology file
•
Check the validity of file.
•
Expands topology from the molecular
information into the atomic
information.
•
Recognize and read topology file (*.
top), the parameter file (*. tpr) and the
coordinates file (*. gro).
•
Generate *. tpr file as input in the
molecular dynamics and energy of
contraction that will be done by mdrun.
Grompp copy any information that required
on topology file.
3.5.4. Genbox
3.5. Gromacs Programs
Genbox can do 3 things:
3.5.1. Pdb2gmx
•
Generate solvent box
Pdb2gmx is a program that is used to
convert pdb file. Pdb2gmx can do some
things such as reading file pdb, adding
hydrogen to molecule structure, and generate
coordinate file a topologi file.
•
Solvate protein
•
Adding extra molecules on random
position
3.5.2. Editconf
Editconf is used to define box water
that will be used for simulation. This
program not only defines the model, but also
set the relative distance between edge of box
and molecules. There are 3 types of box such
as
Genbox remove atom if distance
between solvent and solute is less then
sum of vanderwalls radii of each atom.
3.5.5. Mdrun
Mdrun is main program for
computing chemistry. Not only
performs
molecular
dynamics
simulation, but it can also perform
Brownian
dynamics,
Langevin
dynamics, and energy minimization.
Mdrun can read tpr as input file and
generate three type of file such as
trajectory file, structure file, and
energy file.
4. Result Of Simulation
4.1. Research Method
The testing carried out on different
types of protein. Each protein has different
structure and number of atom. Testing is
based on flowchart of gromacs. This testing
do two process, the first is energy
minimization and the second is molecular
dynamics simulation. Number of step for
energy minimization is 200 numstep and
molecular dynamics is 500 numstep.
Figure 4.1 Mechanism of Unfolded
State [16]
4.2. Analysis of Simulation
From the testing, that was made on 4
different types of protein that can be seen the
difference form of molecule before and after
simulation.
In
molecular
dynamics
simulation, occur changes mechanisms of
protein structure from folded state to
unfolded state. Its mechanism is as follows:
In the molecular dynamics simulation
above, each protein has a different velocity
simulation.
Table 4.1 Simulation Time Table
Protein
Number
of Atom
Simulasi ( minute: sec)
for 500 iterations
Minute:sec
sec
Alpha-Lactalbulmin
7960
34:07
2047
1gg1-kappa d1.3 fv
(Light Chain)
2779
20:07
1207
RibonuleosideDiphosphate
Reductase 2 Alpha
5447
3:30
210
Lysozyme C
1006
1:02
62
carried out on four different types of protein.
From The results of testing, it can be seen
that each protein has a different long time.
2500
2000
1500
1000
500
0
1006
2779
5447
7960
Figure 4.2 Time Graphic Simulation
At the protein Alpha-Lactalbulmin with
number of atom 7960, long simulation time
is 34 minutes 7 seconds. 1gg1 FV-d1.3
Kappa (light chain) with number of atom
2779, long simulation time is 20 minutes 7
seconds.
Ribonuleoside-Diphosphate
Reductase Alpha 2 with number of atom
5447, long simulation time is 3 minutes 30
seconds. And Lysozyme C with the number
of atom 1006, long simulation time is 1
minute 2 seconds.
From the data above show differences
long simulations of each protein. Length of
time the simulation is depicted with a nonlinier graph.
In addition gromacs also help
understand the mechanisms Folding and
unfolding of protein.
Length of time simulation is not only
influenced by the number of atoms but also
the number of chain and water blocks.
6. Reference
In the case of protein RibonuleosideDiphosphate Reductase Alpha 2, although
the number of atom is greater than the
protein 1gg1 FV-d1.3 Kappa (Light Chain)
but the simulation time is more quickly.
Because the number of blocks and the chain
of water in this protein are lower than the
protein 1gg1 FV-d1.3 Kappa (Light Chain).
5. Conclusion
Development of the world’s computing
allows the chemist to perform molecular
dynamics simulation on hundreds of or
thousands of atoms with computer.
Molecular dynamics simulation is a
technique for investigating structure of
atoms based on the interaction with other
atoms.
This writing introduces Gromacs as one
of the applications that are able to perform
molecular dynamics simulation, especially
for protein. At this writing, the testing is
[1]. Allen, Michael P. Introduction to
Molekuler Dynamics Simulastion.
John Von Neuman Institute for
computing.2004.vol23
[2]. DeLano WL.The PyMOL Molecular
Graphics System on World Wide
Web. 2002. http:// www.pymol.org
[3]. Foster,
Bob.Fisika
Jakarta:Erlangga.2004
SMA.
[4]. Jinzhi Lei. Molecular Dynamics and
Protein Folding. Zhou Peiyuan Center
For Applied Mathematics.2004
[5]. Kurniawan, Aulia. Percobaan VIII:
Asam-Amino dan Protein
[6]. Lindahl, Erik.Parallel Molecular
Dynamics:Gromacs. 2 agustus 2006
[7]. Lindahl, Erik dkk,. Gromacs User
Manual. http://www.gromacs.org
[8]. Moleculer
Dynamics.
http://andrykidd.wordpress.com/2009/05/1
1/molecular-dynamics/
[9]. Witoelar, Aree. Perancangan dan
Analisa Simulasi Dinamika Molekul
Ensemble Mikrokanononikal dan
Kanonikal
dengan
Potensial
Lennard Jones. Laporan tugas
akhir.2002
[10]. Simulasi-Dinamika-MolekulProtein-G-Dalam-Water-Box-Pada1000-K.
http://biotata.wordpress.com/2008/12/31/s
imulasi-dinamika-molekul-protein-gdalam-water-box-pada-1000-k/
[11]. Warmada, I Wayan. Grace: salah
satu program grafik 2-dimensi
berbasis GUI di lingkungan Linux.
Lab. Geokomputasi, Jurusan Teknik
Geologi, FT UGM
[12]. http://118.98.171.140/DISPENDIK_MAL
ANGKAB/
[13]. Http://www.compsoc.man.ac.uk/~luck
y/Democritus/Theory/moldyn1.html
[14]. Http://www.ch.embnet.org/MD_tutorial/p
ages/MD.Part1.html
[15]. Http://www.gizi.net
[16]. Http://www.gromacs.org
[17]. Http://ilmu-kimia.netii.net
[18]. Http://ilmukomputer.org/
| 5cs.CE
|
O’Dorney
RESEARCH
Rings of small rank over a Dedekind domain and
their ideals
arXiv:1508.02777v3 [math.NT] 17 Aug 2015
Evan M O’Dorney
Correspondence:
[email protected]
119 Shelterwood Lane, 94506
Danville, CA, USA
Full list of author information is
available at the end of the article
Abstract
The aim of this paper is to find and prove generalizations of some of the beautiful
integral parametrizations in Bhargava’s theory of higher composition laws to the
case where the base ring Z is replaced by an arbitrary Dedekind domain R.
Specifically, we parametrize quadratic, cubic, and quartic algebras over R as well
as ideal classes in quadratic algebras, getting a description of the multiplication
law on ideals that extends Bhargava’s famous reinterpretation of Gauss
composition of binary quadratic forms. We expect that our results will shed light
on the statistical properties of number field extensions of degrees 2, 3, and 4.
Keywords: Dedekind domain; ring extension; Gauss composition; Bhargavology
AMS Subject Classification: Primary 13F05; 11E20; 11R11; 11R16; secondary
11E16; 13B02; 13A15; 11E76
1 Introduction
The mathematics that we will discuss has its roots in the investigations of classical
number theorists—notably Fermat, Lagrange, Legendre, and Gauss (see [1], Ch. I)—
who were interested in what integers are represented by expressions such as x2 +ky 2 ,
for fixed k. It became increasingly clear that in order to answer one such question,
one had to understand the general behavior of expressions of the form
ax2 + bxy + cy 2 .
These expressions are now called binary quadratic forms. It was Gauss who first
discovered that, once one identifies forms that are related by a coordinate change
x 7→ px + qy, y 7→ rx + sy (where ps − qr = 1), the forms whose discriminant
D = b2 − 4ac has a fixed value and which are primitive, that is, gcd(a, b, c) = 1, can
be naturally given the structure of an abelian group, which has the property that
if forms φ1 , φ2 represent the numbers n1 , n2 , then their product φ1 ∗ φ2 represents
n1 n2 . This group law ∗ is commonly called Gauss composition.
Gauss’s construction of the product of two forms was quite ad hoc. Since Gauss’s
time, mathematicians have discovered various reinterpretations of the composition
law on binary quadratic forms, notably:
• Dirichlet, who discovered an algorithm simplifying the understanding and
computation of the product of two forms, which we will touch on in greater
detail (see Example 5.9).
• Dedekind, who by introducing the now-standard notion of an ideal, transformed Gauss composition into the simple operation of multiplying two ideals
√
in a quadratic ring of the form Z[(D + D)/2];
O’Dorney
Page 2 of 39
• Bhargava, who in 2004 astounded the mathematical community by deriving
Gauss composition from simple operations on a 2 × 2 × 2 cube [2].
In abstraction, Bhargava’s reinterpretation is somewhat intermediate between
Dirichlet’s and Dedekind’s: it shares the integer-based concreteness of Gauss’s original investigations, yet it also corresponds to natural constructions in the realm of
ideals. One of the highlights of Bhargava’s method is that it extends to give group
structures on objects beyond binary quadratic forms, hence the title of his paper
series, “Higher composition laws.” It also sheds light on previously inaccessible conjectures about Gauss composition, such as an estimate for the number of forms of
bounded discriminant whose third power is the identity [3].
A second thread that will be woven into this thesis is the study of finite ring
extensions of Z, often with a view toward finite field extensions of Q. Quadratic rings
(that is, those having a Z-basis with just two elements) are simply and classically
parametrized by a single integer invariant, the discriminant. For cubic rings, Delone
and Faddeev prove a simple lemma (as one of many tools for studying irrationalities
of degree 3 and 4 over Q) parametrizing them by binary cubic forms ([4], pp. 101ff).
A similar classification for quartic and higher rings proved elusive until Bhargava,
using techniques inspired by representation theory, was able to parametrize quartic
and quintic rings together with their cubic and sextic resolvent rings, respectively,
and thereby compute the asymptotic number of quartic and quintic rings and fields
with bounded discriminant [5, 6, 7, 8]. The analytic virtue of Bhargava’s method
is to map algebraic objects such as rings and ideals to lattice points in bounded
regions of Rn , where asymptotic counting is much easier. (Curiously enough, the
ring parametrizations seem to reach a natural barrier at degree 5, in contrast to the
classical theory of solving equations by radicals where degree 4 is the limit.)
Bhargava published these results over the integers Z. Since then, experts have
wondered whether his techniques apply over more general classes of rings; by far
the most ambitious extensions of this sort are Wood’s classifications of quartic
algebras [9] and ideals in certain n-ic algebras [10] over an arbitrary base scheme
S. In this paper, all results are proved over an arbitrary Dedekind domain R. The
use of a Dedekind domain has the advantage of remaining relevant to the original
application (counting number fields and related structures) while introducing some
new generality.
We will focus on two parametrizations that are representative of Bhargava’s algebraic techniques in general. The first is the famous reinterpretation of Gauss
composition in terms of 2 × 2 × 2 boxes. Following [2], call a triple (I1 , I2 , I3 ) of
ideals of a quadratic ring S balanced if I1 I2 I3 ⊆ S and N (I1 )N (I2 )N (I3 ) = 1, and
call two balanced triples equivalent if Ii = γi Ii0 for some scalars γi ∈ S ⊗Z Q having
product 1. (If S is Dedekind, as is the most common application, then the balanced
triples of equivalence classes correspond to triples of ideal classes having product
1.) Then:
Theorem 1.1 ([2], Theorem 11) There is a canonical bijection between
• pairs (S, (I1 , I2 , I3 )) where S is an oriented quadratic ring of nonzero discriminant over Z and (I1 , I2 , I3 ) is an equivalence class of balanced triples of ideals
of S;
O’Dorney
Page 3 of 39
• trilinear maps β : Z2 ⊗ Z2 ⊗ Z2 → Z, up to SL2 Z-changes of coordinates in
each of the three inputs (subject to a certain nondegeneracy condition).
Our parametrization is analogous, with one crucial difference. Whereas over Z, the
only two-dimensional lattice is Z2 , over a Dedekind domain R there are as many as
there are ideal classes, and any such lattice can serve as the R-module structure of a
quadratic algebra or an ideal thereof. Using a definition of balanced and equivalent
essentially identical to Bhargava’s (see Definition 5.1), we prove:
Theorem 1.2 (see Theorem 5.3) Let R be a Dedekind domain. There is a canonical bijection between
• pairs (S, (I1 , I2 , I3 )) where S is an oriented quadratic algebra over R and
(I1 , I2 , I3 ) is an equivalence class of balanced triples of ideals of S;
• quadruples (a, (M1 , M2 , M3 ), θ, β) where a is an ideal class of R, Mi are lattices of rank 2 over R (up to isomorphism), θ : Λ2 M1 ⊗ Λ2 M2 ⊗ Λ2 M3 → a3
is an isomorphism, and β : M1 ⊗ M2 ⊗ M3 → a is a trilinear map whose three
partial duals βi : Mj ⊗ Mk → aMi∗ ({i, j, k} = {1, 2, 3}) have image a full-rank
sublattice.
Under this bijection, we get identifications Λ2 S ∼
= a and Ii ∼
= Mi .
In particular R may have characteristic 2, the frequent factors of 1/2 in Bhargava’s
exposition notwithstanding, and by weakening the nondegeneracy condition, we are
able to include balanced triples in degenerate rings.
The second main result of our paper is the parametrization of quartic rings (with
the quadratic and cubic parametrizations as preliminary cases). A key insight is to
parametrize not merely the quartic rings themselves, but the quartic rings together
with their cubic resolvent rings, a notion arising from the resolvent cubic used in
the classical solution of the quartic by radicals.
Theorem 1.3 ([5], Theorem 1 and Corollary 5) There is a canonical bijection
between
• isomorphism classes of pairs (Q, C) where Q is a quartic ring (over Z) and
C is a cubic resolvent ring of Q;
• quadratic maps φ : Z3 → Z2 , up to linear changes of coordinates on both the
input and the output.
Any quartic ring Q has a cubic resolvent, and if Q is Dedekind, the resolvent is
unique.
Our analogue is as follows:
Theorem 1.4 (see Theorems 8.3 and 8.7 and Corollary 8.6) Let R be a Dedekind
domain. There is a canonical bijection between
• isomorphism classes of pairs (Q, C) where Q is a quartic ring (over Z) and
C is a cubic resolvent ring of Q;
• quadruples (L, M, θ, φ) where L and M are lattices of ranks 3 and 2 over R
respectively, θ : Λ2 M → Λ3 L is an isomorphism, and φ : L → M is a quadratic
map.
O’Dorney
Page 4 of 39
Under this bijection, we get identifications Q/R ∼
= L and C/R ∼
= M.
Any quartic ring Q has a cubic resolvent, and if Q is Dedekind, the resolvent is
unique.
1.1 Outline
The remainder of this paper is structured as follows. In section 2, we set up basic
definitions concerning projective modules over a Dedekind domain. In sections 3 and
4, respectively, we generalize to Dedekind base rings two classical parametrizations,
namely of quadratic algebras over Z and of their ideals. In section 5, we prove
Bhargava’s parametrization of balanced ideal triples (itself a generalization of Gauss
composition) over a Dedekind domain. In section 6, we work out in detail a specific
example—unramified extensions of Zp —that allows us to explore the notion of
balanced ideal triple in depth. In sections 7 and 8, we tackle cubic and quartic
algebras respectively, and in section 9, we discuss results that would be useful when
using the preceding theory to parametrize and count quartic field extensions.
2 Modules and algebras over a Dedekind domain
A Dedekind domain is an integral domain that is Noetherian, integrally closed, and
has the property that every nonzero prime ideal is maximal. The standard examples
of Dedekind domains are the ring of algebraic integers OK in any finite extension
K of Q; in addition, any field and any principal ideal domain (PID), such as the
ring C[x] of polynomials in one variable, is Dedekind. In this section, we summarize
properties of Dedekind domains that we will find useful; for more details, see [11],
pp. 9–18.
The salient properties of Dedekind domains were discovered through efforts to
generalize prime factorization to rings beyond Z; in particular, every nonzero ideal
a in a Dedekind domain R is expressible as a product pa1 1 · · · pann of primes, unique
up to ordering. Our motivation for using Dedekind domains stems from two other
related properties. Recall that a fractional ideal or simply an ideal of R is a finitely
generated nonzero R-submodule of the fraction field K of R, or equivalently, a set
of the form aa where a ⊆ R is a nonzero ideal and a ∈ K × . (The term “ideal” will
from now on mean “(nonzero) fractional ideal”; if we wish to speak of ideals in the
ring-theoretic sense, we will use a phrasing such as “ideal a ⊆ R.”) The first useful
property is that any fractional ideal a ⊆ K has an inverse a−1 such that aa−1 = R.
This allows us to form the group I(R) of nonzero fractional ideals and quotient
by the group K × /R× of principal ideals to obtain the familiar ideal class group,
traditionally denoted Pic R. (For the ring of integers in a number field, the class
group is always finite; for a general Dedekind domain this may fail, e.g. for the ring
C[x, y]/(y 2 − (x − a1 )(x − a2 )(x − a3 )) of functions on a punctured elliptic curve.)
The second property that we will find very useful is that finitely generated modules over a Dedekind domain are classified by a simple theorem generalizing the
classification of finitely generated abelian groups. For our purposes it suffices to
discuss torsion-free modules, which we will call lattices.
Definition 2.1 Let R be a Dedekind domain and K its field of fractions. A lattice
over R is a finitely generated, torsion-free R-module M . If M is a lattice, we will
O’Dorney
Page 5 of 39
denote by the subscript MK its K-span M ⊗R K (except when M is denoted by
a symbol containing a subscript, in which case a superscript will be used). The
dimension of MK over K is called the rank of the lattice M .
A lattice of rank 1 is a nonzero finitely generated submodule of K, i.e. an ideal;
thus isomorphism classes of rank-1 lattices are parametrized by the class group
Pic R. The situation for general lattices is not too different.
Theorem 2.2 (see [11], Lemma 1.5, Theorem 1.6, and the intervening Remark) A
lattice M over R is classified up to isomorphism by two invariants: its rank m and
its top exterior power Λm M . Equivalently, every lattice is a direct sum a1 ⊕ · · · ⊕ am
of nonzero ideals, and two such direct sums a1 ⊕· · ·⊕am , b1 ⊕· · ·⊕bn are isomorphic
if and only if m = n and the products a1 · · · am and b1 · · · bn belong to the same ideal
class.
In this paper we will frequently be performing multilinear operations on lattices.
Using Theorem 2.2, it is easy to show that these operations behave much more
“tamely” than for modules over general rings. Specifically, for two lattices M =
a1 u1 ⊕ · · · ⊕ am um and N = b1 v1 ⊕ · · · ⊕ bn vn , we can form the following lattices:
• the tensor product
M
M ⊗N =
ai bj (ui ⊗ vj );
1≤i≤m
1≤j≤n
• the symmetric powers
M
Symk M =
ai1 · · · aik (ui1 ⊗ · · · ⊗ uik )
1≤i1 ≤···≤ik ≤m
and the exterior powers
Λk M =
M
ai1 · · · aik (ui1 ∧ · · · ∧ uik )
1≤i1 <···<ik ≤m
of ranks n+k−1
and
k
• the dual lattice
n
k
respectively;
M ∗ = Hom(M, R) =
M
∗
a−1
i ui ;
1≤i≤m
• and the space of homomorphisms
Hom(M, N ) ∼
= M∗ ⊗ N =
M
∗
a−1
i bj (ui ⊗ vj ).
1≤i≤m
1≤j≤n
A particular composition of three of these constructions is of especial relevance to
the present thesis:
O’Dorney
Page 6 of 39
Definition 2.3 If M and N are lattices (or, more generally, M is a lattice and N
is any R-module), then a degree-k map φ : M → N is an element of (Symk M ∗ )⊗N .
A map to a lattice N of rank 1 is called a form.
In terms of decompositions M = a1 u1 ⊕ · · · ⊕ am um and N = b1 v1 ⊕ · · · ⊕ bn vn ,
a degree-k map can be written in the form
φ(x1 u1 + · · · + xm um ) =
n
X
X
ai1 ,...,im ,j · xi11 · · · ximm vj ,
j=1 i1 +···+im =k
m
where the coefficients ai1 ,...,im ,j belong to the ideals a1−i1 · · · a−i
m bj needed to make
each term’s value belong to N . For example, over R = Z, a quadratic map from Z2
to Z is a quadratic expression
φ(x, y) = ax2 + bxy + cy 2
in the coordinates x, y ∈ (Z2 )∗ on Z2 . Two caveats about this notion are in order:
• Although such a degree-k map indeed yields a function from M to N (evaluated by replacing every functional in M ∗ appearing in the map by its value
on the given element of M ), it need not be unambiguously determined by this
function if R is finite. For instance, if R = F2 is the field with two elements,
the cubic map from F22 to F2 defined by φ(x, y) = xy(x + y) vanishes on each
of the four elements of F22 , though it is not the zero map.
• Also, one must not confuse (Symk M ∗ ) ⊗ N with the space (Symk M )∗ ⊗ N of
symmetric k-ary multilinear functions from M to N . Although both lattices
have rank n m+k−1
and there is a natural map from one to the other (defined
k
by evaluating a multilinear function on the diagonal), this map is not in
general an isomorphism. For instance, the quadratic forms φ : Z2 → Z arising
from a symmetric bilinear form λ((x1 , y1 ), (x2 , y2 )) = ax1 x2 +b(x1 y2 +x2 y1 )+
cy1 y2 are exactly those of the form φ(x, y) = ax2 + 2bxy + cy 2 , with middle
coefficient even.
The image φ(M ) of a degree-k map φ : M → N is the smallest sublattice N 0 ⊆ N
such that φ is a degree-k map from M to N 0 , i.e. lies in the image of the natural
injection (Symk M ∗ ) ⊗ N 0 ,→ (Symk M ∗ ) ⊗ N . It may be computed as follows: if
φ(x1 u1 + · · · + xm um ) =
X
xi11 · · · ximm · vi1 ,...,im ,
i1 +···+im =k
then φ(M ) is the R-span of all the 1-dimensional sublattices ai11 · · · aimm vi1 ,...,im in
N . (It is not the same as the span of the values of φ as a function on M .)
If L ⊆ M are two lattices of rank n, the index [M : L] is the ideal a such that
a · Λn L = Λn M.
Since Λn L and Λn M are of rank 1, this is well defined.
O’Dorney
Page 7 of 39
2.1 Algebras
An algebra of rank n over R is a lattice S of rank n equipped with a multiplication
operation giving it the structure of a (unital commutative associative) R-algebra.
Since R is integrally closed, the sublattice generated by 1 ∈ S must be primitive
(that is, the lattice it generates is maximal for its dimension, and therefore a direct
summand of S), implying that the quotient S/R is a lattice of rank n − 1 and we
have a noncanonical decomposition
S = R ⊕ S/R.
(1)
We will be concerned with algebras of ranks 2, 3, and 4, which we call quadratic,
cubic, and quartic algebras (or rings) respectively.
2.2 Orientations
When learning about Gauss composition over Z, one must sooner or later come to
a problem that vexed Legendre (see [1], p. 42): If one considers quadratic forms up
to GL2 Z-changes of variables, then a group structure does not emerge because the
conjugate forms ax2 ± bxy + cy 2 , which ought to be inverses, have been identified.
Gauss’s insight was to consider forms only up to “proper equivalence,” i.e. SL2 Z
coordinate changes. This is tantamount to considering quadratic forms not simply
on a rank-2 Z-lattice M , but on a rank-2 Z-lattice equipped with a distinguished
generator of its top exterior power Λ2 M . For general lattices over Dedekind domains,
whose top exterior powers need not belong to the principal ideal class, we make the
following definitions.
Definition 2.4 Let a be a fractional ideal of R. A rank-n lattice M is of type a
if its top exterior power Λn M is isomorphic to a; an orientation on M is then a
choice of isomorphism α : Λn M → a. The possible orientations on any lattice M are
of course in noncanonical bijection with the units R× . The easiest way to specify
an orientation on M is to choose a decomposition M = b1 u1 ⊕ · · · ⊕ bn un , where
the ideals bi are scaled to have product a, and then declare
α(y1 u1 ∧ · · · ∧ yn un ) = y1 · · · yn .
An orientation on a rank-n R-algebra S is the same as an orientation on the lattice
S, or equivalently on the lattice S/R, due to the isomorphism between Λn S and
Λn−1 S/R given by
1 ∧ v1 ∧ · · · ∧ vn−1 7→ ṽ1 ∧ · · · ∧ ṽn−1 .
(Here, and henceforth, we use a tilde to denote the image under the quotient map
by R, so that the customary bar can be reserved for conjugation involutions. This is
opposite to the usual convention where ṽ denotes a lift of v under a quotient map.)
O’Dorney
Page 8 of 39
3 Quadratic algebras
Before proceeding to Bhargava’s results, we lay down as groundwork two parametrizations that, over Z, were known classically. These are the parametrizations of
quadratic algebras and of ideal classes in quadratic algebras. The extension of these
to other base rings has been thought about extensively, with many different kinds
of results produced (see [12] and the references therein). Here, we prove versions
over a Dedekind domain that parallel our cubic and quartic results.
Let S be a quadratic algebra over R. Since S/R has rank 1, the decomposition
(1) simplifies to S = R ⊕ aξ for an (arbitrary) ideal a in the class of Λ2 S and some
formal generator ξ ∈ SK . The algebra is then determined by a and a multiplication
law ξ 2 = tξ − u, which allows us to describe the ring as R[aξ]/(a2 (ξ 2 − tξ + u)), a
subring of K[ξ]/(ξ 2 − tξ + u). Alternatively, we can associate to the ring its norm
map
NS/R : S → R,
x + yξ 7→ x2 + txy + uy 2 .
It is evident that this is just another way of packaging the same data, namely two
numbers t ∈ a−1 and u ∈ a−2 . The norm map is more readily freed from coordinates
than the multiplication table, yielding the following parametrization.
Lemma 3.1 Quadratic algebras over R are in canonical bijection with rank-2 Rlattices M equipped with a distinguished copy of R and a quadratic form φ : M → R
that acts as squaring on the distinguished copy of R.
Proof Given M and φ, the distinguished copy of R must be primitive (otherwise
φ would take values outside R), yielding a decomposition M = R ⊕ aξ. Write φ in
these coordinates as
φ(x + yξ) = x2 + txy + uy 2 ;
then the values t ∈ a−1 and u ∈ a−2 can be used to build a multiplication table
on M having the desired norm form (which is unique, as for any fixed coordinate
system, the norm form determines t and u, which determine the multiplication
table).
If there is a second copy of R on which NS/R restricts to the squaring map, it must
be generated by a unit of S with norm 1, multiplication by which induces an automorphism of the lattice with norm form. Hence we can eliminate the distinguished
copy of R and arrive at the following arguably prettier parametrization:
Theorem 3.2 Quadratic algebras over R are in canonical bijection with rank-2
R-lattices M equipped with a quadratic form φ : M → R attaining the value 1.
For our applications to Gauss composition it will also be helpful to have a
parametrization of oriented quadratic algebras. An orientation α : Λ2 R → a can
be specified by choosing an element ξ with α(1 ∧ ξ) = 1. Since ξ is unique up to
translation by a−1 , the parametrization is exceedingly simple.
O’Dorney
Page 9 of 39
Theorem 3.3 For each ideal a of R, there is a canonical bijection between oriented
quadratic algebras of type a and pairs (t, u), where t ∈ a−1 , u ∈ a−2 , up to the action
of a−1 via
s.(t, u) = (t + 2s, u + st + s2 )
One other fact that will occasionally be useful is that every quadratic algebra has
an involutory automorphism defined by x̄ = Tr x − x or, in a coordinate representation
S = R[aξ]/(a2 (ξ 2 − tξ + u)),
by ξ 7→ t − ξ. (The first of these characterizations shows that the automorphism is
well-defined, the second that it respects the ring structure.)
Example 3.4 When R = Q (or more generally any Dedekind domain in which 2
is a unit), then completing the square shows that oriented quadratic algebras are in
√
bijection with the forms x2 − ky 2 , k ∈ Q, each of which yields an algebra S = Q[ k]
√
oriented by α(1 ∧ k) = 1.
√
If we pass to unoriented extensions, then we identify Q[ k] with its rescalings
p
√
Q[f k] ∼
= Q[ f 2 k], f ∈ Q× . The resulting orbit space Q/(Q× )2 parametrizes
quadratic number fields, plus the two nondomains
√
Q[ 0] = Q[]/(2 )
√
and Q[ 1] ∼
= Q ⊕ Q.
Example 3.5 When R = Z, we can almost complete the square, putting a general
x2 + txy + uy 2 in the form
x2 −
D 2
y
4
or x2 + xy −
D−1 2
y .
4
Here D = t2 − 4u is the discriminant, the standard invariant used in [2] to
parametrize oriented quadratic rings. It takes on all values congruent to 0 or 1 mod
4. It also parametrizes unoriented quadratic rings, since each such ring has just
two orientations which are conjugate under the ring’s conjugation automorphism.
The rings of integers of number fields are then parametrized by the fundamental
discriminants which are not a square multiple of another discriminant, with the
exception of 0 and 1 which parametrize Z[]/2 and Z ⊕ Z respectively.
Example 3.6 For an example where discriminant-based parametrizations are inapplicable, consider the field R = F2 of two elements. Any nonzero quadratic form
attains the value 1, and there are three such, namely
x2 ,
xy,
and x2 + xy + y 2 .
They correspond to the three quadratic algebras over F2 , respectively F2 []/2 ,
F2 ⊕ F2 , and F4 .
O’Dorney
Page 10 of 39
4 Ideal classes of quadratic algebras
We can now parametrize ideal classes of quadratic algebras, in a way that partially
overlaps [12]. To be absolutely unambiguous, we make the following definition for
quadratic algebras that need not be domains:
Definition 4.1 Let S be a quadratic algebra over R. A fractional ideal (or just
an ideal ) of S is a finitely generated S-submodule of SK that spans SK over K.
Two fractional ideals are considered to belong to the same ideal class if one is a
×
scaling of the other by a scalar γ ∈ SK
. (This is clearly an equivalence relation.)
The ideal classes together with the operation induced by ideal multiplication form
the ideal class semigroup, and the invertible ideal classes form the ideal class group
Pic S.
The condition in bold means that, for instance, the submodule R ⊕ {0} ⊆ R ⊕ R is
not a fractional ideal. Of course, any ideal that is invertible automatically satisfies
it.
Theorem 4.2 (cf. [12], Corollary 4.2) For each ideal a of R, there is a bijection
between
• ideal classes of oriented quadratic rings of type a, and
• rank-2 lattices M equipped with a nonzero quadratic map φ : M → a−1 · Λ2 M .
In this bijection, the ideal classes that are invertible correspond exactly to the forms
that are primitive, that is, do not factor through any proper sublattice of a−1 · Λ2 M .
Proof Suppose first that we have a quadratic ring S = R⊕aξ, oriented by α(1∧ξ) =
1, and a fractional ideal I of R. Construct a map φ : I → a−1 · Λ2 I by
ω 7→ ω ∧ ξω.
Here ξω ∈ a−1 I so the wedge product lies in a−1 · Λ2 I, and we get a well-defined
×
quadratic map φ, scaling appropriately when I is scaled by an element of SK
. Note
that φ is nonzero because, after extending scalars to K, the element 1 ∈ IK = SK
is mapped to 1 ∧ ξ 6= 0.
It will be helpful to write this construction in coordinates. Let I" = b1#η1 ⊕ b2 η2 be
a b
, that is,
a decomposition into R-ideals, and let ξ act on I by the matrix
c d
ξη1 = aη1 + cη2
(2)
ξη2 = bη1 + dη2
where a, b, c, d belong to the relevant ideals: a, d ∈ a−1 , b ∈ a−1 b1 b−1
2 , and c ∈
a−1 b−1
b
.
Then
we
get
2
1
φ(xη1 + yη2 ) = (xη1 + yη2 ) ∧ (xξη1 + yξη2 )
= (xη1 + yη2 ) ∧ (axη1 + cxη2 + byη1 + dyη2 )
= (cx2 + (d − a)xy − by 2 )(η1 ∧ η2 ) ∈ a−1 b1 b2 (η1 ∧ η2 ) = a−1 Λ2 I.
O’Dorney
Page 11 of 39
(3)
(Now φ appears clearly as a tensor in Sym2 I ∗ ⊗ a−1 · Λ2 M .)
We now seek to reconstruct the ideal I from its associated quadratic form. Given
an ideal a, a lattice M = b1 η1 ⊕ b2 η2 , and a quadratic map φ(xη1 + yη2 ) = (px2 +
qxy + ry 2 )(η1 ∧ η2 ) to a−1 · Λ2 M , we may choose a = 0, b = −r, c = p, and
d = q to recover an action (2) of ξ on R yielding the form φ. By (3), this action is
unique up to adding a constant to a and d, which simply corresponds to a change
of basis ξ 7→ ξ + a. Next, by the Cayley-Hamilton theorem, the formal expression
ξ 2 −qξ+pr annihilates M , so M is a module over the ring S = R[aξ]/(a2 (ξ 2 −qξ+pr))
corresponding to the quadratic form x2 + qxy + pry 2 . The last step is to embed M
into SK , or equivalently, to identify MK with SK . For this, we divide into cases
based on the kind of ring that SK is, or equivalently the factorization type of the
polynomial f (x) = x2 − qx + pr over K.
• If f is irreducible, then SK is a field, and MK is an SK -vector space of dimension 1, isomorphic to SK .
• If f has two distinct roots, then SK ∼
= K ⊕ K. There are three different SK modules having dimension 2 as K-vector spaces: writing I1 and I2 for the two
copies of K within SK , we can describe them as I1 ⊕ I1 , I2 ⊕ I2 , and I1 ⊕ I2 .
But on the first two, every element of SK acts as a scalar. If MK were one of
these, then the quadratic form φ(ω) = ω ∧ ξω would be identically 0, which
is not allowed. So MK ∼
= I1 ⊕ I2 ∼
= SK .
• Finally, if f has a double root, then SK ≡ K[]/2 . There are two SK -modules
having dimension 2 as a K-vector space: K ⊕ K and SK . On K ⊕ K, SK
acts by scalars and we get a contradiction as before. So MK ∼
= SK .
This shows that there is always at least one embedding of M into SK . To show
there is at most one up to scaling, we need that every automorphism of SK as an
SK -module is given by multiplication by a unit. But this is trivial (the image of 1
determines everything else).
It will be convenient to have as well an explicit reconstruction of an ideal from
its associated quadratic form. First change coordinates on M such that p 6= 0. (If
r 6= 0, swap b1 η1 and b2 η2 ; if p = 0 but q 6= 0, translate η2 7→ η2 + tη1 for any
nonzero t ∈ b1 b−1
2 .) Then the ideal
I = b1 + b2
ξ
p
(4)
of the ring S = R[aξ]/(a2 (ξ 2 − qξ + pr)) corresponding to the norm form x2 + qxy +
pry 2 is readily seen to yield the correct quadratic form.
We now come to the equivalence between invertibility of ideals and primitivity
of forms. Suppose first that φ : M → a−1 · Λ2 M is imprimitive, that is, there is an
ideal a0 strictly containing a such that φ actually arises from a quadratic map φ0 :
M → a0−1 ·Λ2 M . Following through the (first) construction, we see that φ and φ0 give
the same ξ-action on I = M but embed it as a fractional ideal in two different rings,
0 ∼
S = R ⊕ aξ and S 0 = R ⊕ a0 ξ. We naturally have SK ∼
= SK
= K[ξ]/(ξ 2 − qξ + pr),
0
and S is a subring of S . Suppose I had an inverse J as an S-ideal. Then since I is
an S 0 -ideal, the product IJ = S must be an S 0 -ideal, which is a contradiction.
O’Dorney
Page 12 of 39
Conversely, suppose that φ is primitive and I has been constructed using (4).
Consider the conjugate ideal
ξ¯
q−ξ
I¯ = b1 + b2 = b1 + b2
p
p
and form the product
ξ
q−ξ
b1 + b2
b1 + b2
p
p
q
−
ξ
ξ
ξ ξ¯
+ b22 2
= b21 + b1 b2 + b1 b2
p
p
p
1
= (pb21 + qb1 b2 + rb22 + ξb1 b2 ).
p
I I¯ =
The first three terms in the parenthesis are all fractional ideals in K. The condition
that φ maps into a−1 · Λ2 I is exactly that these lie in a−1 b1 b2 , and the condition
of primitivity is that they do not all lie in any smaller ideal, that is, their sum is
a−1 b1 b2 . So
I I¯ =
a−1 b1 b2
b1 b2 −1
(a + Rξ) =
· S.
p
p
(5)
We conclude that
−1 ¯
2 −1 ¯
I −1 = ab−1
I
1 b2 pI = α(Λ I)
is an inverse for I.
Note that our proof of the invertibility-primitivity equivalence shows something
more: that any fractional ideal I of a quadratic algebra S is invertible when considered as an ideal of a certain larger ring S 0 , found by “canceling common factors”
in its associated quadratic form. The following relation is worth noting:
Corollary 4.3 If I is an ideal of a quadratic algebra S and S 0 = End I ⊆ SK is
its ring of endomorphisms, then
I I¯ =
α(Λ2 I)
· S0.
α(Λ2 S 0 )
Proof The ring S 0 is the one occurring in the proof that imprimitivity implies noninvertibility, provided that the ideal a0 is chosen to be as large as possible (i.e. equal
to (pb21 + qb1 b2 + rb22 )−1 ), so that I is actually invertible with respect to S 0 . This
S 0 must be the endomorphism ring End I, or else I would be an ideal of an even
larger quadratic ring. (We here need that End I is finitely generated and hence a
×
quadratic ring. This is obvious, as it is contained in x−1 I for any x ∈ SK
∩ I.)
0
2 0
Viewing α, by restriction, as an orientation on S , we have α(Λ S ) = a0 and the
formula is reduced to that for I −1 above.
O’Dorney
Page 13 of 39
Example 4.4 If R = Z (or more generally any PID), then the situation simplifies
to a = Z and M = Z2 , and we recover a bijection between ideal classes and binary
quadratic forms. But the theorem also requires us, when changing coordinates on M ,
to change coordinates on Λ2 M appropriately; that is, ideal classes are in bijection
with GL2 (Z)-orbits of binary quadratic forms φ : Z2 → Z, not under the natural
action but under the twisted action
"
a
c
# !
1
b
. φ (x, y) =
· φ(ax + cy, bx + dy).
ad − bc
d
(Compare [1], p. 142 and [12], Theorem 1.2.)
For an example not commonly encountered in the literature, take the order S =
Z[5i] in the domain Z[i]. Its ideal classes correspond simply to GL2 (Z)-orbits of
quadratic forms px2 + qxy + ry 2 having discriminant q 2 − 4pr = −100. Using the
standard theory of “reduction” of quadratic forms developed by Lagrange (see [1],
pp. 26ff.), we may limit our search to the bounded domain where |q| ≤ r ≤ p and
find that there are precisely three, with three corresponding ideal classes:
φ1 (x, y) = x2 + 25y 2
2
! S = Z[5i]
2
φ2 (x, y) = 2x + 2xy + 13y ! A = Zh5, 1 + ii
φ3 (x, y) = 5x2 + 5y 2
! B = Z[i].
The first two ideals, which correspond to primitive forms, are invertible (indeed
A · iA = S); the third is not. In fact we can build a multiplication table for the ideal
class semigroup.
·
S
A
B
S
S
A
B
A
A
S
B
B
B
B
B
5 Ideal triples
We turn now to one of Bhargava’s most widely publicized contributions to mathematics, the reinterpretation of Gauss’s 200-year-old composition law on primitive
binary quadratic forms in terms of simple operations on a 2 × 2 × 2 box of integers.
In fact, Bhargava produced something rather more general: a bijection ([2], Theorem 1) that takes all 2 × 2 × 2 boxes satisfying a mild nondegeneracy condition, up
to the action of the group
(
Γ=
)
3
(M1 , M2 , M3 ) ∈ (GL2 Z) :
Y
det Mi = 1 ,
i
to triples of fractional ideals (I1 , I2 , I3 ) in a quadratic ring S that are balanced, that
is, satisfy the two conditions
(a) I1 I2 I3 ⊆ S;
O’Dorney
Page 14 of 39
(b) N (I1 )N (I2 )N (I3 ) = 1. Here N (I) is the norm of the ideal I, defined by the
formula N (I) = [A : I]/[A : S] for any Z-lattice A containing both S and I.
(This should not be confused with the ideal generated by the norms of the
elements of I. Even over Z, the two notions differ: 2 · Z[i] is an ideal of norm
2 in the ring Z[2i], but every element of 2 · Z[i] has norm divisible by 4.)
The ideals Ii are unique up to a scaling by constants γi ∈ SQ× of product 1.
Our task will be to generalize this result to an arbitrary Dedekind domain. First,
the definition of balanced extends straightforwardly, provided that we define the
norm of a fractional ideal I properly, as the index of I in S as an R-lattice. The
resulting notion of balanced is a special case of the definition used in [10]:
Definition 5.1 A triple of fractional ideals I1 , I2 , I3 of an R-algebra S is balanced
if
(a) I1 I2 I3 ⊆ S;
(b) the image of Λ2 I1 ⊗ Λ2 I2 ⊗ Λ2 I3 in (Λ2 SK )⊗3 is precisely (Λ2 S)⊗3 .
The objects that we will use on the other side of the bijection are, as one might
expect, not merely 8-tuples of elements from R, because the class group intrudes.
The appropriate notion is as follows:
Definition 5.2 Let a be an ideal class of R. A Bhargava box of type a over R
consists of the following data:
• three rank-2 lattices M1 , M2 , M3 ;
• an orientation isomorphism θ : Λ2 M1 ⊗ Λ2 M2 ⊗ Λ2 M3 → a3 ;
• a trilinear map β : M1 ⊗ M2 ⊗ M3 → a satisfying the following nondegeneracy
condition: each of the three partial duals βi : Mj ⊗ Mk → aMi∗ ({i, j, k} =
{1, 2, 3}) has image a full-rank sublattice.
If we choose a decomposition of each Mi into a direct sum bi1 ⊕ bi2 of ideals,
Q
then θ becomes an isomorphism from i,j bij to a3 (which we may take to be the
identity), while β is determined by eight coefficients
−1 −1
βijk ∈ b−1
1i b2j b3k a.
Thus we stress that, in spite of all the abstraction, our parameter space indeed still
consists of (equivalence classes of) 2 × 2 × 2 boxes of numbers lying in certain ideals
contained in K.
Theorem 5.3 (cf. [2], Theorem 1; [10], Theorem 1.4) For each ideal a of R, there
is a bijection between
• balanced triples (I1 , I2 , I3 ) of ideals in an oriented quadratic ring S of type a,
×
up to scaling by factors γ1 , γ2 , γ3 ∈ SK
with product 1;
• Bhargava boxes of type a.
Remark Two balanced ideal triples may be inequivalent for the purposes of this
bijection even if corresponding ideals belong to the same class (see Example 5.9(d)).
Consequently a Bhargava box cannot be described as corresponding to a balanced
triple of ideal classes.
O’Dorney
Page 15 of 39
Proof The passage from ideals to the Bhargava box is simple and derived directly
from [2]. Given a balanced triple (I1 , I2 , I3 ) in a quadratic ring S with an orientation
α : Λ2 S → a, construct the trilinear map
β : I1 ⊗ I2 ⊗ I3 → a
x ⊗ y ⊗ z 7→ α(1 ∧ xyz).
This, together with the identification θ coming from condition (b) of Definition 5.1,
furnishes the desired Bhargava box. Since each Ii spans SK , the nondegeneracy is
not hard to check.
We seek to invert this process and reconstruct the ring S, the orientation α,
and the ideals Ii uniquely from the Bhargava box. We begin by reconstructing the
quadratic forms φi : Mi → a−1 ·Λ2 Mi corresponding to the ideals Ii . For this we first
use β to map M1 to Hom(M2 ⊗M3 , a), in other words Hom(M2 , aM3∗ ). We then take
the determinant, which lands us in Hom(Λ2 M2 , Λ2 (aM3∗ )) ∼
= a2 · Λ2 M2∗ ⊗ Λ2 M3∗ ,
which can be identified via −θ (note the sign change) with a−1 Λ2 M1 . We thus get a
quadratic form φ01 : M1 → a−1 Λ2 M1 . We claim that if the Bhargava box arose from
a triple of ideals, then this is the natural form φ1 : x 7→ x∧ξx on I1 . For convenience
we will extend scalars and prove the equality as one of forms on M1K ∼
= SK . To deal
with φ01 , we must analyze
β(x) = (y 7→ (z 7→ α(1 ∧ xyz))) ∈ Hom(M2K , M3K∗ ).
∗
Now whereas M2K is naturally identifiable with SK , to deal with M3K∗ ∼
we
= SK
have to bring in the symmetric pairing α(1 ∧ ••) : SK ⊗K SK → K, which one easily
∗
checks is nondegenerate and thus identifies SK
with SK . So we have transformed
β(x) to the element
β 0 (x) = (y 7→ xy) ∈ HomK (SK , SK ).
We then take the determinant det β 0 (x), which is simply the norm N (x) ∈ K ∼
=
2
2
HomK (Λ SK , Λ SK ). This is to be compared to
φ1 (x) = x ∧ ξx = N (x)(1 ∧ ξ) = α−1 (N (x)).
It then remains to check that we have performed the identifications properly, that
is, that the four isomorphisms
KO o
α
Λ2 (M1K ⊗SK M2K )
∧2 (x⊗y7→α(xy•))
α⊗α
Λ2 M1K ⊗ Λ2 M2K
−θ
/ Λ2 M3K∗
are compatible. In particular we discover that the pairing α(1 ∧ ••) is given in the
basis {1, ξ} by the matrix
"
0
1
1
Tr ξ
#
O’Dorney
Page 16 of 39
of determinant −1, explaining the compensatory minus sign that must be placed
on θ.
Q
Now write Mi = bi1 ηi1 ⊕ bi2 ηi2 where θ : i,j bij → a3 may be assumed to be the
identity map, and express β in these coordinates as
β
X
xijk η1i η2j η3k =
i,j,k
X
aijk xijk .
i,j,k
It will be convenient to create the single-letter abbreviations a = a111 , b = a112 ,
c = a121 , continuing in lexicographic order to h = a222 . Then φ1 sends an element
xη11 + yη12 ∈ M1 to the determinant
"
ax + ey
− det
cx + gy
#
bx + f y
= (bc − ad)x2 + (bg + cf − ah − de)xy + (f g − eh)y 2 .
dx + hy
We claim that φ1 6= 0. If not, the linear maps from M2K to M3K∗ corresponding
to every element of M1K are singular. It is not hard to prove that a linear system
with dimension at most 2 of singular maps from K 2 to K 2 has either a common
kernel vector or images in a common line, and to deduce from this that the partial
dual M1K ⊗ M3K → M2K∗ or M1K ⊗ M2K → M3K∗ , respectively, is not surjective, a
contradiction.
Thus M1 can be equipped with the structure of a fractional ideal of some quadratic
ring, with a ξ-action given by the matrix
"
ah + de
bc − ad
eh − f g
bg + cf
#
(6)
where we have added a scalar matrix such that the trace ah + bg + cf + de, and
indeed the entire characteristic polynomial
F (x) = x2 −(ah+bg+cf +de)x+abgh+acf h+adeh+bcf g+bdeg+cdef −adf g−bceh,
(7)
is symmetric under permuting the roles of M1 , M2 , and M3 . In other words, we have
exhibited a single ring S = R[aξ]/a2 F (ξ) of which M1 , M2 , and M3 are modules,
under the ξ-action (6) and its symmetric cousins
"
ah + cf
be − af
ch − dg
bg + de
#
"
on M2 and
ah + bg
ce − ag
#
bh − df
on M3 .
cf + de
The next step is is the construction of the elements τijk that will serve as the products η1i η2j η3k of the ideal generators. Logically, it begins with a “voilà” (compare
[2], p. 235):
τijk =
−a
¯
− a2ijk aīj̄ k̄ − aijk ξ,
i + j + k odd,
a
a2ijk aīj̄ k̄
i + j + k even.
ījk aij̄k aij k̄
ījk aij̄k aij k̄
+
+ aijk ξ,
O’Dorney
Page 17 of 39
Here ī, j̄, k̄ are shorthand for 3 − i, etc., while ξ¯ denotes the Galois conjugate
Tr(ξ) − ξ. Bhargava apparently derived this formula (in the case R = Z) by solving
the natural system of quadratic equations that the τ ’s must satisfy (τa τd = τb τc and
so on). For our purposes it suffices to note that this formula is well-defined over any
Dedekind domain (in contrast to [2] where there is a denominator of 2) and yields
a trilinear map β̃ : M1 ⊗ M2 ⊗ M3 → S, defined by
β̃
X
xijk η1i η2j η3k =
i,j,k
X
τijk xijk ,
i,j,k
with the property that following with the projection α(1 ∧ •) : S → a gives back β.
We claim that β̃, in addition to being R-trilinear, is S-trilinear under the newfound
S-actions on the Mi . This is a collection of calculations involving the action of ξ on
each factor, for instance
(ah + de)τa + (bc − ad)τe = ξτa
(where we have taken the liberty of labeling the τijk as τa , . . . , τh in the same manner
as the aijk ). This is routine, and all the other edges of the box can be dealt with
symmetrically. So, extending scalars to K, we get a map
β̃ : M1K ⊗SK M2K ⊗SK M3K → SK .
Since each Mi is isomorphic to a fractional ideal, each MiK is isomorphic to SK and
thus so is the left side. Also, it is easy to see that β̃ is surjective or else β would be
degenerate. So once two identifications ι1 : M1 → I1 , ι2 : M2 → I2 are chosen, the
third ι3 : M3 → I3 can be scaled such that β̃(x ⊗ y ⊗ z) = ι1 (x)ι2 (y)ι3 (z) and hence
β(x ⊗ y ⊗ z) = α(1 ∧ ι1 (x)ι2 (y)ι3 (z)) is as desired.
We have now constructed a triple (I1 , I2 , I3 ) of fractional ideals such that the map
α(1 ∧ • • •) : I1 ⊗ I2 ⊗ I3 → K coincides with β. Two verifications remain:
• That I1 I2 I3 ⊆ S. Since I1 I2 I3 is the R-span of the eight τijk , this is evident
from the construction of the τijk .
Q
Q
• That i Λ2 (Ii ) = i Λ2 (S), and more strongly that the diagram
Q
2
i Λ (Mi )
N
i ιi
θ
/
N
i
Λ2 (Ii )
α⊗3
'
K
commutes. This is a verification similar to that which showed the correspondence of the forms φi . Indeed, if we had recovered a triple of ideals that
produced the correct β but the wrong θ, then the φ’s as computed from β and
the two θ’s would have to mismatch.
This concludes the proof that each Bhargava box corresponds to at least one balanced triple. We must also prove that two balanced triples (I1 , I2 , I3 ) and (I10 , I20 , I30 )
yielding the same Bhargava box must be equivalent; but here we are helped greatly
O’Dorney
Page 18 of 39
by the results that we have already proved. Namely, since the forms φi associated
to the ideals match, these ideals must lie in the same oriented quadratic ring S and
×
there must be scalars γi ∈ SK
such that Ii0 = γi Ii . We may normalize such that
γ2 = γ3 = 1. Then, for all x ∈ I1 , y ∈ I2 , z ∈ I3 ,
0 = β(xyz) − β(xyz) = α(1 ∧ xyz) − α(1 ∧ γ1 xyz) = α(1 ∧ (1 − γ1 )xyz).
In other words, we have (1 − γ1 )x ∈ K for every x ∈ I1 I2 I3 . Extending scalars, we
get the same for all x ∈ KI1 I2 I3 = SK which implies 1 − γ = 0.
5.1 Relation with the class group
Just as in the case R = Z, we can restrict to invertible ideals and get a new
description of the class group.
Theorem 5.4 (cf. [2], Theorem 1) Let a be an ideal of R, and let G be the set of
rank-2 lattices M equipped with a primitive quadratic form φ : M → a−1 · Λ2 M , up
to isomorphism. Then the relations
• φ1 ∗ φ2 ∗ φ3 = 1 for all (φ1 , φ2 , φ3 ) arising from a Bhargava box;
• φ = 1 if a−1 · Λ2 M is principal and φ attains a generator of it
give G the structure of a disjoint union of abelian groups. These are isomorphic to
the class groups of all quadratic extensions of R of type a under the bijection of
Theorem 4.2.
Proof It is easy to see that a triple (I1 , I2 , I3 ) of invertible ideals is balanced if and
only if I1 I2 I3 = S. The condition that φ attains a generator of a−1 · Λ2 M simply
says that φ matches the form corresponding to the entire ring S itself in Theorem
3.2. Now the theorem is reduced to the elementary fact that the structure of an
abelian group is determined by the triples of elements that sum to 0, together with
the identification of that 0-element (without which any 3-torsion element could take
its place).
After establishing the corresponding theorem in [2] establishing a group law on
quadratic forms, Bhargava proceeds to Theorem 2, which establishes a group law on
the 2×2×2 cubes themselves, or rather on the subset of those that are “projective,”
i.e. correspond to triples of invertible ideals. This structure is easily replicated in
our situation: it is only necessary to note that the product of two balanced triples
of invertible ideals is balanced. In fact, a stronger result holds.
Lemma 5.5 Let (I1 , I2 , I3 ) and (J1 , J2 , J3 ) be balanced triples of ideals of a
quadratic ring S, with each Ii invertible. Then the ideal triple (I1 J1 , I2 J2 , I3 J3 )
is also balanced.
Proof We clearly have
I1 J1 · I2 J2 · I3 J3 = (I1 I2 I3 )(J1 J2 J3 ) ⊆ S,
O’Dorney
Page 19 of 39
establishing (a) of Definition 5.1. For (b), the key is to use Corollary 4.3 to get a
handle on the exterior squares of the Ii Ji . We have End Ii = S; each Si = End Ji
is a quadratic ring with S ⊆ Si ⊆ SK . Then since
End Ji ⊆ End Ii Ji ⊆ End Ii−1 Ii Ji = End Ji ,
we see that End Ii Ji = Si as well. Then
α(Λ2 Ji )
α(Λ2 (Ii Ji ))
α(Λ2 Ii )α(Λ2 Ji )
Si = Ii Ji ·Ii Ji = Ii Ii ·Ji Ji = α(Λ2 Ii )S·
Si =
Si .
α(Si )
α(Si )
α(Si )
Intersecting with K, we get
α(Λ2 (Ii Ji )) = α(Λ2 Ii )α(Λ2 Ji ).
We can now multiply and get
Y
i
α(Λ2 (Ii Ji )) =
Y
α(Λ2 Ii ) ·
i
so (I1 J1 , I2 J2 , I3 J3 ) is balanced.
Y
α(Λ2 Ji ) = R,
i
Corollary 5.6 (cf. [2], Theorems 2 and 12) The Bhargava boxes which belong to
a fixed ring S (determined by the quadratic form (7)) and which are primitive (in
the sense of having all three associated quadratic forms primitive) naturally form a
group isomorphic to (Pic S)2 .
Corollary 5.7 The Bhargava boxes which belong to a fixed ring S naturally have
an action by (Pic S)2 .
It is natural to think about what happens when the datum θ is removed from the
Bhargava box. As one easily verifies, multiplying θ by a unit u ∈ R× is equivalent
to multiplying the orientation α of S by u−1 while keeping the same ideals Ii .
Accordingly, we have the following corollary, which we have chosen to state with a
representation-theoretic flavor:
Corollary 5.8 Balanced triples of ideals (I1 , I2 , I3 ) of types a1 , a2 , a3 in an (unoriented) quadratic extension S of type a, up to equivalence, are parametrized by
GL(M1 ) × GL(M2 ) × GL(M3 )-orbits of trilinear maps
β : M1 ⊗ M2 ⊗ M3 → a,
where Mi is the module R ⊕ ai , satisfying the nondegeneracy condition of Definition
5.2.
These orbits do not have a group structure. Indeed, the identifications cause a
box and its inverse, under the group law of Corollary 5.6, to become identified.
O’Dorney
Page 20 of 39
Example 5.9 When R = Z (or more generally any PID), we can simplify the
notation of a Bhargava box by taking each Mi = Z2 , so that θ is without loss of
∼
generality the standard orientation Λ2 (Z2 )⊗3 → Z, and β is expressible as a box
e
a
f
b
g
c
h
d
of integers. The three forms φi are then obtained by slicing β into two 2×2 matrices
and taking the determinant of a general linear combination as described in [2],
Section 2.1:
"
#
"
#!
a b
e f
φ1 (x, y) = − det x
+y
.
c d
g h
We can now derive a balanced triple of ideals from any box of eight integers
a, b, . . . , h, subject only to the very mild condition that no two opposite faces should
be linearly dependent. We recapitulate the boxes having the greatest significance
in [2] and in the theory of quadratic forms generally:
(a) The boxes
0
1
0
1
1
1
0
1
and
0
1
D/4
1
1
0
(D + 3)/4
1
(for D even and odd respectively), have as all three of their associated
quadratic forms x2 − (D/4)y 2 and x2 + xy − (D − 1)/4 · y 2 respectively,
the defining form of the ring S of discriminant D. They correspond to the
balanced triple (S, S, S). These are the “identity cubes” of [2], equation (3).
(b) The boxes
−b/2
a
0
a
1
0
(−b + 1)/2
1
and
−c
b/2
1
0
−c
(b + 1)/2
1
0
(for b even and odd respectively), have as two of their associated quadratic
forms the conjugates
ax2 + bxy + cy 2
and ax2 − bxy + cy 2
O’Dorney
Page 21 of 39
and as the third associated form the form x2 −(D/4)y 2 or x2 +xy−(D−3)/4·y 2
defining the ring S of discriminant D = b2 − 4ac. These boxes express the fact
that the triple
¯
(S, I, α(Λ2 I)−1 I)
is always balanced (compare Corollary 4.3). If gcd(a, b, c) = 1, we also get that
I and I¯ represent inverse classes in the class group and that, correspondingly,
ax2 + bxy + cy 2 and ax2 − bxy + cy 2 are inverse under Gauss’s composition
law on binary quadratic forms.
(c) The box
f
0
1
0
−h
g
0
d
has as associated quadratic forms
φ1 (x, y) = −dx2 + hxy + f gy 2
φ2 (x, y) = −gx2 + hxy + df y 2
φ3 (x, y) = −f x2 + hxy + dgy 2 .
As Bhargava notes ([2], p. 249), Dirichlet’s simplification of Gauss’s composition law was essentially to prove that any pair of primitive binary quadratic
forms of the same discriminant can be put in the form (φ1 , φ2 ), so that the
multiplication relation that we derive from this box,
φ1 ∗ φ2 = −f x2 − hxy + dgy 2 (or, equivalently, dgx2 + hxy − f y 2 ),
encapsulates the entire multiplication table for the class group.
(d) For some examples not found in the classical theory of primitive forms, we
consider the non-Dedekind domain S = Z[5i], whose ideal class semigroup
was computed above (Example 4.4). Let us find all balanced triples that may
be formed from the ideals
S = Z[5i],
A = Zh5, 1 + ii,
B = Z[i]
of S. We compute
α(Λ2 S) = Z,
α(Λ2 A) = Z,
α(Λ2 B) =
1
Z.
5
For each triple (I1 , I2 , I3 ) of ideal class representatives, finding all balanced
×
triples of ideals in these classes is equivalent to searching for all γ ∈ SK
O’Dorney
Page 22 of 39
satisfying γ · I1 I2 I3 ⊆ S which have the correct norm
hN (γ)i =
α(Λ2 I
1
2
2
1 ) · α(Λ I2 ) · α(Λ I3 )
(the right side is an ideal of Z, so N (γ) is hereby determined up to sign, and
as we are in a purely imaginary field, N (γ) > 0).
Using the class B zero or two times, we get four balanced triples
(S, S, S),
(S, A, iA),
(S, B, 5B),
and
(A, B, 5B),
each of which yields one Bhargava box. We get no balanced triples involving
the ideal class B just once; indeed, it is not hard to show in general that if
two ideals of a balanced triple are invertible, so is the third.
The most striking case is I1 = I2 = I3 = B, for here there are two multipliers
γ of norm 125 that send B 3 = Z[i] into Z[5i], namely 10 + 5i and 10 − 5i (we
could also multiply these by powers of i, but this does not change the ideal B).
The balanced triples (B, B, (10 + 5i)B) and (B, B, (10 − 5i)B are inequivalent
under scaling, although corresponding ideals belong to the same classes. Thus
we get two inequivalent Bhargava boxes with the same three associated forms,
namely
−1
2
1
2
−1
2
1
2
and
−1
−1
2
.
−2
−2
1
2
1
(e) The triply symmetric boxes
c
b
a
b
c
b
d
c
correspond to balanced triples of ideals that all lie in the same class; those that
are projective—that is, whose associated forms are primitive—correspond to
invertible ideal classes whose third power is the trivial class. This correspondence was used to prove estimates for the average size of the 3-torsion of class
groups in [3]. Our work suggests that similar methods may work for quadratic
extensions of rings besides Z.
6 Another example: p-adic rings
Example 6.1 It is instructive to look at the local rings R = Zp , where for simplicity we assume p ≥ 3. Thanks to the large supply of squares, the corresponding
O’Dorney
Page 23 of 39
field K = Qp has but five (unoriented) quadratic extensions, namely those obtained
by adjoining a square root of 0, 1, p, u, and pu where u is an arbitrary non-square
modulo p. The quadratic ring extensions S of R then break up into five classes
according to the corresponding extension SK of K. We will work out one represen√
tative case, namely the oriented ring extensions Sn = Zp [pn u] corresponding to
√
the unique unramified extension L = K[ u] of degree 2.
For any fractional ideal I of Sn , we can pick an element of I of minimal valuation
(recalling that L possesses a unique extension of the valuation on K) and scale it
√
to be 1. Then Sn ⊆ I ⊆ S0 , since S0 = Zp [ u] is the valuation ring, and it is easy
to see that the only possible ideals are the subrings S0 , S1 , . . . , Sn . In particular Sn
is the only invertible ideal class, and the class group Pic S is trivial.
We now enumerate the balanced triples that can be built out of these ideals. A
balanced triple is formed from two sorts of data: three ideal classes Si , Sj , Sk ; and
a scale factor γ such that γSi Sj Sk ⊆ S and
hN (γ)i =
1
α(Λ2 Si )α(Λ2 Sj )α(Λ2 Sk )
.
Computing
√
α(Λ2 Si ) = α(1 ∧ pi u) = pi−n ,
we get that N (γ) has valuation p3n−i−j−k and in particular (since L is unramified)
i + j + k ≡ n mod 2.
(8)
Write 3n−i−j −k = 2s. Then γ = ps γ 0 where γ 0 ∈ S0× . To avoid needless repetition
√
of arguments, we assume i ≤ j ≤ k, and then γSi Sj Sk = ps γ 0 Si . Let γ 0 = a + b u
where a, b ∈ Zp . Since ps γ 0 Si is clearly contained in S0 , the condition for it to lie in
Sn is that the irrational parts of its generators
√
√
ps γ 0 · 1 = ps a + ps b u and ps γ 0 · pi = pi+s bu + pi+s a u
are divisible by pn , that is,
vp (a) ≥ n − s − i
and vp (b) ≥ n − s.
Since a and b cannot both be divisible by p, we must have n − s − i ≤ 0, which can
also be written as a sort of triangle inequality:
(n − j) + (n − k) ≥ n − i.
(9)
If this holds, then the restrictions on γ 0 are now merely that pn−s |b, that is, γ 0 ∈ St×
where t = max{n − s, 0}. But if γ 0 is multiplied by a unit in Si× , then the corresponding balanced triple is merely changed to an equivalent one. So the balanced
O’Dorney
Page 24 of 39
triples are in bijection with the quotient St× /Si× . Since the index of Si× in S0× is
pi−1 (p + 1) (i ≥ 1), we have that there are precisely
Bijk =
i−t
p
i−1
p
i≥t>0
(p + 1)
1
i>t=0
i=t=0
classes of Bhargava boxes whose associated ideals are of the classes Si , Sj , Sk , or
equivalently, whose associated quadratic forms are
pn−i x2 − upn+i y 2 ,
pn−j x2 − upn+j y 2 ,
pn−k x2 − upn+k y 2 .
For beauty’s sake let us examine one other angle of looking at the balanced triples.
If we extend the notation Si (i ∈ Z) to denote the Zp -module generated by 1 and
√
pi u for every i ∈ Z, then Si is an ideal of the ring Sn exactly when −n ≤ i ≤ n. Of
√
course S−i = p−i u · Si so we get no further ideal classes. But the admissible values
of i, j, and k now range in the stella octangula (Figure 1) formed by reflecting the
graph of (9) over the three coordinate planes, as well as the diagonal planes i = j,
i = k, j = k. Indeed, the triples (i, j, k) such that some scaling of (Si , Sj , Sk ) is
balanced are exactly the points of the lattice defined by (8) lying within the stella
octangula. In such a case, one such balanced triple can be given by
(Si , Sj , ps Sk )
or
√
(Si , Sj , ps uSk )
according as (i, j, k) belongs to one or the other of the two tetrahedra making up
the stella octangula.
7 Cubic algebras
The second main division of our paper has as its goal the parametrization of quartic
algebras. We begin with cubic algebras, for there the parametrization is relatively
simple and will also furnish the desired ring structure on the cubic resolvents of
our quartic rings. The parametrization was done by Delone and Faddeev for cubic
domains over Z, by Gan, Gross, and Savin for cubic rings over Z, and by Deligne over
an arbitrary scheme ([9], p. 1074 and the references therein). Here we simply state
and prove the result over a Dedekind domain, taking advantage of the construction
in [5], section 3.9.
Theorem 7.1 (cf. [13], Theorem 1; [9], Theorem 2.1; [14], Proposition 5.1 and the
references therein) Let R be a Dedekind domain. There is a canonical bijection
between cubic algebras over R and pairs consisting of a rank-2 R-lattice M and a
cubic map φ : M → Λ2 M .
Proof Given the cubic ring C, we let M = C/R so a = Λ2 M ∼
= Λ3 C is an ideal
2
class. Consider the map φ̃ : C → a given by x 7→ 1 ∧ x ∧ x . This is a cubic map, and
if x is translated by an element a ∈ R, the map does not change. Hence it descends
O’Dorney
Page 25 of 39
to a cubic map φ : M → a. We will show that each possible φ corresponds to exactly
one ring C.
Fix a decomposition M = a1 ξ˜1 ⊕ a2 ξ˜2 of M into ideals. Any C can be written as
R ⊕ M = R · 1 ⊕ a1 ξ1 ⊕ a2 ξ2 as an R-module, where the lifts ξ1 and ξ2 are unique
up to adding elements of a−1
and a−1
respectively. Then the remaining structure
1
2
of C can be described by a multiplication table
ξ12 = ` + aξ1 + bξ2
ξ1 ξ2 = m + cξ1 + dξ2
ξ22 = n + eξ1 + f ξ2 .
It should be remarked that this is not literally a multiplication table for C, but
rather for the corresponding K-algebra CK = C ⊗R K, which does literally have
{1, ξ1 , ξ2 } as a K-basis. For C to be closed under this multiplication, the coefficients
−1
must belong to appropriate ideals (` ∈ a−2
1 , a ∈ a1 , etc.).
Note that the basis change ξ1 7→ ξ1 + t1 , ξ2 7→ ξ2 + t2 (ti ∈ a−1
i ) diminishes c and
d by t2 and t1 , respectively (as well as wreaking greater changes on the rest of the
multiplication table). Hence there is a unique choice of the lifts ξ1 and ξ2 such that
c = d = 0.
We now examine the other piece of data that we are given, the cubic map φ
describable in these coordinates as
φ(xξ˜1 + y ξ˜2 )
= 1 ∧ (xξ1 + yξ2 ) ∧ (xξ1 + yξ2 )2
= 1 ∧ (xξ1 + yξ2 ) ∧ ((` + aξ1 + bξ2 )x2 + mxy + (n + eξ1 + f ξ2 )y 2 ))
= (bx3 − ax2 y + f xy 2 − ey 3 )(1 ∧ ξ1 ∧ ξ2 ).
Thus, in our situation, specifying φ is equivalent to specifying the four coefficients a,
b, e, and f . It therefore suffices to prove that, for each quadruple of values a ∈ a−1
1 ,
−2
−1
b ∈ a−2
a
,
e
∈
a
a
,
f
∈
a
,
there
is
a
unique
choice
of
values
`,
m,
n,
completing
2
1 2
1
2
the multiplication table. The only conditions on the multiplication table that we
have not used are the associative laws (ξ12 )ξ2 = ξ1 (ξ1 ξ2 ) and ξ1 (ξ22 ) = (ξ1 ξ2 )ξ2 .
Expanding out these equations reveals the unique solution ` = −ae, m = −be,
n = −bf , which indeed belong to the correct ideals. So from the map φ we have
constructed a unique cubic ring C.
Example 7.2 Here we briefly summarize the most important examples over R =
Z, where the cubic map φ : M → Λ2 M reduces to a binary cubic form φ : Z2 → Z,
up to the twisted action of the group GL2 Z by
"
a
c
# !
1
b
· φ(ax + cy, bx + dy).
.φ (x, y) =
ad
−
bc
d
• The trivial ring Z[1 , 2 ]/(21 , 1 2 , 22 ) corresponds to the zero form 0.
O’Dorney
Page 26 of 39
• Rings which are not domains correspond to reducible forms (e.g. Z⊕Z⊕Z corresponds to xy(x + y)), and rings which have nontrivial nilpotents correspond
to forms with repeated roots.
• A monogenic ring Z[ξ]/(ξ 3 + aξ 2 + bξ + c) corresponds to a form x3 + ax2 y +
bxy 2 + cy 3 with leading coefficient 1. Accordingly a form which does not
represent the value 1 corresponds to a ring that is not monogenic; for instance,
the form 5x3 + 7y 3 (which attains only values ≡ 0, ±2 mod 7) corresponds
√
√
√
√
3
3
3
3
to the subring Z[ 52 · 7, 5 · 72 ] of the field Q[ 52 · 7] = Q[ 5 · 72 ], proving
that this ring (which is easily checked to be the full ring of integers in this
field) is not monogenic.
• If a form φ corresponds to a ring C, then the form n · φ corresponds to
the ring Z + nC whose generators are n times as large. Hence the content
ct(φ) = gcd(a, b, c, d) of a form φ(x, y) = ax3 + bx2 y + cxy 2 + dy 3 equals the
content of the corresponding ring C, which is defined as the largest integer n
such that C ∼
= Z + nC 0 for some cubic ring C 0 . The notion of content (which
is also not hard to define for cubic algebras over general Dedekind domains)
will reappear prominently in our discussion of quartic algebras (see section
8.2).
8 Quartic algebras
Our next task is to generalize Bhargava’s parametrization of quartic rings with a
cubic resolvent in [5], and in particular to formalize the notion of a cubic resolvent.
The concept was first developed as part of the theory of solving equations by radicals, in which it was noted that if a, b, c, and d are the unknown roots of a quartic,
then
ab + cd,
ac + bd,
and ad + bc
satisfy a cubic whose coefficients are explicit polynomials in those of the original
quartic. Likewise, if Q ⊇ Z is a quartic ring embeddable in a number field, the
useful resolvent map
x 7→ (σ1 (x)σ2 (x)+σ3 (x)σ4 (x), σ1 (x)σ3 (x)+σ2 (x)σ4 (x), σ1 (x)σ4 (x)+σ2 (x)σ3 (x))
lands in a cubic subring of C ⊕ C ⊕ C, where σ1 , . . . , σ4 are the four embeddings
Q ,→ C. The question then arises of what the proper notion of a resolvent map is
in case Q is not a domain. In section 2.1 of [5], Bhargava defines from scratch a
workable notion of Galois closure of a ring, providing a rank-24 algebra in which
the resolvent can be defined. Alternatively (section 3.9), Bhargava sketches a way
of axiomatizing the salient properties of a resolvent map. It is the second method
that we develop here.
Definition 8.1 (cf. [9], p. 1069) Let R be a Dedekind domain, and let Q be a
quartic algebra over R. A resolvent for Q consists of a rank-2 R-lattice M , an Rmodule morphism θ : Λ2 M → Λ3 (Q/R), and a quadratic map φ : Q/R → M such
that there is an identity of biquadratic maps
x ∧ y ∧ xy = θ(φ(x) ∧ φ(y))
(10)
O’Dorney
Page 27 of 39
from Q × Q to Λ3 (Q/R).
The resolvent (M, θ, φ) is called minimal if φ has full image, that is, φ(Q/R) = M .
It is called numerical if θ is an isomorphism.
Our minimal resolvent corresponds to the ring Rinv in Bhargava’s treatment ([5],
p. 1337), while our numerical resolvent corresponds to Bhargava’s resolvent. The
numerical resolvent is more suited to analytic applications, while the minimal resolvent has the advantage of being canonical (for nontrivial Q), as we prove below.
Example 8.2 For the prototypical example of a resolvent, take Q = R⊕4 and
C = R⊕3 , and let M = C/R. Let θ identify the standard orientations on these
lattices, and let φ be given by the roots
φ(a, b, c, d) = (ab + cd, ac + bd, ad + bc)
of the classical resolvent of the quartic (x − a)(x − b)(x − c)(x − d). It is easy to
check that this is a resolvent, which is both minimal and numerical. Many more
examples can be derived from this (see Example 8.10).
8.1 Resolvent to ring
Our first result is that the resolvent encapsulates the data of the ring:
Theorem 8.3 (cf. [5], Theorem 1 and Proposition 10; [9], Corollary 1.2) Let Q̃ and
M be R-lattices of ranks 3 and 2 respectively. Let θ : Λ2 M → Λ3 Q̃ be a morphism,
and let φ : Q̃ → M be a quadratic map. Then there is a unique quartic ring Q with
an isomorphism Q/R ∼
= Q̃ such that (M, θ, φ) is a resolvent for Q.
Proof Write Q̃ = a1 ξ˜1 ⊕ a2 ξ˜2 ⊕ a3 ξ˜3 as usual. The ring Q will of course be R ⊕
a1 ξ1 ⊕ a2 ξ2 ⊕ a3 ξ3 as an R-module, with a multiplication table
ξi ξj = c0ij +
3
X
ckij ξk
k=1
−1
k
where c0ij ∈ a−1
and ckij ∈ ai−1 a−1
i aj
j ak . The 18 coefficients cij are subject to the
expansion of the relation (10):
X
i
!
!
X
X
X
X
xi ξ˜i ∧
yj ξ˜j ∧
xi yj ckij ξ˜k = θ φ
xi ξ˜i ∧ φ
yj ξ˜j .
j
i
i,j,k
j
(11)
Write
φ(x1 ξ1 + x2 ξ2 + x3 ξ3 ) =
X
1≤i≤j≤3
µij xi xj
O’Dorney
Page 28 of 39
−1
where µij ∈ a−1
i aj M . Then define
−1 −1 −1 −1
λij
k` = θ(µij ∧ µk` ) ∈ a1 a2 a3 ai aj ak a` .
We can now expand both sides of (11) as polynomials in the x’s and y’s times
ξ˜1 ∧ ξ˜2 ∧ ξ˜3 , getting
x1
x2
x3
y1
y2
y3
P
c1 x i y j
X X ij
Pi,j 2ij
λk` xi xj yk y` ,
i,j cij xi yj =
P
3
i≤j
k≤`
i,j cij xi yj
and equate coefficients of each biquadratic monomial xi xj yk y` . Due to the skewsymmetry of each side, all terms involving x2i yi2 or xi xj yi yj cancel, and the remaining 30 equations group into 15 matched pairs. They are summarized as follows,
where (i, j, k) denotes any permutation of (1, 2, 3) and = ±1 its sign:
cjii = λii
ik
ckij = λjj
ii
(12)
cjij − ckik = λjk
ii
ciii − cjij − ckik = λij
ik .
At first glance it may seem that one can add a constant a to cjij and ckij , while
adding 2a to ciii , to derive a three-parameter family of solutions from a single one;
but this is merely the transformation induced by the change of lift ξi 7→ ξi + a for
ξ˜i . So there is essentially only one solution. (It could be normalized by taking e.g.
c112 = c223 = c331 = 0, although we do not use this normalization here, preferring to
save time later by keeping the indices 1, 2, and 3 in complete symmetry.)
The constant terms c0ij of the multiplication table are as yet undetermined. They
must be deduced from the associative law. There are several ways to compute each
c0ij , and to prove that they agree, along with all the other relations implied by
the associative law, is the final step in the construction of the quartic ring Q. Our
key tool is the Plücker relation relating the wedge products of four vectors in a
2-dimensional space:
(a ∧ b)(c ∧ d) + (a ∧ c)(d ∧ b) + (a ∧ d)(b ∧ c) = 0,
or, as we will use it,
0
0
0
0
0
0
cc
aa dd
aa bb
λaa
bb0 λdd0 + λcc0 λbb0 + λdd0 λcc0 = 0.
To give succinct names to these relations among the λ’s, note that aa0 , . . . , dd0 are
four of the six unordered pairs that can be formed from the symbols 1, 2, and 3,
and the relation is nontrivial only when these four pairs are distinct. Consequently
we denote it by Pl(ee0 , ff 0 ), where ee0 and ff 0 are the two pairs that do not appear
in it. Then Pl(ee0 , ff 0 ) as a polynomial in the λ’s is unique up to sign, and we will
never have occasion to fix a sign convention.
O’Dorney
Page 29 of 39
We are now ready to derive the associative law from the Plücker relations. Of
course this is a task that could be left to a computer, but since we will soon be
deriving the Plücker relations from the associative law, we find it advisable to
present the process at least in summary form. Here it is:
[(ξi ξi )ξj − (ξi ξj )ξi ]k = Pl(jk, kk)
[(ξi ξj )ξk − (ξi ξk )ξj ]i = Pl(ij, ik)
[(ξi ξj )ξi − (ξi ξi )ξj ]i
[(ξi ξi )ξj − (ξi ξj )ξi ]j
Pl(ij,kk)
[(ξi ξj )ξk − (ξj ξk )ξi ]k (13)
Pl(ik,jk)
Pl(jj,kk)
[(ξi ξi )ξk − (ξi ξk )ξi ]j
[(ξi ξj )ξj − (ξj ξj )ξi ]j
Pl(ik,jk)
Pl(ij,kk)
[(ξi ξj )ξk − (ξj ξk )ξi ]k
And here is the explanation:
• The notation [ω]i denotes the coefficient of ξi when ω is expressed in terms of
the basis {1, ξ1 , ξ2 , ξ3 }.
• Each of the first two equations is a direct calculation. For instance:
[(ξi ξi )ξj − (ξi ξj )ξi ]k
= [(c0ii + ciii ξi + cjii ξj + ckii ξk )ξj − (c0ij + ciij ξi + cjij ξj + ckij ξk )ξi ]k
= ciii ckij + cjii ckjj + ckii ckjk − ciij ckii − cjij ckij − ckij ckik
= (ciii − cjij − ckik )ckij + ckii (ckjk − ciij ) + cjii ckjj
jj
ii ik
ii jj
= (λij
ik λii − λij λjj + λik λij )
= Pl(jk, kk).
• The two lower diagrams show the instances of the associative law that produce
a summand of c0ii or c0ij , respectively. Each node in the diagrams yields a
formula for c0ii or c0ij (having no denominator, and consequently belonging
−1
to the correct ideal a−2
resp. a−1
i
i aj ); and where two nodes are joined by a
line, the difference between the two corresponding formulas is expressible as
a Plücker relation.
We have now proved all of the associative law except the constant terms; that is, we
now have that (xy)z − x(yz) ∈ R for all x, y, z ∈ Q. Attacking the constant terms
in the same manner as above leads to considerably heavier computations, which
could be performed by computer (compare [5], top of p. 1343). Alternatively, we
may use the following trick. Let i, j, k ∈ {1, 2, 3} be any indices, and let h ∈ {1, 2, 3}
be an index distinct from k. Then using the already-proved ξh -component of the
associative law,
ξi (ξj ξk ) − ξj (ξi ξk ) = [ξh (ξi (ξj ξk )) − ξh (ξj (ξi ξk ))]h
= [(ξh ξi )(ξj ξk ) − (ξh ξj )(ξi ξk )]h
= [((ξh ξi )ξj )ξk − ((ξh ξj )ξi )ξk ]h .
This last is necessarily zero, since it consists of the number (ξh ξi )ξj − (ξh ξj )ξi ∈ R
multiplied by ξk , and thus has no ξh -component.
O’Dorney
Page 30 of 39
8.2 Ring to resolvent
Conversely, we will now study all possible resolvents of a given quartic ring Q.
There is one case in which this problem takes a striking turn: the trivial ring
P
Q = R[a1 1 , a2 2 , a3 3 ]/ i,j (ai aj i j ) where all entries of the multiplication table are zero. Here φ can be an arbitrary map to a 1-dimensional sublattice of M , or
alternatively M and φ can be chosen arbitrarily while θ = 0. For all other quartic
rings, the family of resolvents is much smaller, as we will now prove.
Theorem 8.4 (cf. [5], Corollary 18)
Let Q be a nontrivial quartic R-algebra. Then
(a) Q has a unique minimal resolvent (M0 , θ0 , φ0 );
(b) we have θ0 (Λ2 M0 ) = c · Λ3 (Q/R), where c is the ideal (called the content of
Q) characterized by the following property: For each ideal a ⊆ R, there exists
a quartic R-algebra Q0 such that Q ∼
= R + aQ0 if and only if a|c;
(c) all other resolvents (M, θ, φ), up to isomorphism, are found by extending θ0
linearly to Λ2 M , where M is a lattice with [M : M0 ] | c, and taking φ = φ0 ;
(d) the numerical resolvents arise by taking [M : M0 ] = c in the preceding.
Proof Write Q = R ⊕ a1 ξ1 ⊕ a2 ξ2 ⊕ a3 ξ3 . The multiplication table can be encoded
in a family of ckij ’s, from which the fifteen values λij
k` are determined through (12).
ij
These λk` satisfy the fifteen Plücker relations by (13). It then remains to construct
−1
the target module M , the map θ, and the vectors µij ∈ a−1
i aj M such that their
pairwise exterior products µij ∧ µk` have, via θ, the specified value λij
k` .
The six µij are in complete symmetry at this point, and it will be convenient
to denote µij by µx , where x runs over {11, 12, 13, 22, 23, 33} or, if you prefer,
x
{1, 2, 3, 4, 5, 6}. Likewise we write each λij
k` as λy or simply λxy .
We first tackle the problem over K. Let V be an abstract K-vector space of dimension 2. We construct vectors µ1 , . . . , µn whose exterior products are proportional to
the λ’s as follows. Some λxy is nonzero, without loss of generality λ12 . Let (µ1 , µ2 )
be a basis of V . Then, for 3 ≤ x ≤ 6, take
µx =
−λ2x µ1 + λ1x µ2
λ12
to give the products µ1 ∧ µx and µ2 ∧ µx the desired values. The λxy with 3 ≤ x <
y ≤ 6 have not been used, but their values were forced by the Plücker relations
anyway, so we have a system of µx such that
µx ∧ µy =
λxy
· µ1 ∧ µ2 .
λ12
Moreover, these are the only µx ∈ V with this property, up to GL(V )-transformations.
Now define a quadratic map φ0 : Q/R → V by
φ0 (x1 ξ1 + x2 ξ2 + x3 ξ3 ) =
X
i≤j
µij xi xj
O’Dorney
Page 31 of 39
and a linear map θ0 : Λ2 V → Λ3 (Q/R) ⊗R K by
θ0 (µ1 ∧ µ2 ) = λ12 (ξ1 ∧ ξ2 ∧ ξ3 ).
We have that (V, θ0 , φ0 ⊗ K) is the unique resolvent of the quartic K-algebra Q/K.
Resolvents for Q are now in bijection with lattices M ⊆ V such that
and θ0 (Λ2 M ) ⊆ Λ3 (Q/R).
M ⊇ φ0 (Q/R)
(14)
There is now clearly at most one minimal resolvent, gotten by taking M to be the
image M0 = φ(Q/R). We have
θ0 (Λ2 M0 ) = θ0
X
ai aj ak a` · µij ∧ µk`
i,j,k,`
=
X
ξ1
λij
k` ai aj ak a`
∧ ξ2 ∧ ξ3 = cΛ3 (Q/R),
i,j,k,`
where
X
c=
−1 −1 −1
λij
k` ai aj ak a` a1 a2 a3 .
i,j,k,`
−1 −1 −1 −1
The ideal in which λij
k` is constrained to live is a1 a2 a3 ai aj ak a` ; so c ⊆ R and
there is a unique minimal resolvent, proving (a).
If a ⊇ c for some a ⊆ R, we can replace each of the three ai with aai without changing the validity of the λ-system. This means there is an extension ring
Q0 = R ⊕ aa1 ξ1 ⊕ aa2 ξ2 ⊕ aa3 ξ3 with the same multiplication table as Q, and
we see that Q0 = R + aQ. Conversely, given such a Q0 , we write its multiplication table with respect to the basis Q0 /R = aa1 ξ1 ⊕ aa2 ξ2 ⊕ aa3 ξ3 and get that
−1 −1 −1
−1
λij
a1 a2 a3 a−1
i aj ak a` , so c ⊆ a. This proves (b).
k` ∈ a
Finally, the relation θ0 (Λ2 M0 ) = cΛ3 (Q/R) allows us to rewrite (14) as
M ⊇ M0
and
Λ2 M ⊆ cΛ2 M0 .
Now (c) is obvious. A numerical resolvent occurs when θ0 (Λ2 M ) = Λ3 (Q/R), so
(d) is obvious as well.
Bhargava proved ([5], Corollary 4) that the number of (numerical) resolvents of a
quartic ring over Z is the sum of the divisors of its content. Likewise, we now have:
Corollary 8.5 If c 6= 0, then the numerical resolvents of Q are in noncanonical
bijection with the disjoint union
a
R⊇a⊇c
R/a.
O’Dorney
Page 32 of 39
Proof Here we simply have to count the superlattices M of index c over a fixed
lattice M0 . The classical argument over Z extends rather readily; for completeness,
we give the proof.
Note that we must have M ⊆ c−1 M0 , since M ∧ M0 ⊆ M ∧ M = c−1 Λ2 M0 . Pick
a decomposition c−1 M0 ∼
= d1 ⊕ d2 . Then consider the map π : M → d1 that is the
restriction of projection to the first factor. We have ker π = {0}×ad2 and im π = bd1
for some ideals a, b subject to the familiar behavior of top exterior powers in exact
sequences:
c−1 Λ2 M0 = Λ2 M = ad2 ∧ bd1 = abc−2 Λ2 M0 ,
that is, ab = c. Now if a and b are fixed, the lattice M is determined by a picking
a coset in d2 /ad2 to be the preimage of each point b ∈ im π; this is determined by
an R-module map
bd1 → d2 /ad2
or, since cd1 is necessarily in the kernel,
bd1 /cd1 → d2 /ad2 .
We can identify both the domain and the target of this map with R/a via the standard result that if a and b are ideals in a Dedekind domain R, then a/ab ∼
= R/b.
(Proof: Use the Chinese Remainder Theorem to find a ∈ a that has minimal
valuation with respect to each of the primes dividing b. Then a generates a/ab,
and a 7→ 1 is the desired isomorphism.) Then the desired parameter space is
HomR (R/a, R/a) ∼
= R/a. Letting a vary yields the claimed bijection.
In particular, we have the following.
Corollary 8.6 (cf. [5], Corollary 5) Every quartic algebra over a Dedekind domain
possesses at least one numerical resolvent.
8.3 The cubic ring structure of the resolvent
In contrast to the classical presentation, the resolvent maps we have constructed
take their values in modules, without any explicit connection to a cubic ring. In
fact, there is the structure of a cubic ring already latent in a resolvent:
Theorem 8.7 To any quartic ring Q and resolvent (M, θ, φ) thereof, one can
canonically associate a cubic ring C with an identification C/R ∼
= M.
Remark As stated, this theorem has no content, as one can take the trivial ring
structure on R ⊕ M . However, we will produce a ring structure generalizing the
classical notion of cubic resolvent. This C may be called a cubic resolvent of Q, the
maps θ and φ being suppressed.
O’Dorney
Page 33 of 39
Proof We use the following trick of multilinear algebra (compare [9], p. 1076). First
pick a decomposition Q/R = a1 ξ˜1 ⊕ a2 ξ˜2 ⊕ a3 ξ˜3 . Writing
φ(x1 ξ˜1 + x2 ξ˜2 + x3 ξ˜3 ) =
X
xi xj µij
−1
(µij ∈ a−1
i aj M ),
i≤j
consider the determinant
µ11
∆ = 4 det 12 µ12
1
2 µ13
1
2 µ12
µ22
1
2 µ23
1
2 µ13
1
2 µ23
µ33
= 4µ11 µ22 µ33 + µ12 µ13 µ23 − µ11 µ223 − µ22 µ213 − µ33 µ212
3
−2 −2
∈ a−2
1 a2 a3 Sym M
(the two expressions are equal except when char K = 2, in which case the first
−2 −2
2
⊗−2
becomes purely motivational). Next, θ allows us to map a−2
.
1 a2 a3 to (Λ M )
The Λ2 M -valued pairing ∧ on M gives an identification of M with Λ2 M ⊗ M ∗ , so
we can transform
(Λ2 M )⊗−2 ⊗ Sym3 M ∼
= (Λ2 M )⊗−2 ⊗ Sym3 ((Λ2 M ) ⊗ M ∗ )
∼
= (Λ2 M )⊗−2 ⊗ (Λ2 M )⊗3 ⊗ Sym3 (M ∗ )
∼
= (Λ2 M ) ⊗ Sym3 (M ∗ ).
Thus ∆ yields a cubic map δ : M → Λ2 M , which by Theorem 7.1 is equivalent
to a cubic ring C with an identification C/R ∼
= M . That δ is independent of the
˜
˜
˜
chosen basis (ξ1 , ξ2 , ξ3 ) is a polynomial identity that follows from properties of the
determinant, at least when char K 6= 2.
Two theorems concerning this cubic ring structure we will state without proof,
since they are mere polynomial identities already implied by Bhargava’s work over
Z. The first may be used as an alternative to Theorem 7.1 to determine the multiplicative structure on C; as Bhargava notes, it works in all cases over Z except
when Q has nilpotents.
Theorem 8.8 (cf. [5], equation (30)) Let Q be a quartic ring, and let C be
the cubic ring whose structure is determined by the resolvent map data θ :
Λ2 (C/R) → Λ3 (Q/R) and φ : Q/R → C/R. For any element x ∈ Q and any lift
y ∈ C of the element φ(x) ∈ C/R, we have the equality
x ∧ x2 ∧ x3 = θ(y ∧ y 2 ).
We end this section with a theorem concerning discriminants, which until now
have been conspicuously absent from our discussion, in direct contrast to Bhargava’s presentation. Recall that the discriminant of a Z-algebra Q with a Z-basis
O’Dorney
Page 34 of 39
(ξ1 , . . . , ξn ) is defined as the determinant of the matrix [Tr(ξi ξj )]i,j . In like manner,
define the discriminant of a rank-n R-algebra Q to be the map
disc(Q) : x1 ∧ · · · ∧ xn 7→ det[Tr(xi xj )]i,j .
It is quadratic and thus can be viewed as an element of (Λn Q∗ )⊗2 , a rank-1 lattice
that is not in general isomorphic to R. The discriminants of a quartic ring and its
resolvents are “equal” in precisely the way one might hope:
Theorem 8.9 (cf. [5], Proposition 13)
Let Q, C, θ be as above. The morphism
(θ∗ )⊗2 : (Λ3 (Q/R)∗ )⊗2 →(Λ2 (C/R)∗ )⊗2
carries disc Q to disc C.
Example 8.10 Once again, we recapitulate the situation over Z. Here, once bases
Q/R = Zξ1 ⊕ Zξ2 ⊕ Zξ3 and C/R = Zη1 ⊕ Zη2 have been fixed so that θ is given
simply by η1 ∧ η2 7→ ξ1 ∧ ξ2 ∧ ξ3 , the remaining datum φ of a numerical resolvent
can be written as a pair of ternary quadratic forms, or, even more pictorially, as a
pair of symmetric matrices
a11
(A, B) = 12 a12
1
2 a13
1
2 a12
a22
1
2 a23
1
b11
2 a13
1
1
,
2 a23 2 b12
1
a33
2 b13
1
2 b12
b22
1
2 b23
1
2 b13
1
.
2 b23
b33
where aij , bij ∈ Z. The associated cubic ring is found by applying Theorem 7.1 to
the form 4 det(Ax + By). Some salient examples follow:
• First note that there is a resolvent map of C-algebras from Q0 = C⊕4 to
C0 = C⊕3 given by the roots of the equation-solver’s resolvent
(x, y, z, w) 7→ (xy + zw, xz + yw, xw + yz)
or, more accurately, by its reduction modulo C
φ0 : (x, y, z, 0) 7→ (xy − yz, xz − yz, 0),
supplemented of course by the standard identification
θ0 : Λ2 (C0 /C) → Λ3 (Q0 /C).
Accordingly, if we have a quartic Z-algebra Q ⊆ Q0 and a cubic Z-algebra C ⊆
C0 on which the restrictions of φ0 and θ0 are well-defined, then it automatically
follows that C/Z is a resolvent for Q with attached cubic ring structure C.
• As an example, consider the ring
Q = Z + p(Z ⊕ Z ⊕ Z ⊕ Z) = {(a, b, c, d) ∈ Z⊕4 : a ≡ b ≡ c ≡ d mod p}
O’Dorney
Page 35 of 39
of content p, for each prime p. The minimal resolvent of Q comes out to be
φ0 (Q/Z) = C 0 /Z, where
C 0 = Z + p2 · Z⊕3 .
But C 0 is not a numerical resolvent of Q: it has index p4 in Z⊕3 , while Q has
index p3 in Z⊕4 , so the restriction of θ0 cannot possibly be an isomorphism.
We must enlarge C 0 by a factor of p. Note that any subgroup C such that
Z + p2 · Z⊕3 ⊆ C ⊆ Z + p · Z⊕3
is a ring, since the product of two elements in p · Z⊕3 lies in p2 · Z⊕3 . So any
ring of the form
C = Z + p2 · Z⊕3 + hap, bp, 0i
is a numerical resolvent of Q. Letting [a : b] run over P1 (Z/pZ) yields the p + 1
numerical resolvents predicted by Theorem 8.4.
• Note that some of these resolvents are isomorphic under the automorphism
group of Q, which is simply S4 acting by permuting the coordinates. One
verifies that S4 acts through its quotient S3 , which in turn permutes the
three distinguished points 0, 1, ∞ on P1 (Z/pZ). Accordingly, if we are using
Theorem 8.3 to count quartic rings, the ring Q will appear not p + 1 times
but dp/6e + 1 times (1 time if p = 2 or p = 3). This is no contradiction with
Theorem 8.4, which gives the number of resolvents as maps out of the given
ring Q.
9 Maximality
In order to convert his parametrization of quartic rings into one of quartic fields,
Bhargava needed a condition for a ring to be maximal, i.e. to be the full ring
of integers in a field. In like manner we discuss how to tell if a quartic ring Q
over a Dedekind domain R is maximal in its fraction ring QK using conditions on
a numerical resolvent. The first statement to make is that maximality is a local
condition, i.e. can be checked at each prime ideal.
Proposition 9.1 Let Q be a ring of finite rank n over R. Q is maximal if and
only if Qp = Q ⊗R Rp is maximal over Rp for all (nonzero) primes p ⊆ R.
Remark Here, Rp denotes the localization
Rp =
na
b
o
∈ K : a ∈ R, b ∈ R \ p
(not its completion as with the symbol Zp ).
Proof When Q is a domain, one can use the facts that Q is maximal if and only
if it is normal (integrally closed in its fraction field) and that normality is a local
property. A direct proof is also not difficult.
O’Dorney
Page 36 of 39
Suppose that Q is not maximal, so that there is a larger ring Q0 with QK =
The nonzero R-module Q0 /Q is pure torsion, so there is a prime p such that
(Q /Q)p = Q0p /Qp is nonzero, i.e. Qp embeds into the larger ring Q0p .
Suppose now that for some p, Qp embeds into a larger ring T . We construct an
extension ring Q0 of Q by the formula
Q0K .
0
Q0 = Q[p−1 ] ∩ T.
It is obvious that Q0 is a ring containing Q; it is not so obvious that it is a rank-n
ring, in other words, that it is finitely generated as an R-module. Let X be the
R-lattice generated by any K-basis x1 , . . . , xn of QK . Since Q and T are finitely
generated R- and Rp -modules respectively, we may divide the xi by sufficiently
divisible elements of R to assume Q ⊆ X and T ⊆ Rp X. Then
Q0 ⊆ X[p−1 ] ∩ Rp X.
Note that
(
)
X
Rp X =
ai xi : vp (ai ) ≥ 0
i
(
−1
X[p
)
X
]=
ai xi : vq (ai ) ≥ 0
∀q 6= p
i
(
X[p−1 ] ∩ Rp X =
)
X
ai xi : vq (ai ) ≥ 0
∀q
= X,
i
whence Q0 ⊆ X is finitely generated.
Finally we must show that Q0 6= Q. This is obvious by localization:
Q0p = (Q[p−1 ])p ∩ Tp = QK ∩ T = T 6= Qp .
The local rings Rp are DVR’s, and in particular are PID’s, so we can visualize a
localized numerical resolvent (Qp , Mp , θ, φ) in a simple way: Pick bases Qp /Rp =
Rp hξ˜1 , ξ˜2 , ξ˜3 i and Mp = Rp hη1 , η2 i such that θ(η1 ∧ η2 ) = ξ˜1 ∧ ξ˜2 ∧ ξ˜3 , and write φ
as a pair of matrices
a11
1
(A, B) = 2 a12
1
2 a13
1
2 a12
a22
1
2 a23
1
b11
2 a13
1
1
,
2 a23 2 b12
1
a33
2 b13
1
2 b12
b22
1
2 b23
1
2 b13
1
2 b23
b33
where 1/2 is a purely formal symbol and aij , bij ∈ R are the coefficients of the
resolvent map
φ(x1 ξ˜1 + x2 ξ˜2 + x3 ξ˜3 ) =
X
(aij η1 + bij η2 )xi xj .
1≤i<j≤3
O’Dorney
Page 37 of 39
We will characterize maximality of Qp in terms of the aij and bij . The first simplification is applicable to rings of any rank.
Lemma 9.2 Let R be a DVR with maximal ideal p, and let Q be an R-algebra of
rank n. If Q is not maximal, then there exists k ≥ 1 and a basis x1 , x2 , . . . , xn = 1
of Q such that
Q0 = R p−1 x1 , . . . , p−1 xk , xk+1 , . . . , xn
is a ring.
Proof Let Q1 ) Q be a larger algebra. Since Q1 is a finitely generated submodule
S
of QK = i≥0 p−i Q, it is contained in some p−r Q. Pick r such that
Q1 ⊆ p−r Q but Q1 * p−r+1 Q.
Then Q0 = Q + pr−1 Q1 is a rank-n algebra such that
Q ( Q0 ⊆ p−1 Q.
Choose a basis x̃1 , . . . , x̃k for the R/pR-vector space pQ0 /pQ, and complete it to
a basis x̃1 , . . . , x̃n for Q/pQ. Since 1 ∈
/ pQ0 , we can arrange for x̃n = 1. Then by
Nakayama’s lemma, any lifts x1 , . . . , xn generate Q, and
p−1 x1 , . . . , p−1 xk , xk+1 , . . . , xn
generate Q0 .
Theorem 9.3 Let Q be a quartic algebra over a DVR R with maximal ideal p,
and let φ : Q/R → M , θ : Λ3 (Q/R) → Λ2 M be a resolvent. Then Q is non-maximal
if and only if, under some choice of bases, the matrices (A, B) representing φ satisfy
one of the following conditions:
(a)
(b)
(c)
(d)
p2 divides a11 , and p divides a12 , a13 , and b11 .
p divides a11 , a12 , a22 , b11 , b12 , and b22 .
p2 divides a11 , a12 , and a22 , and p divides a13 and a23 .
p divides all aij .
Proof The basic strategy is to find a suitable extension of the resolvent map to the
ring Q0 in Lemma 9.2, examining the cases where k is 1, 2, and 3.
First assume that Q has content 1 (by which we mean that the content ideal ct(Q)
is the whole of R). Then k is 1 or 2 and Q0 also has content 1. Both Q and Q0 have
unique (minimal and numerical) resolvents (M, θ, φ) and (M 0 , θ0 , φ0 ), where (since
QK = Q0K ) we have M ⊆ M 0 ⊆ MK , and θ and φ are the restrictions of θ0 and φ0 .
Also, since Q has index pk in Q0 , M has index pk in M 0 .
O’Dorney
Page 38 of 39
If k = 1, then we can arrange our coordinates such that
Q/R = hξ˜1 , ξ˜2 , ξ˜3 i, Q0 /R = hπ −1 ξ˜1 , ξ˜2 , ξ˜3 i
M = hη1 , η2 i, M 0 = hη1 , πη2 i.
Now since φ0 : Q0 /R → M 0 is the extension of φ, its corresponding matrix (A0 , B 0 )
is given by a straightforward change of basis:
π −2 a11
1 −1
0
0
(A , B ) = 2 π a12
1 −1
a13
2π
1 −1
a12
2π
a22
1
2 a23
1 −1
π −1 b11
a13
2π
1
1
,
2 a23 2 b12
1
a33
2 b13
1
2 b12
πb22
1
2 πb23
1
2 b13
1
2 πb23
πb33
The entries of this matrix (sans 1/2’s) must belong to R, giving the divisibilities
listed in case (a) above.
If k = 2, then the proof works similarly, except that M 0 takes one of the two forms
hπη1 , πη2 i and η1 , π 2 η2 . We leave it to the reader to write out the corresponding
matrices (A0 , B 0 ) and conclude cases (b) and (c) above, respectively.
We are left with the case that ct(Q) 6= 1, that is, there is a quartic ring Q0 with
Q = R + πQ0 . (A priori we might only have a ring Q00 with Q = R + π k Q00 , k ≥ 1;
but then Q0 = R + π k−1 has the aforementioned property.) Then we may select
bases for Q and Q0 in the form of Lemma 9.2, with k = 3. Since the resolvent is
no longer unique, we must take care in choosing the new target module M 0 of the
resolvent φ0 . Since φ is quadratic and Q0 /R = π −1 (Q/R), a natural candidate is
M 0 = π −2 M , but unfortunately this is too large: we have [M 0 : M ] = p4 but [Q0 /R :
Q/R] = p3 , so θ0 cannot possibly be an isomorphism. However, since ct(Q) 6= 1, we
have φ(Q/R) ( M , so picking a sublattice L ( M of index p containing φ(Q/R), we
get that M 0 = p−2 L yields a workable resolvent. Note that p−2 M ( M 0 ( p−1 M ,
so we can take a basis such that
M = hη1 , η2 i and M 0 = hπ −1 η1 , π −2 η2 i.
We then get
π −1 a11
1 −1
0
0
(A , B ) = 2 π a12
1 −1
a13
2π
yielding condition (d).
1 −1
a12
2π
π −1 a22
1 −1
a23
2π
1 −1
a13
b11
2π
1
1 −1
,
π
a
23
2
2 b12
1
−1
π a33
2 b13
1
2 b12
b22
1
2 b23
1
2 b13
1
,
2 b23
b33
10 Conclusion
We have found the Dedekind domain to be a suitable base ring for generalizing the
integral parametrizations of algebras and their ideals by Bhargava and his forebears.
In each case, ideal decompositions a1 ⊕ · · · ⊕ an fill the role of Z-bases, and elements
of appropriate fractional ideals take the place of integers in the parameter spaces.
We have also shown that the notion of “balanced,” introduced by Bhargava to
describe the ideal triples parametrized by general nondegenerate 2 × 2 × 2 cubes,
O’Dorney
Page 39 of 39
has some beautiful properties and is worthy of further study. We expect that the
methods herein will extend to replicate the other parametrizations in Bhargava’s
“Higher Composition Laws” series and may shed light on the analytic properties of
number fields and orders of low degree over base fields other than Q.
Competing interests
I have no competing interests.
Acknowledgements
A previous version of this paper served as my senior thesis at Harvard College. I thank my thesis advisor, Benedict
Gross, for many helpful discussions and comments. I thank Melanie Wood for clarifications on the relationships
between my work and hers. I thank Arul Shankar for useful discussions, especially for informing me that he and
Wood had been interested in the question answered by Corollary 8.6. I thank Brian Conrad for the suggestion that I
work with Prof. Gross. Finally, I thank Ken Ono for suggesting RMS as a publication venue.
References
1. Cox, D.A.: Primes of the Form x2 + ny 2 : Fermat, Class Field Theory, and Complex Multiplication, 2nd edn.
Pure and Applied Mathematics (Hoboken), p. 356. John Wiley & Sons, Inc., Hoboken, NJ (2013).
doi:10.1002/9781118400722. http://dx.doi.org/10.1002/9781118400722
2. Bhargava, M.: Higher composition laws. I. A new view on Gauss composition, and quadratic generalizations.
Ann. of Math. (2) 159(1), 217–250 (2004). doi:10.4007/annals.2004.159.217
3. Bhargava, M., Varma, I.: The mean number of 3-torsion elements in the class groups and ideal groups of
quadratic orders. Unpublished (2014). http://arxiv.org/abs/1401.5875v1
4. Delone, B.N., Faddeev, D.K.: The Theory of Irrationalities of the Third Degree. Translations of Mathematical
Monographs, Vol. 10, p. 509. American Mathematical Society, Providence, R.I. (1964)
5. Bhargava, M.: Higher composition laws. III. The parametrization of quartic rings. Ann. of Math. (2) 159(3),
1329–1360 (2004). doi:10.4007/annals.2004.159.1329
6. Bhargava, M.: The density of discriminants of quartic rings and fields. Ann. of Math. (2) 162(2), 1031–1063
(2005). doi:10.4007/annals.2005.162.1031
7. Bhargava, M.: Higher composition laws. IV. The parametrization of quintic rings. Ann. of Math. (2) 167(1),
53–94 (2008). doi:10.4007/annals.2008.167.53
8. Bhargava, M.: The density of discriminants of quintic rings and fields. Ann. of Math. (2) 172(3), 1559–1591
(2010). doi:10.4007/annals.2010.172.1559
9. Wood, M.M.: Parametrizing quartic algebras over an arbitrary base. Algebra Number Theory 5(8), 1069–1094
(2011). doi:10.2140/ant.2011.5.1069
10. Wood, M.M.: Parametrization of ideal classes in rings associated to binary forms. J. reine angew. Math. 689,
169–199 (2014). doi:10.1515/crelle-2012-0058
11. Milnor, J.: Introduction to Algebraic K-theory, p. 184. Princeton University Press, Princeton, N.J.; University
of Tokyo Press, Tokyo (1971). Annals of Mathematics Studies, No. 72
12. Wood, M.M.: Gauss composition over an arbitrary base. Adv. Math. 226(2), 1756–1771 (2011).
doi:10.1016/j.aim.2010.08.018
13. Bhargava, M.: Higher composition laws. II. On cubic analogues of Gauss composition. Ann. of Math. (2)
159(2), 865–886 (2004). doi:10.4007/annals.2004.159.865
14. Poonen, B.: The moduli space of commutative algebras of finite rank. J. Eur. Math. Soc. (JEMS) 10(3),
817–836 (2008). doi:10.4171/JEMS/131
Figures
√
Figure 1 Stella octangula showing the range of ideal triples in Zp [pn u] that are balanced
| 0math.AC
|
Mysteries of Visual Experience
Jerome Feldman
DRAFT 3/14/18
‘The most beautiful and profound experience is the feeling of mystery. It underlies religion as
well as all deeper aspirations in art and science.’ Einstein
Introduction
Science is a crowning glory of the human spirit and its applications remain our best hope for
social progress. However, there are limitations to existing science and perhaps to any science.
The general mind-body problem is known to be currently intractable and mysterious (8). This is
one of many deep problems that are generally agreed to be beyond the present purview of
Science, including many quantum phenomena, etc. However, all of these famous unsolved
problems are either remote from everyday experience (entanglement, dark matter) or are hard
to even define sharply (phenomenology, consciousness, etc.).
In this note, we will consider some computational problems in vision that arise every time that
we open our eyes and yet are demonstrably inconsistent with current theories of neural
computation. The focus will be on two famous related phenomena, known as the neural
binding problem and the experience of a detailed stable visual world. I, among many others,
have struggled with these issues for more than fifty years (1, 2, 3). Somewhat paradoxically,
the continuing progress in scientific methods and knowledge reveals that these are both
unsolvable within existing neuroscience. By considering some basic facts about how the brain
processes image input, we will show that, under the standard theory, there are not nearly
enough brain neurons to compute what we experience as vision. Inconsistencies like the ones
shown here have had a profound effect on paradigm change in the sciences. More directly, the
discussions below suggest possible new theories and experiments.
Demonstrations
The visual system can only capture fine detail in a small ( ~1 degree) part of the visual field;
this is about the size of your thumbnail at arm’s length. “The experience of a detailed full-field
stable visual world” refers to our subjective perception of a large high-resolution scene. First,
consider Figure 1. Your vision is best in the center of gaze and the small letters in the center of
the figure are easy to read when you look directly at them, but not when you look to the side.
The letters further from the center are progressively larger and this describes how much
coarser your vision becomes with eccentricity.
Figure 1. Size for Equal Visibility with Eccentricity
You can experience this directly using the line of text in Figure 2. Cover or close one eye and
focus on the + in the center from a distance of about 12 inches. While holding focus, try to
name the letters to the left. You should be able to do much better with the progressively larger
letters to the right of the +. In ordinary viewing, there is no problem because we change our
gaze several times a second.
U
Q
C
G
O + C
O
U
QG
Figure 2. Demonstration of Visibility with Eccentricity
More generally, representing more information requires more hardware, which is why new
phones are marketed as having cameras with more megapixels. This is also true for the
neurons in the brain and this fact will play a major role in the discussion. There is a great deal
known about how the brain processes visual information, largely because other mammals,
particularly primates, have quite similar visual systems. We will focus on primary visual cortex
or V1. Looking ahead, Figure 4B shows a flattened and projected view of the human brain with
V1 on the far left.
Unsurprisingly, the brain realizes its high central resolution using many, densely packed,
neurons. The central portion of the retina in the eye is the fovea and the downstream target of
these foveal neurons in V1 of the brain is called the foveal projection.
Figure 3. Tootell, et al. Experiment (4)
An important aspect of this architecture can be seen in Figure 3. The upper part of the figure
depicts an oscillating radial stimulus, also with more detail in the center, which was presented
to a primate subject. The lower half of the figure shows the parts of visual cortex that
responded strongly to the input. As you can see, by far the most activity is the foveal projection
on the far left, corresponding to the detailed image in the center of the input stimulus (red
arrow). So, vision is most accurate in a small central area of the visual field and this is
achieved by densely packed neurons in the corresponding areas of the brain (4).
However, our visual experience is not at all like this. We experience the world as fully detailed
and there is currently no scientific explanation of this. There is more - we normally move
(saccade) our gaze to new places in the scene about 3 or 4 times per second. These
saccades help us see and act effectively and are not random. Again, our experience does not
normally include any awareness of the saccades or the radically different visual inputs that
they entail. Taken together these unknown links between brain and experience are known as
“the experience of a detailed stable visual world” and this is generally accepted, if not
understood.
There is extensive continuing research on various aspects of visual stability (6, 7, 8, 9, 10, 11).
None of this work attempts to provide a complete solution and it is usually explicit that deep
mysteries remain. Reference 9 is an excellent survey of behavioral findings and reference 11
has current neuroscience results.
We are attempting here to establish a much stronger statement. These stable world
phenomena and a number of others are inconsistent with standard theories of neural
processing (20). The demonstrations below involve combining findings from several distinct
areas of investigation, as an instance of Unified Cognitive Science (12). Before digging in to
the computational details, we consider some consequences of establishing that there is
presently no scientific explanation of such visual experiences.
Why Inconsistency is important
It may seem surprising that we can exhibit such strong results without directly considering the
relation between subjective (1st person) and objective (3rd person) experience, which is one
of the core mysteries of the mind. For at least the cases presented, there is NO theory of
the 1st person experience that is consistent with the known theory, structure, and behavior
of the brain. This also challenges several proposed mind/brain relations as
epiphenomenalism or the notion that the 1st/3rd person link is a brute fact of nature that
cannot be further analyzed. These would not yield inconsistency.
In addition, throughout the history of science, crucial instances of inconsistency have led to
profound reconsiderations and discoveries. One of the best-known cases is the fact that
Rutherford’s planetary model of the atom entailed that electrons rotating around the atomic
nucleus would radiate energy and eventually crash into it. This was one of a number of deep
inconsistencies leading to the development of quantum theory.
A recent important inconsistency in Cognitive Science is the “Word Superiority Effect” (34). A
wide range of experiments established that people were faster and more accurate in
recognizing the letter A in context, e.g., CAT, than the same letter alone. These results conflict
with the naïve assumption that more input should require more processing. This was one of
the inconsistencies resulting in the paradigm shift to massively parallel (connectionist) models
of brain function.
If there really are fundamental inconsistencies between visual experience (the mind) and the
neural theory of the brain, this is a major challenge to the (currently dominant) theories that the
mind is constituted entirely from the activity of the brain. As usual, Dennett is unequivocal: “Our
minds are just what our brains non-miraculously do “(30, Preface). Not only philosophers are
so dogmatic. Stanislas Dehaene, a leading experimentalist, says: “If you had any lingering
doubts that your mental life arises entirely from the activity of the brain, these examples should
lift them” (37, p.153). No one has suggested how this postulated mind/brain interaction would
work, and we will show here that the examples above cannot be explained within existing
theories. There is always the possibility of a conceptual breakthrough, but it would entail
abandoning some of our core beliefs about (at least) neural computation.
There is a plausible functional story for the stable world experience and the related binding
problem to be discussed below. First of all, we do have an integrated (top-down) sense of the
space around us that we cannot currently see, based on memory and other sense data –
primarily hearing and smell. In addition, since we are heavily visual, it is adaptive to use vision
as broadly as possible. In fact, it would be extremely difficult to act in the world using only the
bulls-eye images from Figure 1 and separated information on size, color, etc. The mind
(somehow) encodes a more accurate version of the world than can be directly captured by our
limited neural hardware.
We should not be surprised that our subjective experience sometimes deviates from the
information captured and processed by the visual system. Our senses and the nervous system
in general evolved to help our bodies function effectively in a physical and social world that we
cannot directly observe (36). Given that such mental experiences are evolutionarily
fundamental, what can we know about their physical realization? There is an impressive body
of work suggesting how aspects of subjective experience, closely related to the discussion
here, seem to be needed for animals to deal with external space and to combine various
sensory inputs and goals. At least some aspects of these experiences may well be found in
cephalopods (38) and insects (36).
This ecological requirement also reveals a deep problem in the use of the term “illusion”. The
word is sometimes used to describe a perception that is inconsistent with external reality and
sometimes to describe an experience that is inconsistent with the neural representation even
when the subjective experience is more like external reality. To further confuse the issue,
“illusion” is also used metaphorically, as in the postulated “illusion of Free Will” (31). Everyone
agrees that we all act as if we had Free Will, even determinists who deny that they have this
power. There does seem to be an agreed definition of “illusion” that supports its use in a
serious discussion of the mind.
Computational Limitations
We will now prove that some everyday visual experiences cannot be explained within existing
neuroscience. The basic form of the argument will be computational. There is no way that
brain neurons, as we know them, could represent or compute the substrate of our visual
experience. The constraint of explaining visual experience also rules out many proposed and
speculative theories of neural computation in the human brain, as discussed below. To explore
the details, we turn next to Figure 4 A, B.
Figure 4 Flat map projection of the Human brain (5)
Figure 4A is a standard flattened projection of one hemisphere of the human brain with the
various areas colored. The numbers refer to the traditional Brodmann classification of brain
regions from their anatomical details. Modern methods (27, 35) have further refined this picture
and elaborated the basic functions computed in these different areas. Figure 4B provides more
detail on this functional separation in the visual system, which is at the core of the neural
binding problem, one of our mysteries.
The visual area V1, our main concern for the stable world experience and the subject of Figure
3, is shown as the yellow area on the left of Figure 4A (as area 17) and as the magenta area in
Figure 4B. Notice that V1 is the largest of the visual areas; this will be important for our
discussion.
There are two additional lessons to be gleaned from Figure 4 above. First, 4A shows that the
functionality of the cerebral cortex is basically known (5, 27, 35) – there is no large available
space for neural computation of currently mysterious phenomena. In addition, various aspects
of our visual experience are primarily computed in distinct and often distant interacting circuits.
For example, in Figure 4B color calculation is based in the bright green area V4v and motion
calculation involves several areas: V3, V3A, MT, etc. In spite of this extreme separation of
function, we experience the world as an integrated image with objects that combine all visual
properties and even associate these with other senses like sound when appropriate. The
mystery of how this happens is called the “hard binding problem” (3).
There are two immediate challenges to address in “the experience of the stable visual world”:
apparent stability over saccades and the detailed perception of the full visual field. One popular
idea is to suppose that the perceived full field is pieced together as a mosaic of “bull’s-eye”
views (Figure 3) from many saccades. There are two serious flaws in this story, one temporal
and one spatial, as an explanation of the experience. We only make about 3~4 saccades per
second – this is too slow for stable vision (movies are ~ 20 frames per second). In addition, 3
or 4 such images would not yield nearly enough detailed information to build a detailed full field
view.
In addition, it would require a huge area of visual neurons to encode the detailed full field view
that we subjectively perceive. We can give a quantitative estimate of what is involved. There
are a number of alternative calculations, but they all confirm the basic point that fine resolution
over a large visual field would require brain area several times larger than V1 (Figure 4).
Stan Klein, who has looked extensively at this issue, suggests the following analysis focusing
on the retinal ganglion cells –RGC. The key equations from (14) are:
Thr(Ecc) = Thr(fovea)*(1 + Ecc/E2) or
Sep(Ecc) = Sep(fovea)*(1 + Ecc/E2)
where Thr (or Sep) is threshold (or separation) in minutes.
and Ecc is eccentricity in deg
and E2 is the eccentricity at which Thr or Sep double.
E2 is the number of degrees of eccentricity at which the spacing of V1 neurons or
ganglion cells double. Levi et al. (14) found E2=0.7 deg for cortical cells and is about 1.0
deg for ganglion cells. That is, for ganglion cells the spacing would be s = 0.5 (Ecc + 1)
min so at Ecc=0 the spacing is about 0.5 min and at 20 deg it is about 10 min, 20 times
as much. This calculation suggests that it might require 20 times as much V1 area to
capture the precision of the fovea out to 20 deg of visual angle.
From a slightly different perspective, the cortical magnification factor says that the resolution at
20 degrees eccentricity is 20 times worse than at the foveal projection. This is because of
retinal under-sampling in the periphery; the detailed information is only captured at the fovea of
the retina (13). Also, the dense neural circuits in the V1 foveal projection has about 200,000
cells per square mm, while at 20 degrees out it is more like 4,000 cells per square mm (15).
This is a factor of 50:1 denser in the V1 foveal projection than in the periphery. The V1 foveal
projection occupies about a quarter of region on the left in Figure 4. For the brain to encode
our detailed perception out to 20 degrees would require an area roughly 12 times the size of
V1. There is no way that an area nearly this large could fit into Figure 4. The remarkable recent
advances (27, 35) describing a much more detailed parcellation of human cerebral cortex
provides even stronger evidence against unknown visual areas.
We can also consider the evidence from the hundreds of full-brain scanning experiments that
are exploring which brain locations are active for various vision tasks (16, 17, 18, 35). This
precludes the possibility that a network large enough to capture a detailed image could remain
undetected.
In summary, as long as we believe that more detail requires more neurons, there is no place
in the brain that could encode a basis for the detailed large field image that we experience.
This analysis disproves more than the idea of unknown brain circuitry that underlies our stable
world experience. It also refutes any plausible substrate for other proposals such as complete
“remapping” which suggest that all of the information from one saccade is (somehow) mapped
to the input coming from the next saccade (7, p.557).
The binding problem (3, 41) is a closely related mystery of vision that we can consider, also
based on Figure 4. Although the full computational story is more complex, it is the case that
different visual features are largely computed in separate brain areas. However, we experience
the world as coherent entities combining various properties such as size, shape, color, texture,
motion, etc. (39). Again, there is no place in the brain that could encode a detailed substrate
for what we effortlessly perceive. This also suggests that our subjective perception (somehow)
integrates activity from different brain circuits. Various forms of the binding problem are also
the subject of ongoing research (3, 19)
A Touchstone for alternative brain theories
The discussion above is based on the standard theory (20) that information processing in the
brain is based on complex networks of neurons that communicate over long distances mainly
by electrical spikes and learn mainly through changes in the connections (synapses) between
neurons. This theory also includes a wide range of other chemical and developmental factors,
but none that would affect the basic results above.
However, there are a number of alternative proposals that deny the centrality of standard
neural computation and several of these are being actively discussed (21, 22, 23); two good
sources for a wide range of alternative models are the Journal of Consciousness Studies and
http://consciousness.arizona.edu. One reason for this interest in alternative theories is that
everyone agrees that the current standard theory does not support a reductionist explanation
of historic mind-brain problems like subjective experience and consciousness. The standard
theory continues to yield scientific and clinical progress, so any new proposal should be
consistent with it.
Alternative ideas on the basis for brain information processing include quantum effects (28)
and central roles for the glia, for the neuropil, or for microtubules. All of the suggestions for
some sub-neural substrate for perception suffer from the same problem – the only known
mechanism for the requisite fast long-distance communication in the brain is neural spikes.
More general architectural suggestions include global workspace model (40) and the Tononi
information model (21). Many proposals suggest some unspecified mass action of neural
assemblies, following a long tradition (23). Cohen et al. (26) show how known results on
summary statistics and peripheral vision explain some of people’s ability to get the “gist” of a
scene without capturing all of the detail of subjective perception. After extensive analysis and
modeling of peripheral vision and what she calls "the awareness puzzle", Rosenholtz (42)
concludes that "Perception is inherently something of an illusion".
Edwards (25) suggests another approach – unified perception (and consciousness) is based
on wave patterns in the membranes of individual cells. All of these ideas presuppose that
there is some substrate (NCC) for subjective experience in the brain, but there are also more
radical theories like that of Alva Noe (22). He claims to explain how "we enjoy an experience of
worldly detail that is not present in our brains". Starting from a standard argument that utilities
are the basis for perception, Hoffman (33) suggests that: “Your apple and my apple are
distinct, just as your headache is distinct from my headache. Something in the objective world
triggers us both to perceive an apple, but whatever that thing might be in the objective world, it
is almost certainly nonspatial, nontemporal, and in no way resembles an apple”
Since the deep mind-brain phenomena of most common interest are not well defined, it has
not been straightforward to evaluate any of these suggested alternatives to the standard model
of neural computation. The findings described above could yield concrete touchstone
problems for proposed theories of representation, computation, and communication in the
brain. Both the binding problem and the experience of a detailed stable visual scene are
ubiquitous in daily life, are functionally necessary, and have clear informational requirements.
We could ask proponents of speculative brain models how their theory could account for these
two concrete phenomena. That is, assume your theory is true and show how it helps explain
these (or other) touchstone problems. I have done this informally with several leading
proponents of alternative models and have never heard even a vague claim of adequacy.
The general acceptance of some such touchstone tasks could sharpen the discussion of
information processing in the brain. Of course, the deep mind-brain problem remains a
mystery, but we should expect proposed models of neural computation to address some of the
concrete touchstone problems, like those discussed here.
Experiments
We have shown that existing neuroscience cannot address all the basic questions of
subjective experience. Of course, very productive communities are working on many aspects
of cognitive and perceptual neuroscience. There is a wide range of motivations for these
efforts, but these do not usually include trying to elucidate remaining mysteries, such as the
ones discussed here. There are risks involved in attacking difficult problems and relatively few
such efforts in the current stressful research environment.
Nevertheless, there is some research that is making progress on demystifying some of the
mysteries of subjective experience. In the Fall of 2017, a UC Berkeley interdisciplinary
seminar course explored “Science and Subjectivity “ http://rctn.org/wiki/VS298:_Subjectivity .
This web site for the course contains references and video lecture recordings for a wide range
of research that pushes on the boundary between routine science and remaining mysteries.
Week 3 contains lecture and discussion on the material in this article. The web site also
includes presentations and discussion by Michael Cohen, Jerry Feldman, Rich Ivry, Stan Klein,
Bob Knight, Christof Koch, Ken Nakayama, Brian Odegaard, Bruno Olshausen, Terry Regier,
Shin Shimojo, and Peter Tse. If there is a scientific explanation of these mysteries, it will need
to be an evolutionary story, so it should be fruitful to focus on a wide range of animals (36, 38)
Conclusions
There is general agreement that there are mysteries about the world and our place in it that
are not yet understood. Even radical materialists will concede that there are questions (e.g.
free will) that might never have scientific solutions. Nevertheless, it is not widely understood
that, every time we open our eyes, we experience phenomena that cannot be explained with
existing neuroscience and possibly not with any science.
As thinkers, we have no choice but to acknowledge that we do not know and may never know
the answers to many deep questions about the world and ourselves (24). There are two basic
ways to learn about the physical and social world: investigation and stories. Science is a
uniquely powerful tool of investigation, but is limited in scope at least at present. Of course,
there remains a vast amount that can and should be explored and exploited scientifically.
Stories can provide insights that are not directly testable, and this is certainly also important in
science. The stories in art, mythology, religion, etc. have been and will remain powerful
sources of guidance about how to live.
Initial attempts to convey the ideas in this note have not been very successful. I find that
scientists have a strong negative reaction to my simple demonstrations. On the other hand,
philosophers and humanists welcome any attack on reductionism. However, it seems like both
responses are mainly territorial. It does not appear to be too much of a stretch to view the
(Eastern or Western) religious practitioners as another special interest group. It is certainly true
that all these groups have emotional/spiritual as well as financial/power stakes in the big
questions, but there does not seem to be much support for the agnostic mysterian position "we
simply don't know".
Everyone is entitled to his or her own (religious or other) beliefs, but there is nothing in our
current ignorance that privileges one faith over all others. Belief in the inevitability of complete
scientific answers (8, 21, 30) is one such faith. There are beliefs (e.g. about the age of the
Earth) that contradict established scientific knowledge and cannot be taken seriously.
Somewhat surprisingly, there seems to be little support in cognitive and neuro science for "we
simply don't know". This is despite the fact "we simply don't know and may never know" is the
accepted response in physics for some fundamental questions. What has been most surprising
is how many people prefer believing there must be an (inscrutable) reductionist answer rather
than accepting the agnostic stance. Even philosophers, who consider the mind/brain problem
as a possible feature of nature, try to prove their contention rather than leave it as one
possibility (8). This all might be related to recent findings suggesting an innate human drive to
find answers for unanswerable questions (32).
Nevertheless, in the face of all that is unknown, we (following the physicists) would do well to
appreciate both what is scientifically known and also the mysteries that remain. Ideally, results
like those above will encourage theory and experiment on questions at the boundaries
between the known and unknown.
Acknowledgement
This work was supported in part by the Office of Naval Research under Grant N000141110416
and by a grant from Google.
References
1. Four frames suffice: A provisional model of vision and space. Jerome A. Feldman
THE BEHAVIORAL AND BRAIN SCIENCES (1985) Volume 8 / Issue 02 / July 1985, pp 265- 289
2. Integration of form across saccadic eye movements. Hayhoe, M.; Lachter, J.; Feldman, J. Perception,
Vol. 20, No. 3, 01.12.1991, p. 393-402.
3. The neural binding problem(s), Jerome Feldman Cognitive Neurodynamics. February 2013, Volume
7, Issue 1, pp 1-11
4. Functional anatomy of macaque striate cortex. Tootell RB, Silverman MS, Hamilton SL, Switkes E,
De Valois RL., Spatial frequency. J Neurosci. 1988 May; 8(5):1610-24.
5. A Population-Average, Landmark- and Surface-based (PALS) atlas of human cerebral cortex. Van
Essen DC1. Neuroimage 2005 Nov 15;28(3):635-62.
6. “Visual Stability” Proc. Trans Royal Society B February 2011, David Melcher (ed.)
Volume: 366 Issue: 1564 DOI: 10.1098/rstb.2010.0277
http://rstb.royalsocietypublishing.org/content/366/1564/468
7. Computational models of spatial updating in perisaccadic perception. Hamker F. H., Zirnsak M.,
Ziesche A, Lappe M. 2011. Phil. Trans. R. Soc. B 366, 554–571. doi:10.1098/rstb.2010.0229
(doi:10.1098/rstb.2010.0229)
8. The Hard Problem of Consciousness. Weisberg, J. Internet Encyclopedia of Philosophy.
http://www.iep.utm.edu/hard-con/
9. Transsaccadic processing: stability, integration, and the potential role of remapping. Higgins E,
Rayner K., Atten Percept Psychophys (2015) 77:3–27
10. Brain circuits underlying visual stability across eye movements-converging evidence for a
neuro-computational model of area LIP. Ziesche A, Hamker FH. (2014) Front Comput Neurosci
8:25.
11. Saccadic Corollary Discharge Underlies Stable Visual Perception. James Cavanaugh,
Rebecca A. Berman, Wilsaan M. Joiner, Robert H. Wurtz. Journal of Neuroscience, 6 January
2016, 36(1): 31-42; doi: 10.1523/JNEUROSCI.2054-15.2016
12. Cognitive Science should be unified: comment on Griffiths et al. and McClelland et al Feldman
JA. Trends Cogn Sci. 2010 Aug;14(8):341. doi: 10.1016/j.tics.2010.05.008. Epub 2010 Jun 17.
13. Cortical magnification within human primary visual cortex correlates with acuity thresholds.
Duncan RO1, Boynton GM. Neuron. 2003 May 22;38(4):659-71.
14. Vernier acuity, crowding and cortical magnification. Levi DM, Klein SA, Aitsebaomo AP
(1985) Vision Res. 25(7):963-77.
15. Modeling magnification and anisotropy in the primate foveal confluence. Schira MM1, Tyler
CW, Spehar B, Breakspear, M.PLoS Comput Biol. 2010 Jan 29;6(1):e1000651. doi:
10.1371/journal.pcbi.1000651.
16. Borders of multiple visual areas in humans revealed by functional magnetic resonance
imaging. Sereno MI, Dale AM, Reppas JB, Kwong KK, Belliveau JW, Brady TJ, Rosen BR,
Tootell RB. Science. 1995 May 12;268(5212):889-93.
17. The global landscape of cognition: hierarchical aggregation as an organizational principle of
human cortical networks and functions. Taylor P1,2, Hobbs JN1, Burroni J1 SiegelmannHT1,2.2015
Dec 16;5:18112. doi: 10.1038/srep18112.
18. The economy of brain network organization. Bullmore E1, Sporns O. Nat Rev Neurosci 2012 Apr
13;13(5):336-49. doi: 10.1038/nrn3214.
19. Feature- and Face-Exchange illusions: new insights and applications for the study of the
binding problem. Arthur G. Shapiro1, Gideon P. Caplovitz2 and Erica L. Dixon1. Front. Hum.
Neurosci., 16 October 2014 | http://dx.doi.org/10.3389/fnhum.2014.00804
20. Sense and the single neuron: Probing the physiology of perception.
Parker, AJ and Newsome, WT (1998). Annu. Rev. Neurosci. 21: 227-77.
21. Consciousness: Confessions of a Romantic Reductionist, C. Koch. The MIT Press, (2012),
ISBN 978-0-262-01749-7
22. Out of Our Heads: Why You Are Not Your Brain, and Other Lessons from the Biology of
Consciousness , Alva Noë Hill and Wang 2009
ISBN-13: 978-0809016488
23. From the Neuron Doctrine to Neural Networks. Rafael Yuste, Nature Reviews Neuroscience .
v16, August 2015, pp487-496.
24. Science and the Indian Tradition: When Einstein Met Tagore (India in the Modern World).
David L. Gosling. Routledge; (2007) ISBN-13: 978-0415481342
25. Is Consciousness Only A Property Of Individual Cells? Jonathan CW Edwards
Journal of Consciousness Studies, Volume 12, No.4-5, pp60-76
26. What is the bandwidth of perceptual experience? Cohen et al. Trends in Cognitive Sciences, 2016
DOI: 10.1016/j.tics.2016.03.006
27. . A Multi-modal Parcellation of Human Cerebral Cortex. Glasser, M.F. et al Nature DOI:
10.1038/nature18933
28. Consciousness in the universe: A review of the ‘Orch OR’ theory Hameroff,S., Penrose, R.
Physics of Life Reviews, Volume 11, Issue 1, March 2014, Pages 39-78
29. No-Report Paradigms: Extracting the True Neural Correlates of Consciousness. Tsuchiya,
Naotsugu et al. , Trends in Cognitive Sciences , Volume 19 , Issue 12 , 757 – 770. DOI:
http://dx.doi.org/10.1016/j.tics.2015.10.002
30. Freedom Evolves, Dennett, D.C., Viking Press 2003 ISBN 0-670-03186-0
31. WHO 'S IN CHARGE? Free Will and the Science of the Brain. Michael S.
Gazzaniga. HarperCollins, 2011 ISBN: 978-0-06-190610-7
32. Religion Explained: The Evolutionary Origins of Religious Thought, Pascal Boyer
Basic Books, 2002, ISBN 0-465-00696-5.
33. Public Objects and Private Qualia, Donald Hoffman, in Handbook of Experimental
Phenomenology, L. Albertazzi (Ed) , 2013, Wiley, ISBN 978-1-119-95468.
34. Word Superiority Effect. https://en.wikipedia.org/wiki/Word_superiority_effect
35. The ventral visual pathway: an expanded neural framework for the processing of object
quality. Kravitz DJ1, Saleem KS, Baker CI, Ungerleider LG, Mishkin M.
Trends Cogn Sci. 2013 Jan; 17(1):26-49. doi: 10.1016/j.tics.2012.10.011.
36. What insects can tell us about the origins of consciousness
Andrew B. Barrona and Colin Kleinb
4900–4908 | PNAS | May 3, 2016 | vol. 113 | no. 18
37. Consciousness and the Brain, Dehaene, S. Viking, 2014.
38. Other Minds : The Octopus, the Sea, and the Deep Origins of Consciousness, Godfrey-Smith,
P., 2016, Farrar, Straus, and Giroux
39. Visual stability based on remapping of attention pointers, Cavanagh, P., A. R. Hunt, A. Afraz
and M. Rolfs. Trends in Cognitive Sciences Vol.14 No.4, 147-153
40. Global Workspace Theory https://en.wikipedia.org/wiki/Global_workspace_theory
41. The Binding Problem, Roskies, A. Neuron, ISSN: 0896-6273, Vol: 24, Issue: 1, Page: 7-9
1999
42. What modern vision science reveals about the awareness puzzle: Summary-statistic encoding
plus decision limits underlie the richness of visual perception and its quirky failures Rosenholtz,
R. 2017 arXiv:1706.02764
| 2cs.AI
|
arXiv:1609.07135v2 [math.ST] 28 Nov 2017
Convergence of Regression-Adjusted Approximate Bayesian
Computation
B Y W ENTAO L I
School of Mathematics, Statistics and Physics, Newcastle University, Newcastle upon Tyne NE1
7RU, U.K.
[email protected]
PAUL F EARNHEAD
Department of Mathematics and Statistics, Lancaster University, Lancaster LA1 4YF, U.K.
[email protected]
AND
S UMMARY
We present asymptotic results for the regression-adjusted version of approximate Bayesian
computation introduced by Beaumont et al. (2002). We show that for an appropriate choice of
the bandwidth, regression adjustment will lead to a posterior that, asymptotically, correctly quantifies uncertainty. Furthermore, for such a choice of bandwidth we can implement an importance
sampling algorithm to sample from the posterior whose acceptance probability tends to unity
as the data sample size increases. This compares favourably to results for standard approximate
Bayesian computation, where the only way to obtain a posterior that correctly quantifies uncertainty is to choose a much smaller bandwidth; one for which the acceptance probability tends to
zero and hence for which Monte Carlo error will dominate.
Keywords: Approximate Bayesian computation; Importance sampling; Local-linear regression; Partial information.
1. I NTRODUCTION
Modern statistical applications increasingly require the fitting of complex statistical models,
which are often intractable in the sense that it is impossible to evaluate the likelihood function.
This excludes standard implementation of likelihood-based methods, such as maximum likelihood estimation or Bayesian analysis. To overcome this there has been substantial interest in
likelihood-free or simulation-based methods, which replace calculating the likelihood by simulation of pseudo datasets from the model. Inference can then be performed by comparing these
pseudo datasets, simulated for a range of different parameter values, to the actual data.
Examples of such likelihood-free methods include simulated methods of moments (Duffie &
Singleton, 1993), indirect inference (Gouriéroux & Ronchetti, 1993; Heggland & Frigessi, 2004),
synthetic likelihood (Wood, 2010) and approximate Bayesian computation (Beaumont et al.,
2002). Of these, approximate Bayesian computation methods are arguably the most common
methods for Bayesian inference, and have been popular in population genetics (e.g., Beaumont
et al., 2002; Cornuet et al., 2008), ecology (e.g., Beaumont, 2010) and systems biology (e.g.,
Toni et al., 2009); more recently they have seen increased use in other application areas, such as
econometrics (Calvet & Czellar, 2015) and epidemiology (Drovandi & Pettitt, 2011).
1
2
The idea of approximate Bayesian computation is to first summarize the data using lowdimensional summary statistics, such as sample means or autocovariances or suitable quantiles
of the data. The posterior density given the summary statistics is then approximated as follows.
Assume the data are Yobs = (yobs,1 , . . . , yobs,n ) and modelled as a draw from a parametric model
with parameter θ ∈ Rp . Let K(x) be a positive kernel, where maxx K(x) = 1, and ε > 0 is the
bandwidth. For a given d-dimensional summary statistic s(Y ), our model will define a density
fn (s | θ). We then define a joint density, πε (θ, s | sobs ), for (θ, s) as
´
π(θ)fn (s | θ)K{ε−1 (s − sobs )}
,
−1
Rp ×Rd π(θ)fn (s | θ)K{ε (s − sobs )} dθds
where sobs = s(Yobs ). Our approximation to the posterior density is the marginal density
ˆ
πε (θ | sobs ) = πε (θ, s | sobs ) ds,
(1)
(2)
which we call the approximate Bayesian computation posterior density. For brevity we will often
shorten this to posterior density in the following. We will always call the actual posterior given
the summary the true posterior.
The idea of approximate Bayesian computation is that we can sample from πε (θ | sobs ) without needing to evaluate the likelihood function or fn (s | θ). The simplest approach is via rejection sampling (Beaumont et al., 2002), which proceeds by simulating a parameter value and
an associated summary statistic from π(θ)fn (s | θ). This pair is then accepted with probability
K{ε−1 (s − sobs )}. The accepted pairs will be drawn from (1), and the accepted parameter values will be drawn from the posterior (2). Implementing this rejection sampler requires only the
ability to simulate pseudo data sets from the model, and then to be able to calculate the summary
statistics for those data sets.
Alternative algorithms for simulating from the posterior include adaptive or sequential importance sampling (Beaumont et al., 2009; Bonassi & West, 2015; Lenormand et al., 2013; Filippi
et al., 2013) and Markov chain Monte Carlo approaches (Marjoram et al., 2003; Wegmann et al.,
2009). These aim to propose parameter values in areas of high posterior probability, and thus can
be substantially more efficient than rejection sampling. However, the computational efficiency of
all these methods is limited by the probability of acceptance for data simulated with a parameter
value that has high posterior probability.
This paper is concerned with the asymptotic properties of approximate Bayesian computation. It builds upon Li & Fearnhead (2018) and Frazier et al. (2016), who present results on
the asymptotic behaviour of the posterior distribution and the posterior mean of approximate
Bayesian computation as the amount of data, n, increases. Their results highlight the tension
in approximate Bayesian computation between choices of the summary statistics and bandwidth
that will lead to more accurate inferences, against choices that will reduce the computational cost
or Monte Carlo error of algorithms for sampling from the posterior.
An informal summary of some of these results is as follows. Assume a fixed dimensional
summary statistic and that the true posterior variance given this summary decreases like 1/n
as n increases. The theoretical results compare the posterior, or posterior mean, of approximate
Bayesian computation, to the true posterior, or true posterior mean, given the summary of the
data. The accuracy of using approximate Bayesian computation is governed by the choice of
bandwidth, and this choice should depend on n. Li & Fearnhead (2018) shows that the optimal
choice of this bandwidth will be O(n−1/2 ). With this choice, estimates based on the posterior
mean of approximate Bayesian computation can, asymptotically, be as accurate as estimates
based on the true posterior mean given the summary. Furthermore the Monte Carlo error of
3
an importance sampling algorithm with a good proposal distribution will only inflate the mean
square error of the estimator by a constant factor of the form 1 + O(1/N ), where N is the number
of pseudo data sets. These results are similar to the asymptotic results for indirect inference,
where error for a Monte Carlo sample of size N also inflates the overall mean square error of
estimators by a factor 1 + O(1/N ) (Gouriéroux & Ronchetti, 1993). By comparison choosing
a bandwidth which is o(n−1/2 ) will lead to an acceptance probability that tends to zero as n →
∞, and the Monte Carlo error of approximate Bayesian computation will blow up. Choosing a
bandwidth that decays more slowly than O(n−1/2 ) will also lead to a regime where the Monte
Carlo error dominates, and can lead to a non-negligible bias in the posterior mean that inflates
the error.
While the above results for a bandwidth that is O(n−1/2 ) are positive in terms of point estimates, they are negative in terms of the calibration of the posterior. With such a bandwidth
the posterior density of approximate Bayesian computation always over-inflates the parameter
uncertainty: see Proposition 1 below and Theorem 2 of Frazier et al. (2016).
The aim of this paper is to show that a variant of approximate Bayesian computation can
yield inference that is both accurate in terms of point estimation, with its posterior mean having the same frequentist asymptotic variance as the true posterior mean given the summaries,
and calibrated, in the sense that its posterior variance equals this asymptotic variance, when
the bandwidth converges to zero at a rate slower than O(n−1/2 ). This means that the acceptance
probability of a good approximate Bayesian computation algorithm will tend to unity as n → ∞.
2. N OTATION AND S ET- UP
We denote the data by Yobs = (yobs,1 , . . . , yobs,n ), where n is the sample size, and each observation, yobs,i , can be of arbitrary dimension. Assume the data are modelled as a draw from
a parametric density, fn (y | θ), and consider asymptotics as n → ∞. This density depends on
an unknown parameter θ ∈ Rp . Let B p be the Borel sigma-field on Rp . We will let θ0 denote
the true parameter value, and π(θ) the prior distribution for the parameter. Denote the support of
π(θ) by P. Assume that a fixed-dimensional summary statistic sn (Y ) is chosen and its density
under our model is fn (s | θ). The shorthand Sn is used to denote the random variable with density fn (s | θ). Often we will simplify notation and write s and S for sn and Sn respectively. Let
N (x; µ, Σ) be the normal density at x with mean µ and variance Σ. Let Ac be the complement
of a set A with respect to the whole space. For a series xn we write xn = Θ(an ) if there exist
constants m and M such that 0 < m < |xn /an | < M < ∞ as n → ∞. For a real function g(x),
denote its gradient function at x = x0 by Dx g(x0 ). To simplify notation, Dθ is written as D.
Hereafter ε is considered to depend on n, so the notation εn is used.
The conditions of the theoretical results are stated below.
Condition 1. There exists some δ0 > 0, such that P0 = {θ : |θ − θ0 | < δ0 } ⊂ P, π(θ) ∈
C 2 (P0 ) and π(θ0 ) > 0.
´
´ Ql
Condition 2. The kernel satisfies (i) vK(v) dv = 0; (ii)
k=1 vik K(v) dv < ∞ for any
coordinates (vi1 , . . . , vil ) of v and l ≤ p + 6; (iii) K(v) ∝ K(kvk2Λ ) where kvk2Λ = v T Λv
and Λ is a positive-definite matrix, and K(v) is a decreasing function of kvkΛ ; (iv) K(v) =
α
O(e−c1 kvk 1 ) for some α1 > 0 and c1 > 0 as kvk → ∞.
Condition 3. There exists a sequence an , satisfying an → ∞ as n → ∞, a d-dimensional vector s(θ) and a d × d matrix A(θ), such that for all θ ∈ P0 ,
an {Sn − s(θ)} → N {0, A(θ)}, n → ∞,
4
in distribution. We also assume that sobs → s(θ0 ) in probability. Furthermore, (i) s(θ) and
A(θ) ∈ C 1 (P0 ), and A(θ) is positive definite for any θ; (ii) for any δ > 0 there exists
δ 0 > 0 such that ks(θ) − s(θ0 )k > δ 0 for all θ satisfying kθ − θ0 k > δ; and (iii) I(θ) =
Ds(θ)T A−1 (θ)Ds(θ) has full rank at θ = θ0 .
Let fen (s | θ) = N {s; s(θ), A(θ)/a2n } be the density of the normal approximation to S and
introduce the standardized random variable Wn (s) = an A(θ)−1/2 {S − s(θ)}. We further let
fWn (w | θ) and f˜Wn (w | θ) be the densities for Wn under the true model for S and under our
normal approximation to the model for S.
2/5
Condition 4. There exists αn satisfying αn /an → ∞ and a density rmax (w)
satisfying Condition 2 (ii)–(iii) where K(v) is replaced with rmax (w), such that
supθ∈P0 αn fWn (w | θ) − feWn (w | θ) ≤ c3 rmax (w) for some positive constant c3 .
Condition 5. The following statements hold: (i) rmax (w) satisfies Condition 2 (iv); and (ii)
α
supθ∈Pc0 fWn (w | θ) = O(e−c2 kwk 2 ) as kwk → ∞ for some positive constants c2 and α2 , and
A(θ) is bounded in P.
Conditions 1–5 are from Li & Fearnhead (2018). Condition 2 is a requirement for the kernel function and is satisfied by all commonly used kernels, such as any kernel with compact
support or the Gaussian kernel. Condition 3 assumes a central limit theorem for the summary
statistic with rate an , and, roughly speaking, requires the summary statistic to accumulate information when n. This is a natural assumption, since many common summary statistics are sample
moments, proportions, quantiles and autocorrelations, for which a central limit theorem would
apply. It is also possible to verify the asymptotic normality of auxiliary model-based or composite likelihood-based summary statistics (Drovandi et al., 2015; Ruli et al., 2016) by referring to
the rich literature on asymptotic properties of quasi maximum-likelihood estimators (Varin et al.,
2011) or quasi-posterior estimators (Chernozhukov & Hong, 2003). This assumption does not
cover ancillary statistics, using the full data as a summary statistic, or statistics based on distances, such as an asymptotically chi-square distributed test statistic. Condition 4 assumes that,
in a neighborhood of θ0 , fn (s | θ) deviates from the leading term of its Edgeworth expansion
−2/5
by a rate an . This is weaker than the standard requirement, o(a−1
n ), for the remainder from
Edgeworth expansion. It also assumes that the deviation is uniform, which is not difficult to satisfy in a compact neighborhood. Condition 5 further assumes that fn (s | θ) has exponentially
decreasing tails with rate uniform in the support of π(θ). This implies that posterior moments
from approximate Bayesian computation are dominated by integrals in the neighborhood of θ0
and have leading terms with concise expressions. With Condition 5 weakened, the requirement
of εn for the proper convergence to hold might depend on the specific tail behavior of fn (s | θ).
Additionally, for the results regarding regression adjustment the following moments of the
summary statistic are required to exist.
´
´
Condition 6. The first two moments, Rd sfn (s | θ) ds and Rd ssT fn (s | θ) ds, exist.
3.
A SYMPTOTICS OF APPROXIMATE BAYESIAN COMPUTATION
3·1. Posterior
First we consider the convergence of the posterior distribution of approximate Bayesian computation, denoted by Πε (θ ∈ A | sobs ) for A ∈ B p , as n → ∞. The distribution function is a
random function with the randomness due to sobs . We present two convergence results. One is
5
the convergence of the posterior distribution function of a properly scaled and centered version
of θ, see Proposition 1. The other is the convergence of the posterior mean, a result which comes
from Li & Fearnhead (2018) but, for convenience, is repeated as Proposition 2.
The following proposition gives three different limiting forms for Πε (θ ∈ A | sobs ), corresponding to different rates for how the bandwidth decreases relative to the rate of the central limit
theorem in Condition 3. We summarize these competing rates by defining cε = limn→∞ an εn .
P ROPOSITION 1. Assume Conditions 1–5. Let θε denote the posterior mean of approximate
−3/5
Bayesian computation. As n → ∞, if εn = o(an ) then the following convergence holds, depending on the value of cε .
(i) If cε = 0 then
ˆ
ψ(t) dt → 0,
sup Πε {an (θ − θε ) ∈ A | sobs } −
A∈B p
A
in probability, where ψ(t) = N {t; 0, I(θ0 )−1 }.
(ii) If cε ∈ (0, ∞) then for any A ∈ B p ,
ˆ
Πε {an (θ − θε ) ∈ A | sobs } →
ψ(t) dt,
A
in distribution, where
ˆ
ψ(t) ∝
N [t; cε β0 {v − EG (v)}, I(θ0 )−1 ]G(v) dv, β0 = I(θ0 )−1 Ds(θ0 )T A(θ0 )−1 ,
Rd
and G(v) is a random density of v, with mean EG (v), which depends on cε and Z ∼ N (0, Id ).
(iii) If cε = ∞ then
ˆ
−1
sup Πε {εn (θ − θε ) ∈ A | sobs } −
ψ(t) dt → 0,
A∈B p
A
in probability, where ψ(t) ∝ K{Ds(θ0 )t}.
For a similar result, under different assumptions, see Theorem 2 of Frazier et al. (2016). See also
Soubeyrand & Haon-Lasportes (2015) for related convergence results for the true posterior given
the summaries for some specific choices of summary statistics.
The explicit form of G(v) is stated in the Supplementary Material. When we have the same
number of summary statistics and parameters, d = p, the limiting distribution simplifies to
ˆ
ψ(t) ∝
N {Ds(θ0 )t; cε v, A(θ0 )}K(v) dv.
Rd
The more complicated form in Proposition 1 (ii) above arises from the need to project the summary statistics onto the parameter space. The limiting distribution may depend on the value of the
summary statistic, sobs , in the space orthogonal to Ds(θ0 )T A(θ0 )−1/2 . Hence the limit depends
on a random quantity, Z, which can be interpreted as the noise in sobs .
The main difference between the three convergence results is the form of the limiting density
ψ(t) for the scaled random variable an,ε (θ − θε ), where an,ε = an 1cε <∞ + ε−1
n 1cε =∞ . For case
(i) the bandwidth is sufficiently small that the approximation in approximate Bayesian computation due to accepting summaries close to the observed summary is asymptotically negligible. The
asymptotic posterior distribution is Gaussian, and equals the limit of the true posterior for θ given
the summary. For case (iii) the bandwidth is sufficiently big that this approximation dominates
6
and the asymptotic posterior distribution of approximate Bayesian computation is determined by
the kernel. For case (ii) the approximation is of the same order as the uncertainty in θ, which
leads to an asymptotic posterior distribution that is a convolution of a Gaussian distribution and
the kernel. Since the limit distributions of cases (i) and (iii) are non-random in the space L1 (Rp ),
the weak convergence is strengthened to convergence in probability in L1 (Rp ). See the proof in
Appendix A.
P ROPOSITION 2. (Theorem 3.1 of Li & Fearnhead, 2018) Assume conditions of Proposition
−3/5
−1
1. As n → ∞, if εn = o(an ), an (θε − θ0 ) → N {0, IABC
(θ0 )} in distribution. If εn = o(a−1
n )
or d = p or the covariance matrix of the kernel is proportional to A(θ0 ) then IABC (θ0 ) = I(θ0 ).
For other cases, I(θ0 ) − IABC (θ0 ) is semi-positive definite.
Proposition 2 helps us to compare the frequentist variability in the posterior mean of approximate Bayesian computation with the asymptotic posterior distribution given in Proposition 1.
If εn = o(a−1
n ) then the posterior distribution is asymptotically normal with variance matrix
−2
−1
an I(θ0 ) , and the posterior mean is also asymptotically normal with the same variance matrix. These results are identical to those we would get for the true posterior and posterior mean
given the summary.
For an εn which is the same order as a−1
n , the uncertainty in approximate Bayesian compu−1
tation has rate an . However the limiting posterior distribution, which is a convolution of the
true limiting posterior given the summary with the kernel, will overestimate the uncertainty by a
constant factor. If εn decreases slower than a−1
n , the posterior contracts at a rate εn , and thus will
over-estimate the actual uncertainty by a factor that diverges as n → 0.
In summary, it is much easier to get approximate Bayesian computation to accurately estimate
−3/5
the posterior mean. This is possible with εn as large as o(an ) if the dimension of the summary
statistic equals that of the parameter. However, accurately estimating the posterior variance, or
getting the posterior to accurately reflect the uncertainty in the parameter, is much harder. As
commented in Section 1, this is only possible for values of εn for which the acceptance probability in a standard algorithm will go to zero as n increases. In this case the Monte Carlo sample size,
and hence the computational cost, of approximate Bayesian computation will have to increase
substantially with n.
As one application of our theoretical results, consider observations that are independent and
identically distributed from a parametric density f (· | θ). One approach to construct the summary
statistics is to use the score vector of some tractable approximating auxiliary model evaluated at
the maximum auxiliary likelihood estimator (Drovandi et al., 2015). Ruli et al. (2016) constructs
an auxiliary
Q model from a composite likelihood, so the auxiliary likelihood for a single observation is i∈I f (y ∈ Ai | θ) where {Ai : i ∈ I } is a set of marginal or conditional events for y.
Denote the auxiliary score vector for a single observation by clθ (· | θ) and the maximum auxiliary likelihood estimator for our data set by θbcl . Then the summary statistic, s, for any pseudo
P
data set {y1 , . . . , yn } is nj=1 clθ (yj | θbcl )/n.
For y ∼ f (· | θ), assume the first two moments of clθ (y | θ0 ) exist and clθ (y | θ) is differentiable at θ. Let H(θ) = Eθ {∂clθ (y | θ0 )/∂θ} and J(θ) = varθ {clθ (y | θ0 )}. Then if θbcl is consistent for θ0 , Condition 3 is satisfied with
n1/2 [S − Eθ {clθ (Y | θ0 )}] → N {0, J(θ)}, n → ∞,
in distribution, and with I(θ0 ) = H(θ0 )T J(θ0 )−1 H(θ0 ).
Our results show that the posterior mean of approximate Bayesian computation, using εn =
O(n−1/2 ), will have asymptotic variance I(θ0 )−1 /n. This is identical to the asymptotic variance
7
of the maximum composite likelihood estimator (Varin et al., 2011). Furthermore, the posterior
variance will overestimate this just by a constant factor. As we show below, using the regression
correction of Beaumont et al. (2002) will correct this overestimation and produce a posterior that
correctly quantifies the uncertainty in our estimates.
An alternative approach to construct an approximate posterior using composite likelihood is
to use the product of the prior and the composite likelihood. In general, this leads to a poorly
calibrated posterior density which substantially underestimates uncertainty (Ribatet et al., 2012).
Adjustment of the composite likelihood is needed to obtain calibration, but this involves estimation of the curvature and the variance of the composite score (Pauli et al., 2011). Empirical
evidence that approximate Bayesian computation more accurately quantifies uncertainty than
alternative composite-based posteriors is given in Ruli et al. (2016).
3·2. Regression Adjusted Approximate Bayesian Computation
The regression adjustment of Beaumont et al. (2002) involves post-processing the output of
approximate Bayesian computation to try to improve the resulting approximation to the true posterior. Below we will denote a sample from the posterior of approximate Bayesian computation
by {(θi , si )}i=1,...,N . Under the regression adjustment, we obtain a new posterior sample by using {θi − βbε (si − sobs )}i=1,...,N where βbε is the least square estimate of the coefficient matrix
in the linear model
θi = α + β(si − sobs ) + ei ,
i = 1, . . . , N,
where ei are independent identically distributed errors.
We can view the adjusted sample as follows. Define a constant, αε , and a vector βε as
(αε , βε ) = arg min Eε [kθ − α − β(s − sobs )k2 | sobs ],
α,β
where expectation is with respect the joint posterior distribution of (θ, s) given by approximate
Bayesian computation. Then the ideal adjusted posterior is the distribution of θ∗ = θ − βε (s −
sobs ) where (θ, s) ∼ πε (θ, s). The density of θ∗ is
ˆ
πε∗ (θ∗ | sobs ) =
πε {θ∗ + βε (s − sobs ), s | sobs } ds
Rd
and the sample we get from regression-adjusted approximate Bayesian computation is a draw
from πε∗ (θ∗ | sobs ) but with βε replaced by its estimator.
The variance of πε∗ (θ∗ | sobs ) is strictly smaller than that of πε (θ | sobs ) provided s is correlated with θ. The following results, which are analogous to Propositions 1 and 2, show that this
reduction in variation is by the correct amount to make the resulting adjusted posterior correctly
quantify the posterior uncertainty.
T HEOREM 1. Assume Conditions 1–6. Denote the mean of πε∗ (θ∗ | sobs ) by θε∗ . As n → ∞, if
−3/5
εn = o(an ),
ˆ
∗
∗
sup Πε {an (θ − θε ) ∈ A | sobs } −
N {t; 0, I(θ0 )−1 } dt → 0,
A∈B p
A
in probability, and an (θε∗ − θ0 ) → N {0, I(θ0 )−1 } in distribution. Moreover, if βε is replaced by
βeε satisfying an εn (βeε − βε ) = op (1), the above results still hold.
The limit of the regression adjusted posterior distribution is the true posterior given the summary
−3/5
provided εn is o(an ). This is a slower rate than that at which the posterior contracts, which,
8
as we will show in the next section, has important consequences in terms of the computational
efficiency of approximate Bayesian computation. The regression adjustment corrects both the
additional noise of the posterior mean when d > p and the overestimated uncertainty of the
posterior. This correction comes from the removal of the first order bias caused by ε. Blum (2010)
shows that the regression adjustment reduces the bias of approximate Bayesian computation
when E(θ | s) is linear and the residuals θ − E(θ | s) are homoscedastic. Our results do not
require these assumptions, and suggest that the regression adjustment should be applied routinely
with approximate Bayesian computation provided the coefficients βε can be estimated accurately.
With the simulated sample, βε is estimated by βbε . The accuracy of βbε can be seen by the
following decomposition,
βbε = covN (s, θ)varN (s)−1
s − sε
s − sε −1
1
∗
∗
= βε +
covN
, an (θ − θε ) varN
,
an εn
εn
εn
where covN and varN are the sample covariance and variance matrices, and s is the sample
mean. Since cov(s, θ∗ ) = 0 and the distributions of s − sε and θ∗ − θε∗ contract at rates εn
−1 −1/2 }
b
and a−1
n respectively, the error βε − βε can be shown to have the rate Op {(an εn ) N
as n → ∞ and N → ∞. We omit the proof, since it is tedious and similar to the proof of the
asymptotic expansion of βε in Lemma 4. Thus, if N increases to infinity with n, βbε − βε will be
op {(an εn )−1 } and the convergence of Theorem 1 will hold instead.
Alternatively we can get an idea of the additional error for large N from the following proposition.
P ROPOSITION 3. Assume Conditions 1–6. Consider θ∗ = θ − βbε (s − sobs ). As n → ∞, if
−3/5
εn = o(an ) and N is large enough, for any A ∈ B p ,
ˆ
∗
∗
Πε {an (θ − θε ) ∈ A | sobs } →
ψ(t) dt,
A
in distribution, where
ˆ
h
N t;
ψ(t) ∝
Rd
when cε < ∞,
ˆ
ψ(t) ∝
n
N t;
Rp
i
η
−1
{v
−
E
(v)},
I(θ
)
G(v) dv,
0
G
N 1/2
o
η
0
−1
Ds(θ
)t
,
I(θ
)
K{Ds(θ0 )t0 } dt0 ,
0
0
N 1/2
when cε = ∞, and η = Op (1).
The limiting distribution here can be viewed as the convolution of the limiting distribution obtained when the optimal coefficients are used and that of a random variable, which relates to the
error in our estimate of βε , and that is Op (N −1/2 ).
3·3. Acceptance Rates when ε is Negligible
Finally we present results for the acceptance probability of approximate Bayesian computation, the quantity that is central to the computational cost of importance sampling or Markov
chain Monte Carlo-based algorithms. We consider a set-up where we propose the parameter
value from a location-scale family. That is, we can write the proposal density as the density of a
random variable, σn X + µn , where X ∼ q(·), E(X) = 0 and σn and µn are constants that can
9
depend on n. The average acceptance probability pacc,q would then be
ˆ
qn (θ)fn (s | θ)K{ε−1
n (s − sobs )} dsdθ,
P×Rd
where qn (θ) is the density of σn X + µn . This covers the proposal distribution in fundamental
sampling algorithms, including the random-walk Metropolis algorithm and importance sampling
with unimodal proposal distribution, and serves as the building block for many advanced algorithms where the proposal distribution is a mixture of distributions from location-scale families,
such as iterative importance sampling-type algorithms.
We further assume that σn (µn − θ0 ) = Op (1), which means θ0 is in the coverage of qn (θ).
This is a natural requirement for any good proposal distribution. The prior distribution and θ0 as a
point mass are included in this proposal family. This condition would also apply to many Markov
chain Monte Carlo implementations of approximate Bayesian computation after convergence.
−1
As above, define an,ε = an 1cε <∞ + ε−1
n 1cε =∞ to be the smaller of an and εn . Asymptotic
−1
results for pacc,q when σn has the same rate as an,ε are given in Li & Fearnhead (2018). Here we
extend those results to other regimes.
−1/2
T HEOREM 2. Assume the conditions of Proposition 1. As n → ∞, if εn = o(an ): (i) if
−1
cε = 0 or σn /a−1
n,ε → ∞, then pacc,q → 0 in probability; (ii) if cε ∈ (0, ∞) and σn /an,ε →
−1
r1 ∈ [0, ∞), or cε = ∞ and σn /an,ε → r1 ∈ (0, ∞), then pacc,q = Θp (1); (iii) if cε = ∞ and
σn /a−1
n,ε → 0, then pacc,q → 1 in probability.
The proof of Theorem 2 can be found in the Supplementary Material. The underlying intuition
is as follows. For the summary statistic, s, sampled with parameter value θ, the acceptance probability depends on
s − sobs
1
= [{s − s(θ)} + {s(θ) − s(θ0 )} + {s(θ0 ) − sobs }],
εn
εn
(3)
where s(θ) is the limit of s in Condition 3. The distance between s and sobs is at least Op (a−1
n ),
−1
−1
since the first and third bracketed terms are Op (an ). If εn = o(an ) then, regardless of the value
of θ, (3) will blow up as n → ∞ and hence pacc,q goes to 0. If εn decreases with a rate slower
than a−1
n , (3) will go to zero providing we have a proposal which ensures that the middle term is
op (εn ), and hence pacc,q goes to unity.
Theorem 1 shows that, without the regression adjustment, approximate Bayesian computation
requires εn to be o(a−1
n ) if its posterior is to converge to the true posterior given the summary. In
this case Theorem 2 shows that the acceptance rate will degenerate to zero as n → ∞ regardless
of the choice of q(·). On the other hand, with the regression adjustment, we can choose εn =
−3/5
o(an ) and still have convergence to the true posterior given the summary. For such a choice,
if our proposal density satisfies σn = o(εn ), the acceptance rate will go to unity as n → ∞.
4. N UMERICAL E XAMPLE
Here we illustrate the gain of computational efficiency from using the regression adjustment on
the g-and-k distribution, a popular model for testing approximate Bayesian computation methods
(e.g., Fearnhead & Prangle, 2012; Marin et al., 2014). The data are independent and identically
distributed from a distribution defined by its quantile function,
1 − exp{−γz(x)}
−1
F (x; α, β, γ, κ) =α + β 1 + 0.8
{1 + z(x)2 }κ z(x), x ∈ [0, 1],
1 + exp{−γz(x)}
10
where α and β are location and scale parameters, γ and κ are related to the skewness and kurtosis
of the distribution, and z(x) is the corresponding quantile of a standard normal distribution. No
closed form is available for the density but simulating from the model is straightforward by
transforming realisations from the standard normal distribution.
In the following we assume the parameter vector (α, β, γ, κ) has a uniform prior in [0, 10]4
and multiple datasets are generated from the model with (α, β, γ, κ) = (3, 1, 2, 0.5). To illustrate
the asymptotic behaviour of approximate Bayesian computation, 50 data sets are generated for
each of a set of values of n ranging from 500 to 10, 000. Consider estimating the posterior
means, denoted by µ = (µ1 , . . . , µ4 ), and standard deviations, denoted by σ = (σ1 , . . . , σ4 ), of
the parameters. The summary statistic is a set of evenly spaced quantiles of dimension 19.
The bandwidth is chosen via fixing the proportion of the Monte Carlo sample to be accepted,
and the accepted proportions needed to achieve certain approximation accuracy for estimates
with and without the adjustment are compared. A higher proportion means more simulated parameter values can be kept for inference. The accuracy is measured by the average relative errors
of estimating µ or σ,
4
REµ =
µk − µ k |
1 X |b
,
4
µk
k=1
4
REσ =
1 X |b
σk − σk |
,
4
σk
k=1
for estimators µ
b = (b
µ1 , . . . , µ
b4 ) and σ
b = (b
σ1 , . . . , σ
b4 ). The proposal distribution is normal, with
the covariance matrix selected to inflate the posterior covariance matrix by a constant factor c2
and the mean vector selected to differ from the posterior mean by half of the posterior standard
deviation, which avoids the case that the posterior mean can be estimated trivially. We consider a
series of increasing c in order to investigate the impact of the proposal distribution getting worse.
Figure 1 shows that the required acceptance rate for the regression adjusted estimates is higher
than that for the unadjusted estimates in almost all cases. For estimating the posterior mean, the
improvement is small. For estimating the posterior standard deviations, the improvement is much
larger. To achieve each level of accuracy, the acceptance rates of the unadjusted estimates are
all close to zero. Those of the regression-adjusted estimates are higher by up to two orders of
magnitude, so the Monte Carlo sample size needed to achieve the same accuracy can be reduced
correspondingly.
5. D ISCUSSION
One way to implement approximate Bayesian computation so that the acceptance probability
tends to unity as n increases is to use importance sampling with a suitable proposal from a
location-scale family. The key difficulty with finding a suitable proposal is to ensure that the
location parameter is close to the true parameter, where close means the distance is O(εn ). This
can be achieved by having a preliminary analysis of the data, and using the point estimate of the
parameter from this preliminary analysis as the location parameter (Beaumont et al., 2009; Li &
Fearnhead, 2018).
ACKNOWLEDGMENT
This work was funded by the Engineering and Physical Sciences Research Council, under the
i-like programme grant.
11
0.8
0.6
0.4
0.0
0.2
Acceptance rate
0.8
0.6
0.4
0.2
0.0
Acceptance rate
1.0
(b)
1.0
(a)
2.0
2.5
3.0
3.5
4.0
4.5
2.0
2.5
3.0
c
4.0
4.5
3.5
4.0
4.5
c
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
Acceptance rate
0.8
1.0
(d)
1.0
(c)
Acceptance rate
3.5
2.0
2.5
3.0
3.5
c
4.0
4.5
2.0
2.5
3.0
c
Fig. 1: Acceptance rates required for different degrees of accuracy of approximate Bayesian
computation and different variances of the proposal distribution (which are proportional to c). In
each plot we show results for standard (grey-line) and regression adjusted (black-line) approximate Bayesian computation and for different values of n: n = 500 (dotted), n = 3, 000 (dashed)
and n = 10, 000 (solid). The averages over 50 data sets (thick) and their 95% confidence intervals (thin) are reported. Results are for a relative error of 0.08 and 0.05 in the posterior mean, in
(a) and (b) respectively, and for a relative error of 0.2 and 0.1 in the posterior standard deviation,
in (c) and (d) respectively.
S UPPLEMENTARY M ATERIAL
Proofs of lemmas and Theorem 2 are included in the online supplementary material.
12
A PPENDIX
Proof of Result from Section 3·1
Throughout the data are considered to be random. For any integer l > 0 and a set A ⊂ Rl ,
we use the convention that cA + x denotes the set {ct + x : t ∈ A} for c ∈ R and x ´∈ Rl . For a
non-negative function h(x), integrable in Rl , denote the normalised function h(x)/ Rl h(x) dx
by h(x)(norm) . For a vector x, denote a general polynomial of elements of x with degree up to l by
Pl (x). For any fixed δ < δ0 , let Bδ be the neighborhood {θ : kθ − θ0 k < δ}. Let πδ (θ) be π(θ)
truncated in Bδ . Recall that fen (s | θ) denotes the normal density with mean s(θ) and covariance
matrix A(θ)/a2n . Let t(θ) = an,ε (θ − θ0 ) and v(s) = ε−1
n (s − sobs ), rescaled versions θ and s.
For any A ∈ B p , let t(A) be the set {φ : φ = t(θ) for some θ ∈ A}.
e ε (θ ∈ A | sobs ) to be the normal counterpart of Πε (θ ∈ A | sobs ) with truncated prior,
Define Π
obtained by replacing π(θ) and fn (s | θ) in Πε by πδ (θ) and fen (s | θ). So let
π
eε (θ, s | sobs ) = ´
´
Bδ
πδ (θ)fen (s | θ)K{ε−1
n (s − sobs )}
,
−1
e
d πδ (θ)fn (s | θ)K{εn (s − sobs )} dθds
R
´
e ε (θ ∈ A | sobs ) be the distribution function with denπ
eε (θ | sobs ) = Rd π
eε (θ, s | sobs ) ds and Π
e ε by θeε . Let Wobs = an A(θ0 )−1/2 {sobs − s(θ0 )} and
sity π
eε (θ | sobs ). Denote the mean of Π
−1
T
−1
β0 = I(θ0 ) Ds(θ0 ) A(θ0 ) . By Condition 3, Wobs → Z in distribution as n → ∞, where
Z ∼ N (0, Id ).
e ε is an incorrect model for
Since the approximate Bayesian computation likelihood within Π
sobs , standard posterior convergence results do not apply. However, if we condition on the value
of the summary, s, then the distribution of θ is just the true posterior given s. Thus we can express
the posterior from approximate Bayesian computation as a continuous mixture of these true
−1
−1 e
posteriors. Let π
eε,tv (t, v) = a−d
n,ε πδ (θ0 + an,ε t)fn (sobs + εn v | θ0 + an,ε t)K(v). For any A ∈
e ε as,
B p , we rewrite Π
ˆ ˆ
e ε (θ ∈ A | sobs ) =
e ∈ A | sobs + εn v)e
Π
Π(θ
πε,tv (t, v)(norm) dtdv,
(4)
Rd
t(Bδ )
e ∈ A | s) is the posterior distribution with prior πδ (θ) and likelihood fen (s | θ).
where Π(θ
e ∈ A | sobs + εn v)
Using results from Kleijn & van der Vaart (2012), the leading term of Π(θ
can be obtained and is stated in the following lemma.
d
L EMMA 1. Assume Conditions 3 and 4. If εn = O(a−1
n ), for any fixed v ∈ R and small
enough δ,
ˆ
e
N [t; β0 {A(θ0 )1/2 Wobs + cε v}, I(θ0 )−1 ] dt → 0,
sup Π{an (θ − θ0 ) ∈ A | sobs + εn v} −
A∈B p
A
in probability as n → ∞.
This leading term is Gaussian, but with a mean that depends on v. Thus asymptotically, the
posterior of approximate Bayesian computation is the distribution
´ of the sum of a Gaussian
random variable and β0 cε V , where V has density proportional to π
eε,tv (t, v) dt.
To make this argument rigorous, and to find the distribution of this sum of random variables we
need to introduce several functions that relate to the limit of π
eε,tv (t0 , v). For a rank-p d × p matrix
A, a rank-d d × d matrix B and a d-dimensional vector c, define g(v; A, B, c) = exp[−(c +
13
Bv)T {I − A(AT A)−1 AT }(c + Bv)/2]/(2π)(d−p)/2 . Let
(
N nDs(θ0 )t; an εn v + A(θ0 )1/2 Wobs , A(θ0 ) K(v),
o
gn (t, v) =
N Ds(θ0 )t; v + an1εn A(θ0 )1/2 Wobs , a21ε2 A(θ0 ) K(v),
n n
cε < ∞,
cε = ∞,
Gn (v) be g{v; A(θ0 )−1/2 Ds(θ0 ), an εn A(θ0 )−1/2 , Wobs }K(v), and EGn (·) be´ the expectation
under the density Gn (v)(norm) . In both cases it is straightforward to show that Rp gn (t, v) dt =
|A(θ0 )|−1/2 Gn (v). Additionally, for the case cε = ∞, define v 0 (v, t) = A(θ0 )1/2 Wobs +
an εn v − an εn Ds(θ0 )t and
1 0
1
gn0 (t, v 0 ) = N {v 0 ; 0, A(θ0 )}K Ds(θ0 )t +
v −
A(θ0 )1/2 Wobs .
an εn
an εn
Then with the transformation v 0 = v 0 (v, t), gn (t, v)dv = gn0 (t, v 0 )dv 0 . Let
(
N {Ds(θ0 )t; cε v + A(θ0 )1/2 Z, A(θ0 )}K(v), cε < ∞,
g(t, v) =
K{Ds(θ0 )t}N {v; 0, A(θ0 )},
cε = ∞,
−1/2 , Z}K(v) and E (·) be the expectation under the
G(v) be g{v; A(θ0 )−1/2 Ds(θ0 ), cε A(θ
G
´ 0)
(norm)
−1/2 G(v).
density G(v)
.
When
c
<
∞,
g(t,
v)
dt
=
|A(θ
)|
ε
0
Rp
´ ´
Expansions of Rd t(Bδ ) π
eε,tv (t, v)dtdv are given in the following lemma.
−1/2
Conditions
1–3.
If
εn´ = ´o(an ),
then
´ L´EMMA 2. Assume
πε,tv (t, v) − π(θ0 )gn (t, v)| dtdv → 0 in probability and Rd t(Bδ ) gn (t, v) dtdv =
Rd t(Bδ ) |e
´ ´
Θp (1), as n → ∞. Furthermore, for l ≤ 6, Rd t(Bδ ) Pl (v)gn (t, v) dtdv converges
´
−1/2
to
in distribution when cε < ∞ and converges to
Rd Pl (v)G(v) dv
´ |A(θ0 )|
P
{Ds(θ
)t}K{Ds(θ
)t}
dt
in
probability
when cε = ∞, as n → ∞.
p
0
0
l
R
e ε are asymptotically the same and gives an expansion
The following lemma states that Πε and Π
e
of θε .
−1/2
L EMMA 3. Assume Conditions 1–4. If εn = o(an ), then
e ε (θ ∈ B c | sobs ) are op (1);
(a) for any δ < δ0 , Πε (θ ∈ Bδc | sobs ) and Π
δ
e ε (θ ∈ A ∩ Bδ | sobs ) =
(b) there exists a δ < δ0 such that supA∈Bp Πε (θ ∈ A ∩ Bδ | sobs ) − Π
op (1);
(c) if, in addition, Condition 5 holds, then an,ε (θε − θeε ) = op (1), and θeε = θ0 +
1/2 W
−1
a−1
obs + εn β0 EGn (v) + rn,1 where the remainder rn,1 = op (an ).
n β0 A(θ0 )
Proof of Proposition 1. Lemma 10 in the Supplementary Material shows that Πε {an,ε (θ −
e ε {an,ε (θ − θeε ) ∈ A | sobs } have the same limit, in distribution when cε ∈
θε ) ∈ A | sobs } and Π
(0, ∞) and in total variation form when cε = 0 or ∞. Therefore it is sufficient to only consider
e ε of the properly scaled and centered θ.
the convergence of Π
e ε {an (θ − θeε ) ∈ A | sobs } equals
When an εn → cε < ∞, according to (4), Π
ˆ ˆ
e n (θ − θ0 ) ∈ A + an (θeε − θ0 ) | sobs + εn v}e
Π{a
πε,tv (t0 , v)(norm) dt0 dv.
Rd
t(Bδ )
14
By Lemma 1 and Lemma 3(c), we have
ˆ
N {t; µn (v), I(θ0 )−1 } dt = op (1),
e n (θ − θ0 ) ∈ A + an (θeε − θ0 ) | sobs + εn v} −
sup Π{a
A∈B p
A
where µn (v) = β0 {cε v − an εn EGn (v)} − an rn,1 . Then with Lemma 2, the leading term of
e ε {an (θ − θeε ) ∈ A | sobs } equals
Π
ˆ ˆ
ˆ
e
e
sup Πε {an (θ − θε ) ∈ A | sobs } −
N {t; µn (v), I(θ0 )−1 }gn (t0 , v)(norm) dtdt0 dv = op (1).
A∈B p
Rd
t(Bδ )
A
(5)
The numerator of the leading term of (5) is in the form
ˆ ˆ
ˆ
N {t; cε β0 v + x3 , I(θ0 )−1 }N {Ds(θ0 )t0 ; x1 v + x2 , A(θ0 )}K(v) dtdt0 dv,
Rd
t(Bδ )
A
where x1 ∈ R, x2 ∈ Rd and x3 ∈ Rp . This is continuous by Lemma 9 in the Supplementary
Material. Then since EGn (v) → EG (v) in distribution as n → ∞ by Lemma 2, we have
ˆ ˆ
ˆ
N {t; µn (v), I(θ0 )−1 }gn (t0 , v) dtdt0 dv
Rd t(Bδ ) A
ˆ ˆ ˆ
→
N [t; cε β0 {v − EG (v)}, I(θ0 )−1 ]g(t0 , v) dtdt0 dv,
Rd
Rp
A
in distribution as n → ∞. Putting the above results together, it holds that
ˆ ˆ
e ε {an (θ − θeε ) ∈ A | sobs } →
Π
N [t; cε β0 {v − EG (v)}, I(θ0 )−1 ]G(v)(norm) dvdt,
A
Rp
in distribution, and statement (ii) of the proposition holds.
When cε = 0, since µn (v) does not depend on v, (5) becomes
ˆ
e
e
N {t; −an εn β0 EGn (v) − an rn,1 , I(θ0 )−1 } dt = op (1),
sup Πε {an (θ − θε ) ∈ A | sobs } −
A∈B p
A
and by the continuous mapping theorem (van der Vaart, 2000),
ˆ
N {t; −an εn β0 EGn (v) − an rn,1 , I(θ0 )−1 } − N {t; 0, I(θ0 )−1 } dt = op (1).
Rp
´
e ε {an (θ − θeε ) ∈ A | sobs } − N {t; 0, I(θ0 )−1 } dt = op (1), and stateTherefore supA∈Bp Π
A
ment (i) of the proposition holds.
When cε = ∞, Lemma 1 cannot be applied to the posterior distribution within (4) directly.
e ε {θ ∈ A | sobs } equals,
With transformation v 00 = an εn v, Π
ˆ ˆ
−1 00 (norm) 0 00
e ∈ A | sobs + a−1 v 00 )e
Π(θ
πε,tv (t0 , a−1
dt dv ,
n
n εn v )
Rd
t(Bδ )
e ε {ε−1 (θ − θeε ) ∈ A | sobs } equals
which implies that Π
n
ˆ ˆ
−1 00 (norm)
e n (θ − θ0 ) ∈ an εn A + an (θeε − θ0 ) | sobs + a−1 v 00 }e
Π{a
πε,tv (t0 , a−1
dt0 dv 00 .
n
n εn v )
Rd
t(Bδ )
15
Then Lemma 1 can be applied. Using Lemma 2 and transforming v 00 back to v we have
e ε {ε−1 (θ − θeε ) ∈ A | sobs }
sup Π
n
A∈B p
ˆ
ˆ
ˆ
(an εn )p N {an εn t; µ0n (v), I(θ0 )−1 }gn (t0 , v)(norm) dtdt0 dv = op (1),
−
Rd
t(Bδ )
(6)
A
where µ0n (v) = β0 {A(θ0 )1/2 Wobs + an εn v} − an (θeε − θ0 ).
Let t00 (t, t0 ) = an εn (t − t0 ) + an (θeε − θ0 ) and t00 (A, Bδ ) be the set {t00 (t, t0 ) : t ∈ A, t0 ∈
t(Bδ )}. With transformations v 0 = v 0 (v, t0 ) and t00 = t00 (t, t0 ), since β0 Ds(θ0 ) = Ip , we have
ˆ ˆ
ˆ
(an εn )p N {an εn t; µ0n (v), I(θ0 )−1 }gn (t0 , v) dtdt0 dv
d
R
t(B ) A
ˆ ˆδ
ˆ
(an εn )p N {an εn (t − t0 ); β0 v 0 − an (θeε − θ0 ), I(θ0 )−1 }gn0 (t0 , v 0 ) dtdt0 dv 0
=
Rd t(Bδ ) A
ˆ
ˆ ˆ
1
00
0
−1 0
00
0
e
e
=
N {t ; β0 v − an (θε − θ0 ), I(θ0 ) }gn t −
{t − an (θε − θ0 )}, v dtdt00 dv 0 .
a n εn
Rd t00 (A,Bδ ) A
The idea now is that as n → ∞, an εn → ∞, so the gn0 term in the integral will tend to gn0 (t, v 0 ).
Then by integrating first with respect to t00 and then with respect to v, we get the required result.
To make this argument rigorous, consider the following function,
ˆ ˆ
ˆ
N {t00 ; β0 v 0 , I(θ0 )−1 }N {v 0 ; 0, A(θ0 )}
Rd
t00 (A,Bδ )
A
× K{Ds(θ0 )t + x1 v 0 − x2 t00 + x3 } − K{Ds(θ0 )t} dtdt00 dv 0 ,
where x1 ∈ R, x2 ∈ R and x3 ∈ Rd . This is continuous by Lemma 9 in the Supplementary Material, so by the continuous mapping theorem,
ˆ
ˆ ˆ
ˆ
1 00 0
00
0
−1 0
00 0
N {t ; β0 v , I(θ0 ) }gn t −
sup
t , v dtdt dv −
K{Ds(θ0 )t} dt = op (1).
an εn
A∈B p
A
Rd t00 (A,Bδ ) A
Then using Lemma 2,
ˆ
e ε {ε−1 (θ − θeε ) ∈ A | sobs } −
sup Π
n
A∈B p
ˆ
K{Ds(θ0 )t} dt/
A
K{Ds(θ0 )t} dt = op (1).
Rp
Therefore statement (iii) of the proposition holds.
Proof of Result from Section 3·2
The intuition behind this result is that, as shown above, the joint posterior of (θ, s) under
approximate Bayesian computation can be viewed as a marginal distribution for s times a conditional for θ given s. The latter is just the true posterior for θ given s, and this posterior converges
to a Gaussian limit that depends on s only through its mean. Regression adjustment works because it corrects for the dependence of this mean on s, so if we work with the regression adjusted
parameter θ∗ then the conditional distribution for θ∗ given s will tend to a Gaussian limit whose
mean is the same for all s.
The following lemma gives an expansion of the optimal linear coefficient matrix βε . The order
of the remainder leads to successful removal of the first order bias in the posterior distribution of
approximate Bayesian computation.
16
−3/5
L EMMA 4. Assume Conditions 1–6. Then if εn = o(an
), an εn (βε − β0 ) = op (1).
The following lemma, similar to Lemma 3, says that the approximate Bayesian computation
e ε . Recall that θ∗ is the mean of π ∗ (θ∗ |
posterior distribution of θ∗ is asymptotically the same as Π
ε
ε
´
∗
∗
∗
sobs ). Let π
eε (θ | sobs ) = Rd π
eε {θ + βε (s − sobs ), s | sobs } ds and θeε∗ be the mean of π
eε∗ (θ∗ |
sobs ).
−3/5
L EMMA 5. Assume Conditions 1–6. If εn = o(an ), then
e ε (θ∗ ∈ B c | sobs ) are op (1);
(a) for any δ < δ0 , Πε (θ∗ ∈ Bδc | sobs ) and Π
δ
∗
e ε (θ∗ ∈ A ∩ Bδ | sobs ) =
(b) there exists δ < δ0 such that supA∈Bp Πε (θ ∈ A ∩ Bδ | sobs ) − Π
op (1);
1/2 W
(c) an (θε∗ − θeε∗ ) = op (1), and θeε∗ = θ0 + a−1
obs + εn (β0 − βε )EGn (v) + rn,2
n β0 A(θ0 )
−1
where the remainder rn,2 = op (an ).
Proof of Theorem 1. Similar to the proof of Proposition 1, it is sufficient to only consider the
e ε of the properly scaled and centered θ∗ . Similar to (4),
convergence of Π
(´ ´
e ∈ A + εn βε v | sobs + εn v)e
Π(θ
πε,tv (t0 , v)(norm) dt0 dv,
e ε (θ∗ ∈ A | sobs ) = ´Rd ´t(Bδ )
Π
−1 π
0 −1 −1 (norm) dt0 dv,
e
ε,tv (t , an εn v)
Rd t(Bδ ) Π(θ ∈ A + εn βε v | sobs + an v)e
cε < ∞,
cε = ∞.
Similar to (5), by Lemma 1 and Lemma 5(c), we have
ˆ
e ε {an (θ∗ − θeε∗ ) ∈ A | sobs } −
sup Π
A∈B p
Rd
ˆ
ˆ
N {t; µ∗n (v), I(θ0 )−1 }gn (t0 , v)(norm) dtdt0 dv = op (1),
t(Bδ )
A
(7)
where
µ∗n (v) = {an εn (β0 − βε ) + (cε − an εn )β0 1cε <∞ }{v − EGn (v)} + an rn,2 .
Since
an εn (βε − β0 ) = op (1), by Lemma 9 in the Supplementary Material and the continuous
mapping theorem,
ˆ ˆ
ˆ
N {t; µ∗n (v), I(θ0 )−1 } − N {t; 0, I(θ0 )−1 } gn (t0 , v) dtdt0 dv = op (1).
Rd
t(Bδ )
Rp
Then we have
ˆ
e ε {an (θ∗ − θe∗ ) ∈ A | sobs } −
sup Π
ε
A∈B p
N {t; 0, I(θ0 )−1 } dt = op (1),
A
and the first convergence in the theorem holds.
The second convergence in the theorem holds by Lemma 5(c).
0
Since the only requirement for βε in the above is an εn (βε − βε ) = op (1), the above arguments
will hold if βε is replaced by a p × d matrix βbε satisfying an εn (βbε − βε ) = op (1).
Proof of Proposition 3. Consider θ∗ where βε is replaced by βbε . Let ηn = N 1/2 an εn (βbε −
βε ). Since βbε − βε = Op {(an εn )−1 N −1/2 } as n → ∞, ηn = Op (1) as n → ∞ and let its limit
be η. In this case, if we replace µ∗n (v) in (7) with µ∗n (v) + N −1/2 ηn {v − EGn (v)}, denoted by
µ
b∗n (v), the equation still holds. Denote this equation by (70 ). Limits of the leading term in (70 )
can be obtained by arguments similar as those for (5).
17
When cε < ∞, since for fixed v, µ
b∗n (v) converges to N −1/2 η{v − EG (v)} in distribution, by
following the same line we have
ˆ ˆ
∗
∗
e
e
N [t; N −1/2 η{v − EG (v)}, I(θ0 )−1 ]G(v)(norm) dvdt,
Πε {an (θ − θε ) ∈ A | sobs } →
A
Rp
in distribution as n → ∞.
cε = ∞, by ´Lemma 2 we have EGn (v) → 0 in probability, and
´ When
´
0
0
g
Rd t(Bδ ) n (t , v) dt dv → Rp K{Ds(θ0 )t} dt in probability. Then with transformation
v 0 = v 0 (v, t0 ), for fixed v,
1
1 0
1/2
∗
−1/2
0
v −
A(θ0 ) Wobs − EGn (v)
µ
bn (v) = {an εn (β0 − βε ) + N
ηn } Ds(θ0 )t +
an εn
an εn
→ N −1/2 ηDs(θ0 )t0 ,
in distribution as n → ∞. Recall that gn (t0 , v) dv = gn0 (t0 , v 0 ) dv 0 . Then by Lemma 9 in the Supplementary Material and the continuous mapping theorem,
ˆ ˆ
ˆ
N {t; µ
b∗n (v), I(θ0 )−1 }gn0 (t0 , v 0 ) dtdt0 dv
d
R
t(B ) A
ˆ ˆ δ ˆ
N {t; N −1/2 ηDs(θ0 )t0 , I(θ0 )−1 }K{Ds(θ0 )t0 } dtdt0 dv,
→
Rd
t(Bδ )
A
in distribution as n → ∞. Therefore by (70 ) and the above convergence results,
ˆ ˆ
e ε {an (θ∗ − θeε∗ ) ∈ A | sobs } →
Π
N {t; N −1/2 ηDs(θ0 )t0 , I(θ0 )−1 }K{Ds(θ0 )t0 }(norm) dt0 dt,
A
in distribution as n → ∞.
Rp
S UPPLEMENTARY M ATERIAL
Notations and Set-up
First
and conventions´ are given.
For two sets A and B, the sum of inte´ some limit notations
´
´
grals A f (x) dx + B f (x) dx is written as ( A + B )f (x) dx. For a constant d × p matrix A, let
the minimum and maximum eigenvalues of AT A be λ2min (A) and λ2max (A) where λmin (A) and
λmax (A) are non-negative. Obviously, for any p-dimension vector x, λmin (A)kxk ≤ kAxk ≤
λmax (A)kxk. For two matrices A and B, we say A is bounded by B or A ≤ B if λmax (A) ≤
λmin (B). For a set of matrices {Ai : i ∈ I} for some index set I, we say it is bounded if
λmax (Ai ) are uniformly bounded in i. Denote the identity matrix with dimension d by Id . Notations from the main text will also be used.
The following basic asymptotic results (Serfling, 2009) will be used throughout.
L EMMA 6. (i) For a series of random variables Zn , if Zn → Z in distribution as n → ∞,
Zn = Op (1). (ii) (Continuous mapping) For a series of continuous function gn (x), if gn (x) =
O(1) almost everywhere, then gn (Zn ) = Op (1), and this also holds if O(1) and Op (1) are replaced by Θ(1) and Θp (1).
Some notations regarding the posterior distribution of approximate Bayesian computation are
given. For A ⊂ Rp and a scalar function h(θ, s), let
ˆ ˆ
−d
πA (h) =
h(θ, s)π(θ)fn (s | θ)K{ε−1
n (s − sobs )}εn dsdθ,
A
Rd
18
and
ˆ ˆ
π
eA (h) =
A
Rd
−d
h(θ, s)πδ (θ)fen (s | θ)K{ε−1
n (s − sobs )}εn dsdθ.
e ε (θ ∈ A | sobs ) =
Then Πε (θ ∈ A | sobs ) = πA (1)/πP (1) and its normal counterpart Π
π
eA (1)/e
πP (1).
The following results from Li & Fearnhead (2018) will be used throughout.
L EMMA 7. Assume Conditions 1–4. Then as n → ∞,
(i) if Condition 5 also holds then, for any δ < δ0 , πBδc (1) and π
eBδc (1) are op (1), and
αδ
Op (e−an,ε cδ ) for some positive constants cδ and αδ depending on δ;
eA (1)| /e
πBδ (1) = Op (αn−1 );
eBδ (1){1 + Op (αn−1 )} and supA⊂Bδ |πA (1) − π
(ii) πBδ (1) = π
−1/2
eP (1) and πP (1) are
(iii) if εn = o(an ), π
eBδ (1) and πBδ (1) are Θp (ad−p
n,ε ), and thus π
d−p
Θp (an,ε );
−1/2
−3/5
(iv) if εn = o(an ) and Condition 5 holds, θε = θeε + op (a−1
), θε = θeε +
n,ε ). If εn = o(an
op (a−1
n ).
Proof. (i) is from Li & Fearnhead (2018, Lemma 3) and a trivial modification of its proof
when Condition 5 does no hold; (ii) is from Li & Fearnhead (2018, equation 13 of supplements);
(iii) is from Li & Fearnhead (2018, Lemma 5 and equation 13 of supplements); and (iv) is from
Li & Fearnhead (2018, Lemma 3 and Lemma 6).
Proof for Results in Section 3·1
e ∈ A | sobs + εn v) is the posterior
Proof of Lemma 1. For any fixed v ∈ Rd , recall that Π(θ
distribution given sobs + εn v with prior πδ (θ) and the misspecified model fen (· | θ). By Kleijn &
van der Vaart (2012), if there exist ∆n,θ0 and Vθ0 such that,
(KV1) for any compact set K ⊂ t(Bδ ),
sup log
t∈K
fen (sobs + εn v | θ0 + a−1
1
n t)
− tT Vθ0 ∆n,θ0 + tT Vθ0 t → 0,
2
fen (sobs + εn v | θ0 )
in probability as n → ∞, and
e n kθ − θ0 k > Mn | sobs + εn v)} → 0 as n → ∞ for any sequence of constants
(KV2) E{Π(a
Mn → ∞,
then
ˆ
e n (θ − θ0 ) ∈ A | sobs + εn v} −
sup Π{a
A∈B p
A
N (t; ∆n,θ0 , Vθ−1
) dt → 0,
0
in probability as n → ∞.
For (KV1), by the definition of fen (s | θ),
log
−2
−1
fen (sobs + εn v | θ0 + a−1
N {sobs + εn v; s(θ0 + a−1
n t)
n t), an A(θ0 + an t)}
= log
.
N {sobs + εn v; s(θ0 ), a−2
fen (sobs + εn v | θ0 )
n A(θ0 )}
As xT Ax − y T By = xT (A − B)x + (x − y)T B(x + y), for vectors x and y and matrices A
and B, by applying a Taylor expansion on s(θ0 + xt) and A(θ0 + xt) around x = 0, the right
19
hand side of above equation equals
a−1
n
T
−1
{Ds(θ0 + e(1)
t)t}
A(θ
)
ζ
(v,
t)
−
ζn (v, t)T
0
n
n
2
oT
a−1 n
+ n
D log A(θ0 + e(3)
t,
n t)
2
( p
X
)
Dθi A−1 (θ0 + e(2)
n t)ti
ζn (v, t)
i=1
(1)
(j)
where ζn (v, t) = A(θ0 )1/2 Wobs + an εn v − 12 Ds(θ0 + en t)t and for j = 1, 2, 3, en is a func(j)
tion of t satisfying |en | ≤ a−1
n which is from the remainder of the Taylor expansions. Since
Ds(θ), DA−1 (θ) and D log |A(θ)| are bounded in Bδ when δ is small enough,
sup log
t∈K
fen (sobs + εn v | θ0 + a−1
1
n t)
− tT I(θ0 )β0 {A(θ0 )1/2 Wobs + cε v} + tT I(θ0 )t → 0,
2
fen (sobs + εn v | θ0 )
in probability as n → ∞, for any compact set K. Therefore (KV1) holds with ∆n,θ0 =
β0 {A(θ0 )1/2 Wobs + cε v} and Vθ0 = I(θ0 ).
For (KV2), let rn (s | θ0 ) = αn {fn (s | θ0 ) − fen (s | θ0 )}. Since rn (s | θ0 ) is bounded by a
function integrable in Rd by Condition 4,
ˆ
e
e n kθ − θ0 k > Mn | s + εn v)fen (s | θ0 ) ds
E{Π(an kθ − θ0 k > Mn | sobs + εn v)} −
Π(a
d
R
ˆ
−1
≤αn
|rn (s | θ0 )| ds = o(1).
Rd
Then it is sufficient for the expectation under fen (s | θ0 ) to be o(1). For any constant M > 0,
with the transformation v̄ = an {s − s(θ0 )},
ˆ
e n kθ − θ0 k > Mn | s + εn v)fen (s | θ0 ) ds
Π(a
d
R
´
ˆ
ˆ
e(t, v̄ | v) dt
ktk>Mn π
´
N {v̄; 0, A(θ0 )} dv̄ +
N {v̄; 0, A(θ0 )} dv̄,
≤
e(t, v̄ | v) dt
kv̄k>M
kv̄k≤M
t(Bδ ) π
−1
−1
e
where π
e(t, v̄ | v) = πδ (θ0 + a−1
n t)fn {s(θ0 ) + an v̄ + εn v | θ0 + an t}. For the first term in the
above upper bound, it is bounded ´by a series which does not depend on M and is o(1) as Mn →
∞, as shown below. Obviously t(Bδ ) π
e(t, v̄ | v) dt can be lower bounded for some constant
mδ > 0. Choose δ small enough such that Ds(θ) and A(θ)1/2 are bounded for θ ∈ Bδ . Let λmin
and λmax be their common bounds. When kv̄k < M and Mn is large enough,
supθ∈Bδ kDs(θ)tk
{t : ktk > Mn } ⊂ t :
≥ kan εn v + v̄k .
(8)
2
Then since for any v̄ satisfying kv̄k < M , by a Taylor expansion,
−1
d
(1)
−1
fen {s(θ0 ) + a−1
n v̄ + εn v | θ0 + an t} = an N {Ds(θ0 + en t)t; v̄ + an εn v, A(θ0 + an t)},
π
e(t, v̄ | v) ≤ cN (λ−1
max λmin ktk/2; 0, 1), where c is some positive constant, for t in the right hand
side of (8). Then
´
ˆ
ˆ
e(t, v̄ | v) dt
ktk>Mn π
−1
´
N (λ−1
N {v̄; 0, A(θ0 )} dv̄ ≤ mδ c
max λmin ktk/2; 0, 1) dt,
π
e
(t,
v̄
|
v)
dt
kv̄k≤M
ktk>Mn
t(Bδ )
20
the right hand side of which is o(1) when Mn → ∞. Meanwhile by letting M → ∞, it can be
seen that the expectation under fen (s | θ0 ) is o(1). Therefore (KV2) holds and the lemma holds.
´
The following lemma is used for equations Rp gn (t, v) dt = |A(θ0 )|−1/2 Gn (v) and
´
−1/2 G(v).
Rp g(t, v) dt = |A(θ0 )|
L EMMA 8. For a rank-p d × p matrix A, a rank-d d × d matrix B and a d-dimension vector
c,
N (At; Bv + c, Id ) = N t; (AT A)−1 AT (c + Bv), (AT A)−1 g(v; A, B, c),
(9)
where P = AT A, and
g(v; A, B, c) =
1
(2π)(d−p)/2
1
T
T
−1 T
exp − (c + Bv) (I − A(A A) A )(c + Bv) .
2
Proof. This can be verified easily by matrix algebra.
The following lemma regarding the continuity of a certain form of integral will be helpful
when applying the continuous mapping theorem.
L EMMA 9. Let l1 , l10 , l2 , l20 and l3 be positive integers satisfying l10 ≤ l1 and l20 ≤ l2 . Let A and
B be l1 × l10 and l2 × l20 matrices, respectively, satisfying that AT A and B T B are positive definite. Let g1 (·), g2 (·) and g3 (·) be functions in Rl1 , Rl2 and Rl3 , respectively, that are integrable
and continuous almost everywhere. Assume:
(i) gj (·) is bounded in Rlj for j = 1, 2;
(ii) gj (w) depends on w only through kwk and is a decreasing function of kwk, for j = 1, 2;
and
´ Ql10 +l20 +l
(iii) there exists a non-negative integer l such that Rl3 k=1
wik g3 (w) dw < ∞ for any
coordinates (wi1 , . . . , wil0 +l0 +l ) of w.
1 2
Then the function,
ˆ ˆ ˆ
Pl (w1 , w2 , w3 ) |g1 (Aw1 + x1 w2 + x2 w3 + x3 ) − g1 (Aw1 )| g2 (Bw2 + x4 w3 + x5 )g3 (w3 ) dw3 dw2 dw1 ,
0
where x1 ∈ Rl1 ×l2 , x2 ∈ Rl1 ×l3 , x4 ∈ Rl2 ×l3 , x3 ∈ Rl1 and x5 ∈ Rl2 , is continuous almost everywhere.
Proof. Let mA and mB be the lower bound of A and B respectively. For any
0
(x01, . . . , x05 ) ∈ Rl1 ×l2 × Rl1 ×l3 × Rl2 ×l3 × Rl1 × Rl2 such that the integrand in the target
integral is continuous, consider any sequence (xn1 , . . . , xn5 ) converging to (x01, . . . , x05 ).
It is sufficient to show the convergence of the target function at (xn1 , . . . , xn5 ). Let
VA = {w1 : kAw1 k/2 ≥ sup(xn1 ,xn2 ,xn3 ) kxn1 w2 + xn2 w3 + xn3 k}, VB = {w2 : kBw2 k/2 ≥
sup(xn4 ,xn5 ) kxn4 w3 + xn5 k}, UA = {w1 : kw1 k ≤ 4m−1
A (kx01 w2 k + kx02 w3 k + kx03 k)} and
−1
UB = {w2 : kw2 k ≤ 4mB (kx04 w3 k + kx05 k)}. We have VAc ⊂ UA and VBc ⊂ UB . Then according to the following upper bounds and condition (iii),
|g1 (Aw1 + xn1 w2 + xn2 w3 + xn3 ) − g1 (Aw1 )| ≤ g1 (Aw1 + xn1 w2 + xn2 w3 + xn3 ) + g1 (Aw1 ),
g1 (Aw1 + xn1 w2 + xn2 w3 + xn3 ) ≤ ḡ1 (mA kw1 k/2)1{w1 ∈VA } + sup g1 (w)1{w1 ∈UA } ,
w∈Rl1
g2 (Bw2 + x4 w3 + x5 ) ≤ ḡ2 (mB kw2 k/2)1{w2 ∈VB } + sup g2 (w)1{w2 ∈UB } ,
w∈Rl2
21
where g1 (w) = ḡ1 (kwk) and g2 (w) = ḡ2 (kwk), by applying the dominated convergence theorem, the target function at (xn1, . . . , xn5 ) converges to its value at (x01, . . . , x05 ).
Proof of Lemma 2. The first part holds according to Lemma 5 of Li & Fearnhead (2018). For
the second part, when cε = ∞, by the transformation v 0 = v 0 (v, t),
ˆ
ˆ
ˆ
ˆ
Pl Ds(θ0 )t +
Pl (v)gn (t, v) dtdv =
Rd
Rd
t(Bδ )
t(Bδ )
1 0
1
v −
A(θ0 )1/2 Wobs gn0 (t, v 0 ) dtdv 0 .
a n εn
a n εn
By applying Lemma 9 and the continuous
´ ´mapping theorem in Lemma 6 to the right hand
side of the above when cε = ∞, and to Rd t(Bδ ) Pl (v)gn (t, v) dtdv when cε < ∞, and using
´
−1/2 G(v), the lemma holds.
Rp g(t, v) dt = |A(θ0 )|
Proof of Lemma 3. (a), (b) and the first part of (c) hold immediately by Lemma 7. The second
part of (c) is stated in the proof of Theorem 1 of Li & Fearnhead (2018).
L EMMA 10. Assume conditions 1–5.
e ε {an (θ − θeε ) ∈ A | sobs } have the
(i) If cε ∈ (0, ∞) then Πε {an (θ − θε ) ∈ A | sobs } and Π
same limit in distribution.
(ii) If cε = 0 or cε = 0∞ then
e ε {an,ε (θ − θeε ) ∈ A | sobs } = op (1).
supA∈Bp Πε {an,ε (θ − θε ) ∈ A | sobs } − Π
(iii) If Condition 6 holds then
e ε {an (θ∗ − θeε∗ ) ∈ A | sobs } = op (1).
supA∈Bp Πε {an (θ∗ − θε∗ ) ∈ A | sobs } − Π
Proof. Let λn = an,ε (θε − θeε ), and by Lemma 3(c), λn = op (1). When cε ∈ (0, ∞), for any
A ∈ B p , decompose Πε {an (θ − θε ) ∈ A | sobs } into the following three terms,
h
i
e ε {an (θ − θε ) ∈ A | sobs }
Πε {an (θ − θε ) ∈ A | sobs } − Π
h
i
e ε {an (θ − θeε ) ∈ A + λn | sobs } − Π
e ε {an (θ − θeε ) ∈ A | sobs }
+ Π
e ε {an (θ − θeε ) ∈ A | sobs }.
+Π
For (i) to hold, it is sufficient that the first two terms in the above are op (1). The first term is
e ε {an (θ − θeε ) ∈
op (1) by Lemma 3. For the second term to be op (1), given the leading term of Π
A | sobs } stated in the proof of Proposition 1 in the main text, it is sufficient that
ˆ
v∈Rd
ˆ
−
sup
A+λn
N {t; µn (v), I(θ0 )−1 } dt = op (1).
A
´
´
This holds by noting that the left hand side of the above is bounded by ( A+λn − A )c dt for
some constant c and this upper bound is op (1) since λn = op (1). Therefore (i) holds.
22
e ε {an,ε (θ − θeε ) ∈ A | sobs }
When cε = 0 or ∞, supA∈Bp Πε {an,ε (θ − θε ) ∈ A | sobs } − Π
is bounded by
e ε {an,ε (θ − θε ) ∈ A | sobs }
supA∈Bp Πε {an,ε (θ − θε ) ∈ A | sobs } − Π
ˆ
e
e
ψ(t) dt
+supA∈Bp Πε {an,ε (θ − θε ) ∈ A + λn | sobs } −
A+λn
ˆ
e ε {an,ε (θ − θeε ) ∈ A | sobs } −
+supA∈Bp Π
ψ(t) dt
A
ˆ
ˆ
ψ(t) dt .
+supA∈Bp
ψ(t) dt −
(10)
A
A+λn
With similar arguments as before, the first three
´ terms are op (1). For the fourth term, by transforming t to t + λn , it is upper bounded by Rp |ψ(t − λn ) − ψ(t)| dt which is op (1) by the
continuous mapping theorem. Therefore (ii) holds.
For (iii), the left hand side of the equation has the decomposed upper bound similar to (10),
with θ, θε , θeε and ψ(t) replaced by θ∗ , θε∗ , θeε∗ and N {t; 0, I(θ0 )−1 }. Then by Lemma 5, using
the leading term of Πε {an (θ∗ − θε∗ ) ∈ A | sobs } stated in the proof of Theorem 1, and similar
arguments to those used for the fourth term of (10), it can be seen that this upper bound is op (1).
Therefore (iii) holds.
Proof for Results in Section 3·2
To prove Lemmas 4 and 5, some notation regarding the regression adjusted approximate
Bayesian computation posterior, similar to those defined previously, are needed. Consider transp
p
d
formations t´= t(θ)
´ and v = v(s). For A ⊂ R and the scalar function h(t, v) in R × R , let
π
eA,tv (h) = t(A) Rd h(t, v)e
πε,tv (t, v) dvdt.
Proof of Lemma 4. Since βε = covε (θ, s)varε (s)−1 , to evaluate the covariance matrices, we
need to evaluate πRp {(θ − θ0 )k1 (s − sobs )k2 }/πRp (1) for (k1 , k2 ) = (0, 0), (1, 0), (1, 1), (0, 1)
and (0, 2).
First of all, we show that πBδc {(θ − θ0 )k1 (s − sobs )k2 } is ignorable for any δ < δ0 by showing
αδ
that it is Op (e−an,ε cδ ) for some positive constants cδ and αδ . By dividing Rd into {v : kεn vk ≤
δ 0 /3} and its complement,
ˆ
(s − sobs ) fn (s | θ)K
sup
θ∈Bδc
k2
Rd
θ∈Bδc
+
ε−d
n ds
ˆ
(
≤ sup
s − sobs
εn
sup
ks−sobs k≤δ 0 /3
(s − sobs )k2 K
fn (s | θ)
0
−d
K{λmin (Λ)ε−1
n δ /3}εn
ˆ
Rd
s − sobs
εn
(s − sobs )k2 fn (s | θ) ds.
)
ε−d
n ds
(11)
Rd
By Condition 2(ii), Condition 6 and following the arguments in the proof of Lemma 3 of Li &
αδ
Fearnhead (2018), the right hand side of (11) is Op (e−an,ε cδ ), which is sufficient for πBδc {(θ −
αδ
θ0 )k1 (s − sobs )k2 } to be Op (e−an,ε cδ ).
23
For the integration over Bδ , by Lemma 7 (ii),
π
eBδ ,tv (tk1 v k2 )
πBδ {(θ − θ0 )k1 (s − sobs )k2 }
−k1 k2
= an,ε εn
+
πBδ (1)
π
eBδ ,tv (1)
´
´ k k
)
−1 t)K(v) dvdt
1 v 2 π(θ + a−1 t)r (s
t
+
ε
v
|
θ
+
a
0
n
n
0
obs
n,ε
n,ε
t(B
)
δ
{1 + Op (αn−1 )}
αn−1
π
eBδ ,tv (1)
where rn (s | θ) is the scaled remainder αn {fn (s | θ) − fen (s | θ)}. In the above, the second term
in the first brackets is Op (αn−1 ) by the proof of Lemma 6 of Li & Fearnhead (2018). Then
π
eBδ ,tv (tk1 v k2 )
πBδ {(θ − θ0 )k1 (s − sobs )k2 }
−k1 k2
−1
= an,ε εn
+ Op (αn ) ,
πBδ (1)
π
eBδ ,tv (1)
and the moments π
eBδ ,tv (tk1 v k2 )/e
πBδ ,tv (1) need to be evaluated. Theorem 1 of Li & Fearnhead
(2018) gives the value of π
eBδ ,tv (t)/e
πBδ ,tv (1), and this is obtained by substituting the leading
term of π
eε,tv (t, v), that is π(θ0 )gn (t, v) as stated in Lemma 2, into the integrands. The other
moments can be evaluated similarly, and give
−1
bn β0 {A(θ0 )1/2 Wobs + an εn EGn (v)},
(k1 , k2 ) = (1, 0),
k
k
1/2 W
T )}, (k , k ) = (1, 1),
π
eBδ ,tv (t 1 v 2 )
b−1
β
{A(θ
)
E
(v]
+
a
ε
E
(vv
0
0
n n Gn
1 2
obs Gn
n
=
π
eBδ ,tv (1)
E
(v),
(k
1 , k2 ) = (0, 1),
Gn
T
EGn (vv ),
(k1 , k2 ) = (0, 2),
2 4
+ Op (a−1
n,ε ) + Op (an εn ),
(12)
where bn = 1 when cε < ∞, and an εn when cε = ∞. By Lemma 2, EGn (vv T ) = Θp (1). Since
−2/5
−2/5
αn−1 = o(an ), covε (θ, s) =ε2n β0 varGn (v) + op (an ε2n ) and varε (s) = ε2n varGn (v){1 +
−2/5
op (an )}. Thus
βε = β0 + op (a−2/5
),
n
and the lemma holds.
´ ´
(13)
−d
For A ⊂ Rp and B ⊂ Rd , let π(A, B) = A B π(θ)fn (s | θ)K{ε−1
n (s − sobs )}εn dsdθ and
´ ´
−d
π
e(A, B) = A B π(θ)fen (s | θ)K{ε−1
n (s − sobs )}εn dsdθ. Denote the marginal mean values
of s for πε (θ, s | sobs ) and π
eε (θ, s | sobs ) by sε and seε respectively.
Proof of Lemma 5. For
(a),
write
Πε (θ∗ ∈ Bδc | sobs )
as
π[Rp , {s : θ∗ (θ, s) ∈
d−p
p
d
By Lemma 7, π(R , R ) = πP (1) = Θp (an,ε ). By the triangle inequality,
Bδc }]/π(Rp , Rd ).
c
π[Rp , {s : θ∗ (θ, s) ∈ Bδc }] ≤ π(Bδ/2
, Rd ) + π[Bδ/2 , {s : kβε (s − sobs )k ≥ δ/2}],
(14)
and it is sufficient that the right hand side of the above inequality is op (1). Since its first term is
c (1), by Lemma 7 the first term is op (1).
πBδ/2
−7/5
−7/5
When εn = Ω(an ) or Θ(an ), by (13), βε − β0 = op (1) and so βε is bounded in probability. For any constant βsup > 0 and β ∈ Rp×d satisfying β ≤ βsup ,
δ
π[Bδ/2 , {s : kβ(s − sobs )k ≥ δ/2}] ≤ K ε−1
ε−d
n ,
2βsup
and by Condition 2(iv), the second term in (14) is op (1).
24
−7/5
When εn = o(an ), βε is unbounded and the above argument does not apply. Let δ1 be a
constant less than δ0 such that inf θ∈Bδ1 /2 λmin {A(θ)−1/2 } ≥ m and inf θ∈Bδ1 /2 λmin {Ds(θ)} ≥
m for some positive constant m. In this case, it is sufficient to consider δ < δ1 . By Condition 4,
rn (s | θ) ≤ adn |A(θ)|1/2 rmax [an A(θ)−1/2 {s − s(θ)}].
Using the transformation t = t(θ) and v = v(s), fn (s | θ) = fen (s | θ) + αn−1 rn (s | θ) and applying the Taylor expansion of s(θ0 + xt) around x = 0,
π[Bδ/2 , {s : kβε (s − sobs )k ≥ δ/2}] ≤
ˆ
ˆ
−1/2
1/2
c
N [A(θ0 + a−1
{Ds(θ0 + e(1)
Wobs − an εn v}; 0, Id ]K(v) dvdt
n t)
n t)t − A(θ0 )
t(Bδ/2 )
ˆ
kβε εn vk≥δ/2
ˆ
+c
t(Bδ/2 )
kβε εn vk≥δ/2
−1/2
1/2
rmax [A(θ0 + a−1
{Ds(θ0 + e(1)
Wobs − an εn v}]K(v) dvdt,
n t)
n t)t − A(θ0 )
for some positive constant c. To show that the right hand side of the above inequality is op (1),
consider a function g4 (·) in Rd satisfying that g4 (v) can be written as g 4 (kvk) and g 4 (·) is
−1/2 , C (t) = Ds(θ + ξ ) and c = A(θ )1/2 W
decreasing. Let An (t) = A(θ0 + a−1
n
0
1
0
obs . For
n t)
p
c
each n divide R into Vn = {t : kCn (t)tk/2 ≥ kc + an εn vk} and Vn . In Vn , kAn (t){Cn (t)t −
c − an εn v}k ≥ m2 ktk/2 and in V c , ktk ≤ 2m−1 kc + an εn vk. Then
ˆ
ˆ
g4 [An (t){Cn (t)t − c − an εn v}]K(v) dvdt
t(Bδ/2 )
ˆ
kβε εn vk≥δ/2
(ˆ
≤
kβε εn vk≥δ/2
where
´
c
Vn
ˆ
2
Rp
g 4 (m ktk/2) dt + sup g4 (v)
v∈Rp
)
1 dt K(v) dv,
c
Vn
1 dt is the volume of Vnc in Rp . Then since βε εn = op (1), an εn = op (1) and
´
Vnc
1 dt
vkp , the right hand side of the above inequality is o
is proportional to kc + an εn
p (1). This implies
π(Bδ/2 , {s : kβε (s − sobs )k ≥ δ/2}) = op (1).
e ε (θ∗ ∈ B c | sobs ), since the supTherefore in both cases Πε (θ∗ ∈ Bδc | sobs ) = op (1). For Π
δ
e ε (θ∗ ∈ B c | sobs ) = 0.
port of its prior is Bδ , there is no probability mass outside Bδ , i.e. Π
δ
Therefore (a) holds.
For (b),
e ε (θ∗ ∈ Aθ ∩ Bδ | sobs )
supA∈Bp Πε (θ∗ ∈ Aθ ∩ Bδ | sobs ) − Π
supA∈Bp |π(Rp , {s : θ∗ (θ, s) ∈ Aθ ∩ Bδ }) − π
e(Rp , {s : θ∗ (θ, s) ∈ Aθ ∩ Bδ })|
+ op (1)
π
eBδ (1)
´ ´
−d
π(θ)|rn (s | θ)|K{ε−1
n (s − sobs )}εn dsdθ
−1 Bδ Rd
≤αn
+ op (1).
π
eBδ (1)
=
Then by the proof of Lemma 6 of Li & Fearnhead (2018), (b) holds.
For (c), to begin with, an (θε∗ − θeε∗ ) = an (θε − θeε ) − an βε (sε − seε ). By Lemma 7, an (θε −
θeε ) = op (1). For an βε (sε − seε ), similar to the arguments of the proof of Lemma 4,
π
eBδ,tv (v)
π
eB (v)
−1
sε − sobs = εn
+ Op (αn ) {1 + Op (αn−1 )}, seε − sobs = εn δ,tv
{1 + Op (αn−1 )}.
π
eBδ,tv (1)
π
eBδ,tv (1)
25
−3/5
Then an βε (sε − seε ) = Op (αn−1 an εn ) which is op (1) if εn = o(an ). Therefore the first part
of (c) holds. Since θeε∗ = θeε − βε (e
sε − sobs ), by the expansion of θeε in Lemma 3(c), the above
expansion of seε − sobs and (12), the second part of (c) holds.
Proof for Results in Section 3·3
Proof of Theorem 2. The integrand of pacc,q is similar to that of πRp (1). The expansion of
πRp (1) is given in Lemma 7(ii), and following the same reasoning, pacc,q can be expanded as
´ ´
εdn Bδ Rd qn (θ)fe(sobs + εn v | θ)K(v) dvdθ{1 + op (1)}. With transformation t = t(θ), plugging the expression of qn (θ) and π
eε,tv (t, v) gives that
ˆ
π
eε,tv (t, v)
−1
(rn,ε )−p q(rn,ε
t − cµ )
pacc,q = (an,ε εn )d
dvdt{1 + op (1)},
πδ (θ0 + a−1
t(Bδ )
n,ε t)
where rn,ε = σn /a−1
n,ε and cµ,n = σn (µn − θ0 ). By the assumption of µn , denote the limit of
cµ,n by cµ . Then by Lemma 2, pacc,q can be expanded as
ˆ
−1
pacc,q = (an,ε εn )d
(rn,ε )−p q(rn,ε
t − cµ,n )gn (t, v) dvdt{1 + op (1)}.
(15)
t(Bδ )×Rd
Denote the leading term of the above by Qn,ε .
For (1), when cε = 0, since supt∈Rp gn (t, v) ≤ c1 K(v) for some positive constant c1 , Qn,ε
is upper bounded by (an εn )d c1 almost surely. Therefore pacc,q → 0 almost surely as n → ∞.
When rn,ε → ∞, since q(·) is bounded
in Rp by some positive constant c2 , Qn,ε is upper
´
−p
d
bounded by (rn,ε ) c2 (an,ε εn ) Rp ×Rd gn (t, v) dvdt. Therefore pacc,q → 0 in probability as
´
n → ∞ since Rp ×Rd gn (t, v) dvdt = Θp (1) by Lemma 2.
−1 t(θ) − c
For (2), let t̃(θ) = rn,ε
µ,n and t̃(A) be the set {φ : φ = t̃(θ) for some θ ∈ A}. Since
−1
t̃ = σn (θ − θ0 ) − cµ,n and σn−1 → ∞, t̃(Bδ ) converges to Rp in probability as n → ∞. With
the transformation t̃ = t̃(θ),
(
´
(an εn )d t̃(Bδ )×Rd q(t̃)gn {rn,ε (t̃ + cµ,n ), v} dt̃dv, cε < ∞,
Qn,ε = ´
0
0
0
cε = ∞.
t̃(Bδ )×Rd q(t̃)gn {rn,ε (t̃ + cµ,n ), v } dt̃dv ,
By Lemma 9 and the continuous mapping theorem,
( ´
cd p d q(t̃)g{r1 (t̃ + cµ ), v} dt̃dv,
Qn,ε → ´ε R ×R
Rp ×Rd q(t̃)g{r1 (t̃ + cµ ), v} dt̃dv,
cε < ∞,
cε = ∞,
in distribution as n → ∞. Since the limits above are Θp (1), pacc,q = Θp (1).
´ For (3), when cε = ∞ and r1 = 0, in the above, the limit of Qn,ε in distribution is
Rp ×Rd q(t̃)g(0, v) dt̃dv = 1. Therefore pacc,q converges to 1 in probability as n → ∞.
R EFERENCES
B EAUMONT, M. A. (2010). Approximate Bayesian computation in evolution and ecology. Annual Review of Ecology,
Evolution, and Systematics 41, 379–406.
B EAUMONT, M. A., C ORNUET, J.-M., M ARIN , J.-M. & ROBERT, C. P. (2009). Adaptive approximate Bayesian
computation. Biometrika 96, 983–990.
B EAUMONT, M. A., Z HANG , W. & BALDING , D. J. (2002). Approximate Bayesian computation in population
genetics. Genetics 162, 2025–2035.
B LUM , M. G. (2010). Approximate Bayesian computation: a nonparametric perspective. Journal of the American
Statistical Association 105, 1178–1187.
26
B ONASSI , F. V. & W EST, M. (2015). Sequential Monte Carlo with adaptive weights for approximate Bayesian
computation. Bayesian Analysis 10, 171–187.
C ALVET, L. E. & C ZELLAR , V. (2015). Accurate methods for approximate Bayesian computation filtering. Journal
of Financial Econometrics 13, 798–838.
C HERNOZHUKOV, V. & H ONG , H. (2003). An MCMC approach to classical estimation. Journal of Econometrics
115, 293–346.
C ORNUET, J.-M., S ANTOS , F., B EAUMONT, M. A., ROBERT, C. P., M ARIN , J.-M., BALDING , D. J., G UILLE MAUD , T. & E STOUP, A. (2008). Inferring population history with DIY ABC: a user-friendly approach to approximate Bayesian computation. Bioinformatics 24, 2713–2719.
D ROVANDI , C. C. & P ETTITT, A. N. (2011). Estimation of parameters for macroparasite population evolution using
approximate Bayesian computation. Biometrics 67, 225–233.
D ROVANDI , C. C., P ETTITT, A. N. & L EE , A. (2015). Bayesian indirect inference using a parametric auxiliary
model. Statistical Science 30, 72–95.
D UFFIE , D. & S INGLETON , K. J. (1993). Simulated moments estimation of Markov models of asset prices. Econometrica 61, 929–952.
F EARNHEAD , P. & P RANGLE , D. (2012). Constructing summary statistics for approximate Bayesian computation:
semi-automatic approximate Bayesian computation (with Discussion). Journal of the Royal Statistical Society:
Series B (Statistical Methodology) 74, 419–474.
F ILIPPI , S., BARNES , C. P., C ORNEBISE , J. & S TUMPF, M. P. (2013). On optimality of kernels for approximate
Bayesian computation using sequential Monte Carlo. Statistical Applications in Genetics and Molecular Biology
12, 87–107.
F RAZIER , D. T., M ARTIN , G. M., ROBERT , C. P. & ROUSSEAU , J. (2016). Asymptotic Properties of Approximate
Bayesian Computation. arXiv.1607.06903 .
G OURI ÉROUX , C. & RONCHETTI , E. (1993). Indirect inference. Journal of Applied Econometrics 8, s85–s118.
H EGGLAND , K. & F RIGESSI , A. (2004). Estimating functions in indirect inference. Journal of the Royal Statistical
Society, Series B (Statistical Methodology) 66, 447–462.
K LEIJN , B. & VAN DER VAART, A. (2012). The Bernstein-von-Mises theorem under misspecification. Electronic
Journal of Statistics 6, 354–381.
L ENORMAND , M., JABOT, F. & D EFFUANT, G. (2013). Adaptive approximate Bayesian computation for complex
models. Computational Statistics 28, 2777–2796.
L I , W. & F EARNHEAD , P. (2018). On the asymptotic efficiency of approximate Bayesian computation estimators.
Biometrika , to appear.
M ARIN , J.-M., P ILLAI , N. S., ROBERT, C. P. & ROUSSEAU , J. (2014). Relevant statistics for Bayesian model
choice. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 76, 833–859.
M ARJORAM , P., M OLITOR , J., P LAGNOL , V. & TAVARE , S. (2003). Markov chain Monte Carlo without likelihoods.
Proceedings of the National Academy of Sciences 100, 15324–15328.
PAULI , F., R ACUGNO , W. & V ENTURA , L. (2011). Bayesian composite marginal likelihoods. Statistica Sinica 21,
149–164.
R IBATET, M., C OOLEY, D. & DAVISON , A. C. (2012). Bayesian inference from composite likelihoods, with an
application to spatial extremes. Statistica Sinica 22, 813–845.
RULI , E., S ARTORI , N. & V ENTURA , L. (2016). Approximate Bayesian computation with composite score functions. Statistics and Computing 26, 679–692.
S ERFLING , R. J. (2009). Approximation theorems of mathematical statistics, vol. 162. John Wiley & Sons.
S OUBEYRAND , S. & H AON -L ASPORTES , E. (2015). Weak convergence of posteriors conditional on maximum
pseudo-likelihood estimates and implications in ABC. Statistics & Probability Letters 107, 84–92.
T ONI , T., W ELCH , D., S TRELKOWA , N., I PSEN , A. & S TUMPF, M. P. (2009). Approximate Bayesian computation
scheme for parameter inference and model selection in dynamical systems. Journal of the Royal Society Interface
6, 187–202.
VAN DER VAART, A. W. (2000). Asymptotic Statistics. Cambridge University Press, Cambridge.
VARIN , C., R EID , N. & F IRTH , D. (2011). An overview of composite likelihood methods. Statistica Sinica 21, 5–42.
W EGMANN , D., L EUENBERGER , C. & E XCOFFIER , L. (2009). Efficient approximate Bayesian computation coupled
with Markov chain Monte Carlo without likelihood. Genetics 182, 1207–1218.
W OOD , S. N. (2010). Statistical inference for noisy nonlinear ecological dynamic systems. Nature 466, 1102–1104.
| 10math.ST
|
arXiv:1609.08395v2 [stat.ME] 3 Feb 2017
Appraisal of data-driven and mechanistic
emulators of nonlinear hydrodynamic urban
drainage simulators
Juan Pablo Carbajal1 , João Paulo Leitão1 , Carlo Albert1 , and Jörg
Rieckermann1
1
Swiss Federal Institute of Aquatic Science and Technology,
Eawag, Überlandstrasse 133, 8600 Dübendorf, Switzerland
February 6, 2017
Abstract
Many model based scientific and engineering methodologies, such as system identification, sensitivity analysis, optimization and control, require a
large number of model evaluations. In particular, model based real-time control of urban water infrastructures and online flood alarm systems require
fast prediction of the network response at different actuation and/or parameter values. General purpose urban drainage simulators are too slow for this
application. Fast surrogate models, so-called emulators, provide a solution
to this efficiency demand. Emulators are attractive, because they sacrifice
unneeded accuracy in favor of speed. However, they have to be fine-tuned to
predict the system behavior satisfactorily. Also, some emulators fail to extrapolate the system behavior beyond the training set. Although, there are
many strategies for developing emulators, up until now the selection of the
emulation strategy remains subjective. In this paper, we therefore compare
the performance of two families of emulators for open channel flows in the
context of urban drainage simulators. We compare emulators that explicitly use knowledge of the simulator’s equations, i.e. mechanistic emulators
based on Gaussian Processes, with purely data-driven emulators using matrix factorization. Our results suggest that in many urban applications, naive
data-driven emulation outperforms mechanistic emulation. Nevertheless, we
discuss scenarios in which we think that mechanistic emulation might be favorable for i) extrapolation in time and ii) dealing with sparse and unevenly
sampled data. We also provide many references to advances in the field of
Machine Learning that have not yet permeated into the Bayesian environmental science community.
1
Introduction
For many real-world systems with a nonlinear response, model based tasks
such as sensitivity analysis, learning model parameters from data (i.e. system identification or model calibration), and real-time control, are hampered
by the long runtime of the employed numerical simulators. Even if runtimes
1
are short, these methods require a large number of model runs, which can
take a prohibitive long time. One way of speeding up these tasks is to build
fast surrogate models, so called emulators, to replace the computationally expensive simulators. An emulator is a numerical model that is tailored to approximate the results of a computationally expensive simulator with a huge
reduction in the time needed to run a simulation [O’Hagan, 2006], i.e. it is
a metamodel. These ideas also belong to the technique of Reduced-Order
Models (ROM), specially for models based on Partial Differential Equations
(PDE) [Baur et al., 2014, Quarteroni et al., 2016], and emulation as described
below.
To ground ideas, imagine that the flow at the outlet of a drainage network
is limited using a flow limiting gate or by activating water storage systems
(Fig. 1). The position of the gate and the activation of storage is controlled using a model predictive controller [Xi et al., 2013]. Such a scenario is relevant
in performance optimization of water treatment plants [Fu et al., 2008]. The
signals used to control the flows could be the current intensity and duration
of rain events from several rain gauges within the catchment. The controller
needs to estimate an optimal course of action by predicting the flows induced
by the rain and many possible actuations. This optimization generally requires thousands of model runs, which can take a prohibitive long time when
running a physically detailed simulator of the sewer network, such as a EPA
Storm Water Management Model (SWMM) model [Rossman, 2010]. However, the simulator is just used to estimate the relation between the rain,
the actuation, and the flow. The full details of the simulator might not be
required to obtain an accurate estimation of this relation. Feedback control
might further reduce the required accuracy of the estimated relation.
Emulation and interpolation are equivalent problems. An emulator is
built using the best available simulator to sample the space of actuations
and/or parameters (henceforth the latter will include actuations). The training data is then used to build an interpolation function which should predict
values at unseen parameters with an acceptable degree of accuracy, which is
case dependent. That is, we reconstruct an unknown function F : R|γ|+1 → R,
that takes a parameter vector of size |γ| and a time instant, and generates
the value of the magnitude of interest. When this function is evaluated at the
inputs used for training, the results are the same as the training outputs1 ,
i.e. this is the meaning of interpolation of the training data adopted herein.
Stated in this way, no distinction is needed between the parameters and the
time components in the input. However, knowing that the data is generated
by a dynamical system, we separate time from the other parameter components. Thus, we can find one interpolant in time and one in parameter space,
which might be coupled to each other. This parameter-time coupling emerges
naturally in mechanistic emulation, as will be shown in Sec. 2.2.
When the simulator is based on differential equations, the link between
Gaussian Processes (GP) and linear stochastic differential equations (SDE) [Poggio and Girosi, 1990, Steinke and Schölkopf, 2008, Albert, 2012, Särkkä et al.,
2013, González et al., 2014, Solin, 2014], permits the creation of GP based
emulators that include knowledge about the simulator dynamics; these are
called mechanistic emulators (MEMs). Conceptually, mechanistic emulation
seeks a function that interpolates the training data whithin a class of functions defined by an SDE. The importance of GPs for MEMs stems from the
fact that they are the formal solution of this SDE. Hence when the simulator is linear the emulator gives exact results; while for nonlinear simula1 Herein considered noiseless since they are generated by a deterministic simulator.
2
Figure 1: An imaginary drainage network with flow limiting gate and/or water storage
systems. The interesting singal is the level/flow at the outlet of the network, marked with
a circle. Isometric tiles from www.kenney.nl.
3
tors, the MEM will provide only an approximation. Increasing the accuracy
of this approximation and the efficiency of the methods are two fundamental
challenges in GP based emulation [O’Hagan, 2006, Rasmussen and Williams,
2006, Ch. 8].
Reichert et al. [2011] enumerated four overlapping approaches for developing emulators of dynamic simulators:
i) Gaussian Processes
ii) Basis function decompositions
iii) State space transition function approximation
iv) Stochastic linear model conditioned on data using Kalman smoothing.
In particular approaches i) and iv) are two different implementations of the
same problem [Steinke and Schölkopf, 2008]. Roughly speaking the Kalman
smoothing algorithm used in iv) is an iterative implementation of the conditioning of the GP in i). The iteration in iv) avoids the ill-conditioned covariance matrices [Hansen, 1998] involved in GP when sampling rates are
high [Steinke and Schölkopf, 2008, Reichert et al., 2011] and it is faster than
direct matrix inversion in a serial implementation. The GP approach i) is better suited for parallelization, speedups and energy saving via approximated
computing [Angerer et al., 2015].
Approaches i) (or iv)) and ii) are similar with respect to their implementation. That is, approach ii) can be implemented using GP regression [Rasmussen and Williams, 2006, sec. 2.7]. Therefore, the essential difference
between i) and ii) is that the former explicitly introduces mechanistic knowledge. It will be shown here that approach i) is currently constrained to
linear mechanistic knowledge, while popular methods based on maximum
entropy [Victor and Johannesma, 1986, Christakos, 1998, Harte, 2011] can
handle nonlinear knowledge. The difference between GP based mechanistic
emulation and maximum entropy methods is that in the latter, the mechanistic knowledge is added as constraints on the moments of the predictive
or posterior distribution, while in the former its is added as the dynamics of
the prior model. Adding constraints to predictive distributions require expert
knowledge available at the level of the emerging behavior of the simulator,
while adding dynamic information requires knowledge about the constitutive
elements parts of the simulator. The latter is likely to be readily available
from the development of the simulator itself.
Herein we compare the performance of GP emulators built using approaches i),
which we call mechanistic emulation, and ii) which we call data-driven emulation. The basis function that will be used for data-driven emulation will be
derived solely from the data using matrix factorization, i.e. they will not explicitly include mechanistic knowledge. In this article we use singular value
decomposition (SVD) and nonnegative matrix factorization (NMF) to extract
these bases (Sec. 2.1), but more general basis extraction methods like Proper
Orthogonal Decomposition (POD) could be used [Hesthaven et al., 2016]. We
compare results in an academic emulation problem to highlight the differences between the approaches (Sec. 3.1-3.2). We also provide emulation examples pertinent to the fields of hydrology and urban water management
(Sec. 3.3-3.4). In all of these, data-driven emulation outperforms mechanistic
emulation. The objective of this comparison is to provide intuition about the
suitability of each approach, which is not available to date to the best of our
knowledge, to highlight the need of enhancements of our emulators, and to
motivate research questions in the field of emulation. This is relevant, because, as described above, many applications in the field of urban drainage
4
and flood predictions to date are hampered by slow models. To a lesser degree, this is one of the few NMF applications in hydrology [Alexandrov and
Vesselinov, 2014].
2
Methods and Materials
In the subsequent sections we firstly describe the two emulation approaches
used herein (Sec. 2.1-2.2). Aiming at a wide readership, mathematical detail and rigor are kept at a minimum required level. References are provided
for the interested reader. Secondly, we describe the datasets used for training
and testing the emulators (Sec. 2.3). These include models of two small catchments in Switzerland, that we use to thoroughly evaluate the performance of
the emulators.
We use the word data to refer to the input and output pairs provided by
the simulator being emulated, e.g. in the case of a hydrodynamic simulator, inputs could be time and physical parameters of a sewer network, and
outputs water levels or flows.
2.1
Data-driven emulators
In this approach we make the following assumptions about the simulation
data used to build the emulator:
a) The data contains the most significant dynamic features of the system
response and these can be used as a time varying basis to reproduce the
data.
b) Unseen system responses are well approximated by a linear combination of the features in a), i.e. there exist a solution manifold that can be
well approximated with low dimensional sets [Quarteroni et al., 2016,
Hesthaven et al., 2016].
c) There exists a "smooth" mapping between inputs and the coefficients of
the linear combinations of features.
With these assumptions in mind we define the approximation strategy:
y(t, γ) '
q
X
β i (γ)φ i (t),
(1)
i =1
¢
β i (γ) ∼ GP m i (γ), K i (γ, γ0 )
¡
(2)
where y(t, γ) is the output of the simulator at time t and parameters γ, β i
is a mapping between©theseªparameters and the components of the output in
the basis function set φ i (t) . Following Rasmussen and Williams [2006, Sec.
2.7] this could be generalized to a full GP regression problem. However, as
stated here the problem is simpler and it is justified by the performance it
provides in the examples showcased in Sec. 3.
The procedure to build a data-driven emulator follows:
©
ª
©
ªn
i. Extract the first q ≤ n trn features φ i (t) using the training set yi (t, γ i ) i=trn1 .
This gives a q × n trn matrix B of coefficients (q coefficients for each simulation used for training) and the basis evaluated at the observed time
points Φ i j = φ j (t i ), i = 1, . . . , N, j = 1, . . . , q.
ii. Interpolate the B coefficients from step i. using a GP to obtain a function of the inputs B = f (γ).
5
iii. Evaluate f (γ) on the parameters in test set, use ©eq. (1) toª predict outn
puts, and compare them with the output test set yi (t, γ i ) i=tst1 .
©
ª
The features φ i (t) are extracted from the training data itself via matrix
factorization, which allows us to impose some general constraints on the features, e.g. nonnegativity. Herein we use the singular value decomposition
(SVD) and nonnegative matrix factorization (NMF), which are explained in
later pragraphs.
These methods provide the basis evaluated only at the observed time
points, Φ, and to predict at unobserved times we linearly interpolate them
over time. The implications of this interpolation will be discussed in Sec. 4.
This approach decouples the interpolation in time with the interpolation in
parameter space.
SVD is a robust factorization to calculate the eigenvectors of the covariance of the data matrix, i.e. the principal components. Given the matrix Y ∈
R N×n , SVD calculates the orthogonal matrices Φ ∈ R N×n and W ∈ Rn×n , and
n×n with only Σ 6= 0.
the element-wise nonnegative diagonal matrix Σ ∈ R+
ii
These matrices fulfill the relation
Y = ΦΣW > .
(3)
Using this factorization to calculate the covariance of the data matrix gives
Y Y > = ΦΣΣ> Φ> .
(4)
Showing that Φ are eigenvectors of the covariance matrix, i.e the principal
components of the data. A decomposition of Y in the form of eq. (1) is obtained
by defining B as the follows:
B = ΣW > ,
(5)
Y = ΦB.
(6)
The quality of approximation of the data degrades graciously with decreasing number of principal components. Hence, only the first q < n trn principal
components are used in general. This reduces the size of the representation
of the training data. In the context of ROM and PDE, this decomposion is
also know as Proper Orthogonal Decomposition (POD).
NMF provides an approximate minimum norm positive decomposition of
the data [Kim and Park, 2008]2 . Formally it solves the following problem:
Problem 1 (NMF). Given the matrix Y ∈ R N×n and q ∈ N with q ≤ n, find
N×q
q×n
matrices Φ ∈ R+ and B ∈ R+ such that they minimize
kY − ΦBk2 + akΦk2 + bkBk2 .
(7)
Where k · k is the Frobenius matrix norm and R+ is the set of nonnegative real
numbers.
Intuitively, NMF works as SVD but constraining the principal components
and the mixing coefficients to be nonnegative. The decomposition controls
the norm of the basis and its coefficients using two regularization terms
parametrized with weights a and b, which are manually tuned. The nonnegative basis provided by NMF might be readily interpreted in physical
terms in hydrological applications, yet the method is not widespread in the
community.
2 function nmf_bpas of GNU Octave’s linear-algebra package
linear-algebra/.
6
http://octave.sourceforge.net/
2.2
GP based mechanistic emulator (MEM)
The predictive mean of a GP in x conditioned on data observed at x0 , has the
following structure
y(x) = K (x, x0 )α + m (x),
¡
¢−1 ¡
¢
α = K (x0 , x0 ) + κ I
y(x0 ) − m (x0 ) .
(8)
(9)
where K and m stand for the covariance and the mean function of the prior
GP, respectively. The weights vector α is learned from the data at the conditioning step, requiring the inversion of the matrix obtained by evaluating
the covariance function at the observed inputs, shown in eq. (9). The regularization parameter κ encodes our trust on the prior knowledge and the error,
if any, of the observations y(x0 )3 .
A MEM is the GP in time and simulator parameters associated with an
input-driven linear stochastic ordinary differential equation (SODE) with a
multidimensional state space. Each component of the state space is defined
by an observed simulator’s parameter vector (γ i ). Extra components are reserved for prediction at unseen simulator’s parameter vectors4 .
The process of building a MEM requires the definition of
i. A linear prior model.
ii. The covariance and mean functions of the GP.
iii. A mapping from the parameters of the simulator to the parameters of
the GP.
In the following sections we obtain the mathematical expression of the covariance and mean functions of a first order linear time invariant (LTI) SODE,
and explain the construction of the mapping between parameter spaces. LTI
systems often arise, for example, from a finite element modeling of partial
differential equations.
For a rigorous and more general development, that include time varying
parameters, see Albert [2012]. The development for periodic difference equations was presented in Steinke and Schölkopf [2008] and extended to a more
general setting in González et al. [2014].
A LTI SODE in the vector-valued function
s(t) : R+ → R M ,
(10)
with initial condition s(0) = s 0 ∼ N ( s̄ 0 , Σ0 ), is defined by the equation
ds
(t) = A s(t) + u(t) + ξ(t),
dt
(11)
where
A ∈ R¢M × M , u(t) is a deterministic exogenous actuation, and ξ(t) ∼
¡
N 0, Σξ (t, t0 ) is a Gaussian noise term with covariance function
Σξ (t, t0 ) = Σδ(t − t0 ) Σ ∈ R M × M .
(12)
The general solution of this equation is
s(t) = eA t s 0 +
Z t
0
eA( t−τ) (u(τ) + ξ(τ)) dτ.
(13)
3 In the MEMs used herein, to obtain interpolation of the training data, we set the regularization
parameter to the machine epsilon.
4 typically only one component is reserved for this, but more could be used in a parallel setting.
7
where the first term is the solution of the homogenous ODE, i.e. with u(t)
and ξ both zero. The second term is the response of the ODE to the noisy
actuation.
As mentioned above, this system depends on the simulator’s parameters.
In general this means A(γ), Σξ (γ, γ0 ), u(t, γ), and potentially s 0 (γ).
2.2.1
Covariance function
We calculate the covariance function of the trajectories in eq. (13) using the
formula
h
i
K (t, r) = E (s(t) − E[s(t)]) (s(r) − E[s(r)])> ,
(14)
where s> indicates transposition and E is the ensemble average over initial
conditions and noise realizations.
The ensemble average of the solutions, E[s(t)], is solely determined by the
average of the homogeneous part (depending only on the initial condition)
and the deterministic actuation u, because the contribution of the noise term
ξ vanishes due to its zero mean value:
Z t
E[s(t)] = eA t s̄ 0 +
eA( t−τ) u(τ)dτ.
(15)
0
That is, the mean function is the solution to the deterministic (noise-free)
inhomogeneous ODE.
From eqs. (13) and (15) we obtain the factors in the expectation in eq. (14),
s(t) − E[s(t)] = eA t (s − s̄ 0 ) +
Z t
0
eA( t−τ) ξ(τ)dτ.
(16)
Inserting this in the expression for the covariance function, eq. (14), we obtain four terms. The first term is the covariance of the initial conditions. The
second term is a product of the integral of the noise term. The last two terms
are products of the initial conditions and the noise term; these terms will
vanish due to the independence of the initial condition and the noise term,
and due to the zero mean of the latter. Finally, we obtain:
>
K (t, r) = eA t Σ0 eA r +
Z tZ r
0
0
>
e A t Σ0 e A r +
>
eA( t−τ) Σδ(τ − ρ )eA (r−ρ ) dρ dτ =
Z min( t,r)
0
(17)
>
eA( t−µ) Σ eA (r−µ) dµ.
The second term can be recognized as the property of the covariance under
a linear transformation:
s = Gv
Σs = G (G Σv )> ,
(18)
where the matrix product should be interpreted as the application of the
integral
Z ∞
(Av)(t) =
A(t, µ)v(µ)dµ,
(19)
0
Z ∞
(AB)(t, r) =
A(t, µ)B(µ, r)dµ.
(20)
0
In our case G stands for the Green’s function of the linear operator
d
− A,
dt
8
(21)
which is
G(t, r) = H(t − r)eA( t−r) ,
(22)
with H the Heaviside or Step function. Its transpose (adjoint) is
>
G† (t, r) = H(r − t)eA (r− t) .
(23)
giving
³ ¡
¢† ´
G G Σξ (t, r) =
µZ ∞
¶†
Z ∞
G(t, µ)
G(µ, τ)Σδ(τ − r)dτ dµ =
0
0
Z ∞
G(t, µ)Σ G† (µ, r)dµ =
0
Z ∞
>
H(t − µ)H(r − µ)eA( t−µ) Σ eA (r−µ) dµ =
(24)
0
Z min( t,r)
0
>
eA( t−µ) Σ eA (r−µ) dµ
This implies that if the differential operator has a known Green’s function, then the covariance function of the GP can be calculated by transforming the covariance function Σ of the noisy input ξ. A more general and formal
treatment of this process is described in Kimeldorf and Wahba [1970] and the
relation between differential operators and kernels is summarized in Steinke
and Schölkopf [2008]. This covariance function computes the statistical interactions between the components of the LTI SODE. In other words, it provides the coupling of components of what is know as multi-output GP in the
machine learning community and cokriging in geostatistics [Rasmussen and
Williams, 2006, Sec. 9.1]. In Fricker et al. [2013] different structures of the
coupling between output components were studied, MEMs automatically derive the coupling structure from the available prior knowledge.
2.2.2
Linear prior
To determine the elements in the system matrix A in (11), we select a linear
model for each output in the training data corresponding to a simulator’s
parameter vector, which we call the linear proxy:
L i : R+ × R q → R d ,
(25)
t, θ i 7→ z(t, θ i ).
(26)
The q dimensional parameter vector θ i of the proxy depends on the simulator’s parameters used for the ith simulation, i.e. θ i (γ i ). The proxy’s output
dimension d is equal to dimension of each output in the training data, usually d = 1. The concatenated set of proxies defines the linear prior of the
MEM.
Herein, the proxies will be m dimensional linear time invariant (LTI)
ODEs, i.e. in state space form
ds
(t) = A(θ )s(t) + u(t, θ )
dt
A ∈ R m× m , u ∈ R m
C∈R
z(t) = Cs(t)
d ×m
(27)
(28)
Here all proxies have the same dimension m. Although this is not required
by the method, it is the simplest structure of MEMs [Albert, 2012]. Nevertheless, proxies with different dimensions might be better suited for dynamical systems with bifurcations, e.g. training data containing a mixture of
9
oscillating and converging time series, due to the presence of a Supercritical
Andronov-Hopf bifurcation [Kuznetsov, 2006] in the simulator.
The emulator’s linear prior is constructed by aggregating the proxies and
by coupling them with a noise term with a covariance that depends on the
simulator’s parameters, i.e
dS
(t) = A(Θ)S(t) + U(t, Θ) + ξ(t, Γ),
dt
Z(t) = C (Θ)S(t),
A∈R
mn×mn
, U, ξ ∈ R
mn×1
C ∈R
(29)
(30)
d×mn
.
Where Θ = θ i and Γ = γ i contain the corresponding n parameter vectors.
The matrix A is block diagonal, C is a horizontal concatenation, and U a
vertical concatenation:
©
ª
©
ª
A(θ1 )
A(Θ) =
..
u(t, θ1 )
..
,
.
u(t, θn )
U(t, Θ) =
.
A(θn )
,
C (Θ) = C(θ1 )
£
2.2.3
...
(31)
¤
C(θn ) .
Parameter mapping
To evaluate the matrices in eq. (31) we need a mapping from simulator’s
parameters γ to proxy’s parameters θ , , i.e. Γ 7→ Θ. It can be an ad-hoc
function derived from knowledge about the simulator, as was done in [Albert,
2012, Machac et al., 2016a,b] or it can be learned directly from the data. The
latter is especially useful if proxies with different dimensions are combined
to form the linear prior of the emulator.
A proxy’s parameter θ i can be learned from the data via optimization,
e.g. least squares fit of the data yi (T, γ i ) using z(T, θ i ). Alternatively, Θ
can be left as hyperparameters in the GP and estimated by maximizing the
likelihood. In any case, the obtained pairs (γ i , θ i ) are then used to learn a
mapping between the simulator’s and emulator’s parameter spaces.
2.2.4
Warped MEM
It is possible to add some nonlinear knowledge to a mechanistic emulator
when the simulator’s equations can be approximated with a Wiener model,
i.e. a time independent nonlinear function applied to the states of a linear
dynamical system. To do this we use Warped GPs [Snelson et al., 2003].
This modification does not affect the structure of the emulator, only the data
on which it is trained. In addition to the linear dynamical system parameters defining the prior, the parameters of the nonlinear function need to be
learned as well unless this mapping is explicitly given, e.g. from the linearization of a nonlinear ODE via a change of variables as in the Bernoulli
ODE [Hairer et al., 1993].
10
2.3
Datasets description
Simulated data used to build emulators are provided in a set of scalar signals
yi (t j©, γªi ) with i = 1, . . . , n, all of them sampled at the same time points, i.e.
T = t j , j = 1, . . . , N. For all emulators reported here, the time series in the
datasets were subsampled as much as possible, without deteriorating their
relevant dynamic features, e.g. oscillations and/or peaks. In all cases, the
datasets are randomly separated in a training set of size n trn and a test set
of size n tst , with n trn + n tst = n. Table 1 summarizes the properties of the data
used.
Dataset
# parameters, |γ|
Time samples, N (used/total)
Training set, n trn
Test set, n tst
Nonlinear DS I & II
1
6/40
10
190
Wartegg
2
52/2880
200
700
Adliswil
8
193/601
128
128
Table 1: Summary of datasets used herein.
2.3.1
Nonlinear dynamical system dataset I & II
To illustrate the virtues of MEMs, we first consider a didactical example using data generated by the model
dx
= a(x0 )x + b(x0 ),
dt
x(0) = x0 ,
2
(32)
a(x) = −a 0 e−a1 ( x−sign ( x)) ,
(33)
b(x) = b 0 tanh (x).
(34)
The parameters values are a 0 = 12.616, a 1 = 5, b 0 = 2.
The time evolution of this contrived system is linear. The parameters
defining the evolution depend nonlinearly on the initial condition, i.e. the parameters of the linear system are nonlinear functions of the initial condition.
Therefore an emulator of this system takes time and the initial condition
(|γ| = 1) as inputs. The behavior of this system meets the above assumptions
underlying a MEM with a linear time-invariant prior. Thus an emulator
without coupling noise solves this system exactly.
The first dataset consists of simulated time series with 40 output observations for 200 initial conditions x0 ∈ [−1, 1].
The second dataset is built by mixing the trajectories of the dynamical
system in eq. (32), specifically:
Z
x̂(t, x0 ) = x(t, x0 + α) N (x0 , σ2 )dα,
(35)
α
α
where σ = 0.5. Although the nonlinearities in the dynamical system still
depend only on the initial condition, the smoothing couples neighboring trajectories.
For the mechanistic emulation we will use a 1 dimensional linear proxy:
ds
(t) = θ1 (s 0 )s(t) + θ2 (s 0 ),
dt
11
(36)
and use the knowledge from the first simulator to set s 0 = x0 , θ1 (s 0 ) = a(s 0 ),
and θ2 (s 0 ) = b(s 0 ). Since each proxy is the solution of the system in eq. (32),
no coupling of the components of the MEM is needed in the first case, i.e.
Σ = I. For the second simulator we will couple the components using a 1
dimensional Matérn covariance function and the parameters mapping will
be learned from the data.
2.3.2
Wartegg catchment dataset
This dataset was generated from a SWMM model of a 2.64 km2 urban catchment located in the city of Lucerne in the canton of Lucerne, Switzerland.
This model has been calibrated satisfactorily to observed rainfall-runoff and
used for hydrological studies of the site [Tokarczyk et al., 2015, detailed
model description provided therein].
The dataset consists of 900 time series with 2880 data points, simulating
24 h of water levels in an open outlet during different rain events. To drive
the system into a highly nonlinear behavior we synthetically generated rain
events covering a wide range of return periods. These events were generated
following a block rain model [Gujer, 2007, Ch. 13.2] parametrized by intensity (I) and duration (d), i.e. γ = (I, d) and |γ| = 2. The intensity of the event
spans 30 values in the range I ∈ [10, 100] mm h−1 , with 30 different durations
(d ∈ [10, 240] min) for each intensity.
For the mechanistic emulation we will use a 1 dimensional linear proxy
for the water level:
dh
(t) = θ1 (γ)h(t) + θ2 (γ)R(t, γ)
dt
(37)
where R(t, γ) is the rain event used in the simulation. The values of θ1 (γ)
and θ2 (γ) are obtained by a least squares fit of the data.
Due to the low performance achieved by standard MEMs in this dataset,
the actual MEM presented in Sec. 3.3 uses a Wiener model as proxy. This
is done using Warped GPs (Sec. 2.2.4), with a nonlinearity given by ĥ(t) =
a(γ) tanh(b(γ)h(t) + c(γ)), where the parameters a, b and c are learned from
the data.
2.3.3
Adliswil catchment dataset
This dataset was generated from a SWMM model of a 1.63 km2 urban catchment located in the city of Adliswil in the canton of Zurich, Switzerland. The
dataset was used in Machac et al. [2016a](detailed model description provided therein) to create an emulator which was then used to speed-up the
calibration (identification) of the parameters in the simulator.
The dataset consist of 256 time series with 601 data points, all of them
corresponding to a single rain event, but with different 8 dimensional input parameter vectors (|γ| = 8). The time series are simulations of inflow to
the local WWTP, and the parameters describe the physical properties of the
sewer network.
For the mechanistic emulation we will use the same linear proxy as in Machac
et al. [2016a], a 1 dimensional ODE for the discharge:
dQ
(t) = θ1 (γ)Q(t) + θ2 (γ)R(t)
dt
(38)
where R(t) is the measured rain event.
Two MEMs will be built, the first uses the values of θ1 (γ) and θ2 (γ) using
the relation from Machac et al. [2016a], and the second will obtain them from
a least squares fit of the data.
12
2.4
Performance assessment
To asses the quality of an emulator we compute the following emulation errors on the data reserved for testing:
i. Maximum absolute error (MAE)
¯
¯
e MAE (γ j ) = max ¯ ŷ(t i , γ j ) − y(t i , γ j )¯
i
(39)
where ŷ(t, γ) and y(t, γ) are the emulated and simulated responses, respectively.
ii. Root mean square error (RMSE)
v
u
N ¡
u1 X
¢2
e RMSE (γ j ) = t
ŷ(t i , γ j ) − y(t i , γ j )
N i=1
(40)
Error e MAE measures the error in reproducing extreme values and/or peaks
in the simulated signals, while error e RMSE measures an overall quality of
the emulation.
3
3.1
Results
Nonlinear system dataset
This dataset is used to highlight the value of the mechanistic over datadriven emulation. It is suited for illustration purposes, since the surface
to be reconstructed can be plotted (|γ| = 1).
Fig. 2 shows the response of the system described in eqs. (32)-(34) as a
surface R2 → R. Fig. 2a illustrates the weakness of the SVD emulator, which
does not exploit the simulator’s parameter dependence. Moreover predictions
at unseen time points are provided by linear interpolation of the SVD basis.
To build the MEM we used the exact simulator’s parameter dependence. Although training data is sparse in the time direction, since the MEM encodes
the right time evolution, the reconstruction quality is high, as can be seen
in Fig. 2b. Although the SVD emulator could be improved by using an exponential basis for time interpolation [Franz and Gehler, 2006], i.e. the one
provided by the covariance function of the MEM, the recovered parameter
dependence will still be poor due to the low sampling of the parameter space.
This example shows that mechanistic emulation is ideal for situations
where there is good prior knowledge and the training data is sparse. This is
even more striking when data corresponding to different simulator parameters is sampled at different times, e.g. with adaptive stepsize simulators. In
this case mechanistic emulation can be applied directly, while SVD emulation
becomes complicated as a matrix completion problem needs to be solved before factorization can be applied [see Oh, 2010, for an SVD relevant analysis].
In a similar fashion the Kalman smoothing algorithm described in Reichert
et al. [2011] also requires a first step in which all the data is interpolated into
the same temporal grid, e.g. via interpolation.
3.2
Nonlinear system dataset II
Since the structure of the proxy of a MEM does not change once its dimension
m is set, we can optimize the proxy’s parameters to reduce epistemic biases.
The example shown in Fig. 3 illustrates the effect of mismatches between
13
(a) SVD
(b) MEM
Figure 2: Nonlinear system emulation of didactical model. The colored surface shows the
behaviour of the simulator output. Red dots show the training data. The emulated surface
is shown with a black wireframe. In this contrived scenario a MEM (b) outperforms an
SVD emulator (a). The failure of SVD is mainly due to the linear interpolation in the time
direction.
14
the dynamical system and the proxy, i.e. epistemic bias, and how it can be
mitigated. In Fig. 3a we show the performance of a MEM built with the same
proxy as in Fig. 2b, but used to emulate the nonlinear dynamical system
described at the end of Sec. 2.3.1. Comparing with Fig. 2b, we see that in the
regions where there is no observed data, the emulation is biased. A MEM
with proxies fitted to the data is shown in Fig. 3b. In this case the epistemic
bias is considerably reduced although we did not used mechanistic knowledge
to improve the parameters mapping.
Hence, if the mapping between the simulator parameters and the linear proxy’s parameters cannot be exactly determined from the mechanistic
knowledge, it is better to fit the proxy’s parameters to the data instead of
using and ad-hoc calculation based on on one’s best guess or expert opinion.
The latter can be improved a posteriori by studying the parameters mapping
emerging from the fit.
3.3
Wartegg catchment dataset
Results of these simulations can be seen in Fig. 4. For rains shorter than a
certain duration (about 20 min for the intensity used in the plot, 41 mm h−1 )
the water level response shows the typical wave of runoff in an open channel.
After some critical duration, which corresponds to the catchment’s time of
concentration, the water level becomes constant and remains fixed for the
duration of the rain, defining the triangular region marked in the plot. After
the rain, the water level goes back to a lower fixed level and remains there
for a period of time which is a nonlinear function of the rain duration, e.g.
between 4-6 h for a 4 h event. This shows the nonlinear nature of the storage
involved in the effective discharge of the network. Finally there is a slow
decay in the level fueled by the residual water in the network, the rate of
this decay also depends nonlinearly on the rain’s duration.
Fig. 5 shows the distribution of the test error of a MEM5 and a NMF
emulator with 7 components. Table 2 summarizes the mean of the error distributions. The NMF emulator outperforms the mechanistic emulator. In
previous trials, an SVD emulator was also built and provided results comparable with NMF (not shown here), however it produced negative predictions
just before the steep increase of the water level at the beginning of the rain
event.
Emulator
NMF(1 × 104 )
MEM(1 × 102 )
MAE (m, %)
3.2 × 10−2 , 7.7
22.6 × 10−2 , 53.7
RMSE (m, %)
0.94 × 10−2 , 4.1
4.4 × 10−2 , 17.4
Table 2: Mean emulation errors corresponding to the Wartegg catchment dataset. The
value in parenthesis is the simulation speed-up factor obtained with respect to the original
simulator, e.g. if SWMM simulation takes 3 s, MNF emulator takes 0.3 ms.
Fig. 6 shows the quality of the NMF emulation for three different rain
intensities.
3.4
Adliswil catchment dataset
In [Machac et al., 2016a] a SWMM model of the Adliswil catchment was emulated using a MEM with a prior derived from a simplified version of the
5 Warped MEM, see Sec. 2.3.2. MEMs with linear proxies were unable to reduce average RMSE
below 25%.
15
(a) MEM
(b) MEM fitted proxy
Figure 3: Nonlinear system emulation II. The colored surface shows the surface to reconstruct. In this second contrived scenario the MEM (b) performs the same as the SVD
emulator (a). Red dots show the training data. The emulated surface is shown with a
black wireframe.
16
Figure 4: Simulator output for a fixed intensity (41 mm h−1 ) and varying duration. The
black bold lines label different runoff regimes: free-surface flows, runoff with activated
storage and emptying of storage in the catchment.
Figure 5: Test error distribution of the NMF emulator. The violin plots show the distribution of the logarithm of the error for the maximum absolute error (MAE) and the root
mean square error (RMSE). Light colored plots corresponds to the NMF emulator, solid
colored plots to the MEM. The mean(median) error is indicated with filled(empty) circles.
See Table 2 for their numerical values.
17
Figure 6: NMF emulation quality. The colored surface shows the simulation outputs for
three different rain intensities. The NMF emulated surface is shown with a black wireframe.
18
simulator’s equations, with the aim of running a system identification task
from a single rain event.
Figure 7 shows the distribution of the test error of a MEM and a SVD
emulator with 6 components. Both errors are calculated on the same test set.
Figure 7: Test error distribution of the MEM from Machac et al. [2016a] and a SVD emulator. The violin plots show the error distribution for the mean maximum absolute error
(MAE) and the root mean square error (RMSE). Light colored plots corresponds to the
SVD emulator, solid colored plots to the MEM. The mean(median) error is indicated with
filled(empty) circles. See Table 3 for their numerical values.
Table 3 summarizes the mean of the error distributions. In this example,
the SVD emulator also performs better than its mechanistic counter part.
Although the time interpolation of SVD does not respect the dynamics of
the system, the density of data is enough to provide a good estimation. The
simplicity of the SVD emulator, when compared with the mechanistic one,
makes this approach much more compelling for this application.
4
Discussion
The results presented in the previous sections seem to suggest that MEMs
are not useful for emulating complicated hydrological or hydrodynamic simulators and that they remain as an academic curiosity. However MEMs still
have many properties that can be exploited for better and faster emulation,
which should not be ignored only due to the early state of the method. These
known advantages include the suitability for parallelization and, speedups
and energy saving via approximated computing [Angerer et al., 2015].
19
Emulator
SVD(5 × 104 )
MEM(1 × 103 )
MEM-fit (1 × 103 )
MAE (l s−1 , %)
18.3, 6.2
23.2, 7.9
20.2, 6.9
RMSE (l s−1 , %)
3.19, 2.6
4.94, 4.0
3.42, 2.8
Table 3: Mean emulation errors corresponding to the Adliswil catchment dataset. MEM
refers to an emulator in Machac et al. [2016a], while MEM-fit to an emulator with the
same proxy structure but parameters fitted to the data. The value in parenthesis is the
simulation speed-up factor obtained with respect to the original simulator, e.g. if SWMM
simulation takes 3 s, SVD emulator takes 60 µs.
Table 4 summarizes the steps involved in the two approaches presented
here. There we indicate the relation between the method and the characteristics of the dataset. On one side, although data-driven emulators are computationally and conceptually simpler than MEMs, they are more sensitive to
the sparseness of the data. On the other side, MEM’s performance is limited
by the linearity of the prior which might fail to express our knowledge about
the nonlinear simulator. If good prior knowledge is available, the mechanistic emulation can incorporate correct dependencies, which could be exploited
to reduce the amount of data needed to achieve a desired performance.
Steps 1 and 2 of mechanistic emulation are the most sensitive to the mismatch between prior and simulator. To mitigate this, we keep the model
structure provided by the prior and learn the parameter values from the
data, thus providing the best linear proxy for the dataset (Sec. 3.2). This
generally provides sharper error distributions than using the parameter values obtained directly from the prior [Albert, 2012]. Note that mitigating this
simulator-prior mismatch is not a question of data quantity, since more data
will override the prior and all mechanistic insights it provided, thus rendering mechanistic knowledge unnecessary [Steinke and Schölkopf, 2008]. More
data would also increase the memory and computational resources needed
by the emulator. Increasing the density of time samples will also reduce the
condition number of the covariance matrices and the inversion problem at
the training phase will be ill-posed, unless iterative condition methods are
used [Reichert et al., 2011].
In step 2 of a data-driven emulator we need to factorize the data. If the
data is very sparse the generalization quality of the factorization is expected
to be poor. For data that is sampled at different temporal grids the situation becomes even more delicate since data preprocessing is required to build
up a grid, e.g. via interpolation or matrix completion. Matrix factorization
also provide features only at the observed inputs, therefore an interpolation method is required when emulating unseen time points at step 5. The
Kalman smoothing implementation of the mechanistic emulation used in Reichert et al. [2011] will be similarly affected. Optimizing this interpolation
can be as hard as using a MEM directly (Sec. 3.1).
Equation (15) shows that the mean function of the prior GP is given by
the solution to the noise-free linear ODE obtained by removing the noisy
term of the SDE (11). This suggests a simple improvement of the emulator
in which the mean function is replaced with a better approximator of the
data. This is the underlying idea behind the work of González et al. [2014],
in which the actuation affecting the mean of the GP is replaced by a signal
generated by the nonlinear part of the model applied to a surrogate trajectory. In that work however, the surrogate trajectory, e.g. built with matrix
factorization, does not encode mechanistic knowledge explicitly. This knowl-
20
Mechanistic emulation
Step
Data-driven emulation
1
Define prior
–
2
Obtain emulator parameters
Define/Build parameters
mapping
Conditioning
Factorize the data
3
4
5 (Emulation)
Matrix · vector
Build parameters mapping
–
Matrix · vector +
time interpolation
Table 4: Steps involved in the two emulation approaches described in this article. Datadriven approaches are simpler than mechanistic ones. Step 2 and 5 of data-driven approaches suffer if data is sparse or unevenly sampled. Steps 1 and 2 of the mechanistic
approach suffer from the lack of expressiveness of linear priors.
edge could be introduced via the analytical solution of a nonlinear differential equation, such as the nonlinear Bernoulli ODE, or analytical approximations of more general nonlinear differential equations [Adomian, 1991].
These enhancements would only affect the mean function of the predictive
GP, improving extrapolation quality and thus allowing for more sparse training sets. However, the estimation of uncertainties remains limited by the
covariance function associated with the linear prior. This can be improved by
including the mean function parameters in the conditioning step [Rasmussen
and Williams, 2006, sec. 2.7].
All these observations are derived from more general results on the reducibility and emulation readinness of general simulators and not only valid
for the especific simulators we used here. This is specially true for the datadriven approach, see for example Ch. 5 of Quarteroni et al. [2016].
5
Conclusions
We provided a comparison of mechanistic and data-driven emulation in several examples pertinent to the field of hydrology and urban water management. In all of these, data-driven emulation outperforms mechanistic emulation. The current state of MEMs makes them advantageous to fully datadriven emulators, when the training data is sparse and unevenly sampled.
This is the case when many simulation runs with high temporal resolution
are prohibitively expensive, or when adaptive stepsize simulators are used.
If the only objective of emulation is to obtain a fast tool to replace a simulator, there seem to be no advantage in using mechanistic knowledge besides
the case of sparse and unevenly sampled data mentioned before. The gain
obtained from enhancements of MEMs discussed here, such as Wiener model
proxies, Nonlinear mean functions, and hybrid mechanistic/data-driven emulators should be quantified in relation to the test error of an inexpensive
data-driven emulator.
Acknowledgements
The authors would like to thank Prof. Peter Reichert for his support during
the development of this article. We thank Dr. David Machac for his emulation results used in Figure 7. We thank Dr. Frank Blumensaat for shar-
21
ing with us the SWMM model file of the Wartegg catchment resulting from
many data collection campaings. We thank the developers of GNU Octave,
Sage and Inkscape for their excellent software tools, which were used for this
article. We thank the reviewers for their help improving this manuscript.
Funding The research leading to these results has received funding from
the European Union’s Horizon 2020 research and innovation programme under grant agreement No 641931 (Centaur).
Author contributions JPC developed the software, carried out simulations, data analysis. JPC & JR wrote this manuscript. JPL adapted
SWMM model files for the simulations and helped interpreting the results.
The work was done under the active supervision of CA & JR. All authors
copy-edited this manuscript.
References
Anthony O’Hagan. Bayesian analysis of computer code outputs: A tutorial.
Reliability Engineering & System Safety, 91(10-11):1290–1300, 2006. doi:
10.1016/j.ress.2005.11.025.
Ulrike Baur, Peter Benner, and Lihong Feng. Model order reduction for linear and nonlinear systems: a system-theoretic perspective. Archives of
Computational Methods in Engineering, 21(4):331–358, 2014.
Alfio Quarteroni, Andrea Manzoni, and Federico Negri. Reduced Basis Methods for Partial Differential Equations, volume 92 of UNITEXT. Springer
International Publishing, Cham, 2016. ISBN 978-3-319-15430-5. doi:
10.1007/978-3-319-15431-2.
Yu-Geng Xi, De-Wei Li, and Shu Lin. Model Predictive Control — Status
and Challenges. Acta Automatica Sinica, 39(3):222–236, mar 2013. ISSN
18741029. doi: 10.1016/S1874-1029(13)60024-5.
Guangtao Fu, Soon-Thiam Khu, and David Butler. Multiobjective optimisation of urban wastewater systems using parego: a comparison with nsga
ii. In 11th International Conference on Urban Drainage, 11 ICUD, Edinburgh, Scotland, 2008.
L.A. Rossman. Storm Water Management Model User’s Manual Version 5.0.
Technical Report EPA/600/R-05/040, National Risk Management Research
Laboratory, Office of Research and Development, U.S. Environmental Protection Agency, 2010.
Tomaso Poggio and F. Girosi. Networks for approximation and learning.
Proceedings of the IEEE, 78(9):1481–1497, 1990. ISSN 00189219. doi:
10.1109/5.58326.
Florian Steinke and Bernhard Schölkopf. Kernels, regularization and differential equations. Pattern Recognition, 41(11):3271–3286, 2008. ISSN
00313203. doi: 10.1016/j.patcog.2008.06.011.
Carlo Albert. A mechanistic dynamic emulator. Nonlinear Analysis: Real
World Applications, 13(6):2747–2754, 2012.
22
Simo Särkkä, Arno Solin, and Jouni Hartikainen. Spatiotemporal Learning
via Infinite-Dimensional Bayesian Filtering and Smoothing: A Look at
Gaussian Process Regression Through Kalman Filtering. IEEE Signal
Processing Magazine, 30(4):51–61, jul 2013. ISSN 1053-5888. doi: 10.
1109/MSP.2013.2246292.
Javier González, Ivan Vujačić, and Ernst Wit. Reproducing kernel Hilbert
space based estimation of systems of ordinary differential equations. Pattern Recognition Letters, 45:26–32, 2014.
Arno Solin. Explicit Link Between Periodic Covariance Functions and State
Space Models. Proc. of the Seventeenth Int. Conf. on Artificial Intelligence
and Statistics, 33:904–912, 2014.
C E Rasmussen and C K I Williams. Gaussian Processes for Machine Learning. Adaptive Computation And Machine Learning. MIT Press, 2006.
P. Reichert, G. White, M.J. Bayarri, and E.B. Pitman. Mechanism-based emulation of dynamic simulation models: Concept and application in hydrology. Computational Statistics & Data Analysis, 55(4):1638–1655, 2011.
doi: 10.1016/j.csda.2010.10.011.
Per Christian Hansen. Rank-Deficient and Discrete Ill-Posed Problems. Society for Industrial and Applied Mathematics, jan 1998. ISBN 978-0-89871403-6. doi: 10.1137/1.9780898719697.
C.M. Angerer, R. Polig, D. Zegarac, H. Giefers, C. Hagleitner, C. Bekas, and
A. Curioni. A fast, hybrid, power-efficient high-precision solver for large
linear systems based on low-precision hardware. Sustainable Computing:
Informatics and Systems, 2015. ISSN 2210-5379. doi: 10.1016/j.suscom.
2015.10.001.
J D Victor and P Johannesma. Maximum-entropy approximations of stochastic nonlinear transductions: an extension of the Wiener theory. Biological
cybernetics, 54(4-5):289–300, 1986. ISSN 0340-1200.
George Christakos. Spatiotemporal information systems in soil and environmental sciences. Geoderma, 85(2-3):141–179, 1998. ISSN 00167061. doi:
10.1016/S0016-7061(98)00018-4.
J. Harte. Maximum Entropy and Ecology: A Theory of Abundance, Distribution, and Energetics. Oxford Series in Ecology and Evolution. OUP Oxford,
2011. ISBN 9780191621161.
Jan S. Hesthaven, Gianluigi Rozza, and Benjamin Stamm. Reduced Basis
Methods, pages 27–43. Springer International Publishing, Cham, 2016.
ISBN 978-3-319-22470-1. doi: 10.1007/978-3-319-22470-1_3.
Boian S. Alexandrov and Velimir V. Vesselinov. Blind source separation for
groundwater pressure analysis based on nonnegative matrix factorization.
Water Resources Research, 50(9):7332–7347, 2014. ISSN 1944-7973. doi:
10.1002/2013WR015037.
Jingu Kim and Haesun Park. Toward faster nonnegative matrix factorization: A new algorithm and comparisons. In Proceedings of the 2008 Eighth
IEEE International Conference on Data Mining, ICDM ’08, pages 353–362,
Washington, DC, USA, 2008. IEEE Computer Society. ISBN 978-0-76953502-9. doi: 10.1109/ICDM.2008.149.
23
George S. Kimeldorf and Grace Wahba. A Correspondence Between Bayesian
Estimation on Stochastic Processes and Smoothing by Splines. The Annals
of Mathematical Statistics, 41(2):495–502, apr 1970. ISSN 0003-4851. doi:
10.1214/aoms/1177697089.
Thomas E Fricker, Jeremy E Oakley, and Nathan M Urban. Multivariate
Gaussian Process Emulators With Nonseparable Covariance Structures.
Technometrics, 55(1):47–56, feb 2013.
Y. A. Kuznetsov. Andronov-Hopf bifurcation. Scholarpedia, 1(10):1858, 2006.
revision #90964.
David Machac, Peter Reichert, Jörg Rieckermann, and Carlo Albert. Fast
mechanism-based emulator of a slow urban hydrodynamic drainage simulator. Environmental Modelling & Software, 78:54–67, 2016a. doi:
10.1016/j.envsoft.2015.12.007.
David Machac, Peter Reichert, and Carlo Albert. Emulation of dynamic simulators with application to hydrology. Journal of Computational Physics,
313:352–366, May 2016b.
Edward Snelson, Carl Edward Rasmussen, and Zoubin Ghahramani. Warped
gaussian processes. In In Advances in Neural Information Processing Systems (NIPS, page 2003. MIT Press, 2003.
E. Hairer, S. P. Nørsett, and G. Wanner. Solving Ordinary Differential Equations I (2Nd Revised. Ed.): Nonstiff Problems. Springer-Verlag New York,
Inc., New York, NY, USA, 1993. ISBN 0-387-56670-8.
P. Tokarczyk, J. P. Leitao, J. Rieckermann, K. Schindler, and F. Blumensaat.
High-quality observation of surface imperviousness for urban runoff modelling using uav imagery. Hydrology and Earth System Sciences, 19(10):
4215–4228, 2015. doi: 10.5194/hess-19-4215-2015.
Willi Gujer. Siedlungswasserwirtschaft. Springer, Berlin, 3. edition, 2007.
Matthias O Franz and Peter V Gehler. How to choose the covariance for
gaussian process regression independently of the basis. In Proc. Gaussian
Processes in Practice Workshop, 2006.
Sewoong Oh. Matrix completion: Fundamental limits and efficient algorithms. PhD thesis, Stanford University, 2010.
G Adomian. A review of the decomposition method and some recent results
for nonlinear equations. Computers & Mathematics with Applications, 21
(5):101–127, 1991. ISSN 08981221. doi: 10.1016/0898-1221(91)90220-X.
Octave community. GNU Octave 4.0.2, 2016. URL www.gnu.org/software/
octave/. Last accessed Sept. 1, 2016.
The Sage Development Team. Sage Mathematics Software (Version 7.2),
2016. URL http://www.sagemath.org. Last accessed Sept. 1, 2016.
Inkscape community. Inkscape 0.91, 2016. URL http://www.inkscape.
org/. Last accessed Sept. 1, 2016.
24
| 5cs.CE
|
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
1
Beam Design and User Scheduling for Non-Orthogonal
Multiple Access with Multiple Antennas Based on
Pareto-Optimality
arXiv:1705.06959v1 [cs.IT] 19 May 2017
Junyeong Seo, Student Member, IEEE, Youngchul Sung† , Senior Member, IEEE
Abstract
In this paper, an efficient transmit beam design and user scheduling method is proposed for multi-user
(MU) multiple-input single-output (MISO) non-orthogonal multiple access (NOMA) downlink, based on
Pareto-optimality. The proposed beam design and user scheduling method groups simultaneously-served
users into multiple clusters with practical two users in each cluster, and then applies spatical zeroforcing (ZF) across clusters to control inter-cluster interference (ICI) and Pareto-optimal beam design
with successive interference cancellation (SIC) to two users in each cluster to remove interference to
strong users and leverage signal-to-interference-plus-noise ratios (SINRs) of interference-experiencing
weak users. The proposed method has flexibility to control the rates of strong and weak users and
numerical results show that the proposed method yields good performance.
Index Terms
Non-orthogonal multiple access, multi-user MIMO, scheduling, Pareto-optimal design, SIC
I. I NTRODUCTION
NOMA is a promising technology for 5G wireless networks to increase the spectral efficiency [1].
Unlike conventional orthogonal multiple access (OMA) which serves multiple users based on time,
frequency and/or spatial domains, NOMA exploits the power domain that results from unequal channel
conditions under which users with strong channels are basically limited by degree-of-freedom (DoF) such
as bandwidth not by noise but users with weak channels are limited by additive noise [2, P. 239]. In
†
Corresponding author
The authors are with Dept. of Electrical Engineering, KAIST, Daejeon 305-701, South Korea. E-mail:[email protected],
[email protected].
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
2
NOMA with such channel conditions, the base station (BS) uses superposition coding and allocates less
power to strong-channel users and more power to weak-channel users. Here, less power to strong users
is not so detrimental since strong users are in the DoF-limited regime, but more power to weak users
leverages received SINRs of weak users. The strong interference caused from more power assigned to
weak users through good channels to strong users is eliminated by SIC to maintain high quality channels
for strong users.
Initially, NOMA with a single antenna in both the BS and users was studied [1], [3]–[8]. Recently,
there have been efforts to extend NOMA to multiple-antenna systems. Unlike conventional MU multipleantenna downlink systems which serve as many users as the number of antennas, more users can be
served in multiple-antenna NOMA systems. Although the possibility that multiple-antenna NOMA can
outperform conventional multiple-antenna OMA was shown [9], [10], many important problems need to
be investigated further for multiple-antenna NOMA. In multiple-antenna NOMA systems, typically user
grouping is done first by forming clusters as many as the number of transmit antennas and assigning
multiple users to each cluster, and then multiple users in each cluster share the spatial dimension and are
served in the power domain. Since the performance of multiple-antenna NOMA significantly depends
on the channel conditions of users across clusters and within each cluster, effective scheduling and
user grouping methods should be devised in order to achieve both MU diversity and the NOMA gain
from unequal channel conditions. In addition, the problem of optimal beam design and power allocation
compatible to user scheduling should be solved to maximize the performance.
A. Related Works
There have been several studies on multiple-antenna NOMA for cellular downlink especially for MUMISO downlink which is the main focus of this paper. In [11], the downlink beam design for sum-rate
maximization was considered for one given cluster based on minorization-maximization. In [12]–[14],
user scheduling and clustering is considered with the assumption that two users are assigned to each
cluster. In [12], highly correlated users are chosen as candidates to be clustered, and two users having
the largest channel gain difference are assigned to the same cluster. In [13], strong users are selected
by the semi-orthogonal user selection (SUS) algorithm [15], and then weak users are selected using the
matching user selection algorithm by considering inter-cluster interference. In [14], a fairness-oriented
user selection algorithms was proposed by selecting two paired users based on their NOMA data rate. In
all the works of [12]–[14], the beam design problem for MU-MISO NOMA was simplified by designing
zero-forcing (ZF) beams based on strong users’ channels and allocating the same beam to the weak user
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
3
as the beam of the strong user in each cluster. There also exists some study on multiple-input multipleoutput (MIMO) NOMA. For example, in [16] the impact of user paring is analyzed with fixed power
allocation under the assumption that inter-cluster interference is removed by multiple receive antennas.
Some part of this work was included in [17].
B. Contributions
In this paper, we consider MU-MISO NOMA downlink with practical two users in each cluster and
solve the aforementioned problem for MU-MISO NOMA downlink. The contributions of this paper are
summarized in the below:
• First, we solved the Pareto-optimal beam design and power allocation problem for two-user MISO
broadcast channels (BCs) in which superposition coding is used at the transmitter and the interference
at the strong user is eliminated by SIC while the interference at the weak user is treated as noise. This
work is the basis for successive development in this paper and is valuable as an independent item.
• We proposed an effective user scheduling, beam design and power allocation method for MU-
MISO NOMA downlink based on the above two-user Pareto-optimal design result by exploiting both the
spatial domain provided by multiple transmit antennas and the power domain provided by SIC. The key
advantages of the proposed method compared to the previous methods are that 1) the rates of strong users
and weak users can be controlled arbitrarily under Pareto-optimality, which provides great operational
flexibility to NOMA networks, and 2) beam design of the proposed method is generalized to include
multi-dimensional subspace if available and to yield performance improvement, whereas the same onedimensional beam is always used by both strong and weak users in the same cluster in the previous
methods [12]–[14].
Notations: Vectors and matrices are written in boldface with matrices in capitals. All vectors are
column vectors. For a matrix A, A∗ , AH and AT indicate the complex conjugate, conjugate transpose,
and transpose of A, respectively, and L(A) and L⊥ (A) denotes the linear space spanned by the columns
of A and its orthogonal complement, respectively. ΠA and Π⊥
A are the projection matrices to L(A) and
L⊥ (A), respectively. ||x|| represents the 2-norm of vector x. I denotes the identity matrix. y ∼ CN (µ, Σ)
mean that random vector y is circularly-symmetric complex Gaussian distributed with mean vector µ
and covariance matrix Σ.
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
4
II. S YSTEM M ODEL
We consider a single-cell MU-MISO NOMA downlink system with a BS equipped with Nt transmit
antennas and K single-antenna users. We assume that the K users in the cell are divided into the set
K1 of K/2 strong-channel users and the set K2 of K/2 weak-channel users. We assume that out of the
K users in the cell, 2Kc users are selected and simultaneously served for each scheduling interval with
Kc ≤ Nt (thus 2Nt users can be served simultaneously) and this simultaneous service to 2Kc users is
done by forming Kc clusters with two paired users in each cluster composed of one from the strong
user set K1 and the other from the weak user set K2 . We assume that the BS uses linear precoding and
NOMA is applied to two paired users in each cluster. Under these assumptions, the transmit signal of
the BS for one scheduling interval is given by
q
Kc q
X
(k) (k) (k)
(k) (k) (k)
x =
p1 w̃1 s1 + p2 w̃2 s2
,
(1)
k=1
where
(k)
si
(k)
is the transmit symbol for user i in cluster k from CN (0, 1), w̃i
is the Nt × 1 beamforming
vector for user i in cluster k out of the feasible beamforming vector set W := {w̃ | kw̃k2 ≤ 1}, and
(k)
pi
is the power assigned to user i in cluster k. The total BS transmit power PT is equally divided into
PT /Kc = P for each cluster. Then, the received signals of the two users in cluster k are given by
q
q
(k) (k) (k)
(k) (k) (k)
(k)
(k)H
p1 w̃1 s1 + p2 w̃2 s2
y1 = h̃1
(k)H
+ h̃1
(2)
q
Kc q
X
(k ′ ) (k ′ ) (k ′ )
(k ′ ) (k ′ ) (k ′ )
(k)
+ w1 ,
p1 w̃1 s1 + p2 w̃2 s2
k ′ 6=k
(k)
y2
+
=
(k)H
h̃2
(k)H
h̃2
q
q
(k) (k) (j)
(k) (k) (k)
p1 w̃1 s1 + p2 w̃2 s2
(3)
q
Kc q
X
(k ′ ) (k ′ ) (k ′ )
(k ′ ) (k ′ ) (k ′ )
(k)
p1 w̃1 s1 + p2 w̃2 s2
+ w2 ,
k ′ 6=k
(k)
where h̃i
denotes the actual Nt × 1 (conjugated) channel vector between the BS and user i in cluster
(k)
k, and wi
is the zero-mean additive white Gaussian noise (AWGN) at user i in cluster k from
(k)
CN (0, [ǫi ]2 ).
In the MU-MISO-NOMA system, we have both the spatial domain and the power domain. We consider
the design approach in which two users in each cluster are served in the power domain with SIC while
multiple clusters are served based on the spatial domain. Note that the last two terms in each of the
right-hand sides (RHSs) of (2) and (3) are the ICI and AWGN. In order to control ICI, we apply spatial
ZF across clusters. However, due to lack of spatial dimensions, we cannot remove ICI completely for all
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
5
users. Hence, we remove ICI for the strong users with spatial ZF to keep the strong users not interference(k)
limited. With this approach, the beam vector w̃i
(k)
w̃i
can be expressed as
(k)
wi , i = 1, 2,
= Π⊥
H̃
(4)
k
(k)
for some vector wi , where
(1)
(2)
(k−1)
H̃k := [h̃1 , h̃1 , · · · , h̃1
(k+1)
, h̃1
(Kc )
, · · · , h̃1
].
(5)
Once the ICI is controlled and given, the model (2) - (3) is a two-user MISO BC. Thus, our approach
to the overall design is to first investigate the optimal beam design and power allocation for a two-user
MISO BC with SIC at the strong user’s receiver in Section III, to derive certain performance properties
relevant to selection of two users in a cluster in Section IV, and then to develop an overall user selection
and beam design method for all clusters with controlling ICI in Section V.
III. T WO - USER MISO
BROADCAST CHANNEL WITH
SIC: PARETO - OPTIMAL
DESIGN
In this section, we focus on optimal beam vector design and power allocation for a two-user MISO BC
with SIC at the strong user’s receiver from the perspective of Pareto-optimality. With the cluster index
(k) omitted, the two-user model (2) - (3) for cluster k is given by
√
√
y i = hH
i ( pi w i s i + pj w j s j ) + n i
i, j ∈ {1, 2}, j 6= i,
(6)
where p1 + p2 ≤ P with the total power P allocated to the cluster; ni ∼ CN (0, σi2 ) is the sum of ICI
(k)
h̃ , i = 1, 2
and AWGN; and hi is the effective channel for user i (in cluster k) given by hi = Π⊥
H̃ i
k
H
H
⊥ H
H ⊥
from (2), (3) and (4) since h̃H
i w̃i = h̃i ΠH̃ wi = h̃i [ΠH̃ ] wi = hi wi . The feasible set for wi is
k
k
given by W = {w |
kwk2
≤ 1} since the Pareto-optimal beam wi under the model (6) lies in the linear
(k)
h̃
space spanned by h1 = Π⊥
H̃ 1
k
(k)
h̃
and h2 = Π⊥
H̃ 2
k
(k)
(k)
(k)
wi k = kw̃i k
[18], and hence kwi k = kΠ⊥
H̃
k
for Pareto-optimal beams.
Under the NOMA framework, we assume that user 1 is the strong user and user 2 is the weak user,
i.e., kh1 k2 /σ12 > kh2 k2 /σ22 and that user 1 decodes the interference from user 2 and subtracts it before
decoding its own data while user 2 treats the interference as noise. With this assumption, the rates of the
two users are given by
s1 (w1 , p1 )
R1 (w1 , p1 ) = log2 1 +
σ12
R2 (w1 , w2 , p1 , p2 )
= log2 1 + min
May 22, 2017
s2 (w2 , p2 )
r1 (w2 , p2 )
,
2
s1 (w1 , p1 ) + σ1 r2 (w1 , p1 ) + σ22
(7)
,
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
6
where the signal power and the interference power are respectively given by
2
si (wi , pi ) := pi |hH
i wi |
2
and ri (wj , pj ) := pj |hH
i wj | .
(8)
Note in (7) that for the rate of user 1, the interference from user 2 is not incorporated due to SIC and
the rate of user 2 is bounded by not only the SINR of user 2 but also the required ’SINR’ for user 1
to decode the message of user 2 for SIC before decoding its own data. Then, for the given (effective)
channel vectors (h1 , h2 ), the achievable rate region R of the two-user MISO-NOMA BC is defined as
the union of the rate-tuples that can be achieved by all feasible beam vectors and power allocation:
R :=
[
(R1 (w1 , p1 ), R2 (w1 , w2 , p1 , p2 )).
(9)
2
(w1 ,w2 )∈W
p1 ,p2 : p1 ,p2 ≥0,
p1 +p2 =P
The Pareto boundary of the rate region R is the outer boundary of R for which the rate of any one
user cannot be increased without decreasing the rate of the other user and Pareto-optimality has been
used widely as a general optimal beam design criterion for MU-MISO networks with linear precoding
[20]. A pair of beam vectors (w1 , w2 ) not achieving a Pareto-boundary point is not optimal since both
users’ rates can be increased by a better designed beam pair. Note that the sum-rate optimal point is
the point on the Pareto-boundary where the Pareto-boundary and the minus 45o degree line touch in the
(R1 , R2 ) plane, and the Pareto-optimality provides a general optimality criterion because we can change
the rate operating point arbitrarily and optimally. It is known that the Pareto-boundary can be found by
maximizing R2 for each given feasible R1∗ [20], i.e.,
max
(w1 ,w2 )∈W 2
p1 ,p2 : p1 ,p2 ≥0, p1 +p2 =P
subject to
R2 (w1 , w2 , p1 , p2 )
R1 (w1 , pi ) = R1∗ .
(10)
By exploiting the relationship between the rates and the SINRs in (7), the problem (10) can be rewritten
in terms of SINR as
max
(w1 ,w2 )∈W 2
p1 ,p2 : p1 ,p2 ≥0,
p1 +p2 =P
subject to
γ2 := min
s2 (w2 , p2 )
r1 (w2 , p2 )
,
2
s1 (w1 , p1 ) + σ1 r2 (w1 , p1 ) + σ22
s1 (w1 , p1 )
= γ1∗ ,
σ12
(11)
where γ1∗ is a given feasible target SINR for user 1. An efficient solution to the problem (11) exploits an
efficient parameterization of the beam vectors w1 and w2 . Note that the number of design variables in
w1 and w2 is 2Nt complex numbers. However, one can realize that it is sufficient that both beam vectors
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
7
are linear combinations of h1 and h2 (equivalently, Πh2 h1 and Π⊥
h2 h1 ). A component in the beam vector
not in the span of h1 and h2 does not affect either the signal power or the interference power and thus
does not affect either R1 or R2 [19]. Thus, it is known from [18] that the Pareto-optimal beam vectors
for the problem (11) can be parameterized as [18], [21]
Π⊥
Πh2 h1
h2 h1
+ β1 ⊥
,
kΠh2 h1 k
kΠh2 h1 k
q
Π⊥
Πh1 h2
h1 h2
,
w2 (α2 ) = α2
+ 1 − α22 ⊥
kΠh1 h2 k
kΠh1 h2 k
w1 (α1 , β1 ) = α1
(12)
(13)
where (α1 , β1 ) ∈ F := {(α, β), α, β ≥ 0, α2 + β 2 ≤ 1} and α2 ∈ [0, 1]. Unlike the conventional
parametrization without SIC in which both users use full power [20], [22], in the parameterization (12)
- (13) user 1 may not use full power whereas user 2 uses full power. This is because full power use of
user 2 helps both SIC at user 1 and its own SINR at user 2, but full power use of user 1 is beneficial for
its own rate but detrimental to user 2’s rate since user 2 treats interference as noise. Substituting (12) (13) into (8), we have
2
h
k
s1 (w1 ) = p1 α1 kΠh2 h1 k + β1 kΠ⊥
h2 1
√
√
= p1 kh1 k2 ( θα1 + 1 − θβ1 )2
|hH h1 |2
p1 α21 2
kΠh2 h1 k2
= p1 kh2 k2 α21
q
√
√
s2 (w2 ) = p2 kh2 k2 ( θα2 + 1 − θ 1 − α22 )2
r2 (w1 ) =
r1 (w2 ) = p2 α22
2
|hH
2 h1 |
= p2 kh1 k2 α22 ,
kΠh1 h2 k2
(14)
(15)
(16)
(17)
where the angle parameter θ between two effective channel vectors h1 and h2 is defined as
θ :=
2
|hH
1 h2 |
kh1 k2 kh2 k2
∈ [0, 1].
(18)
Substituting (14) - (17) into the problem (11) and taking square-root operation yield
max
(α1 ,β1 )∈F
α2 ∈[0,1]
0≤p1 ≤P
subject to
γ2 = min
√
(√
P − p1 kh1 kα2
p
,
σ12 (1 + γ1∗ )
)
p
√
√
√
P − p1 kh2 k( θα2 + 1 − θ 1 − α22 )
p
p1 kh2 k2 α21 + σ22
q
√
√
p1 kh1 k( θα1 + 1 − θβ1 ) = γ1∗ σ12 .
(19)
(20)
For later use, we define the following channel quality factor λi and the normalized target SINR value
for user 1, Γ:
λi :=
May 22, 2017
||hi ||2
, i = 1, 2, and Γ := γ1∗ /λ1 .
σi2
(21)
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
8
Here, λi indicates the signal-to-noise ratio (SNR) quality of user i’s channel, whereas θ in (18) is a
measure of the angle between the two users’ channels. Note that the actual target SINR for user 1 γ1∗
is given by γ1∗ = Γλ1 and the feasible range for Γ is Γ ∈ [0, P ], where the maximum P occurs when
2
2
w1 = h1 /||h1 || and p1 = P since γ1 = s1 (w1 , p1 )/σ12 = p1 |hH
1 w1 | /σ1 .
The optimization problem (19) - (20) with fixed power allocation p1 = p2 = 1 was solved in [21]
under the framework of a two-user MISO interference channel. In the MISO interference channel case,
two transmitters in the network neither cooperate nor share transmit power and hence the two transmit
power values p1 and p2 are fixed. On the other hand, in the MISO-BC case the two power values p1
and p2 can be designed at the BS under the constraints p1 , p2 ≥ 0 and p1 + p2 = P to maximize the
performance. Since our solution to the two-user MISO-NOMA-BC case is based on the result from [21],
we briefly introduce the relevant result in [21] for further development in later sections.
A. Background: The fixed power allocation case [21]
In this subsection, we fix p1 = p2 = 1 and follow [21]. First, note that α1 appears only in the constraint
(20) and the denominator in the second term in the RHS of (19). Thus, optimal α1 can be found by
solving the following problem [21]:
min
(α1 ,β1 )∈F
subject to
α1
(22)
q
√
√
kh1 k( θα1 + 1 − θβ1 ) = γ1∗ σ12 ,
(23)
This is because the second term in the RHS of (19) decreases monotonically with respect to α1 and
hence maximizing the second term in the RHS of (19) is equivalent to minimizing α1 . The problem (22)
- (23) can easily be solved based on the relationship between a line segment (23) and the unit-radius ball
F and the solution is given by [21]
0
α∗1 =
p
√
θΓ − (1 − θ)(1 − Γ)
if
Γ≤1−θ
if
Γ > 1 − θ,
(24)
where Γ is defined in (21) and Γ ∈ [0, 1] for p1 = 1. With the optimal α∗1 in (24), the corresponding
optimal β1∗ can be found by the constraint (20), and substituting the optimal α∗1 into the problem (19) (20) yields [21]
kh1 k
,
σ12 (1+γ1∗ )
where a := √
May 22, 2017
q
2
max γ2 = min aα2 , bα2 + c 1 − α2
b :=
α2 ∈[0,1]
√
√ kh22 k ∗ θ2 2 ,
kh2 k (α1 ) +σ2
(25)
√
kh2 k 1−θ
.
kh2 k2 (α∗1 )2 +σ22
and c := √
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
bα2 + c
p
1 − α22
9
aα2
case 3
aα2
case 2
aα2
case 1
α′2
1
α2
p
Fig. 1: aα2 and bα2 + c 1 − α22 (figure adapted from Fig. 1 in [21])
Note that the graph (α2 , aα2 ) of the first term aα2 in the minimum in (25) is a straight line and the graph
p
p
(α2 , bα2 +c 1 − α22 ) of the second term bα2 +c 1 − α22 in the minimum in (25) is a straight line plus a
p
√
quarter circle, as shown in Fig. 1. Note also that bα2 + c 1 − α22 is maximized at α′2 = b/ b2 + c2 with
p
p
√
maximum b2 + c2 , and the intersection of aα2 and bα2 + c 1 − α22 occurs at α′′2 = c/ c2 + (a − b)2 .
There exist three different cases for the solution to (25) depending on the relationship between the two
p
graphs, as shown in Fig. 1. In the first case of a ≤ b, aα2 is below bα2 + c 1 − α22 , the minimum of
p
the two is aα2 , and thus optimal α∗2 is 1. In the second case that aα2 and bα2 + c 1 − α22 intersect
p
between α′2 and 1, i.e., α′2 ≤ α′′2 , the solution to (25) occurs at α∗2 = α′′2 = c/ c2 + (a − b)2 . Finally,
p
in the third case that aα2 and bα2 + c 1 − α22 intersect between 0 and α′2 , i.e., α′2 > α′′2 , the solution
√
to (25) occurs at α∗2 = α′2 = b/ b2 + c2 . Since the condition α′2 ≤ α′′2 can be rewritten as [21]
√
b2
b
c
≤p
⇐⇒ a ≤ b + c2 /b,
2
2
+c
c + (a − b)2
the optimal solution α∗2 to the problem (25) is summarized as [21]
1
if a ≤ b (case 1),
√2 c
if b < a ≤ b + c2 /b (case 2),
α∗2 =
c +(a−b)2
√
√ b
θ
if a > b + c2 /b (case 3),
=
2
2
b +c
and the corresponding optimal value γ2∗ is given by [21]
∗(1)
kh1 k2
2
γ2 = a = σ12 (1+γ1∗ ) ,
2
∗(2)
1k
∗ 2
γ2∗ =
γ2 = a2 (α∗2 )2 = σ2kh
∗ (α2 ) ,
1 (1+γ1 )
kh2 k2
γ ∗(3) = b2 + c2 =
,
2
May 22, 2017
kh2 k2 (α∗1 )2 +σ22
(26)
(27)
case 1,
case 2,
(28)
case 3.
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
10
B. Pareto-optimal design in two-user MISO BC with SIC with power allocation
Now consider the actual problem (19) - (20) of Pareto-optimal beam design and power allocation
for the two-user MISO-NOMA BC. We obtain the solution to this problem by exploiting the result in
Section III-A. Note that once the power allocation values p1 and p2 are fixed, the corresponding optimal
solution can be obtained from the result in Section III-A. Therefore, we represent the optimal solution as
a function of power allocation and then optimize the power allocation. For given p1 ∈ [0, P ], the problem
(22) - (23) changes to
min
α1
subject to
√
(α1 ,β1 )∈F
(29)
q
√
√
p1 kh1 k( θα1 + 1 − θβ1 ) = γ1∗ σ12 ,
and the corresponding solution is given by α∗1 (p1 ) =
0
p
p
θΓ/p − (1 − θ)(1 − Γ/p )
1
1
if
Γ ≤ p1 (1 − θ),
if
Γ > p1 (1 − θ).
(30)
Furthermore, the coefficients a, b, and c defined in (25) are changed to
p
kh1 k
P − p1 p 2
,
σ1 (1 + γ1∗ )
√
p
kh2 k θ
,
P − p1 p
b(p1 ) :=
p1 kh2 k2 (α∗1 (p1 ))2 + σ22
√
p
kh2 k 1 − θ
c(p1 ) :=
P − p1 p
.
p1 kh2 k2 (α∗1 (p1 ))2 + σ22
a(p1 ) :=
(31)
(32)
(33)
Then, the optimal solution to (19) - (20) can be represented as a function of p1 :
1
if p1 ∈ P1 ,
c(p1 )
√2
if p1 ∈ P2 ,
α∗2 (p1 ) =
c (p1 )+[a(p1 )−b(p1 )]2
√
√ 2 b(p1 ) 2
if p1 ∈ P3 ,
= θ
(34)
{p1 |a(p1 ) > b(p1 ) + c2 (p1 )/b(p1 )}, and the corresponding SINR for
2
∗(1)
1k
γ2 (p1 ) = (P − p1 ) σ2kh
∗
(1+γ
1
1)
∗(2)
kh1 k2
γ2 (p1 ) = (P − p1 ) σ2 (1+γ ∗ ) [α∗2 (p1 )]2
1
1
kh2 k2
γ ∗(3) (p ) = (P − p )
∗
2
(35)
b (p1 )+c (p1 )
where P1 := {p1 |a(p1 ) ≤ b(p1 )}, P2 := {p1 |b(p1 ) < a(p1 ) ≤ b(p1 ) + c2 (p1 )/b(p1 )}, and P3 :=
2
1
1 kh2 k2 p1 [α (p1 )]2 +σ
1
2
user 2 γ2∗ (p1 ) is given by γ2∗ (p1 ) =
if p1 ∈ P1 ,
if p1 ∈ P2 ,
if p1 ∈ P3 .
Finally, the original problem (19) - (20) reduces to
max γ2∗ (p1 ),
0≤p1 ≤P
May 22, 2017
(36)
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
∗(i)
where γ2∗ (p1 ) is given by (35). Note that if we knew which γ2
11
in (35) to use for optimization by
directly computing a(p1 ), b(p1 ), c(p1 ) and their relationship, it would be easy to solve the problem (36).
∗(i)
However, the parameters a(p1 ), b(p1 ) and c(p1 ) to determine γ2
to use are functions of the design
variable p1 . Nevertheless, this is possible and the result is given in the following proposition.
Proposition 1: For given h1 , h2 , σ12 , σ22 , P and γ1∗ , we have the following regarding the optimal solution
√
√ √
1
1√
,
popt
to
the
problem
(36).
If
θΓ
<
τ
or
if
θΓ
≥
τ
≥
0
and
P
≥
Γ
+
θΓ
−
τ
)
θΓ
+
(
1
1−θ
λ2 τ
opt
−1 λ−1 + Γ − λ−1 , and Γ and λ are defined in
then popt
i
1 ∈ P2 . Otherwise, p1 ∈ P3 . Here, τ := θ
1
2
(21).
Proof:
See Appendix.
Due to Proposition 1 we know which of the three cases in (35) is applicable to the given combination
opt
of h1 , h2 , σ12 , σ22 , P , and γ1∗ . Once the set Pi to which popt
1 belongs is determined, optimal p1 can be
∗(i)
found by maximizing the corresponding γ2 (p1 ) in (35) with respect to p1 . A closed-form solution from
solving
dγ2∗(i) (p1 )
dp1
= 0 seems complicated but the solution can easily be found by a numerical method. The
proposed algorithm to design Pareto-optimal beam vectors and power allocation is summarized in Table
I. The Pareto-boundary of a two-user MISO-NOMA BC can be computed by sweeping R1∗ = log(1 + γ1∗ )
and computing the corresponding maximum R2∗ = log(1 + γ2∗ ). An example is shown in Fig. 2. It is seen
that power allocation significantly enlarges the achievable rate region over the fixed-power beam-only
design and optimal power allocation is crucial for MISO-NOMA BC.
3
Pareto-boundary with power allocation
Pareto-boundary with P1 =0.5 , P 2 = 1.5
Pareto-boundary with P =1.0 , P = 1.0
2.5
1
2
Pareto-boundary with P =1.5 , P = 0.5
1
2
R2
2
1.5
1
0.5
0
0
1
2
3
4
5
6
R1
Fig. 2: Pareto-boundary: fixed power versus power allocation (kh1 k2 /σ12 = 20, kh2 k2 /σ22 = 3, θ = 0.5
and P = 2)
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
12
TABLE I
Pareto-optimal design for 2-user MISO-BC with SIC
√
√
[ p1 w1 , p2 w2 ] = D(h1 , h2 , σ12 , σ22 , γ1∗ , P )
Input: channel vectors h1 , h2 , noise power σ12 , σ22 , target SINR of user 1 γ1∗ , and cluster total power P .
|hH h2 |2
−1
∗
−1
λ−1
Initialization: λ1 = kh1 k2 /σ12 , λ2 = kh2 k2 /σ22 , θ = kh1 k12 kh
2 , Γ = γ1 /λ1 , and τ = θ
1 + Γ − λ2
2k
if θΓ < τ
∗(2)
obtain popt
maximizing γ2
1
elseif τ ≥ 0 and
P ≥Γ+
1
(
1−θ
√
θΓ −
√
τ)
√
θΓ +
1√
λ2 τ
∗(2)
obtain popt
maximizing γ2
1
else
∗(3)
obtain popt
maximizing γ2
1
endif
Obtain α∗1 , β1∗ and α∗2 using (30), (20) and (34) with p1 = popt
1 , and obtain w1 and w2 from (12) and (13) with
α∗1 , β1∗ and α∗2 .
q
q
Output: popt
P − popt
w
and
1
1
1 w2
IV. T WO - USER MISO
BROADCAST CHANNEL WITH
SIC: P ERFORMANCE S TUDY
In the previous section, we developed a Pareto-optimal beam design and power allocation algorithm
for given channel vectors h1 and h2 . The performance of the Pareto-optimal design is a function of the
two (effective) channel vectors h1 and h2 . In the conventional ZF downlink beamforming with no SIC at
the receivers, two users with orthogonal channel vectors are preferred since non-orthogonality between
the two channel vectors reduces the effective SINR of ZF beamforming [15]. However, in the considered
MISO-NOMA framework in which user 1 intends to decode the interference from user 2 and subtracts
it before decoding its own data whereas user 2 treats the interference from user 1 as noise, orthogonality
between h1 and h2 , i.e., θ ≈ 0, and corresponding orthogonal beam vectors w1 and w2 (see (12) and
(13) with h1 ⊥ h2 ) do not necessarily imply high performance. Intuitively, if θ is small, user 2 receives
less interference from user 1, but user 1 has difficulty in decoding the message of user 2 for SIC under
the NOMA framework. In this section, we investigate more on the two-user MISO-NOMA BC before
proceeding to overall user scheduling in the next section.
To gain some insight into good channel conditions for two-user MISO-NOMA BCs, let us first consider
the fixed power allocation case with p1 = p2 = 1 as described in Section III-A and investigate the impact
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
13
of the angle θ between the two channel vectors when the magnitudes are given. For given ||h1 ||, ||h2 ||
and γ1∗ , the SINR of user 2 γ2∗ in (28) can be rewritten as a function of θ as
∗(1)
λ1
for case 1,
γ2 (= 1+Γλ1 )
∗
∗(2)
λ1
γ2 (θ) =
[α∗2 (θ)]2 )
for case 2,
γ2 (θ) (= 1+Γλ
1
∗(3)
λ2
γ (θ) (=
)
for case 3,
2
where
α∗1 (θ) =
and
(37)
λ2 [α∗1 (θ)]2 +1
0
√
p
θΓ − (1 − θ)(1 − Γ)
1
√ 2 c(θ)
α∗2 (θ) =
c (θ)+(a−b(θ))2
√
θ
if
Γ≤1−θ
if
Γ>1−θ
(38)
for case 1
for case 2
(39)
for case 3.
Regarding optimal θ that maximizes γ2∗ (θ) in (37), we have the following proposition:
Proposition 2: Let λ1 , λ2 and Γ be given. If Γ ∈ [Γ1 , Γ2 ], where
1
−1
Γ1 =
1 + λ−1
−
2 − λ1
2
1
−1
1 + λ−1
+
Γ2 =
2 − λ1
2
q
1
−1 2
−1
−1
(1 + λ−1
1 + λ2 ) − 4λ2 (1 + λ2 )
2
q
1
−1 2
−1
−1
(1 + λ−1
1 + λ2 ) − 4λ2 (1 + λ2 ),
2
(40)
then optimal θ that maximizes γ2∗ (θ) in (37) is given by the region
(
θ|θ0 ≤ θ ≤
z1 z2 + 2Γ(1 − Γ) +
p
4Γ(1 − Γ)[Γ(1 − Γ) + z1 z2 − z22 ]
2
z1 + 4Γ(1 − Γ)
−1
where z1 = λ−1
1 + 1 − Γ, z2 = λ2 + 1 − Γ, and
λ1 1 ,
λ2 1+Γλ1
√
θ0 =
z1 z2 +2Γ(1−Γ)− 4Γ(1−Γ)[Γ(1−Γ)+z1 z2 −z22 ] ,
z12 +4Γ(1−Γ)
if
λ1
1
λ2 1+Γλ1
)
≤ 1 − Γ,
(41)
(42)
otherwise.
If Γ ∈
/ [Γ1 , Γ2 ], optimal θ is given by the region
Proof:
λ2
λ1 (1 + Γλ1 )
(2)
{θ | ∂γ2∂θ(θ) =
{θ |
≤ θ ≤ 1 − Γ}
if Γ ≤
or
0}
if Γ >
−1
λ−1
2 −λ1
1+λ−1
2
−1
λ−1
2 −λ1
1+λ−1
2
and Γ ∈
/ [Γ1 , Γ2 ],
(43)
and Γ ∈
/ [Γ1 , Γ2 ].
See Appendix.
From Proposition 2 we obtain a more insightful corollary as follows:
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
14
Corollary 1: Let λ1 , λ2 and Γ be given. When λ1 = λ2 , optimal θ for the 2-user MISO BC with SIC
is given by the set {θ | θ0 ≤ θ ≤ 1} with θ0 reduced to
1
if
1+Γλ1
θ0 =
2
2 z1
if
z +4Γ(1−Γ)
1
Proof:
1
1+Γλ1
1
1+Γλ1
≤1−Γ
.
(44)
>1−Γ
See Appendix.
Corollary 1 states that in two-user MISO BCs with SIC with the same channel magnitudes λ1 = λ2 and
the same power p1 = p2 = 1, two aligned channel vectors are preferred to two orthogonal channels and
channel alignment beyond a certain angle is all optimal.
Although Proposition 2 and Corollary 1 provide some insight into good channel conditions in the
SIC BC case, the assumptions for Proposition 2 and Corollary 1 are not valid in the actual NOMA
situation in which power allocation is applied. Unfortunately, the optimal power popt
1 was not obtained
in closed form in the previous section and this puts difficulty on analysis of the impact of channel
angle on the performance. Hence, in the actual case, to enable analysis we derive the SINR γ2∗ for user
2 as a function of θ by assuming the simple power allocation method that assigns minimum power
p1,min (= γ1∗ /λ1 = Γ) to achieve the target SINR γ1∗ to user 1 and assigns the rest of power P to user
2. The simple power allocation strategy is based on the assumption that user 1 has a strong channel and
is limited by channel’s DoF such as bandwidth, whereas user 2 with a weak channel is limited by noise
and needs to receive more power. For this simple power allocation method, the optimal γ2∗ is obtained
by substituting p1 = p1,min = Γ into (35) and given after some manipulation in closed form as
γ2∗ =
P −Γ
1
−1 +1
Γ λ−1
1 Γ
"
1
P −Γ
,
−1
Γ θ+λ−1
2 Γ
1+
θ
1−θ
r
−1 θ −1
1+λ−1
2 Γ
−1
1+λ−1
1 Γ
2 #−1
−1
,
if θ ≤ θ1
(45)
if θ > θ1 ,
q
−1 −1
−2 −2
−1 −1
1
where θ1 := 2 −λ2 Γ + λ2 Γ + 4(λ1 Γ + 1) .
Examples of γ2∗ in (45) as a function of θ are shown in Fig. 3 together with the optimal γ2∗ obtained
by running the algorithm in Table I. It is seen that the γ2∗ behavior depends on the relative magnitude of
λ1 and λ2 through the two performance limiting factors: the SIC processing at user 1 and the SINR of
user 2, as seen in (11). In the case of λ1 = λ2 ≫ 0, the performance of user 2 is not limited by noise at
user 2 but is limited by signal-to-interference ratio (SIR). Thus, in this case it is preferred that channel
vectors are aligned and more power is allocated to user 2 under the constraint that the required SINR
for user 1 is satisfied. By doing so, SIC at user 1 is easy and SIR at user 2 is high. This behavior is
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
3.9
4
3.8
3.5
15
1
0.8
γ*2
γ*2
γ*2
0.6
3.7
3
0.4
2.5
3.6
0.2
3.5
0
0.2
0.4
0.6
The optimal power allocation
The simple power allocation
The optimal power allocation
The simple power allocation
The optimal power allocation
The simple power allocation
0.8
1
2
0
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
θ
θ
θ
(a)
(b)
(c)
0.8
1
Fig. 3: γ2∗ in (45) as a function of θ : (a) λ1 = 10, λ2 = 10, Γ = 2, P = 10 , (b) λ1 = 10, λ2 = 1, Γ = 2,
P = 10, and (c) λ1 = 10, λ2 = 0.1, Γ = 2, P = 10.
evident in Fig. 3(a). It is seen that the optimal power control and the simple power allocation strategy
yield similar performance in Fig. 3(a). (The behavior for λ1 = λ2 with power control seems similar to
that stated in Corollary 1.) However, in the medium asymmetric case of λ1 > λ2 as in Fig. 3(b), there
is a trade-off between the two performance limiting factors. When two channels are orthogonal, SIC at
user 1 is difficult. On the other hand, when two channels are aligned, interference from user 1 to user 2
at user 2 is high. Hence, the performance is good when the two user channels are neither too orthogonal
nor too aligned. This behavior is evident in Fig. 3(b). Note that in this case, there is a large gap between
the optimal power control and the simple power allocation.
In addition to the above simple power allocation result, we have another result exploiting the fact
λ1 ≫ λ2 in actual NOMA, given by the following proposition:
→
Proposition 3: For given λ1 and γ1∗ , if θ 6= 0, as λ2 → 0, the optimal power coefficient popt
1
1
p1,min = Γ; the corresponding γ2∗ converges to γ2∗ = P Γ−Γ θ+λ−1
; and the beam vectors converge to
−1
2 Γ
√
√
√
√
p1 w1 = Γ khh11 k and p2 w2 = P − Γ khh22 k . That is, both users use matched-filtering beams.
Proof:
See Appendix.
Note that γ2∗ in Proposition 3 coincides with the second formula in (45). This is because θ1 in (45)
converges to 0+ as λ2 → 0 for given λ1 and Γ, and the second formula in (45) is valid in this case. By
Proposition 3, if λ2 is sufficiently small compared to λ1 , matched filtering beams for both users with
minimum power to user 1 satisfying the target SINR are optimal regardless of the angle between the two
channel vectors. This is because if λ1 ≫ λ2 , the limitation for γ2∗ results from the SINR of user 2 at user
2. Hence, maximum power should be delivered to user 2 with matched filtering beam w2 by assigning
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
16
(2)
(2)
h̃1
h̃1
(3)
h̃1
(1)
h̃1
(2)
(1)
h̃2
(1)
h̃2
Π⊥ (2)
[h̃1
(2)
(3) h̃1
,h̃1 ]
(1)
Π⊥ (2) (3) h̃1
[h̃
,h̃
]
1
1
(2)
(a)
L⊥ (h̃1 )
(1)
h̃1
(2)
Π⊥(2) h̃1
(1)
Π⊥(2) h̃1
h̃1
h̃
(1)
w1
1
(1)
w2
(3)
L⊥ ([h̃1 , h̃1 ])
(b)
Fig. 4: Beam design in the example of Nt = 3: (a) Kc = Nt and (b) Kc = Nt − 1
minimum power to user 1 with matched filtering beam w1 . This behavior is evident in Fig. 3(c). It is
seen in Fig. 3(c) that the simple power allocation method almost achieves the optimal performance for
all θ .
V. T HE P ROPOSED S CHEDULING M ETHOD
FOR
K -U SER MISO-NOMA D OWNLINK
Now, we propose our overall user scheduling/pairing and beam design method for the K -user MISONOMA downlink with Nt BS antennas based on the results in the previous sections. In MISO-NOMA
downlink scheduling and beam design, two major aspects should be taken into account simultaneously
to guarantee good system performance: One is controlling ICI to reduce interference from other clusters
and the other is pairing and beam design for the paired users in each cluster for maximum performance.
Recall that the gain of NOMA lies in the case that strong users are in the DoF(such as bandwidth)-limited
regime and weak users are limited by noise [2, P. 239]. In MU-MISO NOMA, the beneficial situation
for NOMA should be maintained, i.e., the high channel quality for strong users should be maintained
by SIC and proper interference control, and low SINRs of weak users should be leveraged by assigning
more power to weak users. To be consistent with this design principle, ICI should be eliminated for
strong users by proper measures. With these considerations, we propose the following user scheduling,
pairing and beam design method for K -user MISO-NOMA downlink composed of two steps under the
assumption that all channel vector information is available at the BS and the thermal noise variance is
known.
Algorithm 1: Overall User Scheduling and Beam Design
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
17
Step 1: In the first step, we run the semi-orthogonal user selection (SUS) algorithm [15] targeting
selection of Nt users from the strong user set K1 . Then, the SUS algorithm returns Kc (≤ Nt ) users
with roughly orthogonal channel vectors out of the K/2 users in K1 . (Depending on the size |K1 | and
the semi-orthogonality parameter, the SUS algorithm may return users less than Nt especially for large
Nt although we target selecting Nt users [15].) We set these Kc users returned by the SUS algorithm
as the Kc strong users in Kc clusters (one user for each cluster). Let their actual channel vectors be
(1)
(Kc )
h̃1 , · · · , h̃1
.
Step 2: Weak user selection and overall beam design
Initialization: Γ is given.
K2 = {1, . . . , K/2} (the original weak user set),
(46)
S1 : the set of selected strong users from step 1,
(47)
S2 ← φ (the set of selected weak users),
(1)
⊥ h̃(Kc )
Π
h̃
Π⊥
1
H̃Kc
H̃1 1
c =
,
W
,··· ,
(Kc )
(1)
⊥
||ΠH̃ h̃1 ||
||Π⊥
h̃
||
1
H̃
(48)
1
f1 = [ ]
W
k = 1.
(49)
Kc
f 2 = [ ],
and W
(50)
(51)
Iteration:
while k ≤ Kc do
S.1: For each user u ∈ K2 , estimate ICI plus AWGN:
σ̂u2 =
X
l<k
f 1 (l)|2 + |g̃uH W
f 2 (l)|2 ) +
(|g̃uH W
X
l>k
c 2 + ǫ2u ,
P |g̃uH W(l)|
(52)
c
f 1 (l) and W
f 2 (l) are the l-th columns of W,
c W
f 1 and W
f 2 respectively, and g̃u is
where W(l),
W
the actual channel vector of user u in K2 . With obtained σ̂u2 and given Γ, compute the maximum
(k)
SINR γ2∗ (u) of user u when user u is paired with the channel h̃1
described in Section III-B or Section IV by setting λ1 =
2
θ=
May 22, 2017
g̃ )
h̃(k) )H (Π⊥
(Π⊥
H̃k u
H̃k 1
(k) 2
⊥
g̃
k2
h̃
k
kΠ
kΠ⊥
H̃k u
H̃k 1
of user 1 of cluster k, as
h̃(k) k2
kΠ⊥
H̃k 1
σ12
, λ2 =
g̃u k2
kΠ⊥
H̃
k
σ̂u2
, and
.
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
18
S.2: Select the weak user of cluster k as follows:
u∗ = arg max γ2∗ (u)
(53)
S2 ← S2 ∪ {u∗ }
(54)
u∈K2
(k)
h̃2 = g̃u∗
(55)
√
(k)
(k)
(k) √
(k)
h̃ , σ12 , σ̂u2 ∗ , λ1 Γ, P ) by using the twoand design [ p1 w̃1 , p2 w̃2 ] = D(Π⊥
h̃ , Π⊥
H̃ 2
H̃ 1
k
k
user Pareto-optimal design algorithm in Table. I.
S.3: Store the designed beams, remove u∗ in K2 , and repeat until k = Kc :
f 1 ← [W
f 1 , √p1 w̃(k) ], W
f 2 ← [W
f 2 , √p2 w̃(k) ]
W
1
2
K2 ← K2 \ {u∗ }
k ←k+1
end while
Remark 1: Note that the computation of γ2∗ (u) and the Pareto-optimal beam design for two users in
each cluster in steps S.1 and S.2 in the while loop of Step 2 of Algorithm 1 is based on the projected
(k)
(k)
h̃ lie in L⊥ (H̃k ). Thus,
h̃ and Π⊥
effective channels. Note that the projected effective channels Π⊥
H̃ 2
H̃ 1
k
k
(k)
the corresponding Pareto-optimal beams w1
(k)
beams (see (12) and (13)) [19], and hence wi
(k)
and w2
lie in L⊥ (H̃k ) by the property of Pareto-optimal
(k)
wi
= Π⊥
H̃
k
(k)
= w̃i , i = 1, 2 for (4). Hence, there exists
no ICI to all strong users with the proposed beam design.
(k)
h̃
Remark 2: If Kc = Nt , then H̃k has nullity of one, and Π⊥
H̃ 1
k
(k)
h̃ are aligned, as shown
and Π⊥
H̃ 2
k
in Fig. 4(a). In this case, only Pareto-optimal power control is applied for each cluster by the proposed
design method. (Note that the algorithm in Table I is still applicable in case of two aligned input channel
vectors.) On the other hand, if the number Kc of the returned users by the SUS algorithm in Step 1 is
less than Nt (which is often true for large Nt with small K [15]), then H̃k has nullity larger than or
(k)
h̃
equal to two. In this case, the projected effective channels Π⊥
H̃ 1
k
(k)
h̃
and Π⊥
H̃ 2
span a 2-dimensional
k
(2-D) space and the full 2-D Pareto-optimal beam design is applicable for each cluster, as shown in Fig.
4(b). This is another advantage of the proposed method over the previous methods [12]–[14] based on
simple spatial ZF beam design ignoring the case of Kc < Nt .
Remark 3: In the step (52) of computation of ICI and AWGN for each candidate weak user for cluster k,
√
Π⊥ h̃1(k+1)
already designed beam vectors are used up to cluster k−1 and the beam estimates P ⊥H̃1 (k+1) , · · · ,
√
P
Π⊥
H̃
Kc
||Π⊥
H̃
May 22, 2017
Kc
c)
h̃(K
1
c)
h̃(K
||
1
||ΠH̃
k+1
h̃1
||
are used for undesigned clusters k + 1, · · · , Kc .
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
19
(k)
Remark 4: Note that there is length reduction from the actual channel h̃1
(k)
ΠH̃k h̃1
to the effective channel
of each strong user. This reduction is the typical effective gain loss associated ZF, but the loss
(1)
(Kc )
is not significant because the strong channels h̃1 , · · · , h̃1
are semi-orthogonal by the SUS algorithm
[15]. Only the weak users experience ICI whereas ICI to the strong users is completely removed in the
proposed method to be consistent with the NOMA design principle. However, the weak users are selected
by considering all the factors, i.e., the ICI, projection onto L⊥ (H̃k ) and the friendliness with the strong
users to yield good performance.
VI. N UMERICAL
RESULTS
In this section, we provide some numerical results to evaluate the performance of the proposed
scheduling, beam design and power allocation method described in Section V for MU-MISO-NOMA
downlink.
First, we evaluated the gain of the proposed method over conventional SUS-based MU-MISO scheduling [15] and the results are shown in Figs. 5 and 6. The simulation setup for Figs. 5 and 6 is as follows.
The AWGN variance was one for all users. The numbers of transmit antennas were two and four for Fig.
5 and Fig. 6, respectively. Each element of each channel vector in the strong user set K1 with K/2 users
2 ) with σ 2 = 1 and each element of each
was randomly and independently generated from CN (0, σh,1
h,1
channel vector in the weak user set K2 with K/2 users was randomly and independently generated from
2 ) with σ 2 = 0.01. (Hence, we have λ /λ = 100 = 20 dB.) For the conventional SUS-based
CN (0, σh,2
1
2
h,2
MU-MISO scheduling we considered two scheduling intervals. At the first interval, user scheduling out
of K1 was performed by running the SUS algorithm for MU-MISO with Nt transmit antennas and the
scheduled users were served by ZF downlink beamforming [15]. At the second interval, user scheduling
out of K2 was performed by running the SUS algorithm for MU-MISO with Nt transmit antennas and the
scheduled users were served by ZF downlink beamforming. The average sum rates for K1 and K2 were
obtained by averaging the rates of 1000 independent channel realizations. For the overall sum rate of the
conventional SUS method, the two rates of K1 and K2 were averaged. For the proposed NOMA method,
the same 1000 channel realizations used for the conventional method were used but the proposed NOMA
scheduling was performed over the overall user set K1 ∪ K2 in a single scheduling interval. For the SUS
algorithm applied separately to K1 and K2 and to Step 1 of the proposed method, the semi-orthogonality
parameter δ should be chosen [15] and we used optimal δ for each K/2 provided from Fig. 2 of [15].
In addition, for the proposed method, Γ should be chosen and we set Γ appropriately to balance the
rates from the two groups K1 and K2 . It is seen in Figs. 5 and 6 that the proposed MU-MISO NOMA
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
18
18
The proposed alogirhtm with NOMA, K=10000
SUS algorithm with linear precoding, K=10000
The proposed alogirhtm with NOMA, K=2000
SUS algorithm with linear precoding, K=2000
The proposed alogirhtm with NOMA, K=200
SUS algorithm with linear precoding, K=200
14
12
10
8
6
14
12
10
users scheduled from K1
8
6
4
4
2
2
0
5
10
The proposed alogirhtm with NOMA, K=10000
SUS algorithm with linear precoding, K=10000
The proposed alogirhtm with NOMA, K=2000
SUS algorithm with linear precoding, K=2000
The proposed alogirhtm with NOMA, K=200
SUS algorithm with linear precoding, K=200
16
Sum rate [bits / channel use]
Sum rate [bits / channel use]
16
0
20
15
0
20
users scheduled from K2
0
5
PT (dB)
10
15
20
PT (dB)
(a)
(b)
Fig. 5: Sum rate (Nt = 2) : (a) total sum rate and (b) separate sum rates from K1 and from K2
30
30
The proposed alogirhtm with NOMA, K=10000
SUS algorithm with linear precoding, K=10000
The proposed alogirhtm with NOMA, K=2000
SUS algorithm with linear precoding, K=2000
The proposed alogirhtm with NOMA, K=200
SUS algorithm with linear precoding, K=200
25
Sum rate [bits / channel use]
Sum rate [bits / channel use]
25
20
15
10
20
users scheduled from K1
15
10
users scheduled from K2
5
5
0
The proposed alogirhtm with NOMA, K=10000
SUS algorithm with linear precoding, K=10000
The proposed alogirhtm with NOMA, K=2000
SUS algorithm with linear precoding, K=2000
The proposed alogirhtm with NOMA, K=200
SUS algorithm with linear precoding, K=200
0
5
10
15
PT (dB)
(a)
20
0
0
5
10
15
20
PT (dB)
(b)
Fig. 6: Sum rate (Nt = 4) : (a) total sum rate and (b) separate sum rates from K1 and from K2
method outperforms the conventional MU-MISO downlink based on the SUS user scheduling. Note that
the sum rates for both groups K1 and K2 are better than the conventional scheme. It is also seen that
the performance improvement by NOMA reduces as Nt increases from two to four, but there exists
non-trivial gain for NOMA for large K .
Next, we compared the proposed algorithm with an existing algorithm proposed for MU-MISO NOMA
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
21
45
The propoposed algorithm
NOMA-FOUS algorithm
Sum rate [bits / channel use]
40
35
Nt = 8
30
Nt = 4
25
20
Nt = 2
15
10
40
60
80
100
120
140
160
Number of users (K)
Fig. 7: Sum rate (Nt = 2, 4, 8, K = 2000 and PT = 15dB)
downlink. For the comparison baseline we considered the NOMA-FOUS algorithm in [14] of which
superiority over other methods [12], [13] is shown in [14]. Since the simulation setting is different from
that in [14], we slightly modified the NOMA-FOUS algorithm so that the strong user is selected from
K1 and the weak user is selected from K2 , although the original NOMA-FOUS algorithm considers only
2 = 1 and σ 2 = 0.04, and set Γ to make the sum rate of the
one set of users. For comparison, we set σh,1
h,2
weak users of the proposed algorithm larger than the sum rate of the weak users of the NOMA-FOUS
algorithm. Fig. 7 shows the sum rate performance of the two algorithms. It is seen that the proposed
algorithm outperforms the NOMA-FOUS algorithm.
The key feature of our Pareto-optimality-based design is that we have control over the rate operating
point. Hence, we finally investigated the rate balancing property between the two groups K1 and K2
by controlling the strong-user-target-SINR parameter Γ defined in (21) (larger Γ means larger rates for
strong users), and the result is shown in Fig. 8. The simulation parameters are the same as those for
Fig. 6. For reference, the rates of K1 and K2 separately obtained by the conventional MU-MISO SUS
algorithm are shown. It is seen that by abandoning the improvement for weak users but maintaining the
weak-user performance at the level of the conventional SUS method, significant rate gain can be attained
for strong users.
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
22
25
Sum rate [bits / channel use]
20
15
The proposed algorithm : strong users
The proposed algorithm : weak users
SUS algorithm : strong users
SUS algorithm : weak users
10
5
0
1
1.5
2
2.5
3
3.5
4
Γ
Fig. 8: Sum rate versus Γ (Nt = 4, K = 2000 and PT = 20 dB)
VII. C ONCLUSION
In this paper, we have considered the problem of transmit beam design and user scheduling for MUMISO NOMA downlink and proposed an effective beam design and user scheduling method based
on Pareto-optimality by exploiting both the spatial and power domains available in MU-MISO NOMA
downlink. The proposed method with the ability of rate control between strong and weak users provides
great flexibility to NOMA network operation.
A PPENDIX
Proof of Proposition 1:
The set Pi to which popt
belongs is dependent on the relationship among
1
a(p1 ), b(p1 ) and d(p1 ) := b(p1 ) + c2 (p1 )/b(p1 ), given in terms of Γ, θ and λi by
s
r
p
kh1 k2
λ1
a(p1 ) = P − p1
= (P − p1 )
2
∗
1 + Γλ1
σ1 (1 + γ1 )
s
p
kh2 k2 θ
b(p1 ) = P − p1
kh2 k2 p1 [α∗1 (p1 )]2 + σ22
s
λ2 θ
= (P − p1 )
λ2 p1 [α∗1 (p1 )]2 + 1
s
p
kh2 k2 /θ
d(p1 ) = P − p1
kh2 k2 p1 [α21 (p1 )]2 + σ22
s
λ2 /θ
= (P − p1 )
.
λ2 p1 [α∗1 (p1 )]2 + 1
May 22, 2017
(56)
(57)
(58)
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
23
The three sets P1 , P2 and P3 can be rewritten by squaring a(p1 ), b(p1 ) and d(p1 ) and dropping the
¯ 1 )}, and P3 = {p1 |ā >
common factor (P − p1 ) as P1 = {p1 |ā ≤ b̄(p1 )}, P2 = {p1 |b̄(p1 ) < ā ≤ d(p
¯ 1 )}, where
d(p
λ1
λ2 θ
, b̄(p1 ) =
, and
∗
1 + Γλ1
λ2 p1 [α1 (p1 )]2 + 1
λ2 /θ
¯ 1) =
d(p
.
λ2 p1 [α∗1 (p1 )]2 + 1
ā =
(59)
(60)
First, we show popt
/ P1 . Let p1,min denote the minimum p1 to achieve γ1∗ with w1 = h1 /||h1 ||. Then,
1 ∈
p1,min = γ1∗ /λ1 = Γ. Hence, the second condition in (30), i.e., Γ > p1,min (1 − θ) or Γ = p1,min (1 − θ)
√
is satisfied since 0 ≤ θ ≤ 1, and α∗1 (p1,min ) = θ from (30). Hence, we have
ā =
b̄(p1,min ) =
By the NOMA condition λ1 > λ2 , we have
1
λ1
λ1
=
1 + Γλ1
λ2 θ
=
λ2 Γθ + 1
<
1
θλ2
1
λ1
1
+Γ
1
θλ2
(61)
1
.
+Γ
(62)
since 0 ≤ θ ≤ 1 and thus ā > b̄(p1,min ). In case
of Γ = p1,min (1 − θ), we have θ = 0 and thus b̄(p1,min ) = 0 and ā > b̄(p1,min ). Hence, p1,min ∈
/ P1 .
Note that ā is constant over p1 . It can be shown from (30) that the term p1 [α∗1 (p1 )]2 in the denominator
of b̄(p1 ) in (59) is monotone decreasing with respect to p1 , and hence b̄(p1 ) is monotone increasing with
respect to p1 . If ā > b̄(p1 ) for all p1 , P1 is empty. Otherwise, there exists p1 , denoted as p1,a , such that
ā = b̄(p1 ), as p1 increases, given by
p1,a
= {p1 |ā = b̄(p1 )}
λ2 θ
λ1
=
}
1 + Γλ1
λ2 p1 [α∗1 (p1 )]2 + 1
2
q
√
1
−1
= Γ+
.
θΓ − θΓ + λ−1
θ
−
λ
1
2
1−θ
= {p1 |
∗(2)
At p1 = p1,a , we have γ2
∗(1)
= γ2
(63)
from (35) since α∗2 (p1 ) = 1 at the boundary of P1 (p1 ≥ p1,a
∂γ2∗(2) (p1 )
= −1
∂p1
c1 ,
p1 →p−
1,a
∗(2)
P2 such that γ2 (p1 ) >
side) and P2 (p1 < p1,a side), i.e., ā = b̄(p1 ). Furthermore, it can be shown that
where c1 is a non-negative constant with respect to p1 . Hence, there exists p1 ∈
∗(2)
γ2
∗(1)
(p1,a ) = γ2
∗(1)
(p1,a ). Since γ2
is a monotone decreasing function of p1 as seen in (35), optimal
γ2∗ does not occur in P1 , i.e., popt
/ P1 .
1 ∈
¯ 1)
Next, we check the condition that P3 is empty. Since the term p1 [α∗1 (p1 )]2 in the denominator of d(p
¯ 1 ) is monotone increasing with respect
in (60) is monotone decreasing with respect to p1 , and thus d(p
√
¯ 1,min ), then P3 is empty. Since α∗ (p1,min ) = θ from (30) and p1,min = Γ,
to p1 . Therefore, if ā < d(p
1
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
24
the condition is rewritten from (59) and (60) as
λ1
λ2 /θ
<
1 + Γλ1
λ2 Γθ + 1
1 1
1
⇔ θΓ <
+Γ −
=: τ.
θ λ1
λ2
¯ 1,min ) ⇔
ā < d(p
(64)
(65)
opt
In this case, P3 = ∅ and popt
/ P1 .
1 ∈ P2 since p1 ∈
Now assume θΓ ≥ τ . Then, P3 is not empty. Furthermore, we have a sufficient condition for ∀ p1 ∈ P3
as follows:
¯ 1 ) < ā,
d(p
λ1
∀p1 ⇐ λ2 /θ <
1 + Γλ1
1 1
1
⇔
+Γ −
<0
θ α1
λ2
(66)
(67)
⇔ τ < 0,
(68)
¯ 1 ) (see (60)). In this case, popt ∈ P3 .
because λ2 /θ is an upper bound of d(p
1
¯ 1 ), given by
Finally, if θΓ ≥ τ and τ ≥ 0, compute p1 , denoted by p1,b , such that ā = d(p
¯ 1 )}
p1,b = {p1 |ā = d(p
(69)
λ2 /θ
λ1
=
}
1 + Γλ1
λ2 p1 [α∗1 (p1 )]2 + 1
√ 2
1 √
θΓ − τ .
= Γ+
1−θ
= {p1 |
If
(70)
(71)
P < p1,b ,
(72)
¯ 1 ) for p1 ≤ P since d(p
¯ 1 ) is a monotone increasing function of p1 . Hence, in this case,
then ā > d(p
∀p1 ∈ P3 and popt
1 ∈ P3 . On the other hand, if p1,b ≤ P , we have both nonempty P2 = {p1 ≥ p1,b } and
∗(2)
P3 = {p1 < p1,b }. In this case, we compute the derivatives of γ2
∗(3)
and γ2
at point p1,b , which are
given by
∗(2)
∂γ2
∂p1
p1 =p+
1,b
∗(3)
∂γ2
∂p1
p1 =p−
1,b
√
√ √
√
1−θ
1−θ
P − λ2 τ √
= c2 λ2 τ √
√
√ · Γ + λ2 τ · θΓ + 1
θΓ − τ
θΓ − τ
√
√ √
√
1−θ
1−θ
P − λ2 τ √
= c3 λ2 τ √
√
√ · Γ + λ2 τ · θΓ + 1 ,
θΓ − τ
θΓ − τ
(73)
(74)
where c2 and c3 are non-negative constants. The two derivatives have the same sign. If the two derivatives
are positive, then γ2∗ increases as p1 crosses p1,b from the left to the right and hence popt
1 ∈ P2 . Otherwise,
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
25
γ2∗ increases as p1 crosses p1,b from the right to the left and hence popt
1 ∈ P3 . Equivalently, we have
√ √
1 √
1
( θΓ − τ )( θΓ + √ )
(75)
1−θ
λ2 τ
√ √
1
1 √
(
(76)
θΓ
−
τ )( θΓ + √ ).
popt
∈
P
if
P
<
Γ
+
3
1
1−θ
λ2 τ
√
√
√ 2
√ √
1
1
θΓ − τ < Γ + 1−θ
Since p1,b = Γ + 1−θ
( θΓ− τ )( θΓ+ λ21√τ ), the set {P < p1,b } mentioned
√
√ √
1
in (72) is a subset of the set {P < Γ + 1−θ
( θΓ − τ )( θΓ + λ21√τ )}. Thus, the case of P < p1,b
popt
1 ∈ P2 if P ≥ Γ +
is covered by (76). The only two cases for popt
∈ P2 are [θΓ < τ ] or [θΓ ≥ τ ≥ 0 and P ≥
1
√
√ √
1
1√
Γ + 1−θ ( θΓ − τ )( θΓ + λ2 τ )]. Hence, the claim follows.
To prove Proposition 2, we introduce the following lemma.
Lemma 1: Define
N1 := {θ | a ≤ b(θ)}
(77)
N2 := {θ | b(θ) < a ≤ b(θ) + c2 (θ)/b(θ)}
(78)
N3 := {θ | a > b(θ) + c2 (θ)/b(θ)},
(79)
where a, b(θ) and c(θ) are defined just below (25). Then, for given λ1 , λ2 and Γ, every θ in N1 achieves
the same optimal γ2∗ (θ) in (37), if N1 is not empty.
Proof:
fact that
∗(2)
γ2
(θ)
∗(1)
Let us assume that N1 is not empty. For every θ ∈ N1 , γ2∗ (θ) = γ2
∗(1)
λ1
γ2 = 1+λ
1Γ
∗(1)
< γ2 for all
and
∗(2)
γ2 (θ)
=
λ1
∗
2
1+λ1 Γ [α2 (θ)] ,
==
λ1
1+λ1 Γ .
From the
and 0 ≤ α∗2 (θ) < 1 for θ ∈ N2 , it is obvious that
θ ∈ N2 . Hence, γ2∗ (θ) for θ ∈ N2 is less than γ2∗ (θ) for θ ∈ N1 . Furthermore,
∗(3)
for any θ ∈ N3 , we have γ2
(a)
(θ) =
λ2
λ2 [α∗1 (θ)]2 +1
(b)
≤
(c)
λ2 /θ
λ2 [α∗1 (θ)]2 +1 =
(d)
(e)
∗(1)
[b(θ) + c2 (θ)/b(θ)]2 < a2 = γ2
.
Here, step (a) is by (28), step (b) holds because θ ∈ [0, 1], step (c) is by direction computation based on
b(θ) and c(θ), step (d) holds because θ ∈ N3 , and step (e) is by (28). Consequently, we have the claim.
Proof of Proposition 2: The necessary and sufficient condition for N1 defined in (77) being non-empty
is given by
a2 ≤ max b2 (θ),
0≤θ≤1
(80)
where
a2 =
λ1
λ2 θ
λ2 (1 − θ)
, b2 (θ) =
c2 (θ) =
.
∗
2
1 + Γλ1
λ2 [α1 (θ)] + 1
λ2 [α∗1 (θ)]2 + 1
Since b(θ) is maximized at θ =
max0≤θ≤1 b2 (θ) = λ2 (1 +
May 22, 2017
λ2 Γ
λ2 +1 ),
[λ2 (1−Γ)+1]2
λ22 Γ(1−Γ)+[λ2 (1−Γ)+1]2
(81)
and the corresponding maximum value is
the condition (80) becomes
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
26
λ1
λ2 Γ
≤ λ2 (1 +
)
1 + λ1 Γ
λ2 + 1
1q
1
−1 2
−1
−1
−1
−1
(1 + λ−1
1 + λ2 − λ1 −
⇔ Γ1 :=
1 + λ2 ) − 4λ2 (1 + λ2 ) ≤ Γ
2
2
1q
1
−1 2
−1
−1
−1
−1
≤
(1 + λ−1
1 + λ2 − λ1 +
1 + λ2 ) − 4λ2 (1 + λ2 ) =: Γ2 .
2
2
In the case of non-empty N1 , by substituting α∗1 (θ) in (24) or (85) into b2 (θ), b2 (θ) is given by
λ2 θ
if θ ≤ 1 − Γ
(82)
b2 (θ) =
λ
θ
√ 2
√
if
θ
>
1
−
Γ
2
λ2 [ θΓ−
(1−θ)(1−Γ)] +1
When θ ≤ 1 − Γ, b2 (θ) is linear and it can be shown that b2 (θ) is quasi-concave function when θ > 1 − Γ
∗.
If a2 > λ2 (1 − Γ), there doesn’t exists θ satisfying a2 ≤ λ2 θ (for θ ≤ 1 − Γ) and hence the set
N1 = {θ | a2 ≤ b2 (θ)} is given by
p
z1 z2 + 2Γ(1 − Γ) − 4Γ(1 − Γ)[Γ(1 − Γ) + z1 z2 − z22 ]
θ|
z12 + 4Γ(1 − Γ)
p
z1 z2 + 2Γ(1 − Γ) + 4Γ(1 − Γ)[Γ(1 − Γ) + z1 z2 − z22 ]
≤θ≤
z12 + 4Γ(1 − Γ)
(83)
Otherwise, the minimum of θ satisfying a2 ≤ b2 (θ) is the point such that a2 = λ2 θ and N1 becomes
(
)
p
λ1
1
z1 z2 + 2Γ(1 − Γ) + 4Γ(1 − Γ)[Γ(1 − Γ) + z1 z2 − z22 ]
,
θ|
≤θ≤
λ2 1 + λ1 Γ
z12 + 4Γ(1 − Γ)
−1
where z1 = λ−1
1 +1−Γ and z2 = λ2 +1−Γ. ((83) is obtained by solving
λ1
1+Γλ1
≤
√ λ2 θ
√
λ2 [ θΓ− (1−θ)(1−Γ)]2 +1
reducing to a quadratic inequality). In the case of non-empty N1 , by Lemma 1, N1 is optimal and we
obtain (41).
Next, consider the case that N1 is empty. At θ = 1, we have a(1) =
q
1
Γ+ λ1
1
>
q
1
Γ+ λ1
= b(1) +
2
c2 (1)/b(1) by the NOMA assumption λ1 > λ2 and hence θ = 1 ∈ N3 by the definition of N3 in (79).
We also have
lim b(θ) + c2 (θ)/b(θ) = lim
θ→0
∗
θ→0
s
λ2 θ −1
= ∞,
λ2 [α∗1 (θ)]2 + 1
When θ > 1 − Γ, b2 (θ) can be written as f 2 (θ)/g(θ), where f (θ) =
√
(84)
p
√
λ2 θ and g(θ) = λ2 [ θΓ − (1 − θ)(1 − Γ)]2 + 1.
Since f (θ) is the concave function and g(θ) is the convex function (it can be proved easily by taking secondary derivative), we
can conclude that f 2 (θ)/g(θ) is quasi concave [23].
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
27
√
which can easily be seen from α∗1 (0) = 1 − Γ. Thus, θ = 0 ∈ N2 . Furthermore, b(θ) + c2 (θ)/b(θ) =
q
λ2 θ −1
λ2 [α∗ (θ)]2 +1 is a monotone decreasing function of θ since
1
α∗1 (θ) =
0
√θΓ − p(1 − θ)(1 − Γ)
if
θ ≤ θI := 1 − Γ
if
θ > θI
(85)
is a monotone increasing function of θ . Hence, there exists θa such that a(θa ) = b(θa ) + c2 (θa )/b(θa )
to yield N2 = {θ|θ ≤ θa } and N3 = {θ|θ > θa }.
Now recall that
∗(1)
γ2
=
λ1
λ2
λ1
∗(2)
∗(3)
, γ2 (θ) =
[α∗2 (θ)]2 , and γ2 (θ) =
.
∗
1 + λ1 Γ
1 + λ1 Γ
λ2 [α1 (θ)]2 + 1
(86)
If θa ≤ θI , then the optimal θ set for maximizing γ2∗ is given by {θ|θa ≤ θ ≤ θI }. This is because
∗(2)
γ2
∗(3)
(θa ) = γ2
∗(3)
(θa ), because γ2
(θ) is monotone decreasing with respect to θ as seen in (86) since
∗(2)
α∗1 (θ) is a monotone increasing function of θ , and because γ2
(θ) is monotone increasing with respect
to θ for θ ≤ θa since α∗2 (θ) is a monotone increasing function of θ for θ ≤ min{θa , θI } = θa (this can
be shown by substituting α∗1 = 0 for θ ≤ θI into α∗2 (θ) and taking derivative of α∗2 (θ) with respect to θ
and showing the derivative is positive for θ ≤ θa ). Hence, in this case the optimal γ2∗ occurs at θa but
∗(3)
for all θ in {θ|θa ≤ θ ≤ θI }, α∗1 (θ) = 0 and the corresponding optimal γ2∗ = γ2
= λ2 from (86). In
this case, from the assumption θa ≤ θI , θa is computed based on (81) with α∗1 (θ) = 0 as
θa = θ s.t. a = b(θ) + c2 (θ)/b(θ)
= θ s.t.
=
λ1
λ2
=
1 + λ1 Γ
θ
(87)
(88)
λ2
(1 + λ1 Γ)
λ1
(89)
and the condition θa ≤ θI reduces to
λ2
λ2 −1 − λ1 −1
.
(1 + λ1 Γ) ≤ 1 − Γ ⇐⇒ Γ ≤
λ1
1 + λ2 −1
On the other hand, if θa > θI , i.e., Γ >
∗(2)
γ2
λ2 −1 −λ1 −1
,
1+λ2 −1
(90)
then optimal θ exists between θI and θa because
∗(3)
is an increasing function for θ < min{θa , θI } = θI and γ3
is a decreasing function for θ > θa .
Since optimal θ lies in N2 in this case, it is obtained by solving
∗(2)
∂γ2 (θ)
= 0.
∂θ
Therefore, the claim follows.
Proof of Corollary 1:
(91)
With the assumption of λ1 = λ2 , we have Γ1 = 0 and Γ2 = 1 from
(40) and hence the condition Γ ∈ [Γ1 , Γ2 ] reduces to Γ ∈ [0, 1] which is always valid for p1 =
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
28
p2 = 1 (see the definition of Γ in (21)). Furthermore, with the assumption, we have z1 = z2 and
p
4Γ(1 − Γ)[Γ(1 − Γ) + z1 z2 − z22 ] in (41) is given by 2Γ(1 − Γ). From (41), the optimal θ set is given
by {θ | θ0 ≤ θ ≤ 1}, where
θ0 =
1
1+Γλ1
z12
z12 +4Γ(1−Γ)
if
if
1
1+Γλ1
1
1+Γλ1
≤1−Γ
(92)
>1−Γ
Proof of Proposition 3: As λ2 → 0. the threshold τ in Proposition 1 converges to −∞ and thus neither
opt
of the two conditions for popt
1 ∈ P2 in Proposition 1 is satisfied. Hence, by Proposition 1, p1 ∈ P3 .
∗(3)
opt
Since popt
(p1 ) in (35) and is given by
1 ∈ P3 , p1 can be obtained in closed form by maximizing γ2
q
−2
ψ12 − ψ1 ψ22 + 2λ−1
2 ψ1 + λ2
opt
p̄1 = −P + 2Γ +
(93)
2θ(1 − θ)Γ
where ψ1 := θΓ + (1 − θ)(P − Γ) + λ−1
2 and ψ2 := θΓ − (1 − θ)(P − Γ). Using L’Hospital’s rule, we
√
opt
can show that limλ2 →0 p̄1 = Γ(= p1,min ). With p1 = Γ, we have α∗1 = θ from (30) and consequently
√
√
β1∗ = 1 − θ from the the constraint eq. in (29), and α∗2 = θ from (34), and by substituting these
√ h
√
∗(3)
1
values into γ2 (p1 ) in (35) and (12) and (13), we have γ2∗ = P Γ−Γ θ+λ−1
,
p
w
=
Γ kh11 k and
1
1
−1
2 Γ
√
√
p2 w2 = P − Γ khh22 k .
R EFERENCES
[1] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and K. Higuchi, “Non-orthogonal multiple access (NOMA)
for cellular future radio access,” in Proc. IEEE VTC, pp. 15, Jun. 2013.
[2] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge University Press, 2005.
[3] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of non-orthogonal multiple access in 5G systems with
randomly deployed users,” IEEE Signal Process. Lett., vol. 21, pp. 15011505, Jul. 2014.
[4] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple access in 5G systems,” it IEEE Signal Process. Lett.,
vol. 22, pp. 16471651, Mar. 2015.
[5] F. Liu, P. Mahonen, and M. Petrova, “Proportional fairness-based user pairing and power allocation for non-orthogonal
multiple access,” in Proc. IEEE PIMRC, pp. 11271131, Aug. 2015.
[6] J. Choi, “Non-orthogonal multiple access in downlink coordinated two-point systems,” IEEE Commun. Lett., vol. 18, pp.
313316, Jan. 2014.
[7] M. Al-Imari, P. Xiao, M. A. Imran, and R. Tafazolli, “Uplink non-orthogonal multiple access for 5G wireless networks,” in
Proc. IEEE ISWCS, pp. 781785, 2014.
[8] J. So and Y. Sung, “Improving non-orthogonal multiple access by forming relaying broadcast channels,” EEE Commun.
Lett., vol. 20, pp. 18161819, Jul. 2016.
[9] Y. Lan, A. Benjebboiu, X. Chen, A. Li, and H. Jiang, “Considerations on downlink non-orthogonal multiple access (NOMA)
combined with closed-loop SU-MIMO,” in Proc. IEEE ICSPC, pp. 15, 2014.
May 22, 2017
DRAFT
SUBMITTED TO IEEE TRANSACTIONS ON SIGNAL PROCESSING, MAY 22, 2017
29
[10] X. Chen, A. Benjebbour, Y. Lan, A. Li, and H. Jiang, “Impact of rank optimization on downlink non-orthogonal multiple
access (NOMA) with SU-MIMO,” in Proc. IEEE ICCS, pp. 233237, 2014.
[11] M. F. Hanif, Z. Ding, T. Ratnarajah, and G. K. Karagiannidis, “A minorization-maximization method for optimizing sum
rate in the downlink of non-orthogonal multiple access systems,” IEEE Trans. Signal Process., vol. 64, no. 1, pp. 7688, 2016.
[12] B. Kim, S. Lim, H. Kim, S. Suh, J. Kwun, S. Choi, C. Lee, S. Lee, and D. Hong, “Non-orthogonal multiple access in a
downlink multiuser beamforming system,” in Proc. MILCOM, pp. 12781283, 2013.
[13] S. Liu, C. Zhang, and G. Lyu, “User selection and power schedule for downlink non-orthogonal multiple access (NOMA)
system,” in Proc. IEEE ICCW, pp. 25612565, 2015.
[14] A. Sayed-Ahmed and M. Elsabrouty, “User selection and power allocation for guaranteed SIC detection in downlink
beamforming Non-Orthogonal Multiple Access,” in IEEE Wireless Days, 2017, pp. 188193, 2017.
[15] T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast scheduling using zero-forcing beamforming,” IEEE
J. Sel. Areas in Commun., vol. 24, pp. 528541, Mar. 2006.
[16] Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO to non-orthogonal multiple access,” IEEE Trans. Wireless
Commun., vol. 15, pp. 537552, Sept. 2016.
[17] J. Seo and Y. Sung, “A new transceiver architecture for multi-user mimo communication based on mixture of linear and
non-linear reception,” in Proc. SPAWC, (Sapporo, Japan), Jun. 2017.
[18] K. Ho, D. Gesbert, E. Jorswieck, and R. Mochaourab, “Beamforming on the MISO interference channel with multi-user
decoding capability,” arXiv preprint arXiv:1107.0416, 2011.
[19] E. A. Jorswieck, E. G. Larsson, and D. Danev, “Complete characterization of the Pareto boundary for the MISO interference
channel,” IEEE Trans. Signal Process., vol. 56, pp. 52925296, Jul. 2008.
[20] E. A. Jorswieck and E. G. Larsson, “Linear precoding in multiple antenna broadcast channels: Efficient computation of
the achievable rate region,” in Proc. WSA, pp. 2128, Feb. 2008.
[21] J. Lindblom, E. Karipidis, and E. G. Larsson, “Efficient computation of pareto optimal beamforming vectors for the MISO
interference channel with successive interference cancellation,” IEEE Trans. Signal Process., vol. 61, pp. 47824795, Jul. 2013.
[22] J. Lindblom, E. Karipidis, and E. G. Larsson, “Closed-form parameterization of the Pareto boundary for the two-user
MISO interference channel,” in Proc. IEEE ICASSP, pp. 33723375, 2011.
[23] C. Bector, “Programming problems with convex fractional functions,” Operations Research, vol. 16, pp. 383391, Apr.
1968.
May 22, 2017
DRAFT
| 7cs.IT
|
arXiv:1509.00906v3 [math.GT] 13 Sep 2016
SPHERICAL SPACE FORMS REVISITED
DANIEL ALLCOCK
Abstract. We give a simplified proof of J. A. Wolf’s classification
of finite groups that can act freely and isometrically on a round
sphere of some dimension. We slightly improve the classification
by removing some non-obvious redundancy. The groups are the
same as the Frobenius complements of finite group theory.
In chapters 4–7 of his famous Spaces of Constant Curvature [7], J. A.
Wolf classified the spherical space forms: connected Riemannian manifolds locally isometric to the n-sphere S n . By passing to the action of
the fundamental group on the universal covering space, this is equivalent to classifying the possible free isometric actions of finite groups
on S n . Then, by embedding S n in Euclidean space, this is equivalent
to classifying the real representations of finite groups that are “free”
in the sense that no element except the identity fixes any vector except 0. This allowed Wolf to use the theory of finite groups and their
representations.
Our first goal is to give a simplified proof of Wolf’s classification of
the finite groups G that can act freely and isometrically on spheres.
Wolf’s main result here was the list of presentations in theorems 6.1.11
and 6.3.1 of [7]. Our approach to Wolf’s theorem leads to the most
interesting example, the binary icosahedral group, with very little case
analysis and no character theory. (The trick consists of the equalities
(3.2) and (3.3) in the proof of lemma 3.11. These rely on an elementary
property of the binary tetrahedral group, stated in lemma 3.10.)
Our second goal is to remove the redundancy from Wolf’s list; this
modest improvement appears to be new. Some groups appear repeatedly on Wolf’s list because different presentations can define isomorphic
groups. See example 5.1 for some non-obvious isomorphisms. It would
not be hard to just work out the isomorphisms among the groups defined by Wolf’s presentations. But it is more natural to reformulate
the classification in terms of intrinsically defined subgroups. Namely,
Date: 9 September 2016.
1991 Mathematics Subject Classification. 20B10(57S17, 57S25).
Key words and phrases. spherical space form, Frobenius complement.
Supported by NSF grant DMS-1101566.
1
2
DANIEL ALLCOCK
a finite group G acts freely and isometrically on a sphere if and only if
it has one of 6 possible “structures”, in which case it has a unique such
“structure” up to conjugation. See theorems 1.1 and 1.2. In fact we
parameterize the possible G without redundancy, in terms of a fairly
simple set of invariants. Namely: a type I–VI, two numbers |G| and a,
and a subgroup of the unit group of the ring Z/a. In a special case one
must also specify a second such subgroup. See section 5 for details.
The full classification of spherical space forms requires not just the
list of possible groups, but also their irreducible free actions on real
vector spaces, and how their outer automorphism groups permute these
representations. See [7, Thm 5.1.2] for why this is the right data to
tabulate and [7, Ch. 7] for the actual data for each group. We expect
that this data could be described cleanly in terms of our descriptions
of the groups, but have not worked out the details. It would remain
lengthy, because of many cases and subcases to consider.
For many authors the phrase “spherical space form” means a quotient of a sphere by a free action of a finite group of homeomorphisms
or diffeomorphisms, rather than isometries. In this paper we consider
only isometric actions on round spheres. See [6] for the rich topology
and group theory involved in the more general theory.
Expecting topologists and geometers rather than group theorists as
readers, we have made the paper self-contained, with three exceptions.
First, we omit proofs of Burnside’s transfer theorem and the SchurZassenhaus theorem. Second, we use the fact that SL2 (F5 ) is the unique
perfect central extension of the alternating group A5 by Z/2, giving a
citation when needed. Third, we use Aut SL2 (F5 ) = PGL2 (F5 ) ∼
= S5 ,
which is just an exercise.
Finite group theorists study the same groups Wolf did, from a different perspective. We will sketch the connection briefly because our
descriptions of the groups may have some value in this context. A
finite group G is called a Frobenius complement if it acts “freely” on
some finite group H, meaning that no element of G except 1 fixes any
element of H except 1. To our knowledge, the structure of Frobenius
complements (in terms of presentations) is due to Zassenhaus [8]. Unfortunately his paper contains an error, and the first correct proof is
due to Passman [5, Theorems 18.2 and 18.6]. See also Zassenhaus’ later
paper [9]. If G is a Frobenius complement, then after a preliminary reduction one can show that H may be taken to be a vector space over
a finite field Fp , where p is a prime not dividing |G|. Our arguments
apply with few or no changes; see remark 3.12 and also [4], especially
Prop. 2.1.
SPHERICAL SPACE FORMS REVISITED
3
We will continue to abuse language by speaking of free actions on
vector spaces when really we mean that the action is free away from 0.
We will use the standard notation G′ for the commutator subgroup
of a group G, and O(G) for the unique maximal-under-inclusion odd
normal subgroup when G is finite.
We will also use ATLAS notation for group structures [1]. That is,
if a group G has a normal subgroup A, the quotient by which is B,
then we may say “G has structure A.B”. We sometimes write this
as G ∼ A.B. This usually does not completely describe G, because
several nonisomorphic groups may have “structure A.B”. Nevertheless
it is a helpful shorthand, especially if A is characteristic. If the group
extension splits then we may write A : B instead, and if it doesn’t then
we may write A · B. See theorem 1.1 for a some examples. When we
write A : B, we will regard B as a subgroup of G rather than just a
quotient. (In all our uses of this notation, the complements to A turn
out to be conjugate, so there is no real ambiguity in choosing one of
them.)
I am very grateful to the referee for catching a serious error in an
earlier version of this paper.
1. The groups
It is well-known that the groups of orientation-preserving isometries
of the tetrahedron, octahedron (or cube) and icosahedron (or dodecahedron) are subgroups of SO(3) isomorphic to A4 , S4 and A5 . The
preimages of these groups in the double cover of SO(3) are called the
binary tetrahedral, binary octahedral and binary icosahedral groups.
They have structures 2.A4 , 2.S4 and 2.A5 , where we are using another
ATLAS convention: indicating a cyclic group of order n by simply
writing n; here n is 2. These are the only groups with these structures
that we will encounter in this paper. So we abbreviate them (again
following the ATLAS) to 2A4 , 2S4 and 2A5 , and specify that this notation refers to the binary polyhedral groups, rather than some other
groups with structure 2.A4 , 2.S4 or 2.A5 . Alternate descriptions of the
binary tetrahedral and binary icosahedral groups are 2A4 ∼
= SL2 (3) and
2A5 ∼
= SL2 (5).
It is also well-known that the double cover of SO(3) may be identified
with the unit sphere H∗ in Hamilton’s quaternions H. A finite subgroup
of H∗ obviously acts by left multiplication on H∗ . So 2A4 , 2S4 and 2A5
act freely on the unit sphere S 3 .
Similarly, SO(3) contains dihedral subgroups, and their preimages
in H∗ are called binary dihedral. If we start with the dihedral group
4
DANIEL ALLCOCK
of order 2n, then the corresponding binary dihedral group of order 4n
can be presented by
x, y x2n = 1, yxy −1 = x−1 , y 2 = xn .
This group may be identified with a subgroup of H∗ by taking
(1.1)
x 7→ (any primitive 2nth root of unity in R ⊕ Ri)
y 7→ j.
Left multiplication by these elements of H gives a free action on S 3 .
(Replacing the root of unity by its inverse gives an equivalent representation.) Restricting to the case n = 2m−2 , m ≥ 3, one obtains the
quaternion group Q2m of order 2m . Some authors call this a generalized
quaternion group, with “quaternion group” reserved for Q8 .
Now we can state our version of Wolf’s theorems 6.1.11 and 6.3.1,
supplemented by a uniqueness theorem. An n-element of a group means
an element of order n.
∗
Theorem 1.1 (Groups that act freely and isometrically on spheres).
Suppose G is a finite group that acts freely and isometrically on a sphere
of some dimension. Then it has one of the following six structures,
where A and B are cyclic groups whose orders are odd and coprime,
every nontrivial Sylow subgroup of B acts nontrivially on A, and every
prime-order element of B acts trivially on A.
(I) A : B × (a cyclic 2-group T ) , where if T 6= 1 then its involution fixes A pointwise.
(II) A : B × (a quaternionic group T ) .
(III) (Q8 × A) : (Θ × B), where Θ is a cyclic 3-group which acts
nontrivially on Q8 and whose 3-elements centralize A, and |A|
and |B| are prime to 3.
(IV) (Q8 ×A) : (Θ×B) ·2, where Θ, |A| and |B| are as in (III), and
the quotient Z/2 is the image of a subgroup Φ of G, isomorphic
to Z/4, whose 4-elements act by an outer automorphism on Q8 ,
by inversion on Θ and trivially on B.
(V) 2A5 × (A : B) where
|A| and |B| are prime to 15.
(VI) 2A5 × (A : B) · 2, where |A| and |B| are prime to 15, and the
quotient Z/2 is the image of a subgroup Φ of G, isomorphic to
Z/4, whose 4-elements act by an outer automorphism on 2A5
and trivially on B.
Conversely, any group with one of these structures acts freely and isometrically on a sphere of some dimension.
These groups are parameterized in terms of simple invariants in theorem 5.3. A binary dihedral group has type I or II, 2A4 has type III,
2S4 has type IV, and 2A5 has type V.
SPHERICAL SPACE FORMS REVISITED
5
Theorem 1.2 (Uniqueness of structure). The structure in theorem 1.1
is unique, in the following sense:
(1) Groups of different types I–VI cannot be isomorphic.
(2) Suppose G has one of the types I–VI, with respect to some subgroups A, B (and whichever of T , Q8 , Θ, Φ and 2A5 are relevant), and also with respect to some subgroups A∗ , B ∗ (and T ∗ ,
Q∗8 , Θ∗ , Φ∗ and 2A∗5 , when relevant). Then some element of G
conjugates every unstarred group to the corresponding starred
group. In particular, A∗ = A (and Q∗8 = Q8 and 2A∗5 = 2A5 ,
when relevant).
Remark 1.3 (Correspondence with Wolf’s types). Our types correspond
in the obvious way to Wolf’s in theorems 6.1.11 and 6.3.1 of [7]. However, his generator A might not generate our subgroup A and his generator B might not even lie in our subgroup B.
Remark 1.4 (Implied relations). Some useful information is implicit.
For example, B acts trivially on Q8 for types III and IV, because Q8
has no automorphisms of odd order > 3. Also, Θ must act trivially on
A for type IV, because A has abelian automorphism group and Θ ≤ G′ .
In light of these remarks,
one could rewrite G’s structure for type IV
as (Q8 : Θ) × (A : B) · 2. This would be more informative, but hide
the relationship to type III.
In all cases, G has at most one involution. This is obvious except
for type II. Then, T ’s central involution lies in T ′ , which centralizes A
because Aut A is abelian.
At this point we will prove the easy parts of the theorems, namely
that the listed groups do act freely on spheres, and that groups of
different types cannot be isomorphic. The proof that every group acting
freely on a sphere has structure as in theorem 1.1 appears in sections 3–
4, and theorem 1.2’s uniqueness statement is a byproduct of the proof.
Proof of “Conversely. . .” in theorem 1.1. Define H as the normal subgroup of G that is generated by the elements of prime order. We claim:
if H has a free action on a real vector space V , then G acts freely on its
representation W induced from V . To see this, recall that as a vector
space, W is a direct sum of copies of V , indexed by G/H. And H’s
actions on these copies of V are isomorphic to the representations got
by precomposing H → GL(V ) by automorphisms H → H arising from
conjugation in G. In particular, H acts freely on W . We claim that
G also acts freely on W . Otherwise, some prime-order element of G
would have a fixed point. But it would also lie in H, which acts freely.
6
DANIEL ALLCOCK
Now it suffices to determine H and show that it always has a free
action. For types I–II, H is a cyclic group, and for types V–VI it is
2A5 ×(cyclic group of order prime to 30). For types III–IV, H is either
cyclic or 2A4 ×(cyclic group of order prime to 6), according to whether
|Θ| > 3 or |Θ| = 3. These claims are all easy, using the facts that G’s
involution is central (if one exists) and the prime-order elements of B
and (if relevant) Θ act trivially on A.
Cyclic groups obviously admit free actions. For a product of 2A4 or
2A5 by a cyclic group of coprime order, we identify each factor with a
subgroup of H∗ and make 2A4 or 2A5 act on H by left multiplication and
the cyclic group act by right multiplication. If there were a nonidentity
element of this group with a nonzero fixed vector, then there would be
one having prime order, hence lying in one of the factors. But this is
impossible since each factor acts freely.
Proof of theorem 1.2(1). Suppose G has one of the types I–VI. Then
the subgroup H generated by A, B and the index 3 subgroup of Θ (for
types III–IV) is normal in G and has odd order. The quotient G/H
is a cyclic 2-group, a quaternionic group, 2A4 , 2A4 · 2, 2A5 or 2A5 · 2
respectively. None of these has an odd normal subgroup larger than
{1}. Therefore H is all of O(G), and the isomorphism class of G/O(G)
distinguishes the types.
2. Preparation
In this section we suppose G is a finite group. We will establish general
properties of G under hypotheses related to G having a free action on
a sphere. The results before lemma 2.8 are standard, and included for
completeness. Lemma 2.8 is a refinement of a standard result.
Lemma 2.1 (Unique involution). Suppose G has a free action on a
sphere. Then it has at most one involution.
Proof. Choose a free (hence faithful) action of G on a real vector
space V . An involution has eigenvalues ±1, but +1 cannot appear by
freeness. So there can be only one involution, acting by negation.
Lemma 2.2. Suppose G has a free action on a sphere. Then every
abelian subgroup is cyclic, and so is every subgroup of order pq, where
p and q are primes.
Proof. Fix a free action of G on a real vector space V , and let VC be its
complexification. G also acts freely on VC . Otherwise, some nontrivial
element has a nonzero fixed vector, hence has the real number 1 as an
eigenvalue, hence fixes a nonzero real vector.
SPHERICAL SPACE FORMS REVISITED
7
Now suppose A ≤ G is abelian, and decompose VC under A as a
sum of 1-dimensional representations. By freeness, each of these is
faithful. So A is a subgroup of the multiplicative group C − {0}, hence
cyclic, proving the first claim. Since any group of prime-squared order
is abelian, this also proves the p = q case of the second claim.
So suppose p < q are primes and consider a subgroup of G with
order pq; by discarding the rest of G we may suppose without loss that
this subgroup is all of G. Write P resp. Q for a Sylow p-subgroup resp.
q-subgroup. By Sylow’s theorem, P normalizes Q. If it acts trivially
on Q then G is abelian, so suppose P acts nontrivially on Q.
For purposes of this proof, a character of Q means a homomorphism
Q → C∗ . It is standard that any complex representation of Q is the
direct sum of Q’s character spaces, meaning the subspaces on which Q
acts by its various characters. Fix a character χ of Q whose character
space contains a nonzero vector v ∈ VC ; χ is faithful since Q acts freely
on VC . If g ∈ P then g(v) lies in the character space for the character
∗
−1
χ ◦ i−1
means conjugation by g. Since
g : Q → C , where ig : x 7→ gxg
−1
χ is faithful, and P acts faithfully on Q, the various
P characters χ ◦ ig
are all distinct. Therefore the terms in the sum g∈P g(v) are linearly
independent, so the sum is nonzero. But this contradicts freeness since
the sum is obviously P -invariant.
Lemma 2.3 (Sylow subgroups). Suppose all of G’s abelian subgroups
are cyclic. Then its odd Sylow subgroups are cyclic, and its Sylow 2subgroups are cyclic or quaternionic.
Proof. It suffices to treat the case of G a p-group, say of order pn . We
proceed by induction on n, with the cases n ≤ 2 being trivial. So
suppose n > 2.
First we treat the special case that G contains a cyclic group X of
index p. If G acts trivially on X then G is abelian, hence cyclic. So
we may assume that G/X is identified with a subgroup of order p in
Aut X. Recall that Aut X is cyclic of order (p − 1)pn−2 if p is odd, and
Z/2 times a cyclic group of order 2n−3 if p = 2. This shows that some
y ∈ G − X acts on X by the λth power map, where
(
pn−2 + 1
if p is odd
λ=
n−2
−1 or 2
± 1 if p = 2,
with the possibilities 2n−2 ±1 considered only if n > 3. Write X0 for the
subgroup of X centralized by y. Now, hX0 , yi is abelian, hence cyclic.
The index of its subgroup X0 is p, because y p lies in X and centralizes
y. Since y ∈
/ X0 , y generates hX0 , yi. We write pt for the index of X0
8
DANIEL ALLCOCK
in X, which can be worked out from y’s action on X. Namely,
(
2n−2 if p = 2, and λ = −1 or 2n−2 − 1
pt =
p
otherwise
t
We choose a generator x for X such that y p = x−p .
Since xy and y have the same centralizer in X, the same argument
shows that (xy)p also generates X0 . Now,
(xy)p = x(yxy −1)(y 2xy −2 ) · · · (y p−1xy 1−p )y p
2
p−1
= x · xλ · xλ · · · xλ
t
· x−p .
Our two descriptions hy pi and h(xy)p i of X0 ≤ X ∼
= Z/pn−1 must
coincide, so µ := 1 + λ + · · · + λp−1 − pt generates the same subgroup
of Z/pn−1 as pt does. One computes
pt
p
2n−2
2n−2
µ
n−2
p
if λ = pn−2 + 1 (including the case 2n−2 + 1)
0
if λ = 2n−2 − 1
n−2
2
if λ = −1
p
2
Only in the last case do pt and µ generate the same subgroup of Z/pn−1 .
So p = 2, y inverts X, and y 2 is the involution in X. That is, G is
quaternionic. This finishes the proof in the special case.
Now we treat the general case. Take H to be a subgroup of index p.
If p is odd then H is cyclic by induction, so the special case shows
that G is cyclic too. So suppose p = 2. By induction, H is cyclic or
quaternionic. If it is cyclic then the special case shows that G is cyclic
or quaternionic, as desired. So suppose H is quaternionic and take
X to be a G-invariant index 2 cyclic subgroup of H. This is possible
because H contains an odd number of cyclic subgroups of index 2 (three
if H ∼
= Q8 and one otherwise). Now we consider the action of G/X on
X. If some element of G − X acts trivially then together with X it
generates an abelian, hence cyclic, group, and the special case applies.
So G/X is 2 × 2 or 4 and embeds in Aut X. (These cases require
|X| ≥ 8 or 16 respectively, or equivalently n ≥ 5 or 6.) Furthermore,
Aut X contains just three involutions, and only one of them can be
a square in Aut X, namely the (1 + 2n−2 )nd power map. Therefore,
either possibility for G/X yields an element y of G which acts on X
by this map and has square in X. But then hX, yi is neither cyclic nor
quaternionic, contradicting the special case.
Recall that a group is called perfect if its abelianization is trivial.
SPHERICAL SPACE FORMS REVISITED
9
Lemma 2.4 (2A5 recognition). Suppose G is perfect with center Z/2,
and every noncentral cyclic subgroup has binary dihedral normalizer.
Then G ∼
= 2A5 .
Proof. Write 2g for |G|. The hypothesis on normalizers shows that distinct maximal cyclic subgroups of G have intersection equal to Z(G).
So G is the disjoint union of Z(G) and the subsets C − Z(G) where
C varies over the maximal cyclic subgroups of G. We choose representatives C1 , . . . , Cn for the conjugacy classes of such subgroups and
write 2c1 , . . . , 2cn for their orders. The numbers c1 , . . . , cn are pairwise
coprime because each of C1 , . . . , Cn is the centralizer of each of its subgroups of order > 2. We number the Ci so that c1 is divisible by 2, c2 is
divisible by the smallest prime involved in g but not c1 , c3 is divisible
by the smallest prime involved in g but neither c1 nor c2 , and so on.
In particular, ci is at least as large as the ith prime number.
Each conjugate of Ci −Z(G) has 2ci −2 elements, and the normalizer
hypothesis tells us there are g/2ci many conjugates. Therefore (2ci −
2)g/2ci = g 1 − c1i elements of G −Z(G) are conjugate into Ci −Z(G).
Our partition of G gives
n
X
(2.1)
2g = 2 + g
(1 − c1i )
i=1
P
We can rewrite this as g(2−n) = 2− ni=1 g/ci . Since G has no index 2
subgroups, each term g/ci in the sum is larger than 2. Since the right
side is negative, g(2 − n) is also. So n > 2.
In fact n = 3. Otherwise, we would use c1 ≥ 2, c2 ≥ 3, c3 ≥ 5, c4 ≥ 7
to see that the sum on the right side of (2.1) is
(at least 12 ) + (at least 32 ) + (at least 45 ) + (at least 76 ) + · · · > 2
which is a contradiction. Now we rewrite (2.1) as c11 + c12 + c13 =
1 + 2/g. In particular, the left side must be larger than 1, which
requires c1 = 2, c2 = 3, c3 = 5. Then 21 + 13 + 15 = 1 + 2/g gives a
formula for g, namely g = 60, so |G| = 120. So G/Z(G) is nonsolvable
of order 60. A Sylow’s theorem exercise rules out the possibility that
there are 15 Sylow 2-subgroups, so there must be 5, and it follows easily
that G/Z(G) ∼
= A5 . So G has structure 2.A5 . Finally, A5 has a unique
perfect central extension by Z/2, namely the binary icosahedral group
[5, Prop. 13.7].
Theorem 2.5 (Burnside’s transfer theorem). Suppose G is a finite
group, and P is a Sylow subgroup that is central in its normalizer.
Then P maps faithfully to the abelianization G/G′ .
10
DANIEL ALLCOCK
Proof. [3, Thm. 5.13], [2, Thm. 4.3] or [7, Thm 5.2.9].
Corollary 2.6 (Cyclic transfer). Suppose G is a finite group, p is a
prime, and G’s Sylow p-subgroups are cyclic. If some nontrivial p-group
is central in its normalizer, or maps nontrivially to G/G′ , then every
Sylow p-subgroup maps faithfully to G/G′ . In particular, this holds if
p is the smallest prime dividing |G|.
Proof. Suppose that P0 ≤ G is a p-group satisfying either of the two
conditions, and choose a Sylow p-subgroup P containing it. It is cyclic
by hypothesis, so P0 is characteristic in P , so N(P ) lies in N(P0 ). The
automorphisms of P with order prime to p act nontrivially on every
nontrivial subquotient of P . Under either hypothesis, P0 (hence P )
has a nontrivial subquotient on which N(P ) acts trivially. Therefore
the image of N(P ) in Aut P contains no elements of order prime to p.
It contains no elements of order p either, since P is abelian. So P is
central in its normalizer and we can apply Burnside’s transfer theorem.
Since P maps faithfully to G/G′ , so does every Sylow p-subgroup.
For the final statement, observe that Aut P has no elements of prime
order > p. So its normalizer must act trivially on it, and we can apply
the previous paragraph.
Theorem 2.7 (Schur-Zassenhaus). Suppose G is a group, N is a normal subgroup, and |N| and |G/N| are coprime. Then there exists a
complement to N in G, and all complements are conjugate.
As stated, this relies on the odd order theorem. But we only need
the much more elementary case that N is abelian (Theorem 3.5 of [3]).
In determining the structure of his groups, Wolf used a theorem of
Burnside: if all Sylow subgroups of a given group H are cyclic, then
H ′ and H/H ′ are cyclic of coprime order, H ′ has a complement, and
all complements are conjugate. We prefer the following decomposition
H = A : B because of its “persistence” property (4). We only need
this property for the imperfect case (section 4).
Lemma 2.8 (Metacyclic decomposition). Suppose H is a finite group,
all of whose Sylow subgroups are cyclic. Define A as the subgroup
generated by H ′ and all of H’s Sylow subgroups that are central. Then
A has the following two properties and is characterized by them:
(1) A is normal, and A and H/A are cyclic of coprime orders.
(2) Every nontrivial Sylow subgroup of H/A acts nontrivially on A.
Furthermore,
(3) A has a complement B, and all complements are conjugate.
SPHERICAL SPACE FORMS REVISITED
11
(4) Suppose a finite group G contains H as a normal subgroup,
with |G/H| coprime to |H|. Then there is a complement C to
H, such that the decomposition H ∼ A : B in (3) extends to
G ∼ A : (B × C). Furthermore, all complements of H that
normalize B are conjugate under NH (B).
Proof. First we show that H is solvable. If p is the smallest prime
dividing |H|, and P is a Sylow p-subgroup, then corollary 2.6 shows
that H has a quotient group isomorphic to P . The kernel is solvable
by induction, so H is too.
Now let F be the Fitting subgroup of H, i.e., the unique maximal
normal nilpotent subgroup. Being nilpotent, it is the product of its
Sylow subgroups. Since these are cyclic, so is F . Also, H/F acts
faithfully on F , for otherwise F would lie in a strictly larger normal
nilpotent subgroup. As a subgroup of the abelian group Aut F , H/F
is abelian. Therefore the cyclic group F contains H ′, so H ′ is cyclic.
If p is a prime dividing the order of H/H ′ , then corollary 2.6 shows
that every Sylow p-subgroup meets H ′ trivially. It follows that the
orders of H ′ and H/H ′ are coprime. A is obviously normal. Since A
is the product of H ′ with the central Sylow subgroups of H, we see
that A and H/A also have coprime orders. Since A contains H ′ , H/A
is abelian. Having cyclic Sylow subgroups, H/A is cyclic. We have
proven (1).
Because |A| and |H/A| are coprime, the Schur-Zassenhaus theorem
assures us that A has a complement B, and that all complements are
conjugate, proving (3). For (2), suppose a Sylow subgroup of B acts
trivially on A. Then it is central in H, so A contains it by definition,
which is a contradiction.
Next we prove the “persistence” property (4), so we assume its hypotheses. Since A is characteristic in H, it is normal in G. Since all
complements to A in H are conjugate in A, the Frattini argument shows
that NG (B) maps onto G/H. Applying the Schur-Zassenhaus theorem
to NH (B) inside NG (B) yields a complement C to H that normalizes
B. That theorem also shows that all such complements are conjugate
under NH (B)
To prove (4) it remains only to show that B and C commute. If C
acted nontrivially on B, then G′ would contain a nontrivial p-group for
some prime p dividing |B|. Using corollary 2.6 as before, it follows that
G′ contains a Sylow p-subgroup P of B. Since P lies in the commutator
subgroup of G, it must act trivially on A. This contradicts (2). So B
and C commute, completing the proof of (4). (Remark: since NH (B) =
CA (B)×B, we could replace NH (B) in the statement of (4) by CA (B).)
12
DANIEL ALLCOCK
All that remains is to show that (1) and (2) characterize A; suppose
A∗ ≤ H has these properties. By (2), all the Sylow subgroups of H that
act trivially on A∗ lie in A∗ . Since A∗ is cyclic, Aut A∗ is abelian, so
H ′ acts trivially on it. We already saw that H ′ is the product of some
of H’s Sylow subgroups, so H ′ lies in A∗ . The central Sylow subgroups
of H also act trivially on A∗ , so also lie in A∗ . We have shown that A∗
contains A. If A∗ were strictly larger than A, then the coprimality of
|A∗ | and |H/A∗ | would show that A∗ contains a Sylow subgroup of H
that is not in A. But then A∗ is nonabelian by property (2) of A, and
therefore property (1) fails for A∗ .
3. The perfect case
In this section and the next we prove theorems 1.1 and 1.2 inductively.
We suppose throughout that G is a finite group that acts freely on a
sphere of some dimension. Under the assumption that every proper
subgroup has one of the structures listed in theorem 1.1, we will prove
that G also has such a structure. In this section we also assume G is
perfect. This includes base case G = 1 of the induction, which occurs in
case I. The only other perfect group in theorem 1.1 is 2A5 . Theorem 1.2
is trivial for G ∼
= 2A5
= 1 or 2A5 . Therefore it will suffice to prove G ∼
under the assumption G 6= 1.
Lemma 3.1. G’s Sylow 2-subgroups are quaternionic, in particular
nontrivial.
Proof. By lemma 2.3, all the odd Sylow subgroups are cyclic. If the
Sylow 2-subgroups were too, then corollary 2.6, applied to the smallest
prime dividing |G|, would contradict perfectness. Now lemma 2.3 shows
that the Sylow 2-subgroups must be quaternionic.
Lemma 3.2. G’s 4-elements form a single conjugacy class.
Proof. Let T be a Sylow 2-subgroup (quaternionic by the previous
lemma) and U a cyclic subgroup of index 2. We consider the action of
G on the coset space G/U, whose order is twice an odd number. By
lemma 2.1, G contains a unique involution, necessarily central. Since it
lies in every conjugate of U, it acts trivially on G/U. Now let φ be any
4-element. Since its square acts trivially, φ acts by exchanging some
points in pairs. Since G is perfect, φ must act by an even permutation,
so the number of these pairs is even. Since the size of G/U is not divisible by 4, φ must fix some points. The stabilizers of these points are
conjugates of U, so φ is conjugate into U. Finally, the two 4-elements
in U are conjugate since T contains an element inverting U.
SPHERICAL SPACE FORMS REVISITED
13
Lemma 3.3. O(G) = 1.
Proof. Suppose otherwise. Since G’s odd Sylow subgroups are cyclic,
lemma 2.8(1) shows that O(G) has a characteristic cyclic subgroup of
prime order. Because this subgroup has abelian automorphism group,
and G is perfect, G acts trivially on it. Now corollary 2.6 shows that
G has nontrivial abelianization, contradicting perfectness.
Lemma 3.4. Every maximal subgroup M of G has center of order 2.
Proof. First, M contains G’s central involution. Otherwise, adjoining
it to M would yield G by maximality. Then G would have an index 2
subgroup, contrary to perfectness. Next we show that M has no central
subgroup Y of order 4; suppose it did. By the conjugacy of Z/4’s in
G, and the fact that G’s Sylow 2-subgroups are quaternionic, N(Y )
contains an element inverting Y . So N(Y ) is strictly larger than M,
hence coincides with G, and the map G → Aut(Y ) ∼
= Z/2 is nontrivial,
contrary to perfectness.
Finally we show that M has no central subgroup Y of odd prime
order > 1. The previous lemma shows that N(Y ) is strictly smaller
than G. Since M normalizes Y and is maximal, it is Y ’s full normalizer.
Since Y is central in M, we see that N(Y ) acts trivially on Y . Now
corollary 2.6 shows that G/G′ is nontrivial, contradicting perfectness.
The next lemma is where our development diverges from Wolf’s.
Lemma 3.5 (Maximal subgroups). Every maximal subgroup M of G
has one of the following structures, with O(M) a cyclic group.
(I) O(M) : (cyclic 2-group of order > 2)
(II) O(M) : (quaternionic group)
(III) 2A4
(IV) O(M).2A4 · 2 where the elements of M outside O(M).2A4 act
on O(M) by inversion and on the quotient 2A4 by outer automorphisms.
(V) 2A5
(VI) O(M) × 2A5 · 2 where the elements of M outside O(M)×2A5
act on O(M) by inversion and on 2A5 by outer automorphisms.
The G-normalizer of any nontrivial subgroup of O(M) is M. If M1
and M2 are non-conjugate maximal subgroups, then O(M1 ) and O(M2 )
have coprime orders.
Proof. By induction, M has one of the structures in theorem 1.1. In
light of the previous lemma, we keep only those with center of order 2.
Here are the details. In every case, the prime order elements of B are
14
DANIEL ALLCOCK
central in M, so they cannot exist, so B = 1. For types I–II this leaves
O(M) = A and establishes our claimed structure for M. For type I
we must also show that the cyclic 2-group, call it T , has order > 2.
Otherwise, A is central in M, hence trivial, so T is all of M and has
order 2. That is, the center of G is a maximal subgroup of G, which is
a contradiction because no group can have this property.
For type V we have shown M ∼ 2A5 ×A, so A is central, hence trivial.
For type III we know M ∼ (Q8 × A) : Θ, with Θ’s 3-elements centralizing A. It follows that |Θ| = 3, because otherwise Θ’s 3-elements would
also centralize Q8 , hence be central in M. From |Θ| = 3 it follows
that A is central in M, hence trivial. So M ∼ Q8 : 3 with the Z/3
acting nontrivially on Q8 . Since Aut(Q8 ) ∼
= S4 has a unique class of
3-elements, M is determined up to isomorphism, namely M ∼
= 2A4 .
For type VI we know M ∼ (2A5 × A) · 2 and O(M) = A. Also, A
decomposes as the direct sum of its subgroup inverted by the nontrivial
element t of M/(2A5 × A) ∼
= Z/2 and its subgroup fixed pointwise by t.
The latter subgroup is central in M, hence trivial, so t inverts A as
claimed. Also, t’s image in Out 2A5 is nontrivial by the definition of
type VI groups.
Finally, for type IV we have M ∼ (Q8 × A) : Θ · 2. By remark 1.4,
Θ and A commute. So O(M) is cyclic and generated by A and the
index 3 subgroup of Θ, leaving M ∼ (O(M).2A4 ) · 2. By the argument
for type VI, the elements of M mapping nontrivially to Z/2 must invert
O(M) and act on Q8 by outer automorphisms.
We have shown in each case that O(M) is cyclic. So its subgroups
are characteristic in O(M), hence normal in M. By lemma 3.3 and
maximality, M is the full normalizer of any nontrivial subgroup of
O(M). For the last claim of the theorem, suppose a prime p divides
the orders of O(M1 ) and O(M2 ). We have just shown that M1 is the
normalizer of a cyclic group of order p, and that M2 is the normalizer
of another. These cyclic groups are G-conjugate, so M1 and M2 are
also.
Lemma 3.6 (Done if Q16 6≤ G). Suppose the Sylow 2-subgroups of G
have order 8. Then G ∼
= 2A5 .
Proof. Suppose M is a maximal subgroup of G. It cannot have type
IV or VI, because these contain copies of Q16 . If M has type III or V
then M ∼
= 2A4 or 2A5 by lemma 3.5. We claim that in the remaining
cases, M is binary dihedral.
If M has type I then it has structure O(M) : T where T is a cyclic
group of order ≥ 4. Since this is the largest a cyclic 2-group in G can
be, T has order exactly 4. A generator for it must negate O(M), or else
SPHERICAL SPACE FORMS REVISITED
15
M’s center would have order > 2. We have shown that M is binary
dihedral.
Now suppose M has type II, hence structure O(M) : Q where Q ∼
=
Q
,
which
is
binary
dihedral
as
claimed.
So
Q8 . If O(M) = 1 then M ∼
= 8
suppose O(M) 6= 1 and let P be any Sylow subgroup of O(M). Since
Aut P is cyclic, and Q/Q′ ∼
= 2 × 2, some 4-element of Q acts trivially
on P . Since all 4-elements are conjugate, the centralizer of any one
of them has order divisible by |P |. Now fix a particular 4-element φ.
Letting P vary over all Sylow subgroups of O(M) shows that C(φ) has
order divisible by |O(M)|. The only maximal subgroup of G that could
contain a cyclic group of order 4 · |O(M)| is M, up to conjugacy. So
after conjugation we may suppose that O(M) ≤ C(φ) ≤ M. We have
shown that some 4-element φ of Q centralizes O(M). So Q acts on
O(M) via a quotient group of order ≤ 2. This quotient must be Z/2,
acting by negation, because otherwise M would have center larger than
Z/2. It follows that M is binary dihedral.
We have shown that every maximal subgroup of M is binary dihedral,
binary tetrahedral or binary icosahedral. By examining normalizers in
these groups, one checks that every noncentral cyclic subgroup of G
has binary dihedral normalizer. So G ∼
= 2A5 by lemma 2.4.
To prove the perfect case of theorems 1.1 and 1.2, it now suffices to
rule out the case Q16 ≤ G. We devote the rest of this section to this.
Lemma 3.7 (Z/4 normalizers). For any subgroup Φ ∼
= Z/4 of G, N(Φ)
has structure (odd group).(Sylow 2-subgroup of G).
Proof. We claim first that N(Φ) has structure (odd group).(2-group).
Choosing a maximal subgroup M containing N(Φ), it suffices to show
that NM (Φ) has this structure. To prove this one considers each possible structure for M listed in lemma 3.5, and each subgroup Z/4 of it.
To finish the proof we use the facts that some Z/4 is normal in some
Sylow 2-subgroup, and that all Z/4’s are conjugate.
Lemma 3.8 (Q8 normalizers). Suppose G contains a copy of Q16 .
Choose Q ≤ G isomorphic to Q8 and write N for its normalizer. Then
(1) N ∼ O(N).2A4 · 2.
(2) Q lies in a group 2A4 if and only if 3 ∤ |O(N)|.
(3) G has more than one conjugacy class of Q8 subgroups.
(4) G contains a subgroup 2A4 .
Proof. (1) We begin by exhibiting some elements of N. Choose any
subgroup Φ ∼
= Z/4 of Q. There exists a Sylow 2-subgroup of N(Φ)
that contains Q. So Q lies in some Q16 that normalizes Φ. This Q16
16
DANIEL ALLCOCK
contains an element that inverts Φ and exchanges the other two Z/4
subgroups of Q. We have shown that N contains an element that
normalizes our chosen Z/4 and exchanges the other two Z/4’s in Q. It
follows that N acts by S3 on the three Z/4 subgroups of Q.
Now we fix a maximal subgroup M of G that contains N. The
property of N just established
forces M to have type IV or VI. That is,
M ∼ O(M).(2A4 or 2A5 ) ·2 with Q lying in the index 2 subgroup M ′ .
In either case, N coincides with NM (Q), which has the stated structure.
Also, O(N) = O(M), which we will use later in the proof.
(2) “If” follows from the existence of Sylow 3-subgroup, isomorphic
to Z/3, in N. For “only if”, suppose 3 divides |O(N)|. Then N cannot
contains any 2A4 , because all its 3-elements are central in N.
(3) Fix Φ ≤ Q isomorphic to Z/4. By lemma 3.7 there is a subgroup
Q∗ ∼
= Q8 of N(Φ) that is not N(Φ)-conjugate to Q. We claim that
∗
Q is not G-conjugate to Q either; suppose to the contrary that some
g ∈ G conjugates Q∗ to Q. By Φ ≤ Q∗ we see that g sends Φ into Q.
By replacing g by its composition with an element of N that sends Φg
back to Φ, we may suppose without loss that g normalizes Φ. That is,
Q and Q∗ are conjugate in N(Φ), which is a contradiction.
(4) Supposing that G contains no 2A4 , we will show that every subgroup Q∗ ∼
= Q8 of G is conjugate to Q, which contradicts (3). Applying
the argument for (1) to Q∗ , we write N ∗ for its normalizer and M ∗ for
a maximal subgroup containing N ∗ . By (2) and the nonexistence of
2A4 ’s, 3 divides the orders of O(N) and O(N ∗ ). These groups are the
same as O(M) and O(M ∗ ). Since |O(M)| and |O(M ∗ )| have a common
factor, the last part of lemma 3.5 shows that M and M ∗ are conjugate.
So their index two subgroups M ′ and M ∗ ′ are conjugate, both of which
have structure (odd group).(2A4 or 2A5 ). Since Q resp. Q∗ is a Sylow
2-subgroup of M ′ resp. M ∗ ′ , it follows that Q and Q∗ are conjugate.
Lemma 3.9 (Free actions of binary dihedral groups). Suppose H is
a binary dihedral group, U is an irreducible fixed-point-free real representation of H, and α is the natural map R[H] → End(U). Then
α(R[H]) ∼
= H. Furthermore, if J is any binary dihedral subgroup of H,
then α(R[J]) = α(R[H]).
Proof. It is easy to see that the lemma holds for the representations
(1.1), using their description in terms of H. So it will suffice to show
that U is one of them. Let C be an index 2 cyclic subgroup of H.
Under C, U decomposes as a direct sum of 2-dimensional spaces, on
each of which C acts faithfully by rotations. Fix one, say T , and consider the induced representation IndH
C (T ), of dimension 4. Its canonical
image in U is H-invariant, hence all of U. This image is larger than
SPHERICAL SPACE FORMS REVISITED
17
2-dimensional, because a binary dihedral group has no 2-dimensional
free real representations. (Every finite subgroup of O(2) is cyclic or
dihedral.) Therefore V ∼
= IndH
C (T ). And the representations in (1.1)
are exactly the H-representations of this form.
Lemma 3.10 (Free actions of the binary tetrahedral group). Assume
A∼
= 2A4 acts freely on a real vector space V , and write Q for the copy
of Q8 in A. Then the image of R[A] in End(V ) is isomorphic to H and
equal to the image of R[Q]. In particular, every Q-invariant subspace
is also A-invariant.
Proof. Fix an identification of A with
(3.1)
2A4 = ±1, ±i, ±j, ±k, (±1 ± i ± j ± k)/2 ⊆ H∗
We claim that A has a unique irreducible free real representation. It
follows that V is a direct sum of copies of A’s left-multiplication action
on H, which proves the lemma.
For the claim, fix an irreducible free A-module U. Choose a Q-submodule T which is Q-irreducible. In the proof of lemma 3.9 we showed
that (1.1) accounts for all irreducible free actions of binary dihedral
groups. For Q8 , there was only one. So we may identify T with the real
vector space underlying H, with Q = {±1, ±i, ±j, ±k} ⊆ A acting by
left multiplication. The image of the natural map IndA
Q (T ) → U is AA
invariant, hence all of U. By definition, IndQ (T ) is the real vector space
underlying H3 , with i ∈ Q acting by (x, y, z) 7→ (ix, jy, kz), similarly
with i, j, k cyclically permuted, and θ := (−1+i+j+k)/2 ∈ A acting by
(x, y, z) 7→ (z, x, y). Obviously the kernel of IndA
Q (T ) → U contains the
fixed points of the nonidentity elements of A. For θ and i ◦ θ these are
{(x, x, x)} and {(x, jx, −ix)}, which together span an 8-dimensional
subspace, hence the whole kernel. This proves the uniqueness of U:
it is the quotient of IndA
Q (T ) by the subspace generated by the fixed
points of the nonidentity elements of A.
Lemma 3.11 (The final contradiction). G contains no subgroup Q16 .
Proof. Suppose otherwise, and fix an irreducible fixed-point free real
representation V of G. We write α for the natural map R[G] → End V .
By lemma 3.8(4), G has a subgroup A∗ ∼
= 2A4 . We write Q∗ for its
Q8 subgroup. We fix a Q16 -subgroup of G that contains Q∗ . This is
the only group isomorphic to Q16 we will consider, so we just write Q16
for it. Write Q for the other Q8 subgroup of Q16 . We will distinguish
two cases, according to whether there exists some A ∼
= 2A4 containing
Q. In each case we will prove α(R[G]) ∼
H,
which
implies that G is
=
∗
isomorphic to a subgroup of H . We assume this for the moment.
18
DANIEL ALLCOCK
A well-known property of H∗ is that every non-central cyclic subgroup has H∗ -normalizer isomorphic to the “continuous binary dihedral
group”, meaning the nonsplit extension (R/Z) · 2. It follows that the
G-normalizer of any non-central cyclic subgroup is cyclic or binary dihedral. The cyclic case is ruled out by corollary 2.6 and the perfectness
of G. So lemma 2.4 applies, proving G ∼
= 2A5 .
∼
It remains to prove α(R[G]) = H. First suppose that A exists. G is
generated by A and A∗ , because no maximal subgroup contains two
copies of 2A4 , whose Q8 subgroups generate a copy of Q16 . Now fix a
irreducible Q16 -submodule U of V . It is obviously Q- and Q∗ -invariant,
hence A- and A∗ -invariant by lemma 3.10. So it is G-invariant by
hA, A∗ i = G, and the G-irreducibility of V implies U = V . Lemmas
3.9 and 3.10 also give us the equalities
(3.2)
α(R[A]) = α(R[Q]) = α(R[Q16 ]) = α(R[Q∗ ]) = α(R[A∗ ])
and say that this subalgebra of End(U) is isomorphic to H. Since A
and A∗ generate G, α(R[G]) also equals this copy of H, finishing the
proof in the case that A exists.
Now suppose Q lies in no copy of 2A4 , and set N = NG (Q) and Φ =
Q∩Q∗ ∼
= Z/4. By (1) and (2) of lemma 3.8 we have N ∼ (O(N).2A4 )·2
with |O(N)| divisible by 3. By construction Q∗ normalizes Q, and by
Q∗ 6= Q we see that Q∗ contains a 4-element q ∗ which lies in N but
outside its index 2 subgroup O(N).2A4 . So q ∗ inverts O(N). Of course,
q ∗ also inverts Φ. And O(N) commutes with Q (hence Φ) by the known
structure of N. So H = hq ∗ , O(N), Φi = hQ∗ , O(N)i is binary dihedral.
The rest of the argument is similar to the previous case.
Namely, we have G = hH, A∗ i because no maximal subgroup M of
G contains a copy of Q8 which normalizes one copy of Z/3 in M and
is normalized by a different copy of Z/3 in M. Now we choose an
irreducible H-submodule U of V . It is trivially Q∗ -invariant. Then
lemma 3.10 shows that U is A∗ -invariant, hence G-invariant, hence all
of V . Lemmas 3.9–3.10 give the equalities
(3.3)
α(R[H]) = α(R[Q∗ ]) = α(R[A∗ ])
and say that this subalgebra of End(U) is isomorphic to H. By G =
hH, A∗ i, this subalgebra is also α(R[G]), finishing the proof.
Remark 3.12 (Frobenius complements). As mentioned in the introduction, our arguments adapt easily to classify the Frobenius complements,
which are the same groups. Supposing that G acts freely on a vector
space over a finite field Fq , where q is necessarily odd, we extend scalars
to suppose without loss that all elements of G are diagonalizable. Then
lemmas 3.9–3.11 and their proofs still apply with R replaced by Fq ,
SPHERICAL SPACE FORMS REVISITED
19
H (the algebra) by M2 Fq , H (the module) by F2q , and H∗ by SL2 Fq .
(Although the proof of lemma 3.10 looks H-specific, one defines the
Hurwitz integers as the Z-span of the quaternions (3.1), and tensoring
with Fq gives M2 Fq .)
4. The imperfect case
In this section we will complete the proofs of theorems 1.1 and 1.2.
We suppose throughout that G is an imperfect finite group that acts
freely and isometrically on a sphere of some dimension, and that every
proper subgroup has one of the structures listed in the statement of theorem 1.1. We will prove that G also has one of these structures, in fact
a unique one in the sense of theorem 1.2. We will prove theorems 1.1
and 1.2 in three special cases (which don’t actually use imperfectness),
and then argue that these cases are enough.
Lemma 4.1. Suppose G/O(G) is a 2-group. Then G has type I or II
from theorem 1.1, for a unique-up-to-conjugation triple of subgroups
(A, B, T ).
Proof. We will construct A, B and T such that G ∼ A : (B × T )
as in theorem 1.1, along the way observing that the construction is
essentially unique. Obviously we must choose A and B such that
O(G) = A : B. Applying lemma 2.8 to O(G) proves the following.
The required coprimality of |A| and |B|, together with the requirement
that each nontrivial Sylow subgroup of B acts nontrivially on A, can be
satisfied in a unique way. That is, A is uniquely determined, and B is
determined up to conjugacy in O(G). Using the coprimality of |O(G)|
and |G/O(G)| ∼
= T , part (4) of lemma 2.8 shows that the decomposition O(G) = A : B extends to a decomposition G = A : (B × T ) where
T is determined uniquely up to conjugacy in NG (B).
This finishes the proof of uniqueness; it remains to check a few assertions of theorem 1.1. Lemma 2.3 says that T is cyclic or quaternionic.
Lemma 2.2 assures us that every prime-order element of B or T acts
trivially on every prime-order subgroup of A. An automorphism of the
cyclic group A, of order prime to |A| and acting trivially on every subgroup of prime order, must act trivially on all of A. So the prime-order
elements of B and T centralize A.
Lemma 4.2. Suppose G contains a normal subgroup 2A5 . Then G has
type V or VI from theorem 1.1, for a unique-up-to-conjugation tuple of
subgroups (A, B, 2A5 ) or (A, B, 2A5 , Φ).
Proof. We will show that there is a such a tuple, unique up to conjugation, whose 2A5 term is the given normal subgroup. The existence part
20
DANIEL ALLCOCK
of this assertion shows that G ∼ 2A5 ×(A : B) or G ∼ 2A5 ×(A : B) ·2
as in theorem 1.1. It follows that G has a unique normal subgroup isomorphic to 2A5 . So every tuple has this particular subgroup as its 2A5
term. The uniqueness of the tuple up to conjugacy follows.
It remains to prove the lemma for the given 2A5 subgroup. We define
I as the subgroup of G acting on 2A5 by inner automorphisms. Since
Out(2A5 ) = 2, I has index 1 or 2 in G. By its definition, I is generated
by 2A5 and C(2A5 ), whose intersection is the group Z generated by
G’s central involution. We claim that Z is the
full Sylow 2-subgroup
of C(2A5 ). Otherwise, 2A5 /Z × C(2A5 )/Z ≤ G/Z would contain
an elementary abelian 2-group of rank 3. This is impossible because
the Sylow 2-subgroups of G/Z are dihedral. From corollary 2.6 and
the fact that C(2A5 ) has cyclic Sylow 2-subgroups, we get C(2A5 ) =
Z × O(C(2A5 )). It follows that I = 2A5 × O(I).
Obviously we must choose A and B such that A : B = O(I). As
in the previous proof, lemma 2.8 shows that there is an essentially
unique way to satisfy the conditions that |A| and |B| are coprime and
that every Sylow subgroup of B acts nontrivially on A. That is, A
is uniquely determined and B is determined up to conjugacy in O(I).
Also, |A| and |B| are coprime to 15 because G has no subgroup 3 × 3
or 5 × 5 (lemma 2.2). The prime-order elements of B act trivially on
A by the same argument as in the previous proof.
We have shown that I has type V, so if I = G then the proof is
complete. Otherwise, we know G ∼ 2A5 × (A : B) · 2 and we must
construct a suitable subgroup Φ ∼
= Z/4 of G. Because all complements
to A in A : B are conjugate, the Frattini argument shows that N(B)
surjects to G/A. So N(B) contains a Sylow 2-subgroup T ∼
= Q16 of
G. By Sylow’s theorem it is unique up to conjugacy in N(B). Obviously T ∩ 2A5 is isomorphic to Q8 . Now, T ∼
= Q16 contains exactly
two subgroups isomorphic to Z/4 that lie outside T ∩ 2A5 . So we must
take Φ to be one of them. They are conjugate under T ∩ 2A5 , so Φ is
uniquely defined up to a conjugation that preserves each of A, B and
2A5 . It remains only to check that Φ has the properties required for
G to be a type VI group. It acts on 2A5 by an outer automorphism
because it does not lie in I, by construction. To see that Φ commutes
with B, suppose to the contrary. Then some subgroup of B of prime
order p would lie in G′ . Then corollary 2.6 would show that the Sylow p-subgroup of B lies in G′ . So it acts trivially on A, which is a
contradiction.
SPHERICAL SPACE FORMS REVISITED
21
Lemma 4.3. Suppose G contains a normal subgroup Q ∼
= Q8 that lies
in G′ . Then G has type III or IV from theorem 1.1, for a unique-upto-conjugation tuple of subgroups (A, B, Q8 , Θ) or (A, B, Q8 , Θ, Φ).
Proof. Mimicking the previous proof, we will show that there is a
unique-up-to-conjugation tuple of such subgroups whose Q8 term is Q.
In particular, G ∼ (Q8 × A) : (B × Θ) or G ∼ (Q8 × A) : (B × Θ) · 2
as in theorem 1.1. This shows that G has a unique normal subgroup
isomorphic to Q8 , namely Q, which will complete the proof for the
same reason as before.
To show that there is a unique-up-conjugation tuple of subgroups
whose Q8 term is Q, consider the natural map G → Aut Q ∼
= S4 . We
claim the image G contains a 3-element. Otherwise, G ⊆ Aut Q ∼
=
S4 would lie in a Sylow 2-subgroup of Aut Q, which is dihedral of
′
order 8 with commutator subgroup of order 2. Therefore |G | would
have order ≤ 2. But this is a contradiction because Q lies in G′ and
Q∼
= S3
= 2×2. We have shown that the nontrivial 3-subgroup of Out Q ∼
lies in the image of G. (Now we can discard G.)
Let J be the preimage in G of this copy of Z/3. So G ∼ J.(1 or 2).
Obviously we must choose A, B and Θ so that J = (Q × A) : (B × Θ).
By construction, J has a nontrivial map to Z/3. So corollary 2.6 assures
us that J’s Sylow 3-subgroups map faithfully to J’s abelianization. In
particular, J = I.(cyclic 3-group) where |I| is prime to 3. Mimicking
the previous proof shows that I = Q × O(I). So we must choose A and
B so that A : B = O(I).
Continuing to follow the previous proof shows that there is an essentially unique way to satisfy the conditions that |A| and |B| are coprime
and that every Sylow subgroup of B acts nontrivially on A. That is, A
is uniquely determined and B is determined up to conjugacy in O(I).
The prime-order elements of B act trivially on A for the same reason
as before. We have shown
I
z
}|
{
G ∼ Q × (A : B) .(nontrivial cyclic 3-group) .(1 or 2)
|
{z
}
J
Having worked our way to the “middle” of G, we now work outwards
and construct Θ and (if required) Φ. Using the conjugacy of complements to A in A : B, the Frattini argument shows that NJ (B) maps
onto J/(A : B). So NJ (B) contains a Sylow 3-subgroup of J, indeed a
unique one up to conjugacy in NJ (B). So there is an essentially unique
possibility for Θ. As observed above, Θ acts nontrivially on Q. Its
3-elements act trivially on A for the same reason that B’s prime-order
22
DANIEL ALLCOCK
elements do. Finally, Θ commutes with B. Otherwise, some Sylow
subgroup of B would lie in G′ , hence act trivially on A, contrary to B’s
construction.
We have shown that J has type III, so if J = G then we are done.
Otherwise, mimicking the previous proof shows that there exists a
group Φ ∼
= Z/4 in N(B × Θ) that does not lie in J, and that such
a group is unique up to conjugacy in N(B × Θ). Also as before, Φ
commutes with B. By Φ 6≤ J, Φ’s 4-elements act on Q by outer automorphisms. Since the images of Φ and Θ in Out Q ∼
= S3 do not
commute, Φ cannot commute with Θ. Therefore Φ’s 4-elements must
invert Θ. So G has type IV and the proof is complete.
To finish the proofs of theorems 1.1 and 1.2 we will show that G
satisfies the hypotheses of one of lemmas 4.1–4.3. By imperfectness, G
has a normal subgroup M of prime index p. By induction, M has one
of the structures I–VI. We will examine the various cases and see that
one of the lemmas applies.
If M has type V or VI then it contains a unique subgroup 2A5 , which
is therefore normal in G, so lemma 4.2 applies. If M has type III
or IV then it contains a unique normal subgroup Q8 , which is therefore
normal in G. Also, this Q8 lies in M ′ , hence G′ , so lemma 4.3 applies.
Finally, suppose M has type I or II, so M ∼ A : B × (2-group T ) .
If p = 2 then O(G) = O(M) = A : B and G/O(G) is a 2-group, so
lemma 4.1 applies.
So suppose p > 2, and observe that G/O(M) has structure T.p.
Choose a subgroup P of order p in G/O(M). If it acts trivially on T
then we have G/O(M) ∼
= T ×p, so G/O(G) ∼
= T and lemma 4.1 applies.
So suppose P acts nontrivially on T . The automorphism group of any
cyclic or quaternionic 2-group is a 2-group, except for Aut Q8 ∼
= S4 .
Since P acts nontrivially, we must have T ∼
= Q8 and p = 3. The
nontriviality of P ’s action also implies T < G′ . From this and the fact
that Aut A is abelian, it follows that T acts trivially on A. So T is
normal in M. As M’s unique Sylow 2-subgroup, it is normal in G. So
lemma 4.3 applies. This completes the proofs of theorems 1.1 and 1.2.
5. Irredundant enumeration
The uniqueness expressed in theorem 1.2 makes the isomorphism classification of groups in theorem 1.1 fairly simple. Namely, one specifies
such a group G up to isomorphism by choosing its type I–VI, a suitable
number a for the order of A, a suitable subgroup G of the unit group
(Z/a)∗ of the ring Z/a, and a suitable number g for the order of G.
SPHERICAL SPACE FORMS REVISITED
23
In a special case one must also specify a suitable subgroup G0 of G.
Before developing this, we illustrate some redundancy in Wolf’s list.
Example 5.1 (Duplication). Wolf’s presentations of type II in [7, theorem 6.1.11] have generators A, B, R and relations
Am = B n = 1
R2 = B n/2
BAB −1 = Ar
RAR−1 = Al
RBR−1 = B k
where (m, n, r, k, l) are numerical parameters satisfying nine conditions.
It turns out that the six choices (3, 20, −1, −1, ±1), (5, 12, −1, −1, ±1)
and (15, 4, −1, −1, 4 or 11) give isomorphic groups, namely (3×5) : Q8 .
Here one class of 4-elements in Q8 inverts just the Z/3 factor, another
class inverts just the Z/5 factor, and the third class inverts both. Wolf
decomposes this group by choosing an index 2 subgroup (call it H)
with cyclic Sylow 2-subgroup, taking A to generate H ′ , and taking B
to generate a complement to H ′ in H (which always exists). Then he
takes R to be a 4-element outside of H, that normalizes hBi. There
are three ways to choose H, corresponding to the three Z/4’s in Q8 .
After H is chosen, there are unique choices for hAi and hBi, but two
choices for hRi. These choices lead to the six different presentations.
Our form of this group is A : (B × Q8 ) = 15 : (1 × Q8 ) = 15 : Q8 . (Our
A and B are subgroups, while Wolf’s A and B are elements.)
A minor additional source of redundancy is that replacing Wolf’s
B by a different generator of hBi can change the parameter r in the
presentation above.
Given a group G from theorem 1.1, we record the following invariants. First we record its type I–VI, which is well-defined by the easy
part of theorem 1.2, proven in section 1. Second we record g := |G|,
a := |A| and G, where bars will indicate images in Aut A. Finally,
and only if G has type II and its order is divisible by 16, we record
G0 , where G0 is the unique index 2 subgroup of G with cyclic Sylow
2-subgroups.
Aut A is canonically isomorphic to the group of units (Z/a)∗ of the
ring Z/a, with u ∈ (Z/a)∗ corresponding to the uth power map. Therefore we will regard G and G0 as subgroups of (Z/a)∗ . This is useful
when comparing two groups G, G∗ :
Theorem 5.2 (Isomorphism recognition). Two finite groups G and
G∗ , that act freely and isometrically on spheres of some dimensions,
are isomorphic if and only if they have the same type, and g = g ∗ ,
∗
∗
a = a∗ , G = G , and (if they are defined) G0 = G0 .
24
DANIEL ALLCOCK
Proof. First suppose G ∼
= G∗ . We have already mentioned that they
have the same type, and they obviously have the same order. By the
uniqueness of A and A∗ (theorem 1.2), any isomorphism G → G∗
identifies A with A∗ . In particular, a = a∗ , and the action of G on A
∗
corresponds to that of G∗ on A∗ . So we must have G = G , and (when
∗
defined) G0 = G0 .
Now suppose G and G∗ have the same invariants; we must construct
an isomorphism between them. Case analysis seems unavoidable, but
the ideas are uniform and only the details vary.
Type I. Since A and A∗ are cyclic of the same order, we may choose an
isomorphism A ∼
= A∗ . By order considerations we have |B| = |B ∗ | and
|T | = |T ∗ |. Again by order considerations, the identification of G with
∗
G under the canonical isomorphism Aut A = Aut A∗ identifies B with
∗
∗
B and T with T . Because B and B ∗ are cyclic of the same order,
∗
we may lift the identification B ∼
= B to an isomorphism B ∼
= B ∗,
∗
∗
and similarly for T and T . Together with A ∼
= A , these give an
∗
∼
isomorphism G = G .
∗
Type II when 16 divides |G| = |G∗ |. Because G0 is identified with G0 ,
the previous case gives an isomorphism G0 ∼
= G∗0 that identifies A with
∗
∗
A , B with B and the cyclic 2-group T0 := T ∩ G0 with T0∗ := T ∗ ∩ G∗0 .
We will extend this to an isomorphism G ∼
= G∗ . Choose an element φ
∗
∗
of T − T0 . By the identification of T with T and T 0 with T 0 , there
exists an element φ∗ of T ∗ − T0∗ whose action on A∗ corresponds to
φ’s action on A. As elements of T − T0 and T ∗ − T0∗ , φ and φ∗ have
order 4. They commute with B and B ∗ respectively. Their squares are
the unique involutions in T0 and T0∗ , which we have already identified
with each other. So identifying φ with φ∗ extends our isomorphism
∗
∗
G0 ∼
=G .
= G0 to G ∼
Type II when 16 does not divide |G| = |G∗ |. The argument for
type I identifies A with A∗ and B with B ∗ . Both G and G∗ have
Sylow 2-subgroups isomorphic to Q8 . T is the Sylow 2-subgroup of
G, and is elementary abelian of rank ≤ 2 because Q8 /Q′8 ∼
= 2 × 2.
∗
∗
The identification of G with G identifies T with T . Because Aut Q8
∗
acts as S3 on Q8 /Q′8 , it is possible to lift the identification T ∼
= T to
an isomorphism T ∼
= B ∗,
= A∗ , B ∼
= T ∗ . Now our identifications A ∼
T ∼
= T ∗ fit together to give an isomorphism G ∼
= G∗ .
∗
Type III. Choose an isomorphism A ∼
= A . By |G| = |G∗ | we get
∗
|B| = |B ∗ | and |Θ| = |Θ∗ |. The identification of G ≤ Aut A with G ≤
∗
∗
Aut A∗ identifies Θ with Θ and B with B . These identifications can
be lifted to isomorphisms Θ ∼
= Θ∗ and B ∼
= B ∗ by the same reasoning
as before. Because all 3-elements in Aut Q8 ∼
= S4 are conjugate, it
SPHERICAL SPACE FORMS REVISITED
25
is possible to choose an isomorphism Q8 ∼
= Q∗8 compatible with our
∗
isomorphism Θ ∼
= Θ and the homomorphisms Θ → Aut Q8 and Θ∗ →
Aut Q∗8 . Now our identifications A ∼
= Θ∗ and
= B∗, Θ ∼
= A∗ , B ∼
∗
∗
Q8 ∼
=G .
= Q8 fit together to give an isomorphism G ∼
∗
∗
Type IV. Identify A, B and Θ with A , B and Θ∗ as in the previous
case. Choose generators φ and φ∗ for Φ and Φ∗ . Their actions on A and
∗
A∗ correspond, because they act by the unique involutions in G and G
if these exist, and trivially otherwise. The image of φ in Aut Q8 ∼
= S4
normalizes the image of Θ, so together they generate a copy of S3 , and
similarly for their starred versions. Any isomorphism from one S3 in
Aut Q8 to another is induced by some conjugation in Aut Q8 . (One
checks this using Aut Q8 ∼
= S4 .) Therefore it is possible to identify Q8
∗
with Q8 , such that the actions of Θ and Θ∗ on them correspond, and
the actions of φ and φ∗ also correspond. Using this, and identifying φ
with φ∗ , gives an isomorphism G ∼
= G∗ .
Type V. Identify A and B with A∗ and B ∗ as in previous cases, and
2A5 with 2A∗5 however one likes.
Type VI. Identify A and B with A∗ and B ∗ as before, and choose
generators φ and φ∗ for Φ and Φ∗ . As in the type IV case, their actions
on A and A∗ correspond. Next, φ and φ∗ act on 2A5 and 2A∗5 by involutions which are not inner automorphisms. All such automorphisms
of 2A5 are conjugate in Aut(2A5 ). (They correspond to the involutions
in S5 − A5 .) So we may identify 2A5 with 2A∗5 in such a way that the
actions of φ and φ∗ on them correspond. Using this, and identifying φ
with φ∗ , gives an isomorphism G ∼
= G∗ .
We can now parameterize the isomorphism classes of finite groups G
that admit free actions on spheres. First one specifies a type I–VI. Then
one specifies a positive integer a, a subgroup G of (Z/a)∗ , and possibly
a subgroup G0 of G, all satisfying some constraints. Then one chooses
one or more auxiliary parameters, constrained in terms of properties
of G. Together with a, these specify g, hence the isomorphism type
of G. The following theorem is proven by combining theorem 5.2 with
an analysis of what possibilities can actually arise. We will write the
parameters a, G and g in the order one chooses them, rather than in
the order used in theorem 5.2.
The constraints on the choices of parameters are difficult to express
uniformly. But the constraint on one auxiliary parameter, called b, is
uniform. For each type we obtain a positive integer b from the structure
of G, and b must be the product of b and nontrivial powers of all the
primes dividing b. We will express this by saying “b is as above.”
26
DANIEL ALLCOCK
Theorem 5.3 (Irredundant enumeration). Suppose G is a finite group
admitting a free and isometric action on a sphere. Then there is exactly one tuple (I—VI, a, G, g) or (II, a, G, G0 , g) listed below, whose
corresponding group (defined at the end) is isomorphic to G.
(Type I, a, G, g = abt) where
(1) a is odd.
(2) G is a cyclic subgroup of (Z/a)∗ of order prime to a. Define b
and t as the odd and 2-power parts of |G|.
(3) b is as above and t is a power of 2, larger than t if t 6= 1.
(Type II, a, G, g = abt) or (Type II, a, G, G0 , g = abt) where
(1) a is odd.
(2) G is a subgroup of (Z/a)∗ which is the direct product of a cyclic
group of order prime to 2a and an elementary abelian 2-group
of rank ≤ 2. Define b and t as the orders of these factors.
(3) G0 , if specified, is a subgroup of G of index 2 (if t = 4) or
index ≤ 2 (otherwise).
(4) b is as above; and t = 8, unless G0 was specified, in which case t
is a power of 2, larger than 8.
(Type III, a, G, g = 8abθ) where
(1) a is prime to 6.
(2) G is a subgroup of (Z/a)∗ which is the direct product of a cyclic
3-group and a cyclic group of order prime to 6a. Define θ̄ and b
as the orders of these factors.
(3) b is as above and θ is a power of 3, larger than θ̄.
(Type IV, a, G, g = 16abθ) where
(1) a is prime to 6.
(2) G is a subgroup of (Z/a)∗ which is the direct product of a cyclic
group of order prime to 6a and a group of order 1 or 2. Define b
as the order of the first factor.
(3) b is as above and θ is a nontrivial power of 3.
(Type V, a, G, g = 120ab) where
(1) a is prime to 30.
(2) G is a cyclic subgroup of (Z/a)∗ of order prime to 30a. Define b
as its order.
(3) b is as above.
(Type VI, a, G, g = 240ab) where
(1) a is prime to 30.
SPHERICAL SPACE FORMS REVISITED
27
(2) G is a subgroup of (Z/a)∗ which is the direct product of a cyclic
group of order prime to 30a and a group of order 1 or 2. Define b
as the order of the first factor.
(3) b is as above.
Here are instructions for building G. In a sense this is a constructive
version of the proof of theorem 5.2. For all types, start by taking cyclic
groups A, B with orders a, b. Up to isomorphism of the domain, B
has a unique surjection to the subgroup of G ⊆ Aut A of order b. Form
the corresponding semidirect product A : B. Now we consider the six
cases.
The easiest is type V—just set G = 2A5 × (A : B).
For type I, we take a cyclic group T of order t. Just as for B, there is
an essentially unique surjection from T to the subgroup of G of order t.
Then G is the semidirect product A : (B × T ).
For type III one takes a cyclic group Θ of order θ. Just as for B,
there is an essentially unique surjection from Θ to the subgroup of G of
order θ̄. We also take Θ to act nontrivially on Q8 . (Up to conjugacy in
Aut Q8 there is a unique nontrivial action.) Then G is the semidirect
product (Q8 × A) : (B × Θ). B acts trivially on Q8 ; this is forced since
|B| is prime to 6.
A type IV or VI group is got from a type III or V group by adjoining a
suitable 4-element φ. In both cases, φ squares to the central involution,
centralizes B, and acts on A by the nontrivial involution in G (if one
exists) or trivially (otherwise). For type VI, φ acts on 2A5 by an outer
automorphism of order 2, which is unique up to Aut 2A5 . For type IV,
φ inverts Θ and acts on Q8 by an involution that inverts the action
of Θ. Such an automorphism is unique up to an automorphism of Q8
that respects the Θ-action.
One can describe type II groups in terms of type I in a similar way,
but it is easier to build them directly. Take T to be a quaternion group
of order t. First suppose t = 8. Then we did not specify G0 . Up to
automorphism of the domain there is a unique surjection from T to
the subgroup T of G of order t. We take G = A : (B × T ). On the
other hand, suppose t > 8, in which case we did specify G0 . We write
T0 for the index 2 cyclic subgroup of T . Up to automorphism of the
domain, there is a unique surjection from T to T which carries T0 onto
the 2-part of G0 (which has order 1 or 2). And again G = A : (B × T ).
References
[1] Conway, J. H., et. al., Atlas of Finite Groups., Oxford University Press, Eynsham, 1985.
28
DANIEL ALLCOCK
[2] Gorenstein, D., Finite groups. Harper & Row, New York-London, 1968.
[3] Isaacs, I. Martin, Finite group theory. Graduate Studies in Mathematics 92.
American Mathematical Society, Providence, RI, 2008
[4] Meierfrankenfeld, U., Perfect Frobenius complements, Arch. Math. (Basel) 79
(2002) 19–26.
[5] Passman, D., Permutation Groups, W. A. Benjamin, New York, 1968.
[6] Wall, C. T. C., On the structure of finite groups with periodic cohomology,
in Lie groups: structure, actions, and representations, 381–413, Progr. Math.
306, Birkhuser/Springer, New York, 2013.
[7] Wolf, J., Spaces of Constant Curvature, McGraw-Hill, New York, 1967.
[8] Zassenhaus, H., Über endliche Fastkörper, Abh. Math. Sem. Hamburg 11
(1936) 187–220.
[9] Zassenhaus, H., On Frobenius groups. I, Results Math. 8 (1985) 132–145.
Department of Mathematics,, University of Texas, Austin
E-mail address: [email protected]
URL: http://www.math.utexas.edu/~allcock
| 4math.GR
|
arXiv:1706.02179v2 [cs.CV] 8 Jun 2017
Learning to Represent Mechanics via Long-term
Extrapolation and Interpolation
Sébastien Ehrhardt1
Aron Monszpart2
Andrea Vedaldi1
Niloy Mitra2
1
Department of Engineering Science, University of Oxford
{hyenal, vedaldi}@robots.ox.ac.uk
2
Department of Computer Science, University College London
{a.monszpart, n.mitra}@cs.ucl.ac.uk
Abstract
While the basic laws of Newtonian mechanics are well understood, explaining
a physical scenario still requires manually modeling the problem with suitable
equations and associated parameters. In order to adopt such models for artificial
intelligence, researchers have handcrafted the relevant states, and then used neural
networks to learn the state transitions using simulation runs as training data. Unfortunately, such approaches can be unsuitable for modeling complex real-world
scenarios, where manually authoring relevant state spaces tend to be challenging.
In this work, we investigate if neural networks can implicitly learn physical states
of real-world mechanical processes only based on visual data, and thus enable longterm physical extrapolation. We develop a recurrent neural network architecture
for this task and also characterize resultant uncertainties in the form of evolving
variance estimates. We evaluate our setup to extrapolate motion of a rolling ball
on bowl of varying shape and orientation using only images as input, and report
competitive results with approaches that assume access to internal physics models
and parameters.
1 Introduction
Animals can make remarkably accurate and fast predictions of physical phenomena in order to
perform activities such as navigate, prey, or burrow. However, the nature of the mental models used
to perform such predictions remains unclear and is still actively researched [9].
In contrast, science has developed an excellent formal understanding of physics; for example,
mechanics is nearly perfectly described by Newtonian physics. While the constituent laws are simple
and accurate, applying them to the description of a physical scenario is anything but trivial. First, the
scenario needs to be abstracted (e.g., by segmenting the scene into rigid objects, estimating physical
parameters such as mass, linear and angular velocity, etc., deciding which equations to apply, and so
on). Then, prediction still requires the numerical integration of complex systems of equations. It is
unlikely that this is the process of mental modeling followed by natural intelligences.
In an effort to develop model of physics that are more suitable for artificial intelligence, several
authors have looked at the problem of learning physical predictors using deep neural networks. As
a notable example, the recent Neural Physics Engine (NPE) [5] uses a neural network to learn the
state transition function of mechanical systems. The state itself is handcrafted and includes physical
parameters such as positions, velocities, and masses of rigid bodies. While this approach works well,
a limitation is that it does not allow the network to learn its own abstraction of the physical system.
This may prevent the model from learning efficient approximations of physics that are likely required
to scale to complex real-world scenarios.
In this work, we ask whether a representation of the physical state of a mechanical system can
be learned implicitly by a neural network, and whether this can be used to perform more accurate
predictions. Compared to methods such as NPE, learning such a model is more challenging as no
direct observations of the state of the system are available for training. Instead, the state is a hidden
variable that must be inferred while solving a task for which supervision can be provided. As an
example of such a task, we consider here the problem of long-term physical extrapolation.
Our approach to extrapolation is to develop a recurrent neural network architecture that not only
contains an implicit representation of the state of the system, but is also able to evolve it through
time. This differs from methods such as NPE that predict instantaneous variations of the system
state, which are integrated in long-term predictions a-posteriori, after learning is complete. We show
that accounting for the integration process during learning allows the network to learn an implicit
representation of physics. Furthermore, we show that, in relatively complex physical setups, the
resulting predictions can be competitive to a modified version of NPE, even when the inputs to the
extrapolator are visual observations of the physical system instead of a direct knowledge of its initial
state.
Since physical extrapolation is inherently ambiguous, we allow the model to explicitly estimate its
prediction uncertainty by estimating the variance of a Gaussian observation model. We show that this
modification further improves the quality of long-term predictions.
Empirically, we push our model by considering scenarios beyond the “flat” ones considered in most
recent papers, such as objects sliding on planes and colliding, and look for the first time at the case of
an object rolling on a non-trivial 3D shape, namely a bowl of varying shape and orientation, where
both linear and angular momenta are tightly coupled.
As a final benefit of learning with long-term physical predictions, we show that our model is able, with
minimal modifications, to learn not only to extrapolate physical trajectories, but also to interpolate
them. Remarkably, interpolation is still obtained by computing the trajectory in a feed-forward
manner, from the first to the last time step.
The rest of the paper is organized as follows. The relation of our work to the literature is discussed
in section 2. The detailed structure of the proposed neural networks is given and motivated in section 3.
These networks are extensively evaluated on a large dataset of simulated physical experiments
in section 5. A summary of our finding can be found in section 6.
2 Related work
In this work we address the problem of long-term prediction from observation in a physical environment without voluntary perturbation, which is done by an implicit learning of physical laws. Our
work is closely related to a range of recent works in the machine learning community.
Learning intuitive physics. To the best of our knowledge [4] was the first approach to tackle
intuitive physics with the aim to answer a set of intuitive questions (e.g., will it fall?) using physical
simulations. Their simulations, however, used a sophisticated physics engine that incorporates prior
knowledge about Newtonian physical equations. More recently [17] also used static images and a
graphics rendering engine (Blender) to predict movements and directions of forces from a single RGB
image. Motivated by the recent success of deep learning for image processing (e.g., [12, 10]), they
used a convolutional architecture to understand dynamics and forces acting behind the scenes from a
static image and produced a “most likely motion" rendered from a graphics engine. In a different
framework, [14] and [15] also used the power of deep learning to extract an abstract representation
of the concept of stability of block towers purely from images. These approaches successfully
demonstrated that not only was a network able to accurately predict the stability of the block tower
but in addition, it could identify the source of the instability. Other approaches such as [2] or [7] also
attempted to learn intuitive physics of objects through manipulation. These approaches, however, do
not attempt to precisely model the evolution of the physical world.
Learning dynamics. Learning the evolution of an object’s position also implies to learn about
the object’s dynamics regardless of any physical equations. While most successful techniques
used LSTM-s [11], recent approaches show that propagation can also be done using a single crossconvolution kernel. The idea was further developed in [27] in order to generate a next possible image
frame from a single static input image. The concept has been shown to have promising performance
regarding longer term predictions on the moving MNIST dataset in [6]. The work of [19] also shows
that an internal hidden state can be propagated through time using a simple deep recurrent architecture.
2
These results motivated us to propagate tensor based state representations instead of a single vector
representation using a series of convolutions. Adversarial losses have also been used in [18] which
shows good results in video segmentation. In the future we also aim to experiment with approaches
inspired by [27].
Learning physics. The works of [26] and its extension [25] propose methods to learn physical
properties of scenes and objects. However, in [26] the MCMC sampling based approach assumes the
complete knowledge of the physical equations to estimate the correct physical parameters. In [25]
deep learning has been used more extensively to replace the MCMC based sampling but this work
also employs an explicit encoding and computation of physical laws to regress the output of their
tracker. [22] also used physical laws to predict the movement of a pillow from unlabelled data though
their approach was only applied to a fixed number of frames.
In another related approach [8] attempted to build an internal representation of the physical world.
Using a billiard board with an external simulator they built a network which observing four frames
and an applied force, was able to predict the 20 next object velocities. Generalization in this work was
made using an LSTM in the intermediate representations. The process can be interpreted as iterative
since frame generation is made to provide new inputs to the network. This can also be seen as a
regularization process to avoid the internal representation of dynamics to decay over time which is
different to our approach in which we try to build a stronger internal representation that will attempt
to avoid such decay.
Other research attempted to abstract the physics engine enforcing the laws of physics as neural
network models. [3] and [5] were able to produce accurate estimations of the next state of the world.
Although the results look plausible and promising, reported results show in [5] that accurate long-term
predictions are still difficult. Note, that their process is an iterative one as opposed to ours, which
propagates an internal state of the world through time similarly to [20].
Approximate physics with realistic output. Other approaches also focused on learning to generate
realistic future scenarios ([24] and [13]), or inferring collision parameters from monocular videos [16].
In these approaches the authors used physics based losses to produce visually plausible yet erroneous
results. They however show promising results and constructed new losses taking into account
additional physical parameters other than velocity. Note also that in [3] an energy-based loss has
been used. It can be seen as a way to explicitly incorporate a knowledge of physics in the network
while we aim to understand if we can make accurate prediction without explicit physics knowledge.
3 Method
In this section, we propose a new neural network model (see Fig. 1) that performs predictions in
mechanical systems. Let yt be a vector of physical measurements taken at time t, such as the position
of an object whose motion we would like to track. Physical systems satisfy a Markov condition, in the
sense that there exists a state vector ht such that 1) measurements yt = g(ht ) can be predicted from
the value of the state and 2) the state at the next time step ht+1 = f (ht ) depends only on the current
value of the state ht . Uncertainty in the model can be encoded by means of transition p(ht+1 |ht ) and
observation p(yt |ht ) probabilities, resulting in a hidden Markov model.
Approaches such as NPE [5] start from an handcrafted definition of the state ht . For instance, in order
to model a scenario with two balls colliding, one may choose ht to contain the position and velocity
of each ball. In this case, the observation function g may be as simple as extracting the position
components from the state vector. The goal of NPE is then to learn a neural network approximator φ
of the transition function f . In practice, the authors of [5] suggest that it is often easier to predict a
rate of change ∆t for some of the physical parameters (e.g. the balls’ velocities), which can then be
integrated to update the state: ht+1 = f˜(ht , ∆t ) where f˜ is an hand-crafted integrator and the neural
predictor estimates the change ∆t = φ(ht ).
While these approaches have several advantages [5], there are several limitations too. First, approaches
such as NPE require to handcraft the state representation ht . Even in the simple case of the colliding
balls, the choice of state is ambiguous; for example, one could include in the state the radius, mass and
elasticity and friction coefficients. In more complex situations, choosing a good state representation
may be rather difficult. Overall, this choice is best left to learning. Second, the state must be
observable during training in order to learn the transition function. Third, learning does not account
3
Transition
Network
Encoder network
(VGG)
Decoder Network
Input images t = 0 . . . 3
PhysNet
ProbNet
Figure 1: Overview of our proposed pipeline. The first four images of a sequence first pass through a
partially pre-trained feature encoder network to build the concept of physical state. It then recursively
passes through a transition layer to produce long-term predictions about the future states of the
objects. It is then decoded to produce state estimates. While our PhysNet model is trained to regress
the next states, the ProbNet model trained with the log-likelihood loss is also able to handle the
notion of uncertainty thanks to its extended state space.
for the effect of accumulating errors through integration as integration is applied only after learning.
Finally, the initial value of the state h0 must be known in order to initialize the predictor, whereas
in many applications one would like to start from sensory inputs xt such as images of the physical
system [8].
We propose here an approach to address these difficulties. We assume that the state ht is a hidden
variable, to be determined as part of the learning process. Since the ht cannot be observed, the
transition function ht+1 = f (ht ) cannot be estimated directly as in the NPE. Instead, it must be
inferred as a good explanation of the physical measurements yt . Since the evolution of the state ht
cannot be learned by observing measurement yt in isolation, we supervise the system by explaining
sequences y[0,T ) = (y0 , . . . , yT −1 ). This requires to move the integration step inside the network,
which we do by mean of a recurrent neural network architecture. This has the added advantage of
making learning aware of the integration process, which helps improving accuracy.
The model is analogous to a Hidden Markov Model. Recall that such models are often learned
by maximizing the likelihood of the observations after marginalizing the hidden state.1 However,
since we are interested in extrapolating future observations from past ones, we consider instead
long-term extrapolation as supervisory signal. In order to do so we learn: 1) a transition function
ht+1 = φ(ht ) that evolves the state through time, 2) a decoder function that maps the state ht to an
observation yt = φdec (ht ), and 3) an encoder function that estimates the state ht = φenc (x(t−T0 ,t] )
from the T0 most recent sensor readings (alternatively ht = φenc (y(t−T0 ,t] ) can use the T0 most
recent observations).
In the experiments (section 5) we will show that the added flexibility of learning an internal state
representation automatically can still provide a good prediction accuracy even when the complexity
of the physical scenarios increases. The rest of the section discusses the three modules, encoder,
transition, and decoder maps, as well as the loss function used for training. Further technical details
can be found in section 5.
(i) Encoder map: from images to state. The goal of the encoder map is to take T0 consecutive video
frames recording the beginning of the object motion and to produce an estimate h0 = φenc (x(−T0 ,0] )
1
Formally,
the
Markov
model
is
given
by
p(y[0,T )] , h[0,T ) )
=
Q −2
p(h0 )p(y0 |h0 ) Tt=0
p(ht+1 |ht )p(yt+1 |ht+1 ); traditionally, p can be learned as the maximizer of the
log-likelihood maxp Ey [log Eh [p(y, h)]], where we dropped the subscripts for compactness. Learning to
interpolate/extrapolate can be done by considering subsets ȳ ⊂ y of the measurements as given and optimizing
the likelihood of the conditional probability maxp Ey [log Eh [p(y, h|ȳ)]].
4
of the initial state of the physical system. In order to build this encoder, we follow [8] and concatenate
the RGB channels of the T0 images in a single Hi × Wi × 3T0 . The latter is passed to a convolutional
neural network φenc outputting a feature tensor s0 ∈ RH×W ×C , used as internal representation of the
system. We also add to the state a vector p0 ∈ R2 to store the 2D projection of the object location on
the image plane, so that ht = (st , pt ).
(ii) Transition map: evolving the state. The state ht is evolved through time by learning the
transition function φ : ht 7→ ht+1 , where h0 is obtained from the encoder map, so that ht =
φt (φenc (x(−T0 ,0] )). The state st is updated by using a convolutional network st+1 = φs (st ) whereas
pt is updated incrementally as pt+1 = pt + φp (st ), where φp (st ) is estimated using a single layer
perceptron regressor. Hence (st+1 , pt+1 ) = φ(st , pt ) = (φs (st ), pt + φp (st )). We found that
explicitly incorporating an additive update significantly improves the performance of the model.
(iii) Decoder map: from state to probabilistic predictions. Since we added for convenience the
projected object position pt to the state, the decoder map ŷt = φdec (st , pt ) = pt simply extracts and
returns that part of the state. Training optimizes the average L2 distance between ground truth yt and
PT −1
predicted ŷt positions T1 t=0 kŷt − yt k2 .
Since extrapolation is inherently ambiguous, the L2 prediction error increases with time, which
may unbalance learning. In order to address this issue, we allow the model to explicitly and
dynamically express its prediction uncertainty by outputting the mean and variance (µt , Σt ) of a
bivariate Gaussian observation model. The L2 loss is then replaced with the negative log likelihood
PT −1
− T1 t=0 log N (yt ; µt , Σt ).
In order to estimate µt and Σt , the incremental state component pt = (µt , λ1,t , λ2,t , θt ) is extended
to include both the mean as well as the eigenvalues and rotation of the variance matrix Σt =
R(θt )> diag(λ1,t , λ2,t )R(θt ). In order to ensure numerical stability, eigenvalues are constrained
to be in the range [0.01 . . . 100] by setting them as the output of a scaled and translated sigmoid
λi,t = σλ,α (βi,t ), where σλ,α (z) = λ/(1 + exp(−z)) + α.
4 Experimental setup
In our experimental setup (Fig. 2), we consider a sphere rolling inside a 3D (bowl) surface. When the
bowl is a hemisphere we refer to the setup as ‘Bowl,’ and in the more general case as ‘Ellipse’ (see
Table 1).
We use p = (px , py , pz ) ∈ R3 to denote a point in 3D space or a vector (direction). The camera
center is placed at location (0, 0, cz ), cz > 0 and looks downward along vector (0, 0, −1) using
orthographic projection such that the point (px , py , pz ) projects to a pixel (px , py ) in the image.
(0, 1, 1)
(0, 0, cz )
(0, 0, 0)
(a, 0, 1)
(a)
(b)
(c)
Figure 2: We consider the problem of understanding and extrapolating mechanical phenomena with
recurrent deep networks. (a) Experimental setup: an orthographic camera looks at a ball rolling in a
3D bowl. (b) Example of a 3D trajectory in the 3D bowl simulated using Blender 2.77’s OpenGL
renderer. (c) An example of a rendered frame in the ‘Ellipse‘ experiment that is fed to our model as
input.
We model the bowl as the bottom half of an ellipsoid given by x2 /a2 + y 2 + (z − 1)2 = 1 with
its axes aligned to the XYZ axes and its bottom point being at the origin. We vary the ellipsoid
5
shape by sampling a ∈ U [0.5, 1] for the ‘Ellipse’ case and setting a = 1 (i.e., a hemisphere) for the
‘Bowl’ case. The bowl is given a checker board pattern. Finally, the bowl is given a random rotation
γ ∈ U [−π/2, π/2] only about the z-axis to randomly orient it.
We consider a rolling object in the form of a ball with radius ρ = 0.04 with its center of mass at
time t being located at qt = (qxt , qyt , qzt ), so that its center of mass is imaged at pixel (qxt , qyt ) at any
time t. The ball has a fixed color texture attached to its surface, so it appears as a painted object. We
initially position the ball at angles (θ, φ) with respect to the the bowl center, where the elevation θ
is uniformly sampled in the range θ ∈ U [−9π/10, −π/2] and the azimuth φ ∈ U [−π, π]. We set
the minimum elevation to −9π/10 to avoid starting the ball at the bottom of the bowl. In the end,
the ball will be resting on the bowl surface. The ball is either textured with random color patches or
uniformly colored in white in order to study the impact of observing the ball rotation.
We set the initial orientation of the ball by uniformly sampling its xyz Euler angles in [−π, π]. We set
its initial velocity v by first sampling vx , vy uniformly in the range [5, 10], assigning each of vx , vy a
random sign, and then projecting vector (vx , vy , 0) so that the resulting velocity vector is tangential
to the underlying supporting bowl.
α
Note that, while several parameters of the ball state are included in the observation vector y[−T
,
0 ,T )
these are not part of the state of the neural network, which is inferred automatically. The network
itself is tasked with predicting part of these measurements, but their meaning is not hardcoded.
Simulation setup. For efficiency, we extract multiple sub-sequences xα
[−T0 ,T ) form a single longer
simulation (training, test, and validation sets are completely independent). The simulator runs at
120fps for accuracy, but the data is subsampled to 40fps. We use Blender 2.77’s OpenGL renderer
and the Blender Game Engine (relying on Bullet 2 as physics engine). The ball is a textured sphere
with unit mass. We found that changing the friction parameter of the bowl or the ball does not
influence the motion. Therefore, we added translation and rotation damping (both set to 0.1 in
Blender) to the sphere’s animation properties in order to simulate energy loss due to friction. The
simulation parameters were set as: max physics steps = 5, physics substeps = 3, max logic steps = 5,
FPS = 120. Rendering used white environment lighting (energy = 0.7) and no other light source. The
object color was set to a colored checkerboard texture in order to enable the visual perception of
rotation. We used 70% the data for training, 15% for validation, and 15% for test. During training
we start observation at a random time while it is fixed for test. The output images were stored as
128 × 128 color JPEG files.
5 Experiments
5.1 Baselines
Least squares fit. We compare the performance of our methods to two simple least squares baselines:
Linear and Quadratic. In both cases we fit two least squares polynomials to the screen-space
coordinates of the first T = 10 frames, which are not computed but rather given as inputs. The
polynomials are of first and second degree(s), respectively. Note, that being able to observe the first
10 frames is a large advantage compared to the networks, which only see the first T0 = 4 frames.
NPE. NPE [5] training was done using available online code. We used the same training procedure
as reported in [5]. NPE++ additionally takes angle and angular velocities as parameters and also
predicts angular velocity. In the case of the elliptic bowl, both scaling and bowl rotation angle are
given as input to the networks. In this case NPEs methods carry forward the estimated states via the
network.
While the previously mentioned methods start from state inputs, note that our models work with
raw images as direct observation of the world. Physical properties are then deduced from the
observation and then integrated through our Markov model. Thus we do not need a simulator to
estimate parameters of the physical worlds (such as scaling and rotation angle in the NPE case) and
can train our model on changing environment without requiring additional external measurements of
the underlying 3D spaces.
6
5.2 Results
Implementation details. The encoder network φenc is obtained by taking the ImageNet-pretrained
VGG16 network [21] and retaining the layers up to conv5 (for an input image of size (Hi , Wi ) =
(128, 128, 3) this results in a (8, 8, 512) state tensor st ). The filter weights of all layers except conv1
are retained for fine-tuning on our problem. conv1 is reinitialized as filters must operate on images
with 3T0 channels. The transition network φs (st ) uses a simple chain of two convolution layers (with
256 and 512 filters respectively, of size 3 × 3, stride 1, and padding 1 interleaved by a ReLU layer.
Network weights are initialized by sampling from a Gaussian distribution.
Training uses a batch size of 50 using the first 20 or 40 positions (and angular velocity when explicitly
mentioned) of each video sequence using RMSProp [23]. In our methods we start with a learning rate
of 10−5 and decrease learning rates by a factor of 10 when no improvements of the L2 position loss
have been found after 100 consecutive epochs. Training is halted when the L2 loss hasn’t decreased
after 200 successive epochs; 2,000 epochs were found to be usually sufficient for convergence. Note
here than in every cases where we estimate the angular velocity the corresponding L2 loss on the
latter is simply added to the network’s existing loss.
Since during the initial phases of training the network is very uncertain, the model using the Gaussian
log-likelihood loss wasP
found to get stuck on solutions with very high variance Σ(t); to solve this
issue, the regularizer λ t det Σ(t) was added to the loss, setting λ = 0.01.
In all our experiments we used Tensorflow [1] r0.12 on a single NVIDIA Titan X GPU. In the
following we will refer to our model that has been trained optimizing on the L2 loss over positions as
PhysNet otherwise ProbNet for log-likelihhod loss. When also predicting angular velocities the suffix
‘++’ is added to the name of the model.
Table 1: Long term predictions. The PhysNet and ProbNet models observed the T0 = 4 first frames
as input. PhysNet ++ and ProbNet ++ additionally estimate angular velocity at each time step adding
a L2 angular velocity loss to our position loss. All networks have been trained to predict the T = 20
first positions, except for the NPE and NPE++ which were given T0 = 4 states as input and train to
predict state at time T0 + 1. We report here results for time T = 20 and T = 40. For each time we
report on the left L2 position loss and L2 angular loss on the right. Perplexity (loge values shown in
the table) is defined as 2−E[log2 (p(x))] where p is the estimated posterior distribution.
Method
Images
Linear
Quadratic
NPE
NPE++
PhysNet
PhysNet ++
ProbNet
No
No
No
No
Yes
Yes
Yes
ProbNet ++
Yes
20
39.2
164.3
2.6
2.9
3.0
3.5
2.9
(4.5)
3.4
(4.7)
Bowl
L2 (Perplexity)
7.5
18.4
–
0.8
–
15.9
–
1.2
127.5
120.1
6.0
8.1
29.7
15.9
24.2
(21.9)
15.3
(9.2)
40
20
17.9
861.2
–
2.0
–
11.9
–
3.4
61.9
11.7
3.2
5.7
2.5
2.1
2.9
(32.1)
4.0
(4.5)
Ellipse
L2 (Perplexity)
23.3
14.8
–
1.7
–
1.0
–
1.8
20.1
93.1
6.1
17.5
20.6
16.1
21.8
(54.0)
16.7
(9.3)
40
Ellipse (no ball texture)
L2 (Perplexity)
40
–
–
–
–
–
–
–
–
–
–
–
–
–
44.6
–
1.0 16.2
3.8
–
24.0
–
(12.7)
1.3 15.0
3.5
(8.2)
20
–
70.6
–
3.1
4.4
–
3.8
–
–
–
–
5.2
1.6
3.1
(5.0)
4.3
(4.5)
Extrapolation. Table 1 and Fig. 3 compare the baseline predictors and the four networks on the
task of long term prediction of the object trajectory. All methods observed only the first T0 = 4
inputs (whether frames or object states) except for the linear and quadratic baselines, and aimed
to extrapolate the trajectory to 40 time steps. In that sense predictions can be seen as ”long terms”
relative to the number of inputs.
Table 1 reports the average L2 errors at time Ttest = 20 and 40 for the different estimated parameters.
However all of the methods can perform arbitrary long predictions. In particular our methods are only
trained to predict 20 first positions and reveal to be still competitive with NPEs methods at T = 40.
Our networks perform reasonably well compared to the NPEs methods using only images as inputs.
In our scenario we can see the different NPEs as upper limits of our experiments since they do
have access to the state space (complete or not). Furthermore adding angular velocity has shown to
improve performances of our models while it decreases the accuracy of the NPE++ predictions in
that case. Besides probability based losses show that our models were also able to predict uncertainty
in its outputs even in the case of unobserved scenarios.
7
Ellipse
20.0
L2 residual (pixels)
15.0
12.5
10.0
7.5
5.0
2.5
0.0
Linear
Quadratic
NPE++
PhysNet++
ProbNet++
17.5
L2 residual (angular velocity)
17.5
Ellipse
20.0
Linear
Quadratic
NPE
NPE++
PhysNet
PhysNet++
ProbNet
ProbNet++
15.0
12.5
10.0
7.5
5.0
2.5
0
5
10
15
20 25
Times
30
35
0.0
40
0
5
(a)
10
15
20 25
Times
30
35
40
(b)
Figure 3: Quantitative results. error evolution on Ellipse experiments for all time steps up to 40.
Error bars denote 25th and 75th percentiles of the L2 loss. (a) L2 position loss in pixels. (b) L2
angular velocity loss.
In addition training on a dataset where angular velocity was not explicitly seen (Ellipse no-texture
in Table 1) shows that our models can still provide encouraging results in that case. It managed to
accurately deduce angular velocity without seeing the ball spinning.
Interpolation. In order to remove ambiguity of a short term observed motion one can just indicates
the final desired state. In this experiment we concatenate to the first T0 = 4 input frames the last
frame observed at T = 40 and give it as an input to a model with the same architecture as PhysNet.
This model was then trained using the same aforementioned method with the only difference that we
also extract last positions from the first extracted feature. While this idea is fairly simple it shows to
be very efficient in practice as shown in Table 2 as it efficiently removed the motion ambiguity.
Table 2: Interpolation. InterpNet is essentially the same as PhysNet but takes as input a concatenation
of first T0 = 4 frames and last frame at T = 40. All networks have been trained to predict the T = 40
first positions. InterpNet predicts T = 40 positions using the first extracted feature.
Method
InterpNet
PhysNet
10
1.37
2.19
Bowl
L2
20
30
1.84 1.65
3.64 3.94
40
1.02
4.99
10
1.03
1.40
Ellipse
L2
20
30
1.60 1.35
2.39 2.71
40
0.65
3.00
6 Conclusions
In this paper we studied the possibility of abstracting the knowledge of physics using a single neural
network with a recurrent architecture for long term predictions. We compared our model to strong
baselines on the non-trivial motion of a ball rolling on a 3D bowl with different possible shapes. As
opposed to other concurrent approaches we do not integrate physical quantities but implicitly encode
the states in a feature vector that we can propagate through time.
Our experiments on synthetic simulations indicate that we can still make reasonable predictions
without requiring an explicit encoding of the state space. Besides they are also able to estimate
a distribution over such parameters to account for uncertainty in the predictions. While keeping
the same architecture we also show that we were able to remove motion ambiguity by showing the
network its targeted final states. However, the internal state propagation mechanism is still limited by
its ability to make accurate long term predictions outside observed regimes.
In the future we will aim to bring more robustness to our models by enforcing invariance to observed
regimes to enable longer accurate predictions. Besides, a next obvious step will also be to test the
framework on video footage obtained from real-world data in order to assess the ability to do so from
visual data affected by real nuisance factors.
8
References
[1] Abadi et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015.
Software available from tensorflow.org.
[2] Pulkit Agrawal, Ashvin V Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. Learning to
Poke by Poking: Experiential Learning of Intuitive Physics. In Proc. NIPS, pages 5074–5082,
2016.
[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction
networks for learning about objects, relations and physics. In Proc. NIPS, pages 4502–4510,
2016.
[4] Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of
physical scene understanding. PNAS, 110(45):18327–18332, 2013.
[5] Michael B Chang, Tomer Ullman, Antonio Torralba, and Joshua B Tenenbaum. A compositional
object-based approach to learning physical dynamics. arXiv preprint arXiv:1612.00341, 2016.
[6] Bert De Brabandere, Xu Jia, Tinne Tuytelaars, and Luc Van Gool. Dynamic filter networks. In
Proc. NIPS, 2016.
[7] Misha Denil, Pulkit Agrawal, Tejas D Kulkarni, Tom Erez, Peter Battaglia, and Nando de Freitas.
Learning to perform physics experiments via deep reinforcement learning. Deep Reinforcement
Learning Workshop, NIPS, 2016.
[8] Katerina Fragkiadaki, Pulkit Agrawal, Sergey Levine, and Jitendra Malik. Learning visual
predictive models of physics for playing billiards. arXiv preprint arXiv:1511.07404, 2015.
[9] J. B. Hamrick, P. W. Battaglia, T. L. Griffiths, and J. B. Tenenbaum. Inferring mass in complex
scenes by mental simulation. Cognition, 157:61–76, 2016.
[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In IEEE CVPR, pages 770–778, 2016.
[11] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735–
1780, 1997.
[12] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In Proc. NIPS, pages 1097–1105, 2012.
[13] Ladický, Jeong, SoHyeon, Barbara Solenthaler, Marc Pollefeys, Markus Gross, et al. Datadriven fluid simulations using regression forests. ACM Trans. on Graphics (TOG), 34(6):199,
2015.
[14] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by
example. arXiv preprint arXiv:1603.01312, 2016.
[15] Wenbin Li, Aleš Leonardis, and Mario Fritz. Visual stability prediction and its application to
manipulation. arXiv preprint arXiv:1609.04861, 2016.
[16] Aron Monszpart, Nils Thuerey, and Niloy Mitra. SMASH: Physics-guided Reconstruction of
Collisions from Videos. ACM Trans. on Graphics (TOG), 2016.
[17] Roozbeh Mottaghi, Hessam Bagherinezhad, Mohammad Rastegari, and Ali Farhadi. Newtonian
scene understanding: Unfolding the dynamics of objects in static images. In IEEE CVPR, 2016.
[18] Natalia Neverova, Pauline Luc, Camille Couprie, Jakob Verbeek, and Yann LeCun. Predicting
deeper into the future of semantic segmentation. arXiv preprint arXiv:1703.07684, 2017.
[19] Peter Ondruska and Ingmar Posner. Deep tracking: Seeing beyond seeing using recurrent neural
networks. In Proc. AAAI, 2016.
[20] David Silver, Hado van Hasselt, Matteo Hessel, Tom Schaul, Arthur Guez, Tim Harley, Gabriel
Dulac-Arnold, David Reichert, Neil Rabinowitz, André Barreto, and Thomas Degris. The
predictron: End-to-end learning and planning. CoRR, abs/1612.08810, 2016.
9
[21] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image
recognition. In International Conference on Learning Representations, 2015.
[22] Russell Stewart and Stefano Ermon. Label-free supervision of neural networks with physics
and domain knowledge. CoRR, abs/1609.05566, 2016.
[23] T. Tieleman and G. Hinton. Lecture 6.5—RMSProp: Divide the gradient by a running average
of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 2012.
[24] J. Tompson, K. Schlachter, P. Sprechmann, and K. Perlin. Accelerating Eulerian Fluid Simulation With Convolutional Networks. ArXiv e-print arXiv:1607.03597, 2016.
[25] Jiajun Wu, Joseph J Lim, Hongyi Zhang, Joshua B Tenenbaum, and William T Freeman. Physics
101: Learning physical object properties from unlabeled videos. In Proc. BMVC, 2016.
[26] Jiajun Wu, Ilker Yildirim, Joseph J Lim, Bill Freeman, and Josh Tenenbaum. Galileo: Perceiving
physical object properties by integrating a physics engine with deep learning. In Proc. NIPS,
pages 127–135, 2015.
[27] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics:
Probabilistic future frame synthesis via cross convolutional networks. In Proc. NIPS, 2016.
10
| 2cs.AI
|
Exact uniform sampling over catalan structures
arXiv:1803.03945v1 [cs.DM] 11 Mar 2018
Alexandros Angelopoulos, Eleni Bakali
Department of Computer Science,
School of Electical and Computer Engineering,
National Technical University of Athens, Greece.
{angelop,mpakali}@corelab.ntua.gr
March 13, 2018
Abstract
We present a new framework for creating elegant algorithms for
exact uniform sampling of important Catalan structures, such as triangulations of convex polygons, Dyck words, monotonic lattice paths and
mountain ranges. Along with sampling, we obtain optimal coding, and
optimal number of random bits required for the algorithm. The framework is based on an original two-parameter recursive relation, where
Ballot and Catalan numbers appear and which may be regarded as
to demonstrate a generalized reduction argument. We then describe
(a) a unique n × n matrix to be used for any of the problems -the
common pre-processing step of our framework- and (b) a linear height
tree, where leaves correspond one by one to all distinct solutions of
each problem; sampling is essentially done by selecting a path from
the root to a leaf - the main algorithm. Our main algorithm is linear
for a number of the problems mentioned.
keywords uniform sampling, optimal coding, catalan structures, convex
triangulations, ballot numbers
1
Introduction
The consideration of different triangulations of a point set first appears in a
well-studied form in the middle of the 18th century by Leonhard Euler: he
successfully conjectured the closed formula for the number of triangulations
of the convex n-gon, that was what we denote now by Cn−2 , the (n − 2)-th
Catalan number. These numbers, named after the Belgian mathematician
Eugène Charles Catalan, satisfy the following basic relations:
1
Cn+1 =
n
X
Ck Cn−k ,
C0 = 1
(1)
k=0
2n
1
, C0 = 1
Cn =
n+1 n
4n
4n
Cn ≤ 3/2 √
and
lim
√ =1
n→∞ Cn n3/2 π
n
π
(2)
(3)
They occur as the solution of a very large number of counting problems
in combinatorics. [30] gives more than 200 interpretations, more than 60
exercises, in all, a spectacular volume of work centered around the Catalan
numbers.
Some modern views regarding this core problem is the study of algorithms for efficient coding of triangulations, and for uniform generation or
sampling a random triangulation.
Motivation-Applications. Triangulations of polygons are fundamental structures with many applications e.g. in Computational Geometry
[6, 22], VLSI floorplanning [25], and Graph Drawing [21]. Many algorithms like point location, ray-shooting, visibility area computing and computing shortest paths inside simple polygons, are based on triangulations
[31, 13, 14, 18]. In particular sampling (or equivalently uniform generation)
of random triangulations, apart from its theoretical interest, has many practical applications. Interestingly sampling triangulations was first considered
by physicists for its applications in the study of two dimensional quantum
gravity [3]. Sampling plays a central role in testing and verifying the time
complexity of particular implementations of algorithms like the previously
mentioned, that depend on triangulations. A method to achieve time verification is presented in [10]. Finally efficient coding of triangulations, is
essential for storing them after their random generation, since these data
are usually huge and inefficient coding may cause storing to dominate the
running time of the algorithm that uses these data.
Previous work. There has been significant work on uniform sampling
and coding triangulations, especially since the 1990s [11, 9, 7, 8, 24, 28, 27,
26].
We will mostly compare our results and framework to [8]: Devroye et
al. describe a linear-time algorithm in the RAM model of computation, for
the uniform generation of convex triangulations and monotone lattice paths.
This is actually equivalent to a O(n log n) algorithm, when taking into account that this algorithm needs the generation of O(n) random numbers of
O(log n) bits each. The ballot theorem is invoked in their work (see also
[12], Chapter III), as it describes lattice path problems.
Another line of research regarding triangulations is the study of Markov
chains that move along similar triangulations, most commonly considering as
2
similar those that differ by a single edge flip [16, 20, 19, 23]. These Markov
chains converge to the uniform distribution over the set of triangulations,
thus yielding almost uniform sampling algorithms (this is the well known
Markov chain Monte Carlo method for sampling and counting). The best
of the above mentioned algorithms runs in O(n5 ), but has a deviation ǫ > 0
from the uniform distribution (which is a common drawback/characteristic
of the MCMC method).
While for convex graphs it is the Catalan numbers that give the exact number of triangulations, no such formula exists for a general graph.
Exhaustively enumerating the triangulations is not an easy task: already
Cn = Θ(n−3/2 4n ), while the best known bounds for the generic case are
currently set a lower Ω(2.43n ) ([27]) and an upper O(30n ) ([26]). There is
also no general formula for counting triangulations of convex polygons with
forbidden edges, i.e. when the input is a general graph G embedded to the
plane, and valid triangulations are those that use only edges of G. Naturally,
there has been work on counting the triangulations given a specific point
set asymptotically faster than by enumerating all triangulations ([2]), and
approximately counting with favorable compromises ([1]).
1.1
Our contribution
In this article, we present a unified framework to obtain algorithms for exact
uniform sampling and optimal coding of Catalan structures. We will use the
triangulations of convex polygons as our main problem. We will then use the
monotone lattice path problem to show that our framework can be tweaked
to easily get the very algorithms presented in [8]. In a nutshell, we present:
• new framework and elegant algorithm for exact uniform sampling over
convex triangulations and other Catalan structures;
• new recursive relation for ballot and Catalan numbers with an interesting combinatorial interpretation;
• separate pre-processing step of O(n2 ) time, common to all problems
that can be described by the recursive relation,
• optimal coding of the solutions/samples of the problems: for an input
related to size n, we need exactly log(Cn ) bits to encode each solution;
• optimal number of random bits for the main sampling algorithm: it
needs exactly log Cn random bits to run, i.e. to generate uniformly at
random a solution;
• efficient algorithm for sampling, as well as counting triangulations for
a more general family of graphs embedded to the plane in convex
position: Kn,−m is obtained by an embedding of the complete graph
Kn after removing m consecutive span-2 edges.
3
Although our sampling algorithm has total running time O(n2 ), which
is worse than the O(n log n) of [8], our preprocessing step helps to reduce to
the optitmal the number of random bits needed, as well as the length of the
codewords.
An outline - main ideas Given n, we consider an initial instance which
-in some sense- describes the space of all potential solutions; for instance, in
the case of convex triangulations, we consider the complete convex graph
Kn , as all n(n − 3)/2 diagonals are in our disposal to be selected for some
triangulation, which requires only n − 3 such edges. Our goal is to describe1
a binary tree of polynomial height, with its leaves corresponding one by one
to all triangulations of the convex n-gon and then aim at sampling a leaf
uniformly at random.
In order to achieve this, we need a reduction argument, w.r.t. which the
set of potential solutions on some node’s graph can actually be partitioned
into the solutions obtained either by its left or the right child and which
are disjoint. Then, in order to sample a solution uniformly at random, it
would suffice to be able to calculate the size of each of the two subtrees of
any node [17, 29]. Thus, starting from the root, we should be capable of
uniformly selecting a leaf by recursively selecting one of a node’s children
with probability proportional to the size of the corresponding subtree.
Usually, the estimation of the size of subtrees for an arbitrary tree is
computationally hard [4], and this is the reason that the above method
usually fails. The innovation of our approach consists in that we manage
to construct this tree in a systematic way that not only reflects our reduction argument, but also plenty of isomorphisms are revealed so that (a) the
number of all classes of non-isomorphic instances is polynomial in n (in particular quadratic), and (b) it is easy to determine the class in which each
instance belongs. Eventually, it suffices to solve the general problem only
for a polynomial number of non-isomorphic sub-instances, in a preprocessing
step, bottom-up, in time linear to the number of non-isomorphic instances,
thus in total quadratic in n.
In particular, it occurs that all internal nodes of our tree correspond to
(sub)instances of a more general problem, needing a second parameter to
be described and for which no formula nor algorithm was known until now:
compute the number of triangulations of an embedded to the plane almost
complete convex graph Kn,−m , a graph missing only m consecutive span-2
edges (or ears, as denoted in [15, 8]) from what would be a the complete Kn .
As a first simple note, we get that m ≤ n, thus we have the O(n2 ) classes
of non-isomorphic instances.
In the Sections to follow, we will state and prove properties of the central
1
The construction is done only mentally, for the purposes of the analysis; it is not
actually performed by the algorithm.
4
to our work recursive relation, the sampling algorithm scheme it implies and
its application to convex triangulations and mountain ranges problems (the
latest to be trivially linked also to Dyck words, monotone lattice paths, etc.).
2
Our framework
Let an,m denote the number of solutions corresponding to any problem parameterized by n, m, and consider the following recursive relation:
an,0 = an−1,0 + an,1
(4)
an,m = an−1,m−1 + an,m+1
(5)
a0,0 = 1 and an,n = 0, ∀i ≥ 1
(6)
Theorem 1 Consider the ballot numbers
i+1−j i+1+j
, i, j ≥ 0
Ni,j =
i+1+j
j
Along with defining Ni,−1 = 0, the ballot numbers satisfy 4, 5 and 6 for i = n
and j = n − 1 − m.
In other words, an,m = Nn,n−1−m . Note also that:
2n
2n
2
1
an,0 = Nn,n−1 =
=
= Cn
2n n − 1
n+1 n
(7)
We will from now on refer to the relations 4, 5 and 6 as the BC (recursive)
relation or recursion, a naming derived from its close relationship with ballot
and Catalan numbers.
Theorem 2 The generating function for the numbers an,m is
√
(1 − xy)(1 − 1 − 4x − 2xy)
G(x, y) =
2x(1 − y + xy 2 )
Proof. [Proofs of Theorems 1 and 2] The proofs can be found in Appendix
A.
The BC table Immediately, we may define a n × n table/matrix, with
an,m as entries. It can be calculated in O(n2 ) time, bottom-up (Figure 1),
and will be considered as our framework’s pre-processing step. The main
algorithm will simply recall some of its values when needed, in constant
time.
5
n
m
0
1
2
3
4
5
6
7
8
..
.
0
1
1
2
5
14
42
132
429
1430
↑
1
2
0
1
0
3
1
9
4
28
14
90
48
297 165
1001 572
3
4
5
6
7
8
0
1
5
20
75
275
0
1
6
27
110
0
1
7
35
0
1
8
0
1
0
···
..
.
Cn
Figure 1: The BC table. It has (n − 3)(n − 1) + 1 entries. The arrows show
the sequence in which the cells fill up.
The main algorithm scheme Consider now the binary tree implied by
the BC relation (Figure 2), rooted at a node related to an,0 . This means we
are attempting to randomly select one of Cn samples. The BC table allows
us to achieve:
• optimal coding, requiring exactly log(Cn ) = O(n) bits to encode each
solution;
• optimal number of random bits for the main sampling algorithm: generating one random number in the range [0..Cn − 1], which requires
log Cn random bits, suffices.
A common approach is to calculate the sizes of the two sub-trees hanging from some node upon arriving on the very node and branch left or
right with the appropriate probability. This requires a total of (number
of branchings)×(random bits required per branching) random bits. Indeed,
a
≈ 2(n + m),
we may well apply the algorithm of [8], given that n−1,m−1
an,m
the number of random bits needed is O(n log n), as the height of the tree is
linear to n, thus the branchings required to reach a leaf is linear in n, too.
It is evident that establishing the BC table beforehand, and tossing only
log Cn ≈ 4n coins (see (3)) we generate a code uniformly at random, which
uniquely defines the leaf of the tree the algorithm will end upon: we branch
without any more coin tosses, as we know all critical quantities in advance.
Moreover, the scheme allows for down to linear time main algorithms
for sampling an instances, as all leaves’ distance from the root is linear in
n - it actually fluctuates from n to 2n. For instance, we will show that the
6
pr =
pr =
a2,0
a1,0
a2,0
pr =
a1,0
a3,0
a2,0
a3,0
a2,−1
a2,0
pr =
a2,−1
a1,0
a2,−2
pr =
a3,−1
a3,0
a3,−1
a2,0
a3,−1
pr =
a2,0
a1,0
a3,−2
a2,−1
a1,0
a3,−2
a3,−1
a2,−2
a2,−1
a1,0
a3,−3
a2,−2
Figure 2: The universal sampling algorithm scheme: branching with probability analogous to the size of the subtree. The height of the tree is O(n). We
note that typically all an,n nodes are non-existent, as the algorithm branches
towards them with probability 0 - that’s why they are marked red. Also, all
a1,0 labeled nodes have a left child, a0,0 , which is the actual leaf according
to the BC relation.
mountain ranges problem (and, therefore, Dyck words and monotone lattice
paths) has an O(n) main sampling algorithm, which we consider a good
trade-off when we need many samples for the same problem. Remember,
that only a O(n log n) algorithm is proposed in [8].
3
Convex triangulations
As we have already mentioned, for any problem we consider the object which
describes the space of all solutions. In the case of convex triangulations, this
is the complete geometric graph. A geometric graph is defined [5] as a pair
of a point set V on
R2 and a subset of straight line segments withendpoints
V
in V , i.e. E ⊆ 2 . Now let our point set be convex and E = V2 . We may
refer to this special case as the “convex Kn ”, an alternate to the complete
convex geometric graph, taking into account that all convex drawings of Kn
are isomorphic. Some important definitions are the following.
Proper labeling, consecutive vertices and span-2 edges Given a
convex geometrical graph, we will say that its vertex labeling v0 , ..., vn−1 is
proper if its n vertices appear in order, clockwise or counter-clockwise around
its convex hull. We will always assume such a labeling for any convex graph
we refer to, unless noted otherwise. The above being stated, it is evident that
7
for all i ∈ [0..n − 1], vertices vi and vi+1 are consecutive, naturally defining
that vn ≡ v0 . We shall also admit that all edges vi vi+2 , i ∈ [0..n − 1] are
properly defined, as vx ≡ vx mod n , for any x, n.
Let V be a convex point set and assume proper vertex labeling. For
every edge (diagonal) e = vi vj ∈ E, we define its span, denoted by |e|, to be
the minimum distance of its adjacent vertices around the convex hull. The
edges that miss only one vertex, that is edges of the form vi−1 vi+1 , will hold
a key-role in our work, thus we will denote them as span-2 edges. All such
edges can be named after the missed vertex with which they form a triangle:
ei ≡ vi−1 vi+1 .
As an analog to the notion of consecutive vertices, we define two consecutive span-2 edges to be a pair of edges of the form vi−1 vi+1 , vi vi+2 .
Observe that: K3 has no span-2 edges, K4 has two and, for n ≥ 5, Kn
has exactly n of such edges. For n ≥ 5 and any vertex vi , we denote by ei
the span-2 edge vi−1 vi+1 , in other words, the edge that may form a triangle
together with vi .
The Kn,−m graph. We will need to work on specific graphs, that allow for
the desired structure of triangulations to be witnessed. We define Kn,−m to
be the nearly complete convex Kn which misses only m consecutive span-2
edges. Therefore:
• Kn,0 ≡ Kn and when properly defined, Kn,−m has n2 − m edges.
• For n ≥ 5, it is −n ≤ −m ≤ 0, as a graph may miss up to all n span-2
edges.
• The graphs K4,−1 , K4,−2 are properly defined.
• For fixed n, m, all Kn,−m are isomorphic to each other (under rotation).
Proposition 3 Every triangulation of a convex geometric graph on 5 or
more vertices has at least two span-2 edges.
Proof. A triangulation TC on n points requires that all faces but the outer
are triangles. The number of faces f is equal to n − 1 (e.g. use Euler’s
characteristic on plane graphs), therefore there is a total of n − 2 inner
faces/triangles. Since all n non-diagonal edges of TC (the sides of the polygon) are sides of the n − 2 triangles, by the pigeonhole principle, there are
at least 2 triangles which use 2 sides of the n-gon as their sides. Then, those
triangles’ third edge must be an span-2 one and, as soon as n ≥ 5, those
third edges do not coincide.
Proposition 4 A convex geometric graph with n ≥ 5 and only two span-2
edges which appear consecutive in the drawing has no triangulation. Equivalently, T (Kn,−(n−2) ) = 0.
8
Proof.
Two consecutive span-2 edges intersect, therefore together they
do not form any triangulation, unless only one is needed (quadrilateral).
So there must be a third span-2 edge which, together with one of the 2
consecutive ones, is the obligatory second span-2 edge in a triangulation T
of the graph (see Proposition 3). Contradiction, as the graph has no other
such edge.
3.1
The reduction argument
Our reduction argument is primarily based on partitioning the number of
triangulations of some convex graph G into those which include a specific
span-2 edge ei and those which do not (assuming there is at least one span2 edge in G). Let us already point out the tree structure implied by the
partitioning.
v0
v7
v1
e0
e1
e7
v6
v0
K8
e6
e2
v7
v0
v2
v7
v1
v1
e5
e3
e4
v5
v6
K7
v3
v2
v6
K8,−1
v2
v4
v5
v5
v3
v3
v4
v4
Figure 3: Reduction for the complete Kn , working on v0 /e0
Proposition 5 Let G = (V, E) be a convex geometric graph and T (G) the
number of its triangulations. Then, for any span-2 edge ei we have that
T(G | ei included) = T(G \ vi ).
Obviously, the following holds: T (G) = T (G \ vi ) + T (G \ ei ).
Proof. Observe that if ei is selected to participate in a triangulation TC ,
then all edges incident to vi may not participate in TC , as they all cross
ei . Therefore, the number of triangulations where ei is included is exactly
determined by the number of triangulations of G \ vi , as adding the “hat”
9
vi−1 vi vi+1 in any of the latest will yield a distinct triangulation of the original
graph.
Figure 3 demonstrates the above, nevertheless for the special case of an
initial complete graph. However, we easily get that T (Kn,0 ) = T (Kn−1,0 ) +
T (Kn,−1 ), which is essentially the first of the BC relations.
Now, consider a Kn,−1 graph. If we work on v1 /e1 (Figure 4, marked
blue vertex), thus the selection or not of e1 for the next branching of our
tree, we obtain either a K7 ≡ K7,0 (left child), either a K8,−2 . In general
(Figure 5), working on the next available span-2 edge, we reduce Kn,−m is
to either a Kn−1,−m+1 (left child) or a Kn,−m−1 (right child).
v0
v7
v1
v6
v0
v0
K8,−1
v2
v7
v7
v1
v1
v5
v6
v3
K7
v6
K8,−2
v2
v2
v4
v5
v5
v3
v3
v4
v4
Figure 4: Reduction for Kn,−1 , working on the next vertex/span-2 edge
It proves that the key to our reduction is the correct definition of the
next vertex (or span-2 edge) to work on. If, for instance, while on K8,−1
we select to work on v5 /e5 , the nice to our reduction K7 does not pop out.
More importantly, we would miss the recurring of isomorphic graphs, which
give us the number of triangulations they feature per class (Kn,−m ), and not
per specific instance.
Definition 1 (Next vertex/span-2 edge) Given a properly labeled convex geometric graph on n vertices, missing only m consecutive span-2 edges,
ei , ei+1 , ..., ei+m . The next vertex/span-2 edge is the pair vm+1 /em+1 .
Now, let us formally prove our claim.
Theorem 6 For any Kn,−m and −n + 1 ≤ −m ≤ −1, it is
T (Kn,−m ) = T (Kn−1,−m+1 ) + T (Kn,−m−1 )
10
v0
v7
v1
v6
v 0 ≡ r0
v 7 ≡ r6
v0
K8,−3
v2
v1
v5
v 6 ≡ r5
v7
v 1 ≡ r1
v3
K7,−2
v6
K8,−4
v 2 ≡ r2
v2
v4
v 5 ≡ r4
v5
v3
v3
v 4 ≡ r3
v4
Figure 5: Reduction for Kn,−m , working on the next vertex/span-2 edge
Also, it is
T (Kn,−m ) = an−2,m
for all defined quantities.
Proof. Without loss of generality (see 9), we may relabel our initial graph
–if necessary– and have the edges e0 , ..., em−1 as the missing span-2 ones.
The next vertex/span-2 edge is the pair vm /em . The right subtree is rooted
to a Kn,−m−1 graph, as it derives from deleting em , having a consecutive
m + 1 span-2 (but no other) edges missing.
For the left subtree, deleting vm leaves us to examine the edges of a
graph on n − 1 vertices. We have that edges vn−1 v1 , v0 v2 , ..., vm−3 vm−1
are all missing from the induced subgraph, and together they add up to
m − 1 consecutive span-2 edges. Edge vm−2 vm , missing from Kn,−m is not a
potential edge of the smaller graph. Edge vm−1 vm+1 has now become a side
of the (n − 1)-gon, and vm−2 vm+1 and vm−1 vm+2 are now span-2 edges of
the new graph, as are all the (potentially) remaining vm+1 vm+3 , ..., vn−2 v0 .
Thus, we have obtained a Kn−1,−m+1 graph. To complete the proof, we
remind that:
• Proposition 4 gives us T (Kn,−n+2 ) = 0 for pentagons and up; but
T (K4,−2 ) gives the number of triangulations of a quadrilateral missing
all 2 span-2 edges, therefore the relation is satisfied for all n ≥ 4.
• T (K3,0 ) = 1, as K3 is a triangle.
11
• K3,−1 and K2,0 are not properly defined as convex graphs, and not
actually needed, but for sake of completeness, we shall consider their
number of triangulations equal to a1,1 and a0,0 respectively.
Theorem 7 Our main sampling algorithm scheme adopted for convex triangulations yields an algorithm of running time O(n2 ).
Proof. As an outline, the quadratic (and not linear) time is due to the
information that must be stored when branching left. For a complete proof,
refer to the Appendix. Yet, we note that [8] implies a linear algorithm by
sampling monotone lattice paths, first. However, if one wants to preserve
one-to-one relation between the subinstances of the two problems, we are
not aware of any work proposing any similar framework to ours.
4
Mountain ranges
The mountain range problem consider the ways to form a “mountain range”
with n upstrokes and n downstrokes that all stay above a horizontal line. As
with the convex triangulations, we assume the space of all solutions, that is
actually a n×n lattice rotated as to comprehend our node to node movement
as Up or Down and always maintaining direction to the right (w.l.o.g.).
Assume we are on the horizon, or level 0, so we have to move Up. Our
actual choice is whether to go further Up or slant Down. The latest case
brings us to the horizon with 1 upstroke and downstroke to our left, therefore
the remaining mountain ranges are exactly the mountain ranges with n − 1
upstrokes and downstrokes - correlate to the left child of the BC-implied
tree. That is:
rn,0 = rn−1,0 + rn,1
where the remainder rn,1 denotes the case we moved Up. Of course, the
cases yield mutually disjoint mountain ranges.
It is easy enough to go on, explaining that if we are not on the horizon
(m ≥ 1), we are not dictated to move up (unless we cannot - the “height” or
the range is at most n). If we move up, our m increases by one. If we move
down, we “see” again a disjoint set of potential solutions. Since we slanted
down (m decreased) we can symbolize this reduction by
rn,m = rn−1,m−1 + rn,m+1
.
Theorem 8 The main algorithm obtained by our scheme for the mountain
ranges problem is linear.
12
Proof. We need linear in n coin tosses to get the code of the mountain
range; we branch linear in n times and for each node we lie on, we store a
linear amount of information (Ups-Downs) to backtrack to the root.
A
Omitted proofs
Proof. [Proof of Theorem 1] We will consider the 2 main of the BC-relations.
The marginal cases are trivial to prove.
an,0 = an−1,0 + an,1
⇔
Nn,n−1 = Nn−1,n−2 + Nn,n−2
1
2n
2n − 2
2n − 1
1
3
=
+
n n−1
n−1 n−2
2n − 1 n − 2
(2n − 2)!
3(2n − 2)!
(2n)!
=
+
n!(n + 1)!
(n − 1)!n! (n − 2)!(n + 1)!
(2n − 1)(2n)
1
3
=
+
(n + 1)!
(n − 1)! (n − 2)!(n + 1)
2(2n − 1)
1
3
=
+
(n − 1)(n + 1)
(n − 1) (n + 1)
⇔
⇔
⇔
⇔
⇔
4n − 2 = (n + 1) + 3(n − 1),
which holds.
Regarding an,m = an−1,m−1 + an,m+1 we have:
Nn,m = Nn−1,m + Nn,m−1
n−m n+m
n−m+2 n+m
n+1−m n+1+m
=
+
n+1+m
n+m
n+m
m
m
m−1
(n + 1 − m)(n + m)!
(n − m)(n + m)!
(n − m + 2)(n + m)!
=
+
m!(n + 1)!
(n + m)m!n!
(n + m)(m − 1)!(n + 1)!
(n − m)
(n − m + 2)
(n + 1 − m)
=
+
m(n + 1)
(n + m)m (n + m)(n + 1)
1
1
1
2
(n − m + 2)
−
=
−
+
m n+1
m n + m (n + m)(n + 1)
2
1
(n − m + 2)
−
=
n+m n+1
(n + m)(n + 1)
2(n + 1) − (n + m) = n − m + 2,
⇔
⇔
⇔
⇔
⇔
⇔
which holds.
Therefore, we have completed our proof.
Proof. [Proof of Theorem 2] Let G be the generating function for the numbers
13
ai,j , i.e.
G(x, y) =
∞
∞ X
X
ai,j xi y j
i=0 j=0
and let B be the generating function for numbers bi,j = ai+1,j for all i, j ≥ 0.
We will first determine B, and then it is simply:
G = xB + 1.
(8)
Initially, let us break B into sums of fixed j:
B = A0 + A1 + A2 + ... =
∞
X
Ak ,
(9)
k=0
where for all k ∈ N Ak =
P∞
i=0 bi,k x
A0 =
iyk .
From the BC-relation 4 we obtain:
1
A1 + xA0 + 1,
y
(10)
while from the recursive BC-relation 5 we obtain:
∀k > 0 Ak =
1
Ak+1 + xyAk−1 .
y
(11)
We have also showed that the numbers ai,0 are actually Catalan numbers,
in particular bi,0 = ai+1,0 = Ci+1 , for i ≥ 0. So, if Cat is the generating
function of the Catalan numbers, then:
Cat = xA0 + 1 ⇒ A0 =
Cat − 1
.
x
(12)
It is well known that:
Cat(x) =
1−
√
1 − 4x
.
2x
(13)
In order to get the generating function G, plug equations 10,11,12,13 into
equation 9:
1
(A1 + A2 + ...) + xy(A0 + A1 + A2 + ...) + xA0 + 1
y
1
= (G − A0 ) + xyG + xA0 + 1
y
B=
Cat(x)−1
y + (xy − 1)
y + (xy − 1)A0
x
=
2
y − 1 − xy
y − 1 − xy 2
√
y
(xy − 1)(1 − 1 − 4x − 2x)
=
+
.
y − 1 − xy 2
2x2 (y − 1 − xy 2 )
=
14
Finally, from relation 8 he have:
√
(1 − xy)(1 − 1 − 4x − 2xy)
G(x, y) = xB(x, y) + 1 =
.
2x(1 − y + xy 2 )
Observe that we did not directly use the initial condition that an,n =
0, n ≥ 1; we actually did account for it, as it is hidden in the proof that the
numbers an,0 correspond to Catalan numbers.
B
Notes on the main sampling algorithm for convex triangulations
For our convenience, we hereby adopt the notation Tn,−m = T (Kn,−m ).
The left child of a Kn,−m in our reduction is a graph on n − 1 vertices,
in other words, w.r.t. our initial proper vertex labeling, the labeling of the
left child is not proper, as a vertex is missing. Therefore, we must relabel
this graph, and every left child graph of the tree. The following Rule gives
the final ingredient for building the desirable reduction tree.
Rule 9 Let Kr,−m be the left child of a node of a tree rooted at Kn . If
m = 0, then relabel the complete Kr as desired. Else, place r0 opposite to
the first missing span-2 edge and complete with a proper relabeling. For each
new graph, maintain a 1 × n table to indicate the mapping of the current
to the initial labeling; current missing vertices of the parent graph can be
marked (e.g. −1 entries).
B.0.1
Properties of the reduction tree
In all, we have build a triangulation tree with the following properties:
• Every node has 2 children (left and right – binary tree).
• Every node indicating Kr,−m needs O(n) time to be created and coded:
store integers r and m, the 1×n mapping matrix for the vertex labeling,
plus a 2 integer indicator for backtracking to the parent node: if it is a
left child, hold [vm , −1], for a right child hold [vm−1 , vm+1 ] where vm is
the critical vertex of the parent node w.r.t. which the node branched
(store as labeled in the parent node).
• For any node, the triangulations coded in the left subtree are disjoint
to the ones of the right subtree, as they differ at the edge w.r.t. which
the node branched.
• If a node is K3 , then stop and mark one triangulation (green leaf).
Due to the above, all green leaves are left children and there is a total
of exactly Cn−2 green leaves in the tree.
15
• If a node is of the form Kr,−r+2 , then stop and mark no triangulation
(red leaf)
• Considering all the above, from any green leaf, backtracking to the root
is equivalent to obtaining one specific triangulation; this is achieved in
O(n) time.
0
4
\v
1
3
\e
2
0
0
4
1
3
4
2
3
0
0
4
1
3
2
0
4
2
1
1
3
0
4
2
1
3
4
2
1
3
2
T1
0
0
4
1
3
2
T2
0
4
1
3
2
4
1
3
4
2
0
0
0
1
3
4
2
4
1
3
1
3
2
2
T3
0
0
4
1
3
2
T4
0
4
1
3
2
0
4
1
3
2
4
1
3
T5
Figure 6: Full reduction tree for the convex pentagon. Blue color indicates
the working on node. Relabelings are not marked. Each of its 5 triangulations is encoded in a different green leaf.
Figure 6 shows the full reduction tree for the convex K5 . To give a for
instance, consider the green leaf encoding triangulation T3 . T3 includes the
triangle v0 v3 v4 (end of reduction); as a left child of its parent node, which
branched w.r.t. v2 /v0 v3 , it includes v0 v3 as an span-2 edge of the parent K4
subgraph, so triangle v0 v2 v3 is in T4 . The K4 parent node is a left child of
the K5,−1 node which branched w.r.t. v1 , therefore the triangulation does
also include triangle v0 v1 v2 ; finally (though not necessary), the K5,−1 node
being a right child of the K5 which branched w.r.t. v0 suggests that v4 v1 is
missing from the T4 .
Note that this is a small example, and some leaves (T2 , T4 , T5 ) already
16
2
indicate unambiguously the encoded triangulation. Observe also that even
if it is not directly visible that T1 differs from T4 , our simple branching
rule guarantees that T1 includes v4 v1 as an edge, while T4 does not, as they
belong to a different subtree of the root node which branched w.r.t. v0 /v4 v1 .
References
[1] Victor Alvarez, Karl Bringmann, Saurabh Ray, and Raimund Seidel. Counting triangulations and other crossing-free structures approximately. Comput. Geom., 48(5):386–397, 2015.
[2] Victor Alvarez and Raimund Seidel. A simple aggregative algorithm
for counting triangulations of planar point sets and related problems.
In Symposuim on Computational Geometry 2013, SoCG ’13, Rio de
Janeiro, Brazil, June 17-20, 2013, pages 1–8, 2013.
[3] J. Ambjorn, B. Durhus, and T. Jonsson. Quantum Geometry: A Statistical Field Theory Approach. Cambridge Monographs on Mathematical
Physics, Cambridge University Press, Cambridge, 1997.
[4] Eleni Bakali, Aggeliki Chalki, Aris Pagourtzis, Petros Pantavos, and
Stathis Zachos. Completeness results for counting problems with easy
decision. In Algorithms and Complexity - 10th International Conference,
CIAC 2017, Athens, Greece, May 24-26, 2017, Proceedings, pages 55–
66, 2017.
[5] Prosenjit Bose, Ferran Hurtado, Eduardo Rivera-Campo, and David R.
Wood. Partitions of complete geometric graphs into plane trees. Comput. Geom., 34(2):116–125, 2006.
[6] Mark de Berg, Otfried Cheong, Marc J. van Kreveld, and Mark H.
Overmars. Computational geometry: algorithms and applications, 3rd
Edition. Springer, 2008.
[7] Markus Denny and Christian Sohler. Encoding a triangulation as a permutation of its point set. In Proceedings of the 9th Canadian Conference
on Computational Geometry, Kingston, Ontario, Canada, August 1114, 1997, 1997.
[8] Luc Devroye, Philippe Flajolet, Ferran Hurtado, Marc Noy, and
William L. Steiger. Properties of random triangulations and trees. Discrete & Computational Geometry, 22(1):105–117, 1999.
[9] Q. Ding, J. Qian, W. Tsang, and C. Wang. Randomly generating triangulations of a simple polygon. In Computing and Combinatorics, 11th
Annual International Conference, COCOON 2005, Kunming, China,
August 16-29, 2005, Proceedings, pages 471–480, 2005.
17
[10] Peter Epstein, J. Kavanagh, A. Knight, J. May, T. Nguyen, and JörgRüdiger Sack. A workbench for computational geometry. Algorithmica,
11(4):404–428, 1994.
[11] Peter Epstein and Jörg-Rüdiger Sack. Generating triangulations at random. ACM Trans. Model. Comput. Simul., 4(3):267–278, 1994.
[12] W. Feller. An introduction to probability theory and its applications.
Number v. 1. Wiley, 1968.
[13] Leonidas J. Guibas and John Hershberger. Optimal shortest path
queries in a simple polygon. J. Comput. Syst. Sci., 39(2):126–152, 1989.
[14] Leonidas J. Guibas, John Hershberger, Daniel Leven, Micha Sharir, and
Robert Endre Tarjan. Linear-time algorithms for visibility and shortest
path problems inside triangulated simple polygons. Algorithmica, 2:209–
233, 1987.
[15] Ferran Hurtado and Marc Noy. Ears of triangulations and catalan numbers. Discrete Mathematics, 149(1-3):319–324, 1996.
[16] Ferran Hurtado and Marc Noy. Graph of triangulations of a convex polygon and tree of triangulations. Comput. Geom., 13(3):179–188, 1999.
[17] Donald E. Knuth. Estimating the efficiency of backtrack programs.
Technical report, Stanford, CA, USA, 1974.
[18] D. T. Lee. Visibility of a simple polygon. Computer Vision, Graphics,
and Image Processing, 22(2):207–221, 1983.
[19] Lisa McShine and Prasad Tetali. On the mixing time of the triangulation walk and other catalan structures. In Randomization Methods
in Algorithm Design, Proceedings of a DIMACS Workshop, Princeton,
New Jersey, USA, December 12-14, 1997, pages 147–160, 1997.
[20] Michael S. O. Molloy, Bruce A. Reed, and William Steiger. On the mixing rate of the triangulation walk. In Randomization Methods in Algorithm Design, Proceedings of a DIMACS Workshop, Princeton, New
Jersey, USA, December 12-14, 1997, pages 179–190, 1997.
[21] Takao Nishizeki and Md. Saidur Rahman. Planar Graph Drawing, volume 12 of Lecture Notes Series on Computing. World Scientific, 2004.
[22] Joseph O’Rourke. Computational Geometry in C. Cambridge University
Press, New York, NY, USA, 2nd edition, 1998.
[23] Mohammad Tanvir Parvez, Md. Saidur Rahman, and Shin-Ichi Nakano.
Generating all triangulations of plane graphs. J. Graph Algorithms
Appl., 15(3):457–482, 2011.
18
[24] Dominique Poulalhon and Gilles Schaeffer. Optimal coding and sampling of triangulations. Algorithmica, 46(3-4):505–527, 2006.
[25] Sadiq M. Sait and Habib Youssef. VLSI Physical Design Automation Theory and Practice, volume 6 of Lecture Notes Series on Computing.
World Scientific, 1999.
[26] Micha Sharir and Adam Sheffer. Counting triangulations of planar point
sets. Electr. J. Comb., 18(1), 2011.
[27] Micha Sharir, Adam Sheffer, and Emo Welzl. On degrees in random
triangulations of point sets. J. Comb. Theory, Ser. A, 118(7):1979–1999,
2011.
[28] Micha Sharir and Emo Welzl. Random triangulations of planar point
sets. In Proceedings of the 22nd ACM Symposium on Computational
Geometry, Sedona, Arizona, USA, June 5-7, 2006, pages 273–281, 2006.
[29] Alistair Sinclair and Mark Jerrum. Approximate counting, uniform generation and rapidly mixing markov chains. Inf. Comput., 82(1):93–133,
1989.
[30] Richard P. Stanley. Catalan Numbers. Cambridge University Press,
2015. doi:10.1017/CBO9781139871495.
[31] Robert Endre Tarjan and Christopher J. Van Wyk. An o(n log log n)time algorithm for triangulating a simple polygon. SIAM J. Comput.,
17(1):143–178, 1988.
19
| 8cs.DS
|
Noname manuscript No.
(will be inserted by the editor)
Binets: fundamental building blocks for phylogenetic networks
arXiv:1701.08995v1 [q-bio.PE] 31 Jan 2017
Leo van Iersel · Vincent Moulton · Eveline de Swart ·
Taoyang Wu
the date of receipt and acceptance should be inserted later
Abstract Phylogenetic networks are a generalization of evolutionary trees that are used by biologists to
represent the evolution of organisms which have undergone reticulate evolution. Essentially, a phylogenetic
network is a directed acyclic graph having a unique root in which the leaves are labelled by a given set of
species. Recently, some approaches have been developed to construct phylogenetic networks from collections
of networks on 2- and 3-leaved networks, which are known as binets and trinets, respectively. Here we study
in more depth properties of collections of binets, one of the simplest possible types of networks into which a
phylogenetic network can be decomposed. More specifically, we show that if a collection of level-1 binets is
compatible with some binary network, then it is also compatible with a binary level-1 network. Our proofs
are based on useful structural results concerning lowest stable ancestors in networks. In addition, we show
that, although the binets do not determine the topology of the network, they do determine the number of
reticulations in the network, which is one of its most important parameters. We also consider algorithmic
questions concerning binets. We show that deciding whether an arbitrary set of binets is compatible with
some network is at least as hard as the well-known Graph Isomorphism problem. However, if we restrict to
level-1 binets, it is possible to decide in polynomial time whether there exists a binary network that displays
all the binets. We also show that to find a network that displays a maximum number of the binets is NP-hard,
but that there exists a simple polynomial-time 1/3-approximation algorithm for this problem. It is hoped that
these results will eventually assist in the development of new methods for constructing phylogenetic networks
from collections of smaller networks.
Part of this work was conducted while Vincent Moulton was visiting Leo van Iersel on a visitors grant funded by the Netherlands
Organization for Scientific Research (NWO). Leo van Iersel was partially supported by NWO, including Vidi grant 639.072.602, and
partially by the 4TU Applied Mathematics Institute. We thank the editor and the two anonymous referees for their constructive
comments.
Leo van Iersel
Delft Institute of Applied Mathematics
Delft University of Technology
The Netherlands
E-mail: [email protected]
Vincent Moulton and Taoyang Wu
School of Computing Sciences
University of East Anglia
Norwich
United Kingdom
E-mail: [email protected]
Eveline de Swart
Delft Institute of Applied Mathematics
Delft University of Technology
The Netherlands
E-mail: [email protected]
Taoyang Wu
School of Computing Sciences
University of East Anglia
Norwich
United Kingdom
E-mail: [email protected]
2
Leo van Iersel et al.
x
y
z
x
y
z
x
y
y
z
x
z
Fig. 1: An example of two level-1 trinets (left) that display the same set of three binets (right). All arcs are
directed downwards.
Keywords reticulate evolution · phylogenetic network · subnetwork · binet · algorithm
1 Introduction
Phylogenetic networks are a generalization of evolutionary trees which biologists use to represent the evolution
of species that have undergone reticulate evolution. Such networks are essentially directed acyclic graphs having
a unique root in which the leaves are labelled by a set X of species [14]. In contrast to evolutionary trees,
which can only represent speciation events, phylogenetic networks permit the representation of evolutionary
events such as gene transfer and hybridization which are known to occur in organisms such as bacteria and
plants, respectively. Although theoretical properties of evolutionary trees have been studied since at least the
1970’s, phylogenetic networks have been considered from this perspective only more recently, especially the
rooted variants which we will focus on in this paper.
One of the most important open questions concerning phylogenetic networks is how to construct them for
biological datasets [3]. It is now common practice for biologists to construct evolutionary trees from molecular
data, and several computer programs are available for this purpose [7]. However, the problem of constructing
networks from such data is an active area of research, and there are only a limited number of programs
available for biologists to perform this task. A survey of some of these methods and the theory underpinning
phylogenetic networks may be found in [9, 14, 15].
One approach that has been recently developed for constructing phylogenetic networks involves building
them up from smaller networks, using what can be thought of as a divide-and-conquer approach [16]. In
particular, for a set X of species, a network is constructed for every subset of X size 3 (called a trinet), and
then the trinets are puzzled together to build a network (see Figure 1 for an example of a trinet). This approach
constructs and is based on level-1 networks, networks that are slightly more general than evolutionary trees
(see Section 2 for the definition of such networks).
At first sight, it might appear that trinets are the simplest possible networks that could be considered for
building up networks from smaller ones. However, trinets contain even simpler networks called binets, networks
with 2 leaves (see e.g. Figure 1 for a level-1 trinet and the binets that it displays). Note that whereas binets
are the smallest informative building blocks for phylogenetic networks, for rooted phylogenetic trees, these
are 3-leaf trees (see e.g. [5]). Interestingly, even though binets are in themselves very simple, the collection of
binets displayed by a network can still contain some useful information concerning the network. Indeed, in the
aforementioned approach for building level-1 networks from trinets, binets are used in the process of puzzling
together the trinets.
In light of these considerations some obvious questions immediately arise concerning binets. For example, when is a collection of binets displayed by some phylogenetic network (the compatibility problem), and
how much information might we expect to extract concerning a phylogenetic network by just looking at the
collection of binets that it displays? In this paper, we shall address these and related algorithmic questions
concerning binets. It is hoped that these results will be useful in future for developing improved methods for
constructing phylogenetic networks from smaller networks.
We now present a summary of the rest of the paper. After introducing some preliminaries concerning
phylogenetic networks in the next section, we derive a key structural result for networks (Corollary 1) which
is useful in identifying which of the two possible types of binet is displayed on two leaves within a binary
phylogenetic network (that is a network in which all internal vertices have degree 3). Using this theorem, in
Section 4 we show that the collection of level-1 binets displayed by any binary phylogenetic network can always
be displayed by some binary level-1 network (Theorem 3). This reduces the problem of understanding binets
displayed by arbitrary binary networks to level-1 networks. To prove this result, we develop a framework which
also implies that there is a polynomial-time algorithm in |X| for deciding whether or not a collection of level-1
binets with combined leaf-set X can be displayed by some network with leaf-set X, and, if it is, gives a level-1
network that does this (see Section 6). Note that this is related to an algorithm presented in [11].
Binets: fundamental building blocks for phylogenetic networks
3
In Section 5, we turn to the question as to what can be deduced about the features of a phylogenetic
network just by considering the collection of binets that it displays. Note that, as might be expected, there
are networks - even trinets - that display the same set of binets but that are not equivalent. For example, the
two trinets in Figure 1 both display the same set of binets, but they are not equivalent. Even so, we will show
in Theorem 4 that if two level-1 networks both display exactly the same collection of binets, then they must
have the same number of reticulation vertices (indegree-2 vertices). Note that the number of such vertices
corresponds to the number of reticulate evolutionary events, such as hybridization, that took place in the
evolutionary history of the species labelling the leaves of the network. Consequently, the binets displayed by
a network can at least capture a useful course-grained feature of the network in question.
In Sections 6 and 7, we consider some algorithmic questions concerning binets. As we have mentioned
above, it can be decided in polynomial time in |X| as to when a collection of binets with combined leaf-set
X is displayed by some level-1 network on X. However, we show that if we consider arbitrary binets (i.e. not
necessarily binary or level-1) then this decision problem becomes at least as hard as the graph-isomorphism
problem (see Theorem 5), one of the most famous problems whose complexity is still unknown. In addition,
in Section 7 we consider a related problem which, for a given collection of binary level-1 binets, asks for
a network which displays the maximum number of binets in this collection. This is closely related to the
maximum rooted triplet consistency problem for evolutionary trees [5]. We show that the binet problem is
NP-complete (Theorem 6), by giving a reduction from the feedback-arc set problem. However, we also show
that the problem is 1/3-approximable. In fact, given any collection of binary level-1 binets we can always
find some network that displays at least 1/3 of the binets (see Theorem 7). We conclude in Section 8 with
discussion of some possible future research directions, and a brief discussion of a potential application of our
results.
2 Preliminaries
Throughout this paper, X is a non-empty finite set (which usually represents a set of species or organisms).
2.1 Digraphs
A directed graph, or digraph for short, G = (V, E) consists of a finite set V = V (G) of vertices and a set
E = E(G) of arcs, where each arc is an ordered pair (u, v) of vertices in V in which u is said to be a parent
of v, denoted by u = p(v), and v a child of u. All digraphs studied here contain no loops, that is, vertices
that are children of themselves. The in-degree of vertex u is the number of vertices v in V such that (v, u) is
an arc, and the out-degree of u is the number of vertices w with (u, w) being an arc. A root is a vertex with
in-degree 0. A leaf is a vertex of out-degree 0 and the set of leaves is denoted by L(G). Any vertex in G that
is neither a root nor a leaf is referred to as an interior vertex. In addition, an interior vertex is a tree vertex if
it has in-degree 1, and a reticulation vertex if it has in-degree greater than 1.
A directed path or dipath in a digraph is a sequence u0 , u1 , . . . , uk (k ≥ 1) of vertices such that (ui−1 , ui )
is an arc for 1 ≤ i ≤ k. An acyclic digraph is a digraph that does not contain any directed path starting and
ending at the same vertex. If an acyclic digraph G contains a unique root, which is usually designated by
ρ = ρ(G), then it will be referred to as a rooted acyclic digraph.
An acyclic digraph G induces a canonical partial order ≺G on its vertex set V , that is, v ≺G u if there
exists a directed path from u to v. In this case, we shall say that v is below u. When the digraph G is clear
from the context, ≺G will be written as ≺. In addition, we write v u if u = v or u ≺ v. Given a subset U of
the vertex set of an acyclic digraph, we say that u ∈ U is a lowest vertex in U if there is no v ∈ U with v ≺ u.
Let G be the undirected graph obtained from digraph G by ignoring the direction of the arcs in G. Then G
is connected if G is connected, that is, there exists an undirected path between every pair of distinct vertices in
G. Note that a rooted acyclic digraph is necessarily connected (since each connected component of an acyclic
digraph has at least one root). A cut vertex is a vertex of G whose removal disconnects G. Similarly, a cut arc
is an arc of G whose removal disconnects G. A directed graph is biconnected if it contains no cut vertex, and a
biconnected component of G is a maximal biconnected subgraph, which is called trivial if it contains precisely
one arc (which is necessarily a cut arc), and non-trivial otherwise.
4
Leo van Iersel et al.
2.2 Phylogenetic networks
A phylogenetic network N on X is a rooted acyclic digraph whose leaves are bijectively labeled by the elements
in X and which does not contain any vertex with in-degree one and out-degree one. For simplicity, we will
just write L(N ) = X in case there is no confusion about the labeling. To simplify the argument, throughout
this paper we will also assume that all leaves in a phylogenetic network have in-degree one. In addition, a
phylogenetic network is binary if each tree vertex, as well as the root, has out-degree 2, and each reticulation
vertex has in-degree 2 and out-degree 1. Finally, we say a binary phylogenetic network is level -k (k ≥ 0) if
each of its biconnected components contains at most k reticulation vertices. To some extent, the concept of
the level of a phylogenetic network can be regarded as a measure of its ‘distance’ to being a phylogenetic tree.
In particular, a binary phylogenetic network is a phylogenetic tree if and only if it is level-0. A phylogenetic
network is called simple if it contains precisely one non-trivial biconnected component H and no cut arcs other
than the ones leaving H.
Two networks N1 = (V1 , E1 ) and N2 = (V2 , E2 ) on X are said to be isomorphic if there exists a bijection
f : V1 → V2 such that f (x) = x for all x ∈ X, and (u, v) is an arc in N1 if and only if (f (u), f (v)) is an arc
in N2 .
Finally, the cluster of a vertex u, denoted by CN (u) = C(u), is defined as the subset of X consisting of the
leaves below u. Here we will use the convention that C(u) = {u} if u is a leaf.
2.3 Stable ancestors and binets
Given a phylogenetic network N on X and a subset U ⊆ V (N ), a stable ancestor of U in N is a vertex v
in V (N ) \ U such that every path in N from the root to a vertex in U contains v. Note that for two stable
ancestors u and u0 of U , we have either u v or v u. Therefore, there exists a unique lowest vertex in the
set of stable ancestors of U , which will be referred to as the lowest stable ancestor of U in N and denoted by
lsaN (U ) = lsa(U ). Note that for a subset Y of X with |Y | ≥ 2, there exist two elements x and y in Y such
that lsa(Y ) = lsa({x, y}). For simplicity, we also write lsa({x, y}) as lsa(x, y).
The following property of lowest stable ancestors will be useful.
Lemma 1 Suppose that u and v are two vertices in a phylogenetic network such that u ≺ v ≺ lsa(u), then
we have lsa(v) lsa(u).
Proof Since u ≺ v, we know that there exists a dipath P from ρ to u that contains v. By the definition of
lowest stable ancestor, we know that lsa(u) and lsa(v) are contained in P . Hence, either lsa(v) lsa(u) or
lsa(u) ≺ lsa(v). If lsa(u) ≺ lsa(v), then we have v ≺ lsa(u) ≺ lsa(v). Then there exists a dipath P 0 from
ρ to v that does not contain lsa(u) (otherwise lsa(u) would be a stable ancestor of v that is below lsa(v)).
Using that u ≺ v ≺ lsa(u), it follows that there exists a dipath from ρ to u that does not contain lsa(u), a
contradiction. Therefore, lsa(v) lsa(u).
t
u
For Y ⊆ X, the subnet of N on Y , denoted by N |Y , is defined as the subgraph obtained from N by deleting
all vertices that are not on any path from lsa(Y ) to elements in Y and subsequently suppressing all in-degree
1 and out-degree 1 vertices and parallel arcs until no such vertices or arcs exist. A network N 0 is said to be
displayed by network N if N 0 = N |Y for some Y ⊆ X.
Note that, by definition, N |X = N if and only if lsa(X) = ρ(N ). In this case, N is referred to as a
recoverable network. Note that every subnet of N is necessarily recoverable. Moreover, a collection of subnets
is displayed by some network if and only if it is displayed by some recoverable network. Therefore, we assume
all networks in this paper to be recoverable.
A binet is a phylogenetic network with precisely two leaves, while a trinet is a phylogenetic network with
precisely three leaves. Let
B(N ) = {N |Y : Y ⊆ X and |Y | = 2}
be the collection of binets displayed by N . Note that there are precisely three binary level-1 binets on a set
{x, y}, and they can be grouped into two types: the “tree type”, T (x, y), and the “reticulate type” R(x; y)
and R(y; x) (see Figure 2). A collection of binets B on X is a collection of binets such that the union of the
leaf-sets of the binets is equal to X.
Binets: fundamental building blocks for phylogenetic networks
x
y
T (x, y)
x
y
y
R(x; y)
5
x
y
R(y; x)
x
level-2
Fig. 2: The three binary level-1 binets on {x, y} and an example of a level-2 binet.
ρ
lsa(v)
x
r
v
y
p
q
Fig. 3: Example for the proof of Theorem 1. The lowest stable ancestor of v is not equal to the root ρ. An
undirected path Pu between ρ and lsa(v) that does not contain the incoming arc of lsa(v) is indicated in
bold. Path Pu contains one alternating vertex r, for which the lowest stable ancestor is indeed equal to ρ.
3 A structure theorem
In this section we present a key result (Corollary 1) concerning the structure of the non-trivial biconnected
component of a simple network. Note that a similar result has been obtained for a special collection of (nonbinary) phylogenetic networks in [13].
Let G be a directed acyclic graph and let P = v0 , v1 , . . . , vt be an undirected path in the underlying
undirected graph G, then a vertex vi (with 1 ≤ i ≤ t − 1) is called alternating (with respect to P ) if we have
either {(vi−1 , vi ), (vi+1 , vi )} ⊆ E(G) or {(vi , vi−1 ), (vi , vi+1 )} ⊆ E(G). The number of alternating vertices
contained in P is denoted by alt(P ). Using this concept, we now prove the following theorem. See Figure 3
for an example.
Theorem 1 Let N be a binary phylogenetic network on X whose root ρ is in some non-trivial biconnected
component H. Then there exists a lowest vertex in H with lsa(v) = ρ.
Proof Let Γ0 (H) be the set of reticulation vertices v in H for which the distance (length of a shortest directed
path) between ρ and lsa(v) is minimum over all reticulation vertices in H. Note that Γ0 (H) 6= ∅.
We first show that lsa(v) = ρ for all v ∈ Γ0 (H). Suppose this were not the case. Then there exists a vertex
v ∈ Γ0 (H) such that lsa(v) ≺ ρ. Note that lsa(v) necessarily has outdegree 2 and therefore has indegree 1
since N is binary and lsa(v) 6= ρ. Denote the parent of lsa(v) by v ∗ . Since H is biconnected, there exists
some undirected path from ρ to lsa(v) that does not contain the edge e = {lsa(v), v ∗ }. Let Pu = v0 , . . . , vt ,
where v0 = ρ and vt = lsa(v), be such an undirected path for which alt(Pu ) is minimum.
We claim that alt(Pu ) = 1. To see this, note first that since v0 = ρ, vt = lsa(v) and vt−1 6= v ∗ , we know
that (v0 , v1 ) and (vt , vt−1 ) are arcs of N . Hence, alt(Pu ) is odd and strictly positive. Assume for the sake
of contradiction that alt(Pu ) 6= 1, then we have alt(Pu ) ≥ 3. Let vk (1 < k < t) be the second alternating
vertex contained in Pu (when travelling from v0 to vt ).
Now fix a directed path Pd in N from ρ to vk .
If the arc (v ∗ , lsa(v)) is not contained in Pd , then, we can find an undirected path from ρ to lsa(v)
that does not contain e and has fewer alternating vertices than Pu by following Pd until we reach a vertex
in {vk , . . . , vt } and then following Pu to lsa(v). This gives a contradiction.
Now assume that the arc (v ∗ , lsa(v)) is contained in Pd . Then we can find an undirected path from ρ
to lsa(v) that does not contain e and has only one alternating vertex as follows. Follow Pu up to vk and then
follow Pd backward from vk to lsa(v). Since this path has fewer alternating vertices than Pu , we again obtain
a contradiction.
We have thus shown that alt(Pu ) = 1. Denoting this alternating vertex in Pu by r, then r is necessarily
a reticulation by the choice of Pu . Hence, Pu consists of two directed paths: a directed path from ρ to r that
does not contain lsa(v) and a directed path from lsa(v) to r. However, this means that lsa(v) ≺ lsa(r), a
contradiction to the assumption that v ∈ Γ0 (H).
6
Leo van Iersel et al.
Hence, we know that Γ0 (H) is the set of reticulation vertices v of H such that lsa(v) = ρ and that Γ0 (H)
is not empty.
Now fix a vertex v in Γ0 (H) that is lowest over all vertices of Γ0 (H), that is, there does not exist a vertex
u in Γ0 (H) such that v ≺ u. It remains to show that v is lowest over all vertices of H. Assume that this is
not the case. Then the child c of v is also in H. If c were a reticulation then, by Lemma 1, lsa(v) lsa(c).
However, this would imply that lsa(c) = ρ, contradicting the choice of v. Hence, c is a tree vertex.
Since H is biconnected, there exists some undirected path from ρ to c that does not contain v. Let Pu =
w0 , . . . , wt be such a path such that alt(Pu ) is minimum. Note that we have w0 = ρ and wt = c.
Since c is a tree vertex and Pu does not contain its parent v, (wt , wt−1 ) is an arc of N . Together with
(w0 , w1 ) being an arc in N , we know that alt(Pu ) is odd and strictly positive. We now show, using a similar
proof as above, that alt(Pu ) = 1. If this were not the case, then we would have alt(Pu ) ≥ 3. Let wk
(1 < k < t) be the second alternating vertex contained in Pu . We know that (wk , wk−1 ) and (wk , wk+1 ) are
two arcs contained in N . Now fix a directed path Pd in N from ρ to wk .
If the vertex v is not contained in Pd , then we can find an undirected path from ρ to c that does not contain
v and has fewer alternating vertices than Pu by following Pd from ρ it reaches a vertex from {wk , . . . , wt } and
then following Pu up to c. If v is contained in Pd , then we follow Pu from ρ to wk and then follow Pd from wk
to c and obtain an undirected path from ρ to c that does not contain v and has one alternating vertices, which
is less than the number of alternating vertices in Pu . In either case, we obtain a contradiction.
We have thus shown that alt(Pu ) = 1. Denoting this alternating vertex in Pu by r, then r is necessarily
a reticulation by the choice of Pu . Hence, Pu consists of two directed paths: a directed path from ρ to r
that does not contain v and a directed path from c to r. However, this means that v ≺ lsa(r), and hence
lsa(v) lsa(r) in view of Lemma 1. This implies that r ∈ Γ0 (H), a contradiction to the assumption that v
is lowest among Γ0 (H).
t
u
The following is a direct consequence of the above theorem.
Corollary 1 Suppose that N is a simple binary phylogenetic network. Let H be the unique non-trivial biconnected component of N . Then there exists a lowest vertex v of H such that there exist two arc-disjoint directed
paths from the root of N to v.
4 Displaying binets by binary networks
A collection of binary level-1 binets is compatible if there exists some binary network that displays all binets
from the collection. In this section, we study the compatibility of binets. Our main result in this section
(Theorem 3) shows that when studying the compatibility of binets, we can restrict to binary level-1 networks.
We will restrict ourselves throughout this section to thin collections of binets, i.e. collections containing at
most one binet on x and y for all distinct x, y ∈ X. Clearly, any collection of binets that is not thin is not
compatible.
First, we need some new definitions. Given a digraph G, a sink set of G is a proper subset U ⊂ V (G) such
that there is no arc leaving U , that is, there exists no arc (x, y) with x ∈ U and y ∈ V (G) \ U . A bipartition
(or split) of V (G) into nonempty sets A and B, denoted A|B, is called
– Type I if both A and B are sink sets (i.e. there is no arc from any element in A to any element in B or
vice versa);
– Type II if either A or B (but not both) is a sink set; and
– Type III if for all x ∈ A, y ∈ B (x, y) is an arc in G if and only if (y, x) is an arc in G.
We say that A|B is a typed split of G if it is a split of Type I, II or III.
For a collection B of binary level-1 binets on X, we introduce the digraph D(B) with vertex set X and
(x, y) being an arc in D(B) if T (x, y) ∈ B or R(x; y) ∈ B. See Figure 4 for an example.
The following two lemmas show important properties of typed splits that will be used to establish Theorems 2 and 3.
Lemma 2 Suppose that B and B 0 are two thin collections of binary level-1 binets on X with B ⊆ B 0 . Then
each typed split of D(B 0 ) is a typed split of D(B).
Proof Suppose that A|B is a typed split of D(B 0 ). If A|B is of Type I in D(B 0 ), then it is of Type I in D(B)
since D(B) is a subgraph of D(B 0 ). Similarly, if A|B is of Type II in D(B 0 ), then it is of Type I or II in D(B).
If A|B is of Type III in D(B 0 ) then (since B 0 is thin) any binet on x and y with x ∈ A and y ∈ B is T (x, y).
Therefore, A|B is of Type I or III in D(B).
t
u
Binets: fundamental building blocks for phylogenetic networks
Binets: fundamental building blocks for phylogenetic networks
x4
x4
x1
x2
7
7
x6
x3
x1
x5
x5
x2
x3
x6
D(B)
N
Fig.
Fig. 4:
4: A
A digraph
digraph D(B)
D(B) representing
representing the
the set
set of
of binets
binets B
B=
= {T
{T(x
(x44,, xx33),
), R(x
R(x44;; xx11),
), R(x
R(x44;; xx22),
), TT(x
(x55,, xx11),
), R(x
R(x55;; xx22),
),
R(x
R(x55;; xx33),
), TT(x
(x66,, xx22),
), R(x
R(x66;; xx11),
), R(x
R(x66;; xx33)}
)} and
and aa network
network N
N displaying
displaying B.
B. The
The direction
direction of
of the
the arcs
arcs in
in N
N is
is
downward,
downward, and
and omitted.
omitted.
Lemma
Lemma 33 Suppose
Suppose that
that B
B is
is aa thin
thin collection
collection of
of binary
binary level-1
level-1 binets
binets on
on X.
X. If
If B
B is
is displayed
displayed by
by aa binary
binary
network,
then
D(B)
has
a
typed
split.
network, then D(B) has a typed split.
Proof
Proof Suppose
Suppose that
that B
B is
is displayed
displayed by
by aa binary
binary network.
network. Then
Then B
B is
is displayed
displayed by
by aa binary
binary recoverable
recoverable network
network N
N..
Let
Let B
B00 be
be the
the set
set of
of binary
binary level-1
level-1 binets
binets contained
contained in
in B(N
B(N).
). Then
Then we
we have
have B
B✓
⊆B
B00 ✓
⊆ B(N
B(N).
). By
By Lemma
Lemma 2,
2, it
it
suffices
suffices to
to show
show that
that D(B
D(B00)) has
has aa typed
typed split.
split.
Consider
Consider the
the root
root ⇢ρ of
of N
N,, which
which is
is equal
equal to
to lsa(X)
lsa(X) since
since N
N is
is recoverable.
recoverable. Denote
Denote the
the two
two children
children of
of ⇢ρ
by
by uu11 and
and uu22.. We
We consider
consider two
two cases.
cases.
The
The first
first case
case is
is that
that at
at least
least one
one arc
arc incident
incident with
with ⇢ρ is
is aa cut
cut arc.
arc. Then
Then the
the other
other arc
arc incident
incident with
with ⇢ρ is
is
also
also aa cut
cut arc.
arc. Then
Then let
let A
A=
= C(u
C(u11)) and
and B
B=
= C(u
C(u22).
). Note
Note that
that A|B
A|B is
is aa split
split because
because neither
neither A
A nor
nor B
B is
is empty.
empty.
In
In addition,
addition, for
for all
all xx 2
∈ A,
A, yy 2
∈B
B we
we have
have N
N||{x,y}
= TT(x,
(x, y)
y) and
and hence
hence A|B
A|B is
is aa Type
Type III
III split
split with
with respect
respect to
to
{x,y} =
D(B
D(B00).
).
In
In the
the second
second case,
case, both
both arcs
arcs incident
incident with
with ⇢ρ are
are not
not cut
cut arcs.
arcs. Hence,
Hence, the
the root
root ⇢ρ is
is contained
contained in
in aa non-trivial
non-trivial
biconnected
component
H
containing
u
and
u
.
By
Corollary
1,
there
exists
a
lowest
vertex
v
biconnected component H containing u11 and u22 . By Corollary 1, there exists a lowest vertex v in
in H
H with
with two
two
arc-disjoint
paths
P
,
P
from
⇢
to
v.
Since
v
is
a
lowest
vertex
in
H,
we
know
that
v
is
a
reticulation
arc-disjoint paths P11 , P22 from ρ to v. Since v is a lowest vertex in H, we know that v is a reticulation vertex
vertex
and
and the
the arc
arc leaving
leaving vv is
is aa cut
cut arc.
arc. Let
Let B
B=
= C(v)
C(v) and
and A
A=
=X
X \\ B.
B. Then
Then B
B is
is clearly
clearly nonempty.
nonempty. In
In addition,
addition,
A
A is
is nonempty,
nonempty, as
as otherwise
otherwise lsa(X)
lsa(X) v,
v, aa contradiction
contradiction to
to the
the fact
fact that
that lsa(X)
lsa(X) =
= ⇢ρ (as
(as N
N is
is recoverable).
recoverable).
Therefore,
Therefore, A|B
A|B is
is aa split.
split.
Consider xx 2
∈A
A and
and yy 2
∈B
B and
and the
the subnetwork
subnetwork N
N||{x,y}
There is
is at
at least
least one
one directed
directed path
path from
from ⇢ρ to
to x,
x,
Consider
{x,y}.. There
and each
each such
such path
path contains
contains at
at least
least one
one arc
arc of
of P
P11 or
or P
P22.. Hence,
Hence, in
in the
the process
process of
of obtaining
obtaining N
N||{x,y}
from N
N,,
and
{x,y} from
the paths
paths P
P11,, P
P22 do
do not
not become
become parallel
parallel arcs.
arcs. Therefore,
Therefore, N
N||{x,y}
contains two
two arc-disjoint
arc-disjoint paths
paths from
from ⇢ρ to
to vv
the
{x,y} contains
and we
we can
can conclude
conclude that
that N
N||{x,y}
6= TT(x,
(x, y).
y). Therefore,
Therefore, ifif N
N||{x,y}
∈B
B⇤∗,, that
that is,
is, N
N||{x,y}
is level-1,
level-1, then
then
and
{x,y} is
{x,y} 6=
{x,y} 2
N||{x,y}
= R(x;
R(x; y).
y). This
This implies
implies that
that there
there is
is no
no arc
arc (y,
(y, x).
x). Therefore,
Therefore, A|B
A|B is
is aa Type
Type II or
or Type
Type II
II split
split of
of
N
{x,y} =
D(B⇤∗).
).
t
u
D(B
t
u
Note that
that the
the condition
condition that
that B
B is
is displayed
displayed by
by aa binary
binary network
network in
in the
the above
above lemma
lemma can
can not
not be
be weakened
weakened
Note
to that
that B
B is
is displayed
displayed by
by aa network.
network. For
For example,
example, consider
consider the
the binet
binet collection
collection B
B and
and network
network N
N in
in Figure
Figure 4.
4.
to
Although network
network N
N displays
displays B,
B, digraph
digraph D(B)
D(B) has
has no
no typed
typed split
split (as
(as can
can be
be easily
easily checked).
checked).
Although
We now
now introduce
introduce two
two operations,
operations, which
which can
can be
be used
used to
to combine
combine two
two phylogenetic
phylogenetic networks
networks into
into aa new
new
We
one.
Suppose
that
N
and
N
are
two
phylogenetic
networks
with
disjoint
leaf
sets.
Let
T
(N
,
N
)
be
the
one. Suppose that N11 and N22 are two phylogenetic networks with disjoint leaf sets. Let T (N11, N22) be the
phylogenetic
network
obtained
from
N
and
N
by
adding
a
new
vertex
v
and
two
arcs
from
v
to
the
roots
of
N11
phylogenetic network obtained from N11 and N22 by adding a new vertex v and two arcs from v to the roots of N
and
N
.
In
addition,
the
network
R(N
;
N
)
is
obtained
by
taking
a
binet
R(y
;
y
),
with
y
,
y
∈
/
L(N
)∪L(N
),
and N22. In addition, the network R(N11; N22) is obtained by taking a binet R(y11; y22), with y11, y22 2
/ L(N11)[L(N22),
and
replacing
y
by
the
root
of
N
,
for
i
=
1,
2.
See
Figure
5
for
examples.
and replacing yii by the root of Nii, for i = 1, 2. See Figure 5 for examples.
For aa binet
binet set
set B
B on
on X
X and
and aa subset
subset A
A✓
⊆ X,
X, we
we define
define
For
B|AA =
= {S
{S 2
∈B
B :: L(S)
L(S) ✓
⊆ A}.
A}.
B|
The next
next theorem
theorem can
can be
be used
used to
to determine
determine in
in polynomial
polynomial time
time whether
whether aa collection
collection of
of binary
binary level-1
level-1 binets
binets
The
is
displayed
by
some
binary
level-1
network.
See
Section
6
for
more
details.
is displayed by some binary level-1 network. See Section 6 for more details.
8
Leo van Iersel et al.
x
x
p
p
y
q
y
q
R(R(x; y); R(p; q))
T (R(x; y), R(p; q))
Fig. 5: Examples of two networks built recursively by using the operations introduced in the text.
Theorem 2 Suppose that B is a thin collection of binary level-1 binets on X. If there exists a typed split A|B
of D(B) such that B|A and B|B are both displayed by some binary level-1 network, then B is displayed by a
binary level-1 network. Moreover, if B is displayed by a binary level-1 network, then there exists at least one
typed split of D(B) and, for each typed split A|B of D(B), B|A and B|B are both displayed by some binary
level-1 network.
Proof First suppose that there exists a typed split A|B of D(B) such that B|A and B|B are displayed by binary
level-1 networks NA and NB , respectively.
If A|B is a Type I or Type III split of D(B), then consider the network N = T (NA , NB ). Then N is a
binary level-1 phylogenetic network on X and
B ⊆ {T (x, y) : x ∈ A, y ∈ B} ∪ B(NA ) ∪ B(NB ) = B(N ),
and so B is displayed by N .
If A|B is a Type II split of D(B), then without loss of generality we may assume that B is a sink set in
D(B). Now consider the network N = R(NA ; NB ). Then N is a binary level-1 phylogenetic network on X and
B ⊆ {R(x; y) : x ∈ A, y ∈ B} ∪ B(NA ) ∪ B(NB ) = B(N ),
and so B is displayed by N .
Now suppose that B is displayed by a binary level-1 network N . By Lemma 3, there exists a typed split
A|B of D(B). Then B|A ⊆ B(N |A ) and B|B ⊆ B(N |B ).
t
u
We now prove the main result of this section.
Theorem 3 Suppose that B is a thin collection of binary level-1 binets on X. Then B is displayed by a binary
level-1 network if and only if it is displayed by a binary network.
Proof Suppose that B is displayed by a binary network. We claim that B is also displayed by a binary level-1
network. We shall establish this claim by induction on |X|.
If |X| = 2, then B contains at most one binet, which has leaf set X. Therefore we know that B is displayed
by a binary level-1 network.
Now assume that |X| > 2, and the claim holds for all sets X 0 with 2 ≤ |X 0 | < |X|. Let N be a binary
network on X with B ⊆ B(N ). By Lemma 3, there exists a typed split A|B of D(B). Note that B|A ⊆ B(N |A )
and B|B ⊆ B(N |B ). Therefore, by induction, each of B|A and B|B is displayed by a binary level-1 network. By
Theorem 2, it follows that B is displayed by a binary level-1 network.
t
u
5 Binets determine the number of reticulations of a binary level-1 network
In this section we show that, that although the collection of binets displayed by a level-1 network does not
necessarily determine the network (see Figure 1), it does in fact determine the number of reticulations in the
network. We begin by showing that it suffices to consider level-1 networks in which all cycles (in the underlying
undirected graph) have length 3.
First, we introduce some further notation. A semi-cycle C of an acyclic directed graph is the union of two
non-identical, internally-vertex-disjoint, directed paths from s to t, with s = s(C) and t = t(C) two distinct
vertices that are referred to as the source and terminal of C, respectively. The length of a semi-cycle is the
number of distinct vertices that it contains.
We now show that we may restrict to networks in which all semi-cycles have length 3.
Binets: fundamental building blocks for phylogenetic networks
9
Ai
Ai
vi
Bi
Ai
vi
x
Bi
(a) The parent of x is not in a semicycle.
vi
x
Bi
(b) The parent of x is the terminal of
a semi-cycle.
x
(c) The parent of x is a non-terminal
non-source vertex of a semi-cycle.
Fig. 6: The three cases for the location of x in each of the networks N1 and N2 in the proof of Theorem 4.
A1
v1
v2
x
B1
N1
A2
v1
x
B2
A1
N2
(a) Case 1: the parent of x is not in
a semi-cycle in N1 but is the terminal of a semi-cycle in N2 .
v2
x
B1
N1
A2
x
B2
N2
(b) Case 2: the parent of x is not
in a semi-cycle in N1 but is the
non-terminal non-source vertex of
a semi-cycle in N2 .
Fig. 7: The two possible ways to add leaf x to N10 and N20 in the proof of Theorem 4 such that the obtained
networks N1 and N2 have different numbers of reticulation vertices.
Lemma 4 If N is a binary level-1 network, then there exists a binary level-1 network N 0 in which every semicycle has length 3, such that B(N 0 ) = B(N ) and N and N 0 have the same number of reticulation vertices.
Proof Consider a semi-cycle of N with source s and terminal t and length at least 4. Let (u1 , v1 ), . . . ,(uk , vk ),
(t, w) be the arcs leaving the semi-cycle. Then k ≥ 2. Let N ∗ be a network obtained from a binary tree
on {v1 , . . . , vk } by replacing vi by the subgraph of N rooted at vi , for i = 1, . . . , k. Let Nw be the subgraph
of N rooted at w. Then we construct N 0 from N by replacing the subgraph of N rooted at s by the network R(N ∗ ; Nw ). It is straightforward to see that N 0 is a binary level-1 network with the required properties.
t
u
We now establish the main result of this section.
Theorem 4 If N1 and N2 are rooted binary level-1 phylogenetic networks on X with B(N1 ) = B(N2 ) then N1
and N2 have the same number of reticulation vertices.
Proof The proof is by induction on the number of leaves |X|. The induction basis for |X| = 2 is clear. Now
suppose that N1 and N2 are two non-isomorphic rooted binary level-1 phylogenetic networks on |X| ≥ 3 with
B(N1 ) = B(N2 ) but with different numbers of reticulation vertices. We add an outdegree-1 root to each of N1
and N2 with an arc to the original root. By Lemma 4, we may assume that all semi-cycles in N1 and N2 have
length 3.
Choose an arbitrary leaf x ∈ X and let X 0 = X \ {x}. Let N10 and N20 be the networks obtained from N1 |X 0
and N2 |X 0 , respectively, by adding an outdegree-1 root with an arc to the original root. Then N10 and N20 have
the same number of reticulation vertices by induction.
Since all semi-cycles in N1 and N2 are assumed to have length 3, there are three cases for the location of x
in each of the networks N1 , N2 , illustrated in Figure 6.
If the parent of x is in a semi-cycle in Ni , let vi be the source of this semi-cycle, and let vi be the parent
of x otherwise. Let Bi := C(vi ) \ {x} and Ai := X 0 \ Bi (recall that C(vi ) denotes the cluster of vi ).
We now consider the different ways in which we could add x to both networks. Since N1 and N2 have
different numbers of reticulation vertices, there are two cases to consider (after eliminating symmetric cases),
as illustrated in Figure 7.
The first case is that the parent of x is not in a semi-cycle in N1 but is the terminal of a semi-cycle in N2 .
First suppose that B1 ∩ B2 6= ∅. Then choose an arbitrary vertex y ∈ B1 ∩ B2 . Then N1 |{x,y} = T (x, y)
10
Leo van Iersel et al.
while N2 |{x,y} = R(y; x), a contradiction. Hence, we may assume that B1 ∩ B2 = ∅. Then B1 = A2 and
B2 = A1 . Clearly, B1 , B2 6= ∅. Take y ∈ B1 = A2 and z ∈ B2 = A1 . Then N1 |{x,y} = T (x, y) and hence
N2 |{x,y} = T (x, y), from which we can deduce that N2 |{z,y} = T (z, y). In addition, N2 |{z,x} = R(z; x) and
hence N1 |{z,x} = R(z; x), from which we can deduce that N1 |{z,y} = R(z; y). This leads to a contradiction
since N2 |{z,y} = T (z, y).
The second case is that the parent of x is not in a semi-cycle in N1 but is the non-terminal non-source
vertex of a semi-cycle in N2 . First suppose that B1 ∩ B2 6= ∅. Then choose an arbitrary vertex y ∈ B1 ∩ B2 .
Then N1 |{x,y} = T (x, y) while N2 |{x,y} = R(x; y), a contradiction. Hence, we may assume that B1 ∩ B2 = ∅.
Then, as in the previous case, B1 = A2 6= ∅ and A1 = B2 6= ∅. Take y ∈ B1 = A2 and z ∈ B2 = A1 . Then,
similar to the previous case, N1 |{x,y} = T (x, y) and hence N2 |{x,y} = T (x, y), from which we can deduce that
N2 |{z,y} = T (z, y). In addition, N2 |{z,x} = R(x; z) and hence N1 |{z,x} = R(x; z), from which we can deduce
that N1 |{z,y} = R(y; z). This again leads to a contradiction since N2 |{z,y} = T (z, y).
t
u
6 Complexity of Binet Compatibility
A direct consequence of Theorem 2 is that there exists a simple polynomial-time algorithm to decide whether
there exists a binary level-1 network displaying a given collection B of binary level-1 binets (see [11] for
a related algorithm). In particular, a sink set of D(B) can be found in polynomial-time by computing the
strongly connected components of D(B) [17] and checking for each of them whether it is a sink set. This can
be used to find a typed split, if it exists. If such a split does not exist, then B is not compatible. Otherwise,
we can try to construct networks for B|A and B|B recursively, and combine them as described in the proof of
Theorem 2. This algorithm is similar to the Aho algorithm for deciding whether a set of rooted trees can be
displayed by some rooted tree [1].
From Theorem 3, it now follows that the following problem can also be solved in polynomial time.
Binet Compatibility (BC)
Input: a set B of binary level-1 binets.
Question: is B compatible, i.e., does there exist a binary network N with B ⊆ B(N )?
We show now that the assumption that all binets in B are binary and level-1 is essential. Indeed, for general
binets, the compatibility problem is at least as hard as the well-known graph isomorphism problem (GI) [8,
18], which is not known to be solvable in polynomial time. This is even true when the given binet set is thin
(contains at most one binet for each pair of leaves).
Theorem 5 Deciding whether there exists a phylogenetic network displaying a given thin set B of binets is
GI-hard.
Proof We reduce from DAG-isomorphism, which is known to be GI-complete [18]. Let G1 , G2 be two directed acyclic graphs, which form an instance of the DAG-isomorphism problem. For i = 1, 2, we add vertices ρi , ui , vi , wi , ri , a new leaf labelled x, an arc from wi to each indegree-0 vertex of Gi and from each
outdegree-0 vertex of Gi to ri and arcs (ρi , ui ), (ui , vi ), (ρi , vi ), (vi , wi ) and (ri , x). In G1 , we add a new leaf
labelled y and an arc (u1 , y). In G2 , we add a new leaf labelled z and an arc (u2 , z). We have thus transformed G1 into a binet B1 and G2 into a binet B2 . The third binet is B3 = T (y, z). See Figure 8 for an
illustration.
We claim that G1 and G2 are isomorphic if and only if there exists a network displaying B1 , B2 and B3 .
First assume that G1 and G2 are isomorphic. Then we can construct a network displaying B1 , B2 and B3
as follows. Take B1 and subdivide the arc (u1 , y) by a new vertex u01 and add leaf z with an arc (u01 , z). The
obtained network clearly displays B1 and B3 and it also displays B2 since G1 and G2 are isomorphic.
Now assume that there exists some network N displaying B1 , B2 and B3 . Then N |{x,y} = B1 . Hence, N
contains a cycle (in the underlying undirected graph) containing a reticulation v, such that x and the image
of G1 are below the arc leaving v, while y is below some other arc leaving the cycle. Since N |{y,z} = T (y, z),
leaf z is not below v in N . Therefore, deleting v, x and the parent of x from the subgraph of N rooted at v
gives G1 .
Similarly, N contains a cycle containing a reticulation v 0 , such that x and the image of G2 are below the
arc leaving v 0 , while z is below some other arc leaving the cycle. Since N |{y,z} = T (y, z), leaf y is not below v 0
in N . Therefore, deleting v 0 , x and the parent of x from the subgraph of N rooted at v 0 gives G2 .
Moreover, v = v 0 since N |{y,z} = T (y, z). Hence, G1 and G2 are isomorphic.
t
u
Binets: fundamental building blocks for phylogenetic networks
y
11
z
z
G1
y
z
y
G1
G2
B3
x
x
x
B1
B2
N
Fig. 8: The three binets constructed in the proof of Theorem 5, and a network N that displays all three binets
if the digraphs G1 and G2 are isomorphic.
7 Maximum Binet Compatibility
If a collection of binets is not compatible, the question arises whether it is possible to find a largest compatible
subset of the binets, in polynomial time. Here we show that this is unlikely to be the case. The decision version
of this problem is defined as follows.
Maximum Binet Compatibility (MBC)
Input: a set B of binary level-1 binets and an integer k.
Question: does there exist a compatible subset B 0 of B with |B 0 | ≥ k?
We now establish the complexity of this problem (see Theorem 6). Recall from Section 5 that s(C) and t(C)
denote the source and terminal of a semi-cycle C, respectively.
Lemma 5 If the binet R(x; y) is displayed by a binary level-1 network N , then lsa(x, y) is the source of a
semi-cycle C in N . In addition, y is below t(C) and x is not below t(C).
Proof Let u = lsa(x, y). Note that u is not a reticulation vertex, as otherwise the child of u would be a stable
ancestor of x and y that is below u. Hence, u has two children, denoted by u1 and u2 .
Observe that neither (u, u1 ) nor (u, u2 ) is a cut arc, since otherwise we would have N |{x,y} = T (x, y), while
by the assumption of the lemma N |{x,y} = R(x; y). Hence, u is the source of a semi-cycle C. Let v := t(C) be
the terminal of C. If neither x nor y is below v, then N |{x,y} = T (x, y), a contradiction. If both x and y are
below v, then v is a stable ancestor of x and y, a contradiction to lsa(x, y) = u. Therefore, precisely one of x
and y is below v. If x is below v and y is not, then N |{x,y} = R(y; x), a contradiction. Therefore, y is below v
and x is not.
t
u
In view of the last lemma, for each binet R(x; y) = N |{x,y} , there exists a unique semi-cycle CN (x; y)
containing lsa(x, y).
Lemma 6 If the two binets R(x; y) and R(y; z) are both displayed by a binary level-1 network N , then
s(CN (y; z)) ≺ t(CN (x; y)).
Proof Let C1 = CN (x; y) and C2 = CN (y; z). By Lemma 5, y ≺ t(C1 ) but y is not below t(C2 ), from which
we know that C1 6= C2 . Since s(C1 ) and s(C2 ) are stable ancestors of y in view of Lemma 5, we have either
s(C1 ) ≺ s(C2 ) or s(C2 ) ≺ s(C1 ) but not both.
Note that if s(C1 ) ≺ s(C2 ), then s(C1 ) ≺ t(C2 ) and hence y ≺ s(C1 ) ≺ t(C2 ), a contradiction. Thus
s(C2 ) ≺ s(C1 ), from which it follows that s(C2 ) ≺ t(C1 ).
t
u
Given a digraph G, let R(G) be the collection of binets {R(x; y) | (x, y) ∈ E(G)} induced by G. Note that
R(G) is a binet set on V (G), i.e., the leaves of the binets in R(G) correspond to the vertices of G.
Proposition 1 Let G be a digraph. Then G is acyclic if and only if R(G) is compatible.
12
Leo van Iersel et al.
xn
xn−1
xn−2
x1
Fig. 9: A level-1 network N∗ on X = {x1 , . . . , xn }.
Proof Let n = |X|, with X the vertex set of G. Suppose first that G is acyclic, then there exists a topological
sorting of G, that is, the vertices of G can be ordered as x1 , . . . , xn so that (xi , xj ) ∈ E(G) implies i < j.
Hence, the network N∗ in Figure 9 displays R(G) since N∗ displays each binet R(xi ; xj ) with i < j.
Conversely, suppose that R(G) is compatible. By Theorem 3, there exists a binary level-1 network N with
R(G) ⊆ B(N ). It remains to show that G is acyclic. If not, then there exists a directed cycle (x1 , x2 , . . . , xm )
for some m ≥ 3. Denote xm+1 = x1 . In view of Lemma 5, let Ci = CN (xi ; xi+1 ) be the semi-cycle in N
containing lsa(xi , xi+1 ) for 1 ≤ i ≤ m. Then Lemma 5 implies x1 ≺ s(Cm ) and that x1 is not below t(C1 ).
On the other hand, by Lemma 6 we have
s(Cm ) ≺ t(Cm−1 ) ≺ s(Cm−1 ) ≺ . . . ≺ s(C2 ) ≺ t(C1 ).
Together with x1 ≺ s(Cm ), it follows that x1 ≺ t(C1 ), a contradiction.
t
u
A set of binets B on X is said to be dense if for each pair of distinct elements x and y in X, there exists
precisely one binet on {x, y} in B. Hence, a dense set of binets is always thin.
Theorem 6 The problem MBC is NP-complete, even if the given set of binets is dense.
Proof We reduce from the NP-hard problem Feedback Arc Set in Tournaments (FAST) [2, 6], which is defined
as follows. Given a tournament, i.e. a digraph G = (V, E) with either (a, b) ∈ E or (b, a) ∈ E (but not both)
for each pair of distinct elements a and b in V , and given a positive integer k 0 , does there exist a subset F ⊆ E
of at most k 0 arcs whose removal makes G acyclic. If such an arc set exists, then we call it a feedback arc set
of G.
The reduction is as follows. For each instance (G, k 0 ) of FAST, consider the corresponding instance (R(G), k)
of MBC with k = |R(G)| − k 0 . Since the set R(G) of binets induced by G can be constructed in polynomial
time, it suffices to show that G contains a feedback arc set with size at most k 0 if and only if there exists a
compatible subset of R(G) of size at least k.
First assume that there exists a feedback arc set E 0 of G with size at most k 0 . That is, |E 0 | ≤ k 0 , and the
digraph G∗ obtained from G by deleting the arcs in E 0 is acyclic. Consider the set of binets B 0 = {R(x; y) :
(x, y) ∈ E \ E 0 }. This set contains at least k binets. In addition, since B 0 = R(G∗ ), it follows by Proposition 1
that B 0 is compatible.
Now assume that there exists a compatible binet set B 0 ⊆ R(G) with |B 0 | ≥ k. Consider the set E 0 =
{(x, y) : R(x, y) ∈ R(G) \ B 0 } of arcs of G. Then by Proposition 1, it follows that E 0 is a feedback arc set.
Moreover, |E 0 | ≤ k 0 , which completes the proof.
t
u
We complete the section by showing that there exists a polynomial time 1/3-approximation algorithm for
the MBC problem, which follows directly from the next theorem and its proof.
Theorem 7 Suppose that B is a set of binary level-1 binets on X. Then there exists a binary level-1 network
N such that |B(N ) ∩ B| ≥ |B|/3.
Proof If at least a third of the binets in B are tree type, then take N to be any binary tree on X and we are
done. Hence we may assume that at least two thirds of the binets are reticulate type.
Impose an arbitrary ordering on the elements in X, that is, write X = {x1 , . . . , xn }. Let B1 = B ∩
{R(xi ; xj ) : 1 ≤ i < j ≤ n} and B2 = B ∩ {R(xj ; xi ) : 1 ≤ i < j ≤ n}. Without loss of generality, we may
Binets: fundamental building blocks for phylogenetic networks
13
assume that |B1 | ≥ |B2 | (as the other case can be established in a similar way). Since at least two thirds of
the binets are reticulate type, and each of those is contained in either B1 or B2 (but not both), we know that
|B1 | ≥ |B|/3. Now consider the network N∗ in Figure 9, then clearly we have B1 ⊆ B(N∗ ). Thus we have
|B(N∗ ) ∩ B| ≥ |B1 | ≥ |B|/3, from which the theorem follows.
t
u
8 Discussion
In this paper we have developed some combinatorial results concerning collections of level-1 binets. Several
interesting questions arise from these results. For example, we have shown that the collection of level-1 binets
displayed by a binary phylogenetic network can be displayed by some level-1 network, but is there some
canonical level-1 network that could be used to display such a collection? In addition, can we count the
number of binary level-1 networks that display a dense compatible collection of binets? We have also seen that
the collection of binets displayed by a binary level-1 network determine its reticulation number. Therefore it
is natural to ask which properties of a phylogenetic network in general are determined by its binets?
We have also studied some algorithmic questions concerning binets. Concerning the maximum binet compatibilty problem, note that the constant 1/3 is sharp in Theorem 7. For example, consider the binet collection
{R(x; y), T (x, y), R(y; x)}. However, can a better bound be achieved by restricting to thin collections of binets,
and can improved approximation algorithms also be found?
In another direction, it would be interesting to know whether similar results to those proven in this paper
might hold for higher level networks. For example, what can be said about properties of collections of level-2
binets, and does Theorem 4 hold also for higher level networks? Also, we could try to generalize some of our
results to k-nets, i.e. networks on k leaves, k ≥ 2. For example, does Theorem 3 hold for trinets? In general,
it would be interesting to know what additional information the collection of k-nets displayed by a network
might contain for k ≥ 3. Note that it has been shown that trinets do not completely determine rooted networks
in general [12]. However, do they determine properties of networks such as the number of reticulations?
Similarly, it would be interesting to extend some of our algorithmic results to higher-level networks and
k-nets. For example, it is known that the compatibility problem is NP-complete for collections of level-1 trinets
[11]. However, to date the maximum trinet compatibility problem has not been studied.
Eventually, it is hoped that new results in these directions could be useful for developing novel methods
to construct phylogenetic networks from higher-level networks and k-nets. For example, using our results it
may be possible to develop approaches to build a consensus network for a collection of phylogenetic trees
or networks. Note that consensus networks have already proven themselves useful in the unrooted setting,
where they are used to summarize key features displayed by a collection of trees or networks (see e.g. [10]). A
consensus method based on binets could work by breaking each of the given networks down into a collection of
binets, and then developing methods to pool together the information contained in the resulting binets so as
to construct some consensus network, or at least some constraints that any such network should satisfy. Note
that similar approaches have been developed to build consensus trees for a collection of phylogenetic trees by
breaking each of the trees down into a collection of triplets (see e.g. [4, Section 2]). Probably it would be of
some interest to first consider how to construct a level-1 consensus network for a collection of level-1 networks
by breaking each of them down into level-1 binets. This is already likely to be quite challenging in view of our
result concerning NP-completeness of Maximum Binet Compatibility.
References
1. Aho, A.V., Sagiv, Y., Szymanski, T.G., Ullman, J.D.: Inferring a tree from lowest common ancestors with an application to
the optimization of relational expressions. SIAM Journal on Computing 10(3), 405–421 (1981)
2. Alon, N.: Ranking tournaments. SIAM Journal on Discrete Mathematics 20(1), 137–142 (2006)
3. Bapteste, E., van Iersel, L.J.J., Janke, A., Kelchner, S., Kelk, S., McInerney, J.O., Morrison, D.A., Nakhleh, L., Steel, M.,
Stougie, L., Whitfield, J.: Networks: expanding evolutionary thinking. Trends in Genetics 29(8), 439–441 (2013)
4. Bryant, D.: A classification of consensus methods for phylogenetics. DIMACS series in discrete mathematics and theoretical
computer science 61, 163–184 (2003)
5. Byrka, J., Guillemot, S., Jansson, J.: New results on optimizing rooted triplets consistency. Discrete Applied Mathematics
158(11), 1136–1147 (2010)
6. Charbit, P., Thomassé, S., Yeo, A.: The minimum feedback arc set problem is NP-hard for tournaments. Combinatorics,
Probability and Computing 16(01), 1–4 (2007)
7. Felsenstein, J.: Inferring phylogenies. Sinauer Associates Sunderland (2004)
8. Goldberg, M.: The graph isomorphism problem. In: J.L. Gross, J. Yellen (eds.) Handbook of Graph Theory, pp. 68–78. CRC
Press (2003)
9. Gusfield, D.: ReCombinatorics: The Algorithmics of Ancestral Recombination Graphs and Explicit Phylogenetic Networks.
MIT Press (2014)
14
Leo van Iersel et al.
10. Holland, B., Huber, K., Moulton, V., Lockhart, P.: Using consensus networks to visualize contradictory evidence for species
phylogeny. Molecular Biology and Evolution 21(7), 1459–1461 (2004)
11. Huber, K., van Iersel, L.J.J., Moulton, V., Scornavacca, C., Wu, T.: Reconstructing phylogenetic level-1 networks from
nondense binet and trinet sets. Algorithmica (2015). DOI 10.1007/s00453-015-0069-8
12. Huber, K., van Iersel, L.J.J., Moulton, V., Wu, T.: How much information is needed to infer reticulate evolutionary histories?
Systematic biology 64, 102–111 (2015)
13. Huber, K., Moulton, V., Wu, T.: Closed sets in phylogenetic networks. Preprint (2016)
14. Huson, D., Rupp, R., Scornavacca, C.: Phylogenetic Networks: Concepts, Algorithms and Applications. Cambridge University
Press (2010)
15. Morrison, D.: An introduction to phylogenetic networks. RJR Productions (2011)
16. Oldman, J., Wu, T., van Iersel, L.J.J., Moulton, V.: Trilonet: Piecing together small networks to reconstruct reticulate
evolutionary histories. Molecular biology and evolution 33(8), 2151–2162 (2016)
17. Tarjan, R.: Depth-first search and linear graph algorithms. SIAM journal on computing 1(2), 146–160 (1972)
18. Zemlyachenko, V.N., Korneenko, N.M., Tyshkevich, R.I.: Graph isomorphism problem. Journal of Soviet Mathematics 29(4),
1426–1481 (1985)
| 8cs.DS
|
The infinite simple group V of Richard J. Thompson:
presentations by permutations
arXiv:1511.02123v2 [math.GR] 28 Aug 2017
Collin Bleak & Martyn Quick
School of Mathematics & Statistics, University of St Andrews,
St Andrews, Fife, KY16 9SS, United Kingdom
[email protected], [email protected]
August 29, 2017
Abstract
We show that one can naturally describe elements of R. Thompson’s finitely presented
infinite simple group V , known by Thompson to have a presentation with four generators and
fourteen relations, as products of permutations analogous to transpositions. This perspective
provides an intuitive explanation towards the simplicity of V and also perhaps indicates a
reason as to why it was one of the first discovered infinite finitely presented simple groups: it
is (in some basic sense) a relative of the finite alternating groups. We find a natural infinite
presentation for V as a group generated by these “transpositions,” which presentation bears
comparison with Dehornoy’s infinite presentation and which enables us to develop two small
presentations for V : a human-interpretable presentation with three generators and eight
relations, and a Tietze-derived presentation with two generators and seven relations.
Mathematics Subject Classification (2010).
Primary: 20F05; Secondary: 20E32, 20F65.
Keywords. Thompson’s groups, simple groups, presentations, generators and relations, permutations, transpositions.
1
Introduction
In this article, we investigate R. Thompson’s group V from a mostly-unexplored perspective.
As a consequence we derive new, and hopefully elegant, presentations of this well-known group
and introduce a simple and dextrous notation for handling computations in V .
Recall that the group V first appears in Thompson’s 1965 notes [23] and is given there as
one of two “first-examples” of infinite finitely presented simple groups (along with its simple
subgroup T , called “C” in those notes). Since then, it has been the focus of a large amount of
subsequent research (see, for example, [4, 5, 6, 13, 15, 18, 24] for a small part of that research).
Thompson’s group V arises in various other settings, for example, Birget [2, 3] investigates
connections to circuits and complexity while Lawson [19] considers links to inverse monoids and
étale groupoids.
We shall demonstrate that one can consider V as a symmetric group acting, not on a finite
set, but instead on a Cantor algebra (the algebra of basic clopen sets in a Cantor space). We
focus upon certain well-known properties of a finite symmetric group, namely being generated
by transpositions and being transitive in its natural action. Reflecting these two fundamental
properties, a finite symmetric group possesses a Coxeter-type presentation, with generating
set T corresponding to a set of appropriate transpositions and relations t2 = 1 for all t ∈ T ,
(tu)2 = 1 when t, u ∈ T correspond to transpositions of disjoint support and (tu)3 = 1 when
1
t 6= u but the corresponding transpositions have intersecting support. If we exploit the fact
that these generators have order 2, this third type of relation can be rewritten as t−1 ut = u−1 tu
and indeed in the symmetric group this conjugate equals another transposition v, namely that
whose support satisfies supp v = (supp t)u.
In the context of a Cantor algebra, the analogues of transpositions are piecewise affine maps
which “swap” a pair of basic open sets. We shall observe that Thompson’s group V is generated
by such transpositions of the standard Cantor algebra and hence derive an infinite Coxeter-like
presentation for V , as appears in Theorem 1.1 below.
As is well known, the standard Cantor algebra admits a natural tree-structure where the
nodes correspond to the basic open sets in Cantor space C and these nodes are indexed by finite
words in the alphabet X = {0, 1}. Consequently, we label our transpositions by two incomparable words α and β from X ∗ . Indeed, for such α and β, we write tα,β for the element of V that is
the transposition defined in Equation (2.1). We shall also write sα,β and (α β) for symbols representing elements in two abstract groups whose presentations we give in Theorems 1.1 and 1.2,
respectively. We specifically use different notations for each of these elements so as to distinguish
between the elements of each abstract group and the actual transformations of Cantor space.
The thrust of our work is to demonstrate that the two abstract groups are isomorphic to V and
that under the isomorphisms these three elements sα,β , (α β) and tα,β correspond. The first two
of our families of relations appearing in Theorem 1.1 reflect that these tα,β act as transpositions
so have order 2, commute when their supports are disjoint, and conjugate in a manner analogous
to transpositions in symmetric groups when their supports intersect appropriately.
Passing from the setting of actions of finite permutation groups on finite sets to the setting
of corresponding actions of infinite groups on Cantor algebras has further implications for the
resulting presentation. Namely, due to the self-similar nature of Cantor space, each generating
transposition can be factorised. To be precise, each tα,β satisfies what we call a split relation:
tα,β = tα0,β0 tα1,β1 . This provides the third family of relations that are seen in Theorem 1.1. They
have the consequence that not only is every element of V a product of our transpositions tα,β , but
also we can re-express any such product as one that involves an even number of transpositions.
Thus one can simultaneously view R. Thompson’s group V as an infinite analogue of both the
finite alternating groups and of the finite symmetric groups.
It follows quite easily from the presentation in Theorem 1.1 that any transposition tγ,δ can be
obtained by conjugation using only those tα,β with |α|, |β| 6 3, for example, and this motivates
an effort to find a finite presentation involving permutations and their relations, where these
permutations involve only the nodes in the first three levels of the tree. Theorem 1.2 provides
this presentation (involving three generators and eight relations). Note here that we depart
slightly from the Coxeter-style of presentation: we exploit the presence of the symmetric group
of degree 4 acting upon X 2 to reduce further the presentation, at the cost of employing a “threecycle” as a generator. Of note, this human-interpretable presentation is much smaller than the
currently known finite presentation for V (given by Thompson [23] and discussed in detail in
Cannon, Floyd and Parry’s survey [12]), which has four generators and fourteen relations.
As a technical exercise, we further reduce the presentation in Theorem 1.2 to a 2-generator
and 7-relation presentation, found in Theorem 1.3. The resulting presentation is small, but not
so readily interpretable by humans.
Our infinite presentation in Theorem 1.1 bears comparison with Dehornoy’s infinite presentation for V (see [13, Proposition 3.20]). Dehornoy’s presentation highlights different aspects
as to why the group V can be considered as a fundamental object in group theory, and even in
mathematics, bearing out, as it does, the connection of V to systems with equivalences under
associativity and commutativity.
Our viewpoint of V as a form of a symmetric or alternating group perhaps hints at why
V arose as one of first two known examples of an infinite simple finitely presented group. Permuting sets is a basic activity, and the Cantor algebra represents a fundamental way to pass from
2
a finite to an infinite context, thus it seems natural that researchers eventually noticed V . To
give further background, we note there are many generalisations of V to infinite simple finitely
presented groups, all of which owe their simplicity to the same fundamental idea (similar to
the reason why the alternating groups are simple). One such family is the Higman–Thompson
groups Gn,r for which V = G2,1 , see [15]. (The group Gn,r is simple for n even. When n is
odd, one must pass to the commutator subgroup of index 2, reflecting the observation that the
corresponding split relations in Gn,r do not change the parity of any decomposition as a product
of transpositions.) Other families include the Brin–Thompson groups nV for which V = 1V ,
see [6], and the groups nVm,r that generalise the previous two families, see [20], and where we
have similar simplicity considerations, see [7]. The finite presentability of these groups comes
from the much stronger fact that they are all in fact F∞ groups. (There is a beautiful argument of the F∞ nature of these groups given in [24], which applies to many of these “relatives”
of V . In many specific cases, F∞ arguments already exist for individual groups and for classes
of groups in these families. See, for example, [1, 8, 9, 17].) The ideas of this paper ought to
apply to all of these groups of “Thompson type” in aiding in the discovery of natural and small
presentations. On the other hand, the infinite family of finitely-presented infinite simple groups
arising from the Burger-Mozes construction and following related work (see, e.g., [10, 11, 21])
are of an entirely different nature, and the methods employed here do not seem appropriate to
that context.
We mention here a debt to Matatyu Rubin and Matthew G. Brin. Rubin indicated to Brin
a proof of the simplicity of V , which uses the generation of V by transpositions with restricted
support on Cantor space. Brin set this proof out briefly in his paper [6] and developed the ideas
to extend the proof to the groups nV , which he carries out in the short paper [7]. It is not a
stretch to say that the current article would not exist without that thread of previous research.
A note on content
The first two sections of this article are intended for the interested mathematician and provide
structure and insights into these sorts of groups. The outline of the proofs of the theorems are
found towards the end of Section 2. The sections that follow are more technical and verify the
details required for those proofs.
Statement of results and some notation
Let X = {0, 1}. We write X ∗ for all finite sequences x1 x2 . . . xk where k > 0 and each xi ∈ X.
In particular, we assume that X ∗ contains the empty word ε. We view the elements of X ∗
as representing the nodes on the infinite binary rooted tree with edges between nodes if they
are represented by words which differ by a suffix of length 1. (Figure 1 illustrates this tree
together with the nodes labelled by elements of X ∗ .) Similarly, we give the standard definition
of the Cantor set C as X ω , the set of all infinite sequences x1 x2 x3 . . . of elements of X under
the product topology (starting with X endowed with the discrete topology). Thus points in C
correspond to boundary points of the infinite binary rooted tree.
If α ∈ X ∗ and β ∈ X ∗ ∪ X ω , then we write αβ for the concatenation of the two sequences.
We denote by αC the set of elements of C with initial prefix α. This set is a basic open set
in the topology on C and is itself homeomorphic to C. We shall write α β to indicate that
α is a prefix of β (including the possibility that the two sequences are equal). This notation
then means that β = αγ for some γ ∈ X ∗ ∪ X ω . Moreover, when β ∈ X ∗ , then α and β
represent nodes on the infinite binary rooted tree such that β lies on a path descending from α
(see Figure 2(i)), and therefore βC ⊆ αC.
We also write α ⊥ β to denote that both α 6 β and β 6 α. Then we shall say that α and β
are incomparable. In this case, the paths to α and to β from the root separate at some node
above both α and β (as shown in Figure 2(ii)), so that αC ∩ βC = ∅ (for such α, β ∈ X ∗ ).
3
ε
0
1
00
01
000
..
.
001 010
..
.
..
.
..
.
..
.
10
11
011 100
..
.
..
.
..
.
..
.
101 110
..
.
..
.
..
.
111
..
.
..
.
..
.
..
.
Figure 1: The infinite binary rooted tree with nodes labelled by elements of X ∗
..
.
..
.
..
.
..
.
..
.
..
.
α
α
β
β
|
{z
βC
}
Figure 2: (i) α β (and the paths representing elements of βC); and (ii) α ⊥ β
4
If m is a positive integer, we shall also use X m to denote the collection of all finite sequences
x1 x2 . . . xm of length m with xi ∈ X for each i. We write |α| for the length of the sequence
α ∈ X ∗.
Motivated by the well-known fact (see, for example, [13]) that R. Thompson’s group V has
a partial action on the set of finite binary rooted trees and equally on the set X ∗ , we shall use
the notation γ · (α β), for α, β, γ ∈ X ∗ , defined by
if γ = αδ for some δ ∈ X ∗ ,
βδ
αδ
if γ = βδ for some δ ∈ X ∗ ,
γ · (α β) =
(1.1)
γ
if
both
γ
⊥
α
and
γ
⊥
β,
undefined otherwise.
Thus γ · (α β) is undefined precisely when γ ≺ α or γ ≺ β and when it is defined it represents a
prefix substitution replacing any occurrence of α by β and vice versa.
The notation appearing in Equation (1.1) above is motivated via our anticipated isomorphism
between V and the abstract group defined by the presentation in Theorem 1.2. The map tα,β is
taken to the element (α β) and the above formula reflects the effect of applying tα,β to a point
in the partial action of V on X ∗ . Note also that we are choosing to write our maps on the right
since in our opinion this enables one to more conveniently compose a number of maps and such
a convention is consistent with denoting an element of V by tree-pairs with the domain on the
left and the codomain on the right. Nevertheless, we still view maps in V as being given by
prefix substitutions of the infinite sequences that are elements of the Cantor space so as to be
consistent with other work on these groups.
Our results are as follows:
Theorem 1.1 Let A to be the set of all symbols sα,β where α, β ∈ X ∗ with α ⊥ β. Then
R. Thompson’s group V has an infinite presentation with generating set A and relations
s2α,β = 1
s−1
γ,δ sα,β sγ,δ = sα·(γ δ),β·(γ δ)
(1.2)
sα,β = sα0,β0 sα1,β1
where α, β, γ and δ range over all sequences in X ∗ such that α ⊥ β, γ ⊥ δ, and α · (γ δ) and
β · (γ δ) are defined.
Our primary finite presentation for the group V has three generators a, b and c, but, as mentioned above, it is most naturally expressed in terms of a “permutation-like” notation extending
the transpositions in the infinite presentation. Our generators a, b and c then correspond to
permutations that we denote (00 01), (01 10 11) and (1 00), respectively, and the relations are
similarly expressed in terms of elements (α β) that we define fully in Section 2. This “humanreadable” presentation is as follows:
Theorem 1.2 R. Thompson’s group V has a finite presentation with three generators (00 01),
(01 10 11) and (1 00) and eight relations
4
R1. (00 01)2 = (01 10 11)3 = (00 01) (01 10 11) = 1;
R2. (1 01)(1 00) = (00 01);
R3. (1 00) = (10 000) (11 001);
R4. [(00 010), (10 111)] = [(00 011), (10 111)] = 1;
R5. [(000 010), (10 110)] = 1.
5
We shall provide words in terms of the generators a, b and c to express these relations later
in Equation (2.3). Observe that the Relations R1 tells us that (00 01) and (01 10 11) satisfy the
relations of the symmetric group S4 , so the subgroup that they generate is isomorphic to some
quotient of S4 . In fact, it will turn out that this subgroup is isomorphic to S4 .
The element (α β) will correspond to the element of Thompson’s group V that maps a point
of the Cantor set that has prefix α to a point with prefix β and vice versa. Relations R4 and R5
then reflect the fact that certain elements of V commute because they have disjoint support.
We show that V is generated by the two elements u and v described by the tree-pairs in
Figure 3. Transforming the presentation in Theorem 1.2 to one using these two generators via
Tietze transformations (as described in Corollary 5.2) and reducing the nine resulting relations
using the Knuth–Bendix algorithm, as implemented in the GAP package KBMAG [14, 16], results
in the surprising two generator and seven relation presentation given below:
Theorem 1.3 R. Thompson’s group V has a finite presentation with two generators u and v
and the seven relators
u6 , v 3 , (u3 v)4 ,
v −1 u(u2 v −1 )2 u3 vu−1 v −1 u3 vu(uvu2 (uv −1 u3 v)3 )2 uv −1 u3 v −1 ,
uv −1 u3 v −1 u−2 v −1 uvu2 v −1 u−1 vu2 v −1 uvu−1 (u−1 v −1 )2 u3 vu−1 ,
v(uv −1 u3 v −1 )2 u−1 v −1 u3 v −1 u−1 v −1 u3 v,
uvu3 vuv −1 u−2 v −1 u(u2 v)2 (u2 v −1 )2 u3 vu−2 v −1 u3 v.
This reduction to seven relations caught the authors by surprise, but perhaps it is not so
unexpected in view of the deficiency (as defined, for example, in [22, §14.1]) of the presentations
in Theorems 1.2 and 1.3 both being −5.
2
Further preliminaries and the proofs of the main theorems
This section contains the heart of the mathematics within the article. We present all the remaining preliminaries required to fully understand the statements of the theorems listed in the
introduction, in particular, unpacking the presentations that we use. We then provide the proofs,
subject to deferring technical calculations to the sections that follow.
R. Thompson’s group V
One view of Thompson’s group V is as a group of certain homeomorphisms of the Cantor set C,
namely those that are finite products of the elements tα,β , for α, β ∈ X ∗ with α ⊥ β, defined as
follows
βy if x = αy for some y ∈ C;
(2.1)
xtα,β = αy if x = βy for some y ∈ C;
x
otherwise
u
v
−−→
1 2 3
−−→
2 1 5
4 5
1 2 3 4
1 4 2 3
3 4
Figure 3: Two elements given by tree-pairs that generate for R. Thompson’s group V
6
1
−−−→
1
4
2
2
3
4
3
Figure 4: The map t100,11 denoted using tree-pairs
(see Brin [6, Lemma 12.2]). Note that the map tα,β has the effect of swapping those elements
of C that have an initial prefix α with those that have an initial prefix β and fixing all other
points in C. A general element of V is often denoted by a pair of finite binary rooted trees
representing the domain and codomain of the map. We label the leaves of these two trees by the
numbers 1, 2, . . . , n (for some n) and this then specifies that our element of V has the effect of
substituting the prefix from X ∗ corresponding to the node in the domain tree labelled i by the
member of X ∗ corresponding to the node in the codomain tree with the same label (for each i).
For example, Figure 4 provides such tree-pairs for the map t100,11 as just defined.
From the definition, it is visible that t2α,β = 1. Equations of this type (as α and β range over
all incomparable pairs from X ∗ ) will form our family of order relations.
If we shift our attention to conjugation, it is a straightforward calculation in V , along the
lines of the familiar one concerning conjugation of permutations demonstrated to undergraduates
in a first course on group theory, to verify that
t−1
γ,δ tα,β tγ,δ = tα·(γ δ),β·(γ δ)
whenever α · (γ δ) and β · (γ δ) are both defined. We call this resultant family of relations in V
our conjugacy relations. At this point, we also note that we will use an exponential notation for
conjugation, so the left-hand side of the above relation will also be denoted by tα,β tγ,δ in what
follows.
Our final family of relations exploit the action of our elements tα,β on basic open sets of the
Cantor set C and the self-similar structure of this space. Specifically, if α ∈ X ∗ , then the basic
open set αC splits into two subsets, namely the set of all elements of C with initial prefix α0 and
those with initial prefix α1. In view of this, we obtain the equation
tα,β = tα0,β0 tα1,β1
(2.2)
when α, β ∈ X ∗ with α ⊥ β. We refer to the family of these relations as split relations.
Deriving the presentations for V
One of the presentations that we use in this article is that found by R. Thompson and discussed
in Cannon–Floyd–Parry (see [12, Lemma 6.1]). This presentation has four generators A, B,
C and π0 and fourteen relations. We state these relations when we need them at the start of
Section 4 below.
As described in Theorem 1.1, the first of our new presentations involves the order relations,
conjugacy relations and split relations of V just described. To be precise, we define P∞ to be the
group having the infinite presentation with generating set A = { sα,β | α, β ∈ X ∗ , α ⊥ β } and
the relations listed in Equation (1.2). Of course, we already know that that V is generated by
the maps tα,β and that these satisfy the order, conjugacy and split relations. This ensures that
there is a surjective homomorphism φ : P∞ → V given by sα,β 7→ tα,β for α, β ∈ X ∗ with α ⊥ β.
When establishing Theorem 1.1, we shall be observing that φ is actually an isomorphism.
We now describe our primary finite presentation for V , which has three generators a, b and c
and eight relations but, more importantly, can be readily understood by a human. The majority
7
of our calculations will take place with this presentation and so, as indicated above, we develop
a helpful notation that is parallel to that used in finite permutation groups. It is a consequence
of Higman [15] that all the relations that hold in V can be detected as consequences of products
using tree-pairs of some bounded size. This motivates our presentation which employs essentially
a finite subcollection of the order, conjugacy and split relations from P∞ and involving only some
swaps (α β) all satisfying |α|, |β| 6 3. (To aid reducing the number of relations required we
encode a copy of the symmetric group of degree 4, corresponding to acting on X 2 , within our
presentation.)
Accordingly, the three generators a, b and c of the group P3 that we define will represent cyclic
permutations of basic open sets. As stated above, we shall write (00 01), (01 10 11) and (1 00)
for a, b and c, respectively. This reflects the fact that, under the isomorphism that we shall
establish between P3 and V , the element a corresponds to the map t00,01 that interchanges the
basic open sets 00C and 01C, b corresponds to the product t01,10 t01,11 (inducing a 3-cycle of
the sets 01C, 10C and 11C), and c corresponds to t1,00 . We extend this notation by defining
elements (α β) that we will refer to as swaps below and a formula for each in terms of a, b and c
will be extracted from the definitions we make. The element (α β) will correspond to tα,β under
the isomorphism. It is these swaps that appear in our list of relations found in Theorem 1.2
above.
Once we have defined the swaps below, we can translate the Relations R1–R5 into words in
a, b and c. The result is the following restatement of Theorem 1.2:
Theorem 2.1 R. Thompson’s group V has a finite presentation with three generators a, b and c
and the following eight relations:
a2 = b3 = (ab)4 = 1,
ba
−1 cacaab−1 a
c = abcacaa ab
[ab
−1 cac
,a
cac = a,
−1
b−1 cacab ab a
−1 cacab ab−1 a
[abcac , ab
,
bca
[abca
] = 1,
−1
bcacab ab a
,a
] = 1,
(2.3)
] = 1.
The careful reader will doubtless have observed that the eighth relation can be shortened
by conjugating by a. The presentation as listed is a direct consequence of the interpretation of
Theorem 1.2 in terms of a, b and c. No effort has been made in its statement to reduce the
length of the relations.
Having found this nice presentation for V , we felt obligated to reduce the relations employing
the available technology, specifically the Knuth–Bendix Algorithm. This algorithm shortens the
above relations, although the results are no longer particularly transparent. To carry out these
reductions, we used the implementation of the algorithm found in the freely available KBMAG
package [16] in GAP [14] in the following way. Denote the eight relators corresponding to the
equations in (2.3) by r1 , r2 , . . . , r8 . We can construct a rewriting system associated to each of the
groups Qi = h a, b, c | r1 , r2 , . . . , ri−1 , ri+1 , . . . , r8 i for i = 4, 5, . . . , 8 in sequence. The systems
that KBMAG constructs are not confluent, but nevertheless enable us to replace each ri by a
Tietze-equivalent (in the group Qi ) shorter relation. This process is repeated until the resulting
relations stabilise. As a consequence, the normal closure, in the free group on {a, b, c}, of the
following eight relations is identical to that of our original list:
a2 = b3 = (ab)4 = 1,
c−1 (ac)2 a = 1,
(cab−1 aba)2 cb(cabab−1 a)2 = 1,
a(cb)2 a(b−1 c)2 bcabcb−1 cab−1 acb−1 (cb)2 ab−1 = 1,
ab−1 cbc(ab−1 )2 cbcb−1 a(b−1 c)2 babcb−1 cab−1 = 1,
ca(b−1 c)2 bacabacbc(b−1 ca)2 b(cb−1 )2 (acb)2 cb−1 cab−1 = 1.
8
(2.4)
We observe this mechanical process produces considerably shorter relations than our original
eight in Equation (2.3).
The presentation in Theorem 1.3 is deduced in a manner that similarly depends upon the
use of the Knuth–Bendix Algorithm. One first applies Tietze transformations to pass to a
2-generator presentation employing generators u and v and relations deduced from the list (2.4).
We shall describe these Tietze transformations by expressing u and v in terms of a, b and c
(adding extraneous generators), and expressing a, b and c in terms of u and v (removing extraneous generators). The relevant formulae are
b ab−1 a
u = a(ab ab )caca
−1
and
and
a = u3 ,
b = v,
c = (u3 )vu
−2 vu3
v=b
(u3 )vu
−1 vu3 v
.
These formulae can be deduced by direct calculation in V . This application of Tietze transformations is expanded upon a little in Section 5 and Corollary 5.2 provides the intermediate step
to the theorem. (This corollary is established by purely theoretical means and does not rely
upon computer calculation.)
We then employ the same relation reduction process using KBMAG as described earlier
and this shows that two of the nine relations resulting from the Tietze transformations are
extraneous. In this manner we have deduced Theorem 1.3 from Theorem 1.2.
We now proceed to formally define the swaps (α β) for α, β ∈ X ∗ with α ⊥ β and |α|, |β| 6 3
in terms of our generators a, b and c in order to present the group P3 . To start off, we define
swaps (α β) for α, β ∈ X 2 as follows:
(00 01) = a,
(00 10) = ab ,
(01 10) = aba , (01 11) = ab
−1 a
(00 11) = ab
−1
, (10 11) = abab
(2.5)
Here, and in all that follows, we shall also adopt the convention that the swap (β α) coincides
with (α β) whenever the latter has already been defined. We write T2 for the set of swaps (α β)
with α, β ∈ X 2 .
The swaps in T2 and their effect when conjugating will be of fundamental importance in our
calculations. Accordingly we spend a little time expanding upon the above definitions before
we define the remaining swaps. In the Relations R1, we have assumed that a and b satisfy the
relations of the symmetric group S4 and the formulae on the right-hand side of Equation (2.5) are
those that correspond to transpositions in S4 . Consequently, when we multiply and conjugate
elements of T2 , they behave in exactly the same way as transpositions do. In particular, we can
view individual elements of T2 , and, by extension, products of such swaps, as transformations
of the set X 2 . Indeed, our notation γ · (α β), as defined in Equation (1.1), when α, β, γ ∈ X 2 ,
is precisely the formulae for these maps. Our assumption of Relations R1 justifies our using
products of swaps from T2 as maps X 2 → X 2 , as we shall do explicitly, for example, in Lemma 3.4
below. Similarly, we have written (01 10 11) for the generator b, since it follows from the
Relations R1 that b is equal to the product (01 10) (01 11), which induces a 3-cycle on X 2 .
To define the remaining swaps (α β), where |α|, |β| 6 3, we need one further piece of notation.
If x ∈ X, we define x̄ to be the other element in X; that is,
(
1 if x = 0,
x̄ =
0 if x = 1.
Then for any x, y, z ∈ X, we make our definitions in the following order:
(0 1) = (00 10) (01 11);
9
(2.6)
(1 01) = (1 00)(00 01) ,
(1 00) = c,
(0 1x) = (1 0x)(0 1) ;
(1 00x) = (00 1x)(1 00) ,
(1 01x) = (1 00x)(00 01) ,
(0 1xy) = (1 0xy)(0 1) ;
(00 01x) = (1 01x)(1 00) ,
(01 00x) = (00 01x)(00 01) ,
(1x 1x̄y) = (0x 0x̄y)(0 1) ,
(1x 0yz) = (0ȳ 0yz)(0ȳ 1x) ,
(2.7)
(2.8)
(2.9)
(0x 1yz) = (1x 0yz)(0 1) ;
(000 001) = (1 000)(1 001) ,
(000 010) = (1 000)(1 010) ,
(000 011) = (1 000)(1 011) ,
(001 011) = (1 001)(1 011) ,
(2.10)
(xy0 xy1) = (000 001)(00 xy) .
Finally, for distinct κ, λ ∈ X 2 , fix a product ρκλ of swaps from T2 that moves 00 to κ and 01
to λ when viewed, as described above, as a map X 2 → X 2 . Define
(κx λy) = (00x 01y)ρκλ
(2.11)
for (x, y) ∈ {(0, 0), (0, 1), (1, 1)}. In this way, we have now defined all swaps (α β) where
|α|, |β| 6 3.
Having made these definitions, it is a straightforward matter to convert the relations R1–
R5 into the list (2.3) of actual words expressed in the generators a, b and c, completing the
translation of Theorem 1.2 into Theorem 2.1.
Proofs of the main theorems
We now provide the proofs of the main theorems (that is, Theorems 1.1 and 1.2), subject to
information that we shall establish in the sections of the paper that follow. Here we link the
groups V , P3 and P∞ . Specifically, we build homomorphisms between these groups as indicated
in the following diagram:
/
❥❥4 P3
' ❥❥❥❥❥❥❥
i0
τ
Tietze
/ / P̄ ❚
3 ❚❚❚ θ
❚❚❚❚
❚* *
hs00,01 ,g s01,10 , s10,11 , s1,00 i
hĀ, B̄, C̄, π̄0 i
j❚j ❚❚❚ ψ
❚❚❚❚
❚❚
(2.12)
i1
V oo
φ
P∞
❥G
❥❥❥❥
t ❥❥❥
❥
We already know that φ : P∞ → V is a surjective homomorphism. We now describe the
other parts of the hexagon of maps.
The majority of the work in Sections 3–5 involves the presentation for the group P3 , where
we establish information about the swaps (α β) defined above. In Section 3, we verify that
these swaps satisfy the order relations, conjugacy relations and split relations of R. Thompson’s
group V , providing we restrict to those involving only swaps (α β) with |α|, |β| 6 3. Then in
Section 4, we establish that four specific elements Ā, B̄, C̄ and π̄0 in P3 satisfy the fourteen
relations listed in Cannon–Floyd–Parry [12] that define the group V . In that section, we ensure
that we only rely upon the consequences of our Relations R1–R5 established in Section 3. This
then guarantees the existence of a surjective homomorphism ψ : V → hĀ, B̄, C̄, π̄0 i.
In the final section, we establish that hĀ, B̄, C̄, π̄0 i coincides with the group P3 (see Proposition 5.1), from which it follows that the natural inclusion map i0 is also surjective.
10
Amongst the generators for P3 are the elements a = (00 01) and b = (01 10 11), which satisfy
the relations of the symmetric group S4 (Relations R1). Of course, S4 also enjoys a presentation
in terms of transpositions involving only order and conjugacy relations. Accordingly, we apply
Tietze transformations to convert the presentation for P3 into one for a group P̄3 with generators
(00 01), (01 10), (10 11) and (1 00) and some order, conjugacy and split relations (specifically
translations of R2–R5, together with the new ones to replace R1). Thus we have on the one
hand, an isomorphism τ : P3 → P̄3 and, on the other, a surjective homomorphism θ : P̄3 →
hs00,01 , s01,10 , s10,11 , s1,00 i (a subgroup of P∞ ), since the relations defining P̄3 all hold in P∞ .
We can now deduce that P3 = hĀ, B̄, C̄, π̄0 i is not trivial, since successively composing the
appropriate maps in (2.12) sends, for example, a = (00 01) to the non-identity element t00,01
in V . Hence, from simplicity of V , we conclude P3 ∼
= V and therefore, subject to the work in
Sections 3–5, establish Theorem 1.2.
Finally, it is relatively straightforward to observe that if α is a non-empty sequence in X ∗ ,
then there is a product w involving only the swaps (00 01), (01 10), (10 11) and (1 00) such that
00 · w = α. From this, one quickly deduces, principally relying upon the conjugacy relations,
that one can conjugate s00,01 by some element of hs00,01 , s01,10 , s10,11 , s1,00 i to any generator sα,β
where (α, β) 6= (0, 1). This ensures that the inclusion i1 is surjective. Hence P∞ ∼
= V also and
we have established Theorem 1.1.
The remaining sections are perhaps technical, but carry out the deferred work just as described above. We hope that these sections will quickly impress the reader with the utility of
the permutation notation (α β) in performing calculations within R. Thompson’s group V .
Remarks
(i) If our goal had been simply to establish Theorem 1.1, then our work in the next two
sections would have been greatly reduced. Indeed, it is actually rather easy to show that
the group P∞ defined by all the relations listed in Theorem 1.1 is isomorphic to V (for
example, the relations that Dehornoy [13] lists are particularly straightforward to deduce).
However, from the viewpoint of the high transitivity of the action of V on the Cantor
set C, one expects that it would be enough to restrict to relations involving swaps (α β)
with |α| and |β| bounded. Indeed, this is what leads to our presentation for the group P3
and the small set of relations, R1–R5, where we are relying upon a small subset involving
only swaps (α β) with |α|, |β| 6 3. Establishing that this set of relations is sufficient is the
delicate business of Sections 3 and 4.
(ii) Now that we have established that the groups P3 , P∞ and V are all isomorphic, it is safe
to use the notation (α β) as a convenient notation for the map tα,β as defined earlier. We
can then perform computations within R. Thompson’s group V employing this notation,
for example, along the lines of those in the sections that follow, and we hope this will be
of use to those working with elements in this group.
3
Verification of relations to level 3
Our principal aim is to establish that all the relations holding in R. Thompson’s group V can
be deduced from Relations R1–R5. In this section, we complete the first stage of our technical
calculations by establishing essentially a subset of the infinitely many relations in the list (1.2),
namely those order relations, the conjugacy relations and the split relations involving only
swaps (α β) with |α|, |β| 6 3. It will turn out that these are enough to then deduce the fourteen
relations for V found in [12] as we shall see in Section 4.
Accordingly, in this section, we shall verify all relations of the form
(α β)2 = 1
11
(α β)(γ δ) = (α·(γ δ) β ·(γ δ))
(α β) = (α0 β0) (α1 β1)
whenever any swap (κ λ) appearing above satisfies |κ|, |λ| 6 3. (So, for example, the split
relation (α β) = (α0 β0) (α1 β1) needs only to be verified for |α|, |β| 6 2 in order that the swaps
on the right-hand side of the equation fulfil this requirement.) The main focus throughout the
section will be in establishing all the conjugacy relations σ τ = υ and we will consider these
relations when we select σ, τ and υ from the following sets:
T2 = { (α β) | (α, β) = (0, 1) or α, β ∈ X 2 }
T3 = { (α β) | α, β ∈ X 3 }
Tmn = { (α β) | α ∈ X m , β ∈ X n },
where 1 6 m < n 6 3
We start our verification by first noting that (1 00) is a conjugate of (00 01) by Relation R2
and hence has order dividing 2 by the first relation in R1. From this and the definitions of the
swaps, we now know that (α β)2 = 1 whenever α ⊥ β with |α|, |β| 6 3. We will use this fact
throughout the rest of this section.
We step through the various relations, essentially introducing “longer” swaps through the
stages. Accordingly, we start with relations involving only swaps from T12 ∪ T2 , then introducing
swaps from T13 , and so on. We need to take various side-trips from this general direction in
order to successfully establish all the relations we want.
Relations involving T12 and T2 only
With careful analysis, one can soon establish which conjugacy relations involve only swaps
selected from T12 ∪ T2 . These are, for x, y ∈ X, the following three relations:
(x x̄y)(0 1) = (x̄ xy)
(x̄y x̄ȳ)
(x x̄y)
(x x̄ȳ)
(x x̄y)
(3.1)
= (x x̄ȳ)
(3.2)
= (x̄y x̄ȳ)
(3.3)
These are actually very easy to verify. For example, Equation (3.1) simply follows from our
definition of (0 1x) (for x ∈ X) in (2.7), while we can establish Equation (3.2) using our
definition of (1 01) in (2.7) and then conjugating, if necessary, by (0 1) and using the now
established Equation (3.1). We now have the first step in our verification and results along the
lines of the following lemma will occur throughout our progress.
Lemma 3.1 All conjugacy relations of the form σ τ = υ where σ, υ ∈ T12 and τ ∈ T2 can be
deduced from Relations R1–R5.
This lemma contributes now to establishing Equation (3.3), since we observe that, for
the case x = 1 and y = 0, the equation is Relation R2 and that the general equation can
then be deduced by conjugating by a product of elements from T2 . Specifically, conjugating
by (0 1) gives (0 11)(0 10) = (10 11) and then subsequently the two equations now established
by (00 01) and (10 11), respectively, establishes the final cases.
Relations involving T12 , T13 and T2 only
When we introduce swaps from T13 , in addition to those from T12 and T2 , the relations that
need to be verified are, for x, y, z ∈ X, the following:
(x x̄yz)(0 1) = (x̄ xyz)
(x̄y x̄ȳ)
(x x̄yz)
12
= (x x̄ȳz)
(3.4)
(3.5)
(x x̄yz)(x x̄y) = (xz x̄y)
(3.6)
Both Equations (3.4) and (3.5) follow quickly from the definitions in (2.8). They establish:
Lemma 3.2 All conjugacy relations of the form σ τ = υ where σ, υ ∈ T13 and τ ∈ T2 can be
deduced from Relations R1–R5.
If we expand the definition of (1 00z) from Equation (2.8), we obtain (1 00z)(1 00) = (00 1z)
for z ∈ X. Now conjugate by an appropriate product of elements from T2 , using Lemmas 3.1
and 3.2 similarly to the argument used for Equation (3.3), to establish Equation (3.6).
First batch of relations involving T2 and T23 only
We now turn to relations involving swaps from T23 . There are additional relations that we
shall establish later involving only swaps from T2 and T23 , but for now we are concerned with
the following equations (in particular, the first of our split relations) for x, y ∈ X, for distinct
κ, λ, µ ∈ X 2 , and for any τ ∈ T2 for which the first is defined:
(κ λx)τ = (κ·τ (λx)·τ )
(x x̄y) = (x0 x̄y0) (x1 x̄y1)
(µ λx)
(κ λx)
= (κ µ)
(3.7)
(3.8)
(3.9)
We shall need the following lemma to be able to manipulate the initial swaps from T23 from
which the others are built, as given in (2.9). We shall then establish Equation (3.7) in the form
of Lemma 3.4 below.
Lemma 3.3
(i) (1 00), (10 000) and (11 001) all commute;
(ii) (01 00x)(1 00) = (01 1x);
Proof: (i) Using Relation R3 and the fact that all the swaps involved have order dividing 2, we
conclude that (10 000) and (11 001) commute. It then follows, again using R3, that (1 00) also
commutes with these elements.
(ii) Pull apart the formula for (01 00x) using the definitions in (2.9), and then apply Relation R2 and Equation (3.6) as follows:
(01 00x)(1 00) = (1 01x)(1 00) (00 01) (1 00) = (1 01x)(1 01) = (01 1x).
Lemma 3.4 All conjugacy relations of the form σ τ = υ where σ, υ ∈ T23 and τ ∈ T2 can be
deduced from Relations R1–R5.
Proof: The first half of the proof deals with the case when σ = (κ λ0) for distinct κ, λ ∈ X 2 .
To establish this, we must first verify that the swap (00 010) commutes with (10 11). However,
to achieve this, we actually work first with the swap (10 000). Indeed,
(10 000) (01 11) = (1 00) (11 001) (01 11)
by Rel. R3
= (1 00) (01 11) (01 001)
by the def. of (11 001) in (2.9)
= (01 001) (01 11) (1 00)
by Lem. 3.3(ii) twice
= (01 11) (11 001) (1 00)
by the def. in (2.9)
= (01 11) (10 000)
by Lem. 3.3(i) and Rel. R3.
13
As the definition of (10 000) in Equation (2.9) is (00 010)(00 01) (01 10) , we conclude that (00 010)
commutes with (01 11)(01 10) (00 01) = (10 11).
So we now turn to the required relation when σ = (κ λ0) for distinct κ, λ ∈ X 2 . As described
earlier, we view τ as a permutation of X 2 . As such a map, suppose τ maps κ and λ to µ and ν,
respectively. Then by the definition in Equation (2.9) there are products ρ1 and ρ2 of elements
from T2 such that
(κ λ0) = (00 010)ρ1
and
(µ ν0) = (00 010)ρ2 .
Specifically, ρ1 and ρ2 are products that, when viewed as permutations of X 2 , map 00 and 01
to κ and λ and to µ and ν, respectively. Then ρ1 τ ρ−1
2 ∈ h(10 11)i since it fixes both 00 and 01.
Hence, by our previous calcuation, (00 010) commutes with ρ1 τ ρ−1
2 , so
(κ λ0)τ = (00 010)ρ1 τ = (00 010)ρ2 = (µ ν0).
(3.10)
Thus, we have established Equation (3.7) in the case when x = 0.
To deduce the equation for x = 1, we proceed similarly and first need to establish that
(00 011) commutes with (10 11). We calculate as follows:
(10 000)(01 000) = (10 000)(1 00) (01 10) (1 00)
(01 10) (1 00)
by Lem. 3.3(ii)
= (10 000)
by Lem. 3.3(i)
= (01 000)(1 00)
by Eq. (3.10)
= (01 10)
by Lem. 3.3(ii).
Conjugate by (01 10) and use Equation (3.10) again to establish the formula (01 000)(10 000) =
(01 10). Now we find
(11 001) (01 10) = (10 000) (1 00) (01 10)
by Rel. R3
= (10 000) (01 000) (1 00)
by Lem. 3.3(ii)
= (01 10) (10 000) (1 00)
as just established
= (01 10) (11 001)
by Rel. R3.
Thus [(11 001), (01 10)] = 1, and then using the definition of (11 001) in (2.9) and the Relations R1, we deduce [(00 011), (10 11)] = 1. We then proceed as in the first half of the proof,
using this new equation, and we establish Equation (3.7) when x = 1, completing the proof of
the lemma.
Equation (3.8) now follows by conjugating Relation R3 by an appropriate product of elements
from T2 using Lemmas 3.1 and 3.4.
The establishment of Equation (3.9) requires an intermediate observation first. We use
Lemma 3.4 to tell us that (11 001) commutes with (01 10), so
(01 10) (11 001) = (11 001) (01 10)
= (10 000) (1 00) (01 10)
by Rel. R3
= (10 000) (01 000) (1 00)
by Lem. 3.3(ii)
= (10 000) (01 000) (10 000) (11 001)
by Rel. R3.
Hence (01 000)(10 000) = (01 10). By a similar sequence of calculations we establish
(10 000) (01 11) = (01 11) (10 000)
by Lem. 3.4
= (01 11) (1 00) (11 001)
by Rel. R3
= (1 00) (01 001) (11 001)
by Lem. 3.3(ii)
= (10 000) (11 001) (01 001) (11 001)
by Rel. R3,
so (01 001)(11 001) = (01 11). We now have two equations that, once we conjugate by a product
of elements from T2 , yield the general form of Equation (3.9) using Lemma 3.4 and Relations R1.
14
Relations involving T12 , T2 and T23
The relations that involve swaps from T12 , T23 and T2 and that definitely include at least one
swap from each of the first two sets are, for x, y, z, t ∈ X, the following:
(xy x̄z)(x x̄z̄) = (x̄z x̄z̄y)
(x x̄z)
(xy x̄zt)
(3.11)
= (xt x̄zy)
(3.12)
To establish Equation (3.11), first calculate
(00 1x)(1 01) (1 00) = (00 1x)(1 00) (00 01) = (1 00x)(00 01) = (1 01x)
using Relation R2 and the definitions in (2.8). Hence
(00 1x)(1 01) = (1 01x)(1 00) = (00 01x)
by the definitions in (2.9). This is Equation (3.11) in the case when x = 1 and z = 0. The
general equation then follows by conjugating by a suitable product of elements from T2 and
using Lemmas 3.1 and 3.4.
To establish Equation (3.12), we first deal with the case when x = 1 and z = 0. We calculate
(1y 00t)(1 00) = (01 00t)(01 1y) (1 00) = (01 00t)(1 00) (01 00y)
= (01 1t)(01 00y) = (1t 00y)
using the definitions in (2.9), Equation (3.11) (twice) and (3.9). Now conjugate by an appropriate
product of elements from T2 , using Lemmas 3.1 and 3.4, to conclude that Equation (3.12) holds.
Intermediate relations
Our current goal at this stage is to complete the establishment of relations involving swaps from
T2 and T23 to supplement Equations (3.7)–(3.9) already obtained. However, in order to achieve
this, we need some intermediate relations involving swaps from T3 of the form (κ0 κ1) for some
κ ∈ X 2 , specifically for x, y ∈ X and distinct κ, λ ∈ X 2 :
(κ λx)(κ λx̄) = (λ0 λ1)
(λ0 λ1)
(κ λx)
(3.13)
= (κ λx̄)
(3.14)
[(x x̄y), (x̄ȳ0 x̄ȳ1)] = 1
(3.15)
We start by establishing a helpful lemma, analogous to Lemmas 3.3 and 3.4, concerning the
swaps of the form (κ0 κ1).
Lemma 3.5
(i) (000 001)(1 00) = (10 11);
(ii) (000 001) commutes with (01 10), with (01 11), and with (10 11).
(iii) All conjugacy relations of the form σ τ = υ, where σ, υ ∈ T3 have the form (κ0 κ1) for
some κ ∈ X 2 and τ ∈ T2 , can be deduced from Relations R1–R5.
Proof: (i) This follows using the definitions in (2.8) and (2.10):
(000 001)(1 00) = (1 000)(1 001) (1 00) = (1 000)(1 00) (00 11)
= (00 10)(00 11) = (10 11).
15
(ii) First we know, by Lemma 3.4, that (10 11) commutes with (01 00x) for any x ∈ X.
Conjugating by (1 00), using part (i) and Lemma 3.3(ii), establishes that (000 001) commutes
with (01 1x).
Then we perform the following calculation:
(000 001) (10 11) = (1 00) (10 11) (1 00) (10 11)
by part (i)
= (10 000) (11 001) (10 11) (1 00) (10 11)
by Rel. R3
= (10 11) (11 000) (10 001) (1 00) (10 11)
by Lem. 3.4
= (10 11) (1 00) (10 001) (11 000) (10 11)
by Eq. (3.12)
= (10 11) (1 00) (10 11) (11 001) (10 000)
by Lem. 3.4
= (10 11) (1 00) (10 11) (1 00)
by Lem. 3.3(i) and Rel. R3
= (10 11) (000 001)
by part (i).
Thus (000 001) commutes with (10 11).
(iii) This follows by the same argument as used in Lemma 3.4, noting that if a product ρ of
elements from T2 fixes 00 then it lies in the subgroup generated by (01 10), (01 11) and (10 11)
and so commutes with (000 001) by part (ii).
For Equation (3.13), start with the equations (01 1x)(01 1x̄) = (10 11) and then conjugate
by (1 00). Use Lemmas 3.3(ii) and 3.5(i) to conclude
(01 00x)(01 00x̄) = (000 001)
for any x ∈ X. The required equation now follows by conjugating by a product of elements
from T2 and using Lemmas 3.4 and Lemma 3.5(iii).
To establish Equation (3.14), first use Lemma 3.4 to conclude that (10 000)(10 11) = (11 000).
Now conjugate by (1 00) and use Equation (3.12) and Lemma 3.5(i) to establish the formula
(10 000)(000 001) = (10 001). Conjugating by an appropriate product of elements from T2 and
use of Lemmas 3.4 and 3.5(iii) establishes (κ λ0)(λ0 λ1) = (κ λ1), which is sufficient to verify
Equation (3.14).
We deduce Equation (3.15) by starting with [(00 01), (10 11)] = 1 and conjugating by (1 00)
to yield [(1 01), (000 001)] = 1, using Relation R2 and Lemma 3.5(i). Conjugating by an
appropriate product of elements from T2 and using Lemmas 3.1 and 3.5(iii) establishes the
required relation.
Remaining relations involving T2 and T23
We now establish all the relations remaining that involve just swaps from T2 and T23 . Careful
analysis shows that the ones we are currently missing are, for x, y ∈ X and distinct κ, λ, µ, ν ∈
X 2 , the following:
[(κ λ0), (µ λ1)] = 1
(3.16)
[(κ λx), (µ νy)] = 1
(3.17)
Note that in establishing these equations we are essentially establishing that swaps from T23 that
have disjoint support (or, more accurately, corresponding to maps in V with disjoint support)
commute.
Equation (3.16) simply follows from Lemma 3.3(i) using Lemma 3.4.
Use of Lemma 3.4 deduces [(κ λ0), (µ ν1)] = [(κ λ1), (µ ν1)] = 1 for all distinct κ, λ, µ, ν ∈
X 2 from Relations R4. Consequently, in the case of Equation (3.17), it remains to verify the
relation in the case when x = y = 0. First observe that
(11 011)(000 001) = (11 011)(10 000) (10 001) (10 000) = (11 011)
16
by use of Equation (3.13) and then repeated use of the cases of Equation (3.17) that we already
have; that is, [(11 011), (000 001)] = 1. Now
(000 001)(10 010) = (000 001)(1 01) (11 011)
by Eq. (3.8)
= (000 001)(11 011)
by Eq. (3.15)
= (000 011)
as just established.
Thus, (10 010) and (000 001) commute. This means that when we conjugate the relation
[(10 010), (11 001)] = 1, which is an instance of Equation (3.17) that we already know, by the
swap (000 001), we obtain
[(10 010), (11 000)] = 1,
with use of Equation (3.14). We now make use of Lemma 3.4 in our usual way to establish the
missing case of Equation (3.17), namely when x = y = 0.
Relations involving T12 , T13 , T2 and T23
We now establish all the relations we require that involve swaps from T12 , T13 , T2 and T23 . In
view of the relations that we have already obtained, we can assume that at least one swap
from T13 and at least one from T23 occur within our relation. The relations we need to establish
are therefore, for x, y, z ∈ X, the following:
[(x x̄yz), (x̄ȳ x̄y z̄)] = 1
(x x̄ȳz)
(x x̄y)
(3.18)
= (x̄y x̄ȳz)
(3.19)
(x x̄y)(x̄y x̄ȳz) = (x x̄ȳz)
(x x̄ȳ)
(x x̄yz)
(3.20)
= (x̄ȳ x̄yz)
(3.21)
For Equation (3.18), take the equation [(00 10), (01 11)] = 1, conjugate by (1 00) and use the
definition in (2.8) and Lemma 3.3(ii) to conclude [(1 000), (01 001)] = 1. The required equation
then follows, as usual, by use of Lemmas 3.2 and 3.4.
For Equation (3.19), we calculate as follows:
(1 00)(1 01z) = (1 00)(00 01) (1 00z) (00 01)
by the definition in (2.8)
(1 00z) (00 01)
by the definition in (2.7)
(1 00) (00 1z) (1 00) (00 01)
by the definition in (2.8)
= (1 01)
= (1 01)
= (01 1z)(1 00) (00 01)
using Rels. R2 and R1
(00 01)
= (01 00z)
by Eq. (3.11)
= (00 01z)
by Lem. 3.4.
The required equation now follows using Lemmas 3.1, 3.2 and 3.4.
For Equation (3.20), our main calculation is
(1 00)(00 01z) = (1 00)(1 00) (1 01z) (1 00) = (1 00)(1 01z) (1 00)
= (00 01z)(1 00) = (1 01z),
obtained by exploiting the definition of (00 01z) in (2.9) and Equation (3.19) above. The required
equation again follows by Lemmas 3.1, 3.2 and 3.4.
Equation (3.21) follows from the definition of (1 01z) as in Equation (2.8) with use of
Lemmas 3.1, 3.2 and 3.4.
17
First relations involving T2 and T3 only
We now introduce swaps from T3 into the relations we verify. The first step is to establish that
swaps from T3 behave well when we conjugate by one from T2 and then our other split relation.
All other relations involving just swaps from T2 and T3 will have to wait until we have established
some more intermediate relations. Accordingly, we start with the following for x, y ∈ X, distinct
κ, λ ∈ X 2 and τ ∈ T2 for which Equation (3.22) is defined:
(κx λy)τ = ((κx)·τ (λy)·τ )
(3.22)
(κ λ) = (κ0 κ1) (λ0 λ1)
(3.23)
As with those from T23 , we begin with a lemma to manipulate the basic T3 -swaps from the
definition in (2.10). In particular, we will establish Equation (3.22) as part (iv) in the lemma.
Lemma 3.6
(i) (000 01x)(1 00) = (10 01x), for any x ∈ X;
(ii) (001 011)(1 00) = (11 011);
(iii) (000 010), (000 011) and (001 011) each commute with (10 11).
(iv) All conjugacy relations of the form σ τ = υ, where σ, υ ∈ T3 and τ ∈ T2 , can be deduced
from Relations R1–R5.
(v) (001 010)(1 00) = (11 010);
Proof: (i) We calculate as follows:
(000 01x)(1 00) = (1 000)(1 01x) (1 00)
by the definition in (2.10)
= (1 000)(1 00) (00 01x)
(00 01x)
by Eq. (3.21)
= (00 10)
by the definition in (2.8)
= (10 01x)
by Eq. (3.9).
Part (ii) is established in exactly the same way.
(iii) We perform the following calculations:
(000 01x)(10 11) (1 00)
= (10 01x)(1 00) (10 11) (1 00)
by part (i)
(10 000) (11 001) (10 11) (1 00)
by Rel. R3
(11 000) (10 001) (1 00)
= (11 01x)
by Lem. 3.4
= (11 01x)(11 000) (1 00)
by Eq. (3.17) twice
= (10 01x)
(11 000) (10 000) (11 001)
= (11 01x)
by Rel. R3
= (11 01x)(10 000) (10 11) (11 001)
by Eq. (3.9)
= (11 01x)(10 11) (11 001)
by Eq. (3.17)
(11 001)
= (10 01x)
by Lem. 3.4
= (10 01x)
by Eq. (3.17).
Thus (000 01x)(10 11) = (10 01x)(1 00) = (000 01x), again by part (i). A similar argument,
using (ii), shows that (001 011) commutes with (10 11).
(iv) When σ (and υ) has the form (κ0 κ1), this result was established in Lemma 3.5(iii).
All remaining swaps in T3 are defined by conjugating one of (000 010), (000 011) or (001 011)
18
by some product of elements from T2 . The result then follows by the same argument, but now
relying upon part (iii).
(v) appears to be similar to the first two parts of the lemma, but actually requires a different
argument, based on what we have just established:
(001 010)(1 00) = (000 011)(00 01) (1 00) = (000 011)(1 00) (1 01)
= (10 011)(1 01) = (11 010),
by part (iv), Relation R2, part (i) and Equation (3.12).
Equation (3.23) is now established by taking the equation (1 01) = (10 010) (11 011), which
is an instance of Equation (3.8), and conjugating by (1 00) to yield
(00 01) = (000 010) (001 011),
using Relation R2 and Lemma 3.6(i) and (ii). The required relation then follows by conjugating
by a product of elements from T2 and using the Relations R1 and Lemma 3.6(iv).
Further intermediate relations involving T23 and T3
Our principal direction of travel at this stage is to complete the verification of those relations
involving swaps from T2 and T3 to supplement those in Equations (3.22) and (3.23). However,
to achieve this we need some intermediate results making use of swaps from T23 , specifically, for
x ∈ X, distinct κ, λ, µ ∈ X 2 and distinct α, β ∈ X 3 for which κ ⊥ α, β, the following:
(κ α)(κ β) = (α β)
(α β)
(κ α)
(3.24)
= (κ β)
(3.25)
[(κ λx), (µ0 µ1)] = 1
(3.26)
For Equation (3.24), first note that Equation (3.13) deals with the case when α and β share
the same two-letter prefix. For the remaining cases, first observe
(10 000)(10 010) (1 010)
= (10 000)(10 010) (11 011) (1 010)
by Eqs. (3.16) and (3.17)
= (10 000)(1 01) (1 010)
by Eq. (3.8)
= (10 000)(01 10) (1 01)
by Eq. (3.6)
(1 01)
= (01 000)
by Lem. 3.4
= (1 000)
by Eq. (3.21).
Hence (10 000)(10 010) = (1 000)(1 010) = (000 010), using the definition in Equation (2.10). A
similar argument establishes (11 000)(11 011) = (1 000)(1 011) = (000 011). Equally we apply a
variant of the argument to obtain further equations:
(10 011)(10 000) (1 011) = (10 011)(1 00) (1 011)
(00 011) (1 00)
arguing as before
= (10 011)
by Eq. (3.21)
= (00 10)(1 00)
by Eq. (3.9)
= (1 000)
by the definition in (2.8).
Thus we obtain (10 011)(10 000) = (1 000)(1 011) = (000 011). Similarly we determine the equation
(11 011)(11 001) = (001 011). Thus given any choice of x, y ∈ X, we have obtained one example
of a relation
(κ λx)(κ µy) = (λx µy)
19
(for some choice of distinct κ, λ and µ.) We can then obtain all examples by use of Lemmas 3.4
and 3.6(iv). This completes the establishment of Equation (3.24).
Equation (3.14) is already Equation (3.25) in the case when α and β share the same twoletter prefix. For the remaining cases, use Equation (3.24) to tell us (1x 000)(1x 01y) = (000 01y)
for any x, y ∈ X. Conjugate this equation by (1 00) and use Equation (3.12) and parts of
Lemma 3.6 to establish (10 00x)(00x 01y) = (10 01y). All cases of Equation 3.25 now follow by
conjugating by an appropriate product of elements from T2 .
For Equation (3.26), start with the fact that (10 11) commutes with (000 01x) for any x ∈
X (Lemma 3.6(iii)). Conjugate by (1 00) and use Lemmas 3.5(i) and 3.6(i) to conclude
[(000 001), (01 01x)] = 1. Then conjugating by a product of swaps from T2 establishes the
required equation.
Remaining relations involving only T2 , T23 and T3
Having established the intermediate relations, we can now establish all remaining relations
involving swaps only from T2 and T3 . We obtain those also involving swaps from T23 at the same
time. When one analyses the relations required, we find that they are, for x, y ∈ X, distinct
κ, λ ∈ X 2 and distinct α, β, γ, δ ∈ X 3 , the following:
[(κ0 κ1), (λx µy)] = 1
(κ0 κ1)
(κx λy)
(3.27)
= (κx̄ λy)
(3.28)
[(κ α), (β γ)] = 1
for κ ⊥ α, β, γ
(α β)(α γ) = (β γ)
(3.29)
(3.30)
[(α β), (γ δ)] = 1
(3.31)
We establish Equation (3.27) by choosing ν to be the element in X 2 \ {κ, λ, µ}, using Equation (3.24) to tell us (λx µy) = (ν λx)(ν µy) and then, as Equation (3.26) says that both (ν λx)
and (ν µy) commute with (κ0 κ1), we establish Equation (3.27).
By Lemma 3.4, (1x 01y)(10 11) = (1x̄ 01y) for any x, y ∈ X. Conjugate by (1 00) and
use Lemma 3.5(i) and parts of Lemma 3.6 to conclude (00x 01y)(000 001) = (00x̄ 01y). Equation (3.28) then follows.
Our final equation involving swaps from T23 and T3 is Equation (3.29). We need to establish
this in a number of stages. If x ∈ X, we know that (10 11) commutes with the swap (000 01x) by
Lemma 3.6(iii). If we conjugate by (1 00) and use Lemma 3.5(i) and Lemma 3.6(i), we conclude
[(10 01x), (000 001)] = 1. We then deduce Equation (3.29) in the case when β and γ share the
same two-letter prefix in our now established manner.
The second case of the equation is when α and β share their two-letter prefix. Equation (3.17)
tells us that [(1x 000), (1x̄ 01y)] = 1 for any x, y ∈ X. Now conjugate by (1 00) and use
Equation (3.12) and parts of Lemma 3.6 to conclude [(10 00x), (00x̄ 01y)] = 1. Equation (3.29)
when α and β share their two-letter prefix then follows.
It remains to deal with the case when α, β and γ have distinct two-letter prefixes. One
particular case is our Relation R5: [(10 110), (000 010)] = 1. Conjugating by (110 111), using
Equation (3.14) and (3.27), we deduce [(10 111), (000 010)] = 1. Similarly, conjugating what we
now have, [(10 11x), (000 010)] = 1 for any x ∈ X, by (000 001) and use (3.26) and (3.28) to
now conclude that [(10 11x), (00y 010)] = 1 for all x, y ∈ X. Finally use the same argument,
conjugating by (010 011), to conclude [(10 11x), (00y 01z)] = 1 for all x, y, z ∈ X. The remaining
case of Equation (3.29) now follows.
One case of Equation (3.30), namely when α and γ share the same two-letter prefix, has
already been established as Equation (3.28). For the case when α and β share the same twoletter prefix, start with the equation (10 11)(1x 01y) = (1x̄ 01y), for x, y ∈ X, as given by
Equation (3.9). Conjugate by (1 00) and use Lemma 3.5(i) and parts from Lemma 3.6 to
20
conclude (000 001)(00x 01y) = (00x̄ 01y). From this the general formula (κ0 κ1)(κx λy) = (κx̄ λy)
follows for distinct κ, λ ∈ X 2 .
For the case when α, β and γ have distinct two-letter prefixes, say α = κx, β = λy and
γ = µz, choose ν to be the other element of X 2 . Then
(α β)(α γ) = (κx λy)(κx µz) = (κx λy)(ν κx) (ν µz) (ν κx)
= (ν λy)(ν µz) (ν κx)
= (λy µz)(ν κx) = (λy µz) = (β γ),
by Equations (3.24) (used three times) and (3.29).
Part of Equation (3.31) has already been established as Equation (3.27), but we shall deal
with our required relation in full generality. Indeed, first assume that α, β, γ and δ have between
them at most three distinct two-letter prefixes. Let ν ∈ X 2 be different from those two-letter
prefixes. Then write (γ δ) as (ν γ) (ν δ) (ν γ), by Equation (3.24), and observe this commutes
with (α β) using Equation (3.29).
The case when α, β, γ and δ have four distinct two-letter prefixes can then be deduced as
follows. Suppose as α = κx for some κ ∈ X 2 and x ∈ X. By the previous case, (α β) commutes
with both (κx̄ γ) and (κx̄ δ), and hence it also commutes with (γ δ) = (κx̄ γ)(κx̄ δ) , using
Equation (3.30).
Relations involving T3 , at least one of T12 , T13 and T23 , and possibly T2
We now establish the final batch of relations for this section. These involve swaps from T3 and
at least one of T12 , T13 and T23 , and are the following for x, y, z, t ∈ X and distinct α, β, γ ∈ X 3
satisfying x 6≺ α, β, γ:
(x̄y0 x̄y1)(x x̄y) = (x0 x1)
[(x α), (β γ)] = 1
(α β)
(x α)
(3.32)
(3.33)
= (x β)
(3.34)
(α β)(x α) = (x β)
(3.35)
(x̄yz x̄ȳt)(x x̄y) = (xz x̄ȳt)
(3.36)
Equation (3.32) follows from Lemma 3.5(i) by our standard T2 -conjugation argument.
Equation (3.33) follows by taking the relation [(00 1y), (1ȳ 01z)] = 1 and the relation
[(00 1y), (010 011)] = 1, which hold by Lemmas 3.4 and 3.6(iv) respectively, and then conjugating by (1 00) and proceeding as in previous arguments to conclude that [(x κy), (κȳ λz)] = 1
and [(x κy), (λ0 λ1)] = 1 for any x, y, z ∈ X and any distinct κ, λ ∈ X 2 with x 6≺ κ, λ.
To establish Equation (3.34), conjugate the already established equations (00 1y)(10 11) =
(00 1ȳ) and (00 1y)(1y 01z) = (00 01z), for y, z ∈ X, by (1 00). This yields (1 00y)(000 001) =
(1 00ȳ) and (1 00y)(00y 01z) = (1 01z), which now yields the two forms of Equation (3.34):
(x κy)(κ0 κ1) = (x κȳ) and (x κy)(κy λz) = (x λz) for any x, y, z ∈ X and distinct κ, λ ∈ X 2 with
x 6≺ κ, λ.
For Equation (3.35), we shall show (000 001)(1 00x) = (1 00x̄) and (00x 01y)(1 01y) = (1 00x)
for any x, y ∈ X. Four of these occurrences are found in the definitions in (2.10), while the other
two are deduced by conjugating (10 11)(00 10) = (00 11) and (11 010)(00 010) = (00 11) by (1 00).
The required equation then follows.
Finally, for Equation (3.36), from Lemma 3.6: (00z 01t)(1 00) = (1z 01t) for any z, t ∈ X.
Then as in the previous equations we conjugate by products of elements from T2 .
We have now established all required relations of the form σ τ = υ where σ, τ and υ come
from the sets T2 , T12 , T13 , T23 and T3 . This is the first stage in establishing the existence of the
homomorphism ψ in Diagram (2.12) and, in particular, the verification of Theorem 1.2.
21
4
Verifying the Cannon–Floyd–Parry relations
In this section we describe how to verify that all the relations that hold in R. Thompson’s
group V can be deduced from those assumed in Relations R1–R5. We shall rely upon the work
in the previous section. One might wonder whether it is possible to proceed more directly to
show, for example, that all relations holding in V can be deduced from the infinitely many in
the presentation in Theorem 1.1. It is a consequence of our results that this can be done, but
our own attempt to do so resulted in overly long arguments replicating those already found in
Section 6 of [12]. We have chosen the more direct method of verifying the finite set of relations
known already to define V .
In their paper (see [12, Lemma 6.1]), Cannon–Floyd–Parry provide the following presentation
for V . It has generators A, B, C and π0 and relations
CFP1.
CFP2.
CFP3.
CFP4.
CFP5.
CFP6.
CFP7.
[AB −1 , X2 ] = 1;
[AB −1 , X3 ] = 1;
C1 = BC2 ;
C2 X2 = BC3 ;
C1 A = C22 ;
C13 = 1;
π12 = 1;
CFP8.
CFP9.
CFP10.
CFP11.
CFP12.
CFP13.
CFP14.
π1 π3 = π3 π1 ;
(π2 π1 )3 = 1;
X3 π1 = π1 X3 ;
π1 X2 = Bπ2 π1 ;
π2 B = Bπ3 ;
π 1 C3 = C3 π 2 ;
(π1 C2 )3 = 1;
where the elements appearing here are defined by the following formulae Cn = A−n+1 CB n−1 ,
Xn = A−n+1 BAn−1 both for n > 1, π1 = C2−1 π0 C2 and πn = A−n+1 π1 An−1 for n > 2.
Recall from Section 2 that P3 is the group presented by generators a = (00 01), b =
(01 10) (01 11) and c = (1 00) subject to relations R1–R5. Define four new elements of P3 by
Ā = (0 1) (0 10) (10 11);
B̄ = (10 11) (10 110) (110 111);
C̄ = (10 11) (0 10);
π̄0 = (0 10),
and then new elements C̄n , X̄n and π̄n for n > 1 defined in terms of these four by same formulae
used when defining the relations for V .
It is a consequence of the relations involving swaps from the sets T2 , T12 , T13 , T23 and T3
established in the previous section (i.e., Equations (3.1)–(3.36)) that the elements Ā, B̄, C̄ and
π̄0 satisfy CFP1–CFP14. To verify this is a sequence of calculations. Below we present the verification of CFP1 for these elements. For the entertainment of the reader, Relation CFP2 required
the longest calculation whilst the others are more straightforward. The following formulae are
useful for this work.
Lemma 4.1 The following formulae hold in G:
(i) ĀB̄ −1 = (00 01) (01 10) (0 10);
(ii) X̄2 = (0 11) (00 01) (00 010) (010 011) (0 11);
(iii) X̄3 = (0 111) (00 01) (00 010) (010 011) (0 111);
(iv) C̄2 = (0 10) (0 111) (110 111);
(v) C̄3 = (0 110) (10 111) (0 100) (0 101) (10 110) (110 111);
(vi) π̄1 = (10 110);
(vii) π̄2 = (0 11) (00 010) (0 11);
(viii) π̄3 = (0 111) (00 010) (0 111).
22
Proof: We verify the two formulae, (i) and (ii), required to verify CFP1. Below, we principally
rely upon the conjugacy relations, although a split relation is applied in one step. The other
formulae listed are established similarly.
(i) We calculate
ĀB̄ −1 = (0 1) (0 10) (10 11) · (110 111) (10 110) (10 11)
= (0 1) (0 10) (100 101) (11 100)
= (0 1) (00 01) (00 11) (0 10)
= (00 10) (01 11) (00 01) (00 11) (0 10)
= (00 01) (01 10) (0 10).
(ii) We start with the definition of Ā and B̄:
X̄2 = Ā−1 B̄ Ā
= (10 11) (0 10) (0 1) · (10 11) (10 110) (110 111) · (0 1) (0 10) (10 11)
= (10 11) (0 10) (00 01) (00 010) (010 011) (0 10) (10 11)
= (0 11) (00 01) (00 010) (010 011) (0 11).
We now verify that the elements Ā, B̄, C̄ and π̄0 of our P3 satisfy the relation CFP1:
[ĀB̄ −1 , X̄2 ] = (0 10) (01 10) (00 01) · (0 11) (010 011) (00 010) (00 01) (0 11)
· (00 01) (01 10) (0 10) · (0 11) (00 01) (00 010) (010 011)
· (0 11)
= (0 11) (10 11) (10 111) (110 111) (010 011) (00 010) (00 01)
· (0 11) (00 01) (01 10) (0 10) (0 11) (00 01) (00 010)
· (010 011) (0 11)
= (0 11) (10 11) (10 111) (110 111) (010 011) (00 010) (00 01)
· (110 111) (10 111) (10 11) (00 01) (00 010) (010 011) (0 11)
= (0 11) (11 101) (100 101) (010 011) (00 010) (00 01) (100 101)
· (11 101) (00 01) (00 010) (010 011) (0 11)
= (0 11) (11 101) (100 101) (010 011) (00 010) (100 101)
· (11 101) (00 010) (010 011) (0 11)
= (0 11) (11 101) (010 011) (00 010) (11 101) (00 010) (010 011)
· (0 11)
= (0 11) (11 101) (010 011) (11 101) (010 011) (0 11)
=1
(by first collecting (0 11) to the left, then conjugating some swaps by (0 11), some by (10 11),
some by (00 01), some by (100 101), then single swaps by (00 010) and by (11 101), and finally
exploiting the fact our swaps have order 2).
Once we have established the fourteen relations CFP1–CFP14, it follows that there is indeed
a surjective homomorphism ψ : V → hĀ, B̄, C̄, π̄0 i, as indicated in the Diagram (2.12) and used
in the proof of Theorems 1.1 and 1.2.
5
Final details for the proofs
In this section, we complete the technical details relied upon in the proofs given in Section 2.
23
We first need to show that the subgroup hĀ, B̄, C̄, π̄0 i coincides with the group P3 . Indeed
observe this subgroup contains all the following elements:
π̄0 = (1 00) = c
C̄ π̄0 = (10 11)
(C̄ π̄0 )ĀC̄ = (10 11)(0 1) = (00 01) = a
π̄0B̄
π̄0B̄
−1 C̄
−1 C̄
= (0 10)(110 111) (10 110) (0 10) = (10 110)
C̄ π̄0 B̄ = (110 111)
and
(10 110)(110 111) (10 11) (0 10) (10 11) = (01 10).
The Relations R1 ensure that b ∈ h(00 01), (01 10), (10 11)i, and so we now conclude that this
subgroup generated by Ā, B̄, C̄ and π̄0 is the whole group P3 . This establishes the following result
and means that we have now completed all the details required for the proofs of Theorem 1.1
and 1.2.
Proposition 5.1 The elements Ā, B̄, C̄ and π̄0 , defined earlier, generate the group P3 .
Finally, we deduce a 2-generator presentation for V from Theorem 1.2. As noted in Section 2,
the following is the intermediate step used to establish Theorem 1.3. Although the latter depends
on computer calculation, the following is established by purely theoretical methods in line with
our proofs of Theorems 1.1 and 1.2.
Corollary 5.2 R. Thompson’s group V has a finite presentation with two generators and nine
relations.
Proof: We work in the group P3 . Define u and v to be the elements
u = (00 01) (10 110) (10 111)
and
v = b = (01 10 ).
(When interpreted via the isomorphism from P3 to V , obtained by composing the maps specified
in the diagram (2.12), these two elements correspond to the element of V given by tree-pairs in
Figure 3 in the Introduction.)
If we rely upon the relations that hold in P3 (i.e., simply calculating within R. Thompson’s
group V , as, by this stage, we have completed all steps in establishing Theorem 1.2), then we
can obtain a formula for u as a product of the generators a, b and c, for example,
b ab−1 a
u = w(a, b, c) = a(ab ab )caca
−1
and the following formulae:
a = (00 01) = u3 ,
(10 000) = (u3 )vu
−2 vu3
,
vu−1 vu3 v
(11 001) = (u3 )
3
3
Therefore c = γ(u, v) = (u3 )vu vu (u3 )vu vu v . If r1 (a, b, c), r2 (a, b, c), . . . , r8 (a, b, c) denote
the words in a, b and c that define our relations (see the Equations (2.4) in Section 2, or
alternatively, Equations (2.3)), then applying Tietze transformations shows that
V ∼
= u, v r1 u3 , v, γ(u, v) = r2 u3 , v, γ(u, v) = · · · =
r8 u3 , v, γ(u, v) = 1, u = w u3 , v, γ(u, v) .
−2
−1
This establishes the corollary.
24
As we described in Section
2, Theorem 1.3 is deduced from this presentation. We produce
the formulae ri u3 , v, γ(u, v) , for i = 1, 2, . . . , 8, by taking ri to be the formulae in (2.4).
We then apply the process of producing equivalent relations for this 2-generator presentation
by repeatedly using the Knuth–Bendix Algorithm as described in Section 2. It is during this
process that we discover that two of the relations can be omitted since the relevant word reduces
to the identity. The remaining seven relations reduce to those listed in Theorem 1.3.
Acknowledgements: The authors wish to acknowledge partial support by EPSRC grant
EP/H011978/1. We further wish to recognise our gratitude to the authors of the computer
packages that greatly assisted when we performed calculations and also helped to reduce our
presentations: GAP, KBMAG and the Java VTrees applet written by the first author and Roman
Kogan.
References
[1] James Belk and Bradley Forrest, “Rearrangement groups of fractals,” 2015 (preprint).
arXiv:1510.03133
[2] Jean-Camille Birget, “Circuits, the groups of Richard Thompson, and coNP-completeness,”
Internat. J. Algebra Comput. 16 (2006), no. 1, 35–90.
[3] Jean-Camille Birget, “Factorizations of the Thompson–Higman groups, and circuit complexity,” Internat. J. Algebra Comput. 18 (2008), no. 2, 285–320.
[4] Collin Bleak, Francesco Matucci and Max Neunhöffer, “Embeddings into Thompson’s
group V and coCF groups,” J. Lond. Math. Soc. (2) 94 (2016), no. 2, 583–597.
[5] Collin Bleak and Olga, Salazar-Dı́az, “Free products in R. Thompson’s group V ,” Trans.
Amer. Math. Soc. 365 (2013), no. 11, 5967–5997.
[6] Matthew G. Brin, “Higher dimensional Thompson groups,” Geom. Dedicata 108 (2004),
163–192.
[7] Matthew G. Brin, “On the baker’s map and the simplicity of the higher dimensional Thompson groups nV ,” Publ. Math. 54 (2010), no. 2, 433–439.
[8] Kenneth S. Brown, “Finiteness properties of groups,” J. Pure Appl. Algebra 44 (1987),
no. 1–3, 45–75.
[9] Kenneth S. Brown and Ross Geoghegan, “An infinite-dimensional torsion-free FP∞ group,”
Invent. Math. 77 (1984), no. 2, 367–381.
[10] Marc Burger and Shahar Mozes, “Finitely presented simple groups and products of trees,”
C. R. Acad. Sci. Paris Sér. I Math. 324 (1997), no. 7, 747–752.
[11] Marc Burger and Shahar Mozes, “Groups acting on trees: from local to global structure,”
Inst. Hautes Études Sci. Publ. Math. No. 92 (2001), 113–150.
[12] J. W. Cannon, W. J. Floyd & W. R. Parry, “Introductory notes on Richard Thompson’s
groups,” Enseign. Math. (2) 42 (1996), no. 3–4, 215–256.
[13] Patrick Dehornoy, “Geometric presentations for Thompson’s groups,” J. Pure Appl. Algebra 203 (2005), no. 1–3, 1–44.
[14] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.7.8, 2015,
http://www.gap-system.org.
25
[15] Graham Higman, Finitely presented infinite simple groups, Notes on Pure Mathematics,
No. 8 (1974), Department of Pure Mathematics, Department of Mathematics, I.A.S. Australian National University, Canberra, 1974.
[16] Derek Holt, KBMAG, Knuth–Bendix on Monoid and Automatic Groups, Version 1.5, 2009,
http://www.warwick.ac.uk/~mareg/kbmag.
[17] Dessislava H. Kochloukova, Conchita Martı́nez-Pérez and Brita E. A. Nucinkis, “Cohomological finiteness properties of the Brin–Thompson–Higman groups 2V and 3V ,” Proc.
Edinb. Math. Soc. (2) 56 (2013), no. 3, 777–804.
[18] Mark V. Lawson, “Orthogonal completions of the polycyclic monoids,” Comm. Algebra 35
(2007), no. 5, 1651–1660.
[19] Mark V. Lawson, “Subgroups of the group of homeomorphisms of the Cantor space and
a duality between a class of inverse monoids and a class of Hausdorff étale groupoids,” J.
Algebra 462 (2016), 77–114.
[20] Conchita Martı́nez-Pérez and Brita E. A. Nucinkis, “Bredon cohomological finiteness conditions for generalisations of Thompson groups,” Groups Geom. Dyn. 7 (2013), no. 4,
931–959.
[21] Diego Rattaggi, “A finitely presented torsion-free simple group,” J. Group Theory 10
(2007), no. 3, 363–371.
[22] Derek J. S. Robinson, A Course in the Theory of Groups, Second Edition, Graduate Texts
in Mathematics 80, Springer-Verlag, New York, 1996.
[23] Richard J. Thompson, Handwritten widely circulated notes, 1965.
[24] Werner Thumann, “Operad groups and their finiteness properties,” Adv. Math. 307 (2017),
417–487.
26
| 4math.GR
|
The Karyotype Ontology: a computational representation
for human cytogenetic patterns
Jennifer D. Warrender and Phillip Lord∗
arXiv:1305.3758v1 [cs.CE] 16 May 2013
School of Computing Science, Newcastle University, Newcastle-upon-Tyne, UK
ABSTRACT
The karyotype ontology describes the human chromosome
complement as determined cytogenetically, and is designed as an
initial step toward the goal of replacing the current system which is
based on semantically meaningful strings. This ontology uses a novel,
semi-programmatic methodology based around the tawny library to
construct many classes rapidly. Here, we describe our use case,
methodology and the event-based approach that we use to represent
karyotypes.
The ontology is available at http://www.purl.org/ontolink/
karyotype/. The clojure code is available at http://code.
google.com/p/karyotype-clj/.
complement of a human, as determined cytogenetically by the
banding patterns which are revealed after staining and fixation of
metaphase cells. The ISCN specification has a long history. Initially,
it was developed to address the need for an explicit nomenclature
“to enable communication between workers in the field”. The
early versions date from around 1960, when the emphasis was
on human-to-human communication, and for small numbers of
karyotypes.
ISCN strings represents a number of key concepts:
• The autosomal chromosomes are represented by a number 1 to
22.
• The sex chromosomes are represented by X or Y.
1
INTRODUCTION
A karyotype describes the chromosome complement of the
individual. It can be easily assayed cytogenetically and, therefore,
has been widely used as a mechanism for understanding the
underlying genetic complement of cells and organisms. It remains
of vital diagnostic importance, as well as a key tool for a large
research community. Human karyotypes are normally represented
using a string, as defined by the International System for human
Cytogenetic Nomenclature 2009 (ISCN2009) (Shaffer et al., 2009)
— here we call these ISCN strings. Unlike similar string-based
representations such as InCHI (McNaught, 2006), ISCN strings
lack a formal interpretation, and do not have good computational
properties. For example, they cannot be represented in ASCII as
they include meaningful underlining. Even the ISCN specification
has no electronic representation and is not searchable.
In this paper, we describe our work in developing an ontological
representation for karyotypes. Currently, karyotypes have only
been represented as experimental entities, or results of medical
procedures. The purpose of our ontology is to provide a strong
computational and formal interpretation for a karyotype. This will
enable semantic (and syntactic) checking of karyotypic information
at the point of generation; it will allow the development of a
knowledge base of karyotypes which is open to rich querying, and
finally a web-capable interchange format as many different groups
around the world generate this information.
2
WHAT IS AN ISCN STRING
whom
correspondence
[email protected]
should
• Bands at different resolutions are represented numerically such
as 1p11.1.
• Changes from the base karyotypes are represented: del
represents a deletion, add represents additional material.
In addition to these concepts, there are many more that can be
represented in an extended karyotype: these include chromosomal
groups, mosaicism, ploidy level and so forth. The full specification,
describes many parts of human cytogenetics, including both the
biology and the experimental techniques used. As would be
expected for a specification with a long history, not all parts are
regularly used.
Banding Patterns used to describe chromosome locations are
defined cytogenetically by the appearance of the chromosome,
during a part of division, following staining with a dye; this staining
process is normally lethal to the cell. The original banding pattern
described in the Paris Conference 1971 report, represented the
results of three whole chromosome banding techniques: Quinacrine(Q-), Giemsa- (G-), and Reverse-banding (R-).
The main components of stained chromosomes are:
• A band is a part of the chromosome that is distinguishable
from its adjacent segments, appearing darker or lighter. Bands
proximal to the centromere are labelled as 1, then 2 and so on.
ISCN defines a string format, initially designed for writing and
printing, which provides a representation of the chromosome
∗ To
• Chromosomal structural components are represented: The long
and short arm are represented by q and p; centromeres are
represented by cen or more specifically, p10 for the part of
the centromere facing the short arm or q10 for part facing the
long arm; and telomeres are represented by ter.
be
• A region is an area of a chromosome lying between two
landmarks. Regions adjacent to the centromere are labelled as
1 in each arm, then 2 and so on.
addressed:
1
Jennifer D. Warrender and Phillip Lord4
• A landmark is a consistent and distinct morphological
feature of a chromosome. They are used as a delimiters for
regions.
Bands are represented numerically such that bands are numbered
from the centromere outward. The band name is a combination of:
the chromosome number, the arm symbol, the region number, and
the band number within that region. For example, the band 1q42
is found on the long arm of chromosome 1 and is the second band,
proximal to the centromere in region 4. Broadly speaking there is a
correlation between the cytogenetic bands and the underlying DNA
sequence of the chromosome; however, the very different scales
(1q42 is 12.4Mb long) of these two measurements means that this
relationships is approximate.
Cytogenetic banding also comes at several resolutions: highresolution banding involves the staining of chromosomes during
prophase, prometaphase, or interphase when the chromosomes
are less condensed and spread over a large area; low-resolution
uses the highly-condensed metaphase chromosome. The level of
resolution is determined by the number of bands seen in a haploid
set and ranges approximately from 300 to 850. High-resolution
banding techniques result in existing bands being subdivided into
sub-bands. Whenever a band is subdivided, a decimal place is
placed after the original band number, and the sub-band number is
appended to the band name with proximal sub-bands bring labelled
1, then 2 and so on. For example, the sub-bands of band 1q42 will
be: 1q42.1, 1q42.2, and 1q42.3, such that sub-band 1q42.1
is more proximal to the centromere. However when a sub-band is
subdivided, then no additional decimal is added. For example, the
sub-bands of sub-band 1q42.1 are: 1q42.11, 1q42.12, and
1q42.13.
Typical queries that we might wish to make of a collection of
karyotypes include:
“portions of reality”; for instance, a chromosome in a live cell
cannot meaningfully be said to have bands. These distinctions can
be represented ontologically, however the result is a ontology with
many duplicated hierarchies: chromosome 1, stained chromosome
1, and the visualisation of chromosome 1.
As these distinctions are not required for our application, we
have instead followed a pragmatic approach (Lord and Stevens,
2010). We have developed a lightweight ontology with specific
computational goals. Our desire for computational support and
inferencing, as well as a web-capable interchange format, has lead
us to adopt OWL2 as our representation format.
While avoiding a realist approach has reduced some duplication,
karyotype ontology still requires a considerable number of highly
similar concepts, which is intrinsic to the problem domain.
Trivially, for instance, the human karyotype requires 24 individual
chromosome concepts, with similar logical and textual definitions.
In turn, each chromosome has a complex band pattern (over
850 bands in total), including band intervals for use at different
resolutions. Developing this type of ontology would be complex,
time consuming and difficult to maintain using conventional tools.
Therefore, we developed the tawny library, which allows fully
programmatic development of OWL ontologies (Lord, 2013). This
library allows expansion of arbitrary patterns; this is similar to the
capabilities of OPPL(Egana Aranguren et al., 2009), populous (Jupp
et al., 2010) or safe macros (Mungall et al., 2010). However, it
additionally provides us with the ability to define unit tests to
provide computationally checkable expressions of the requirements
for inferencing, a semi-literate programming environment, and
the ability to make arbitrary syntactic extensions. The syntax is
modelled after and highly similar to Manchester syntax, therefore
it is presented here without further explanation; a fuller description
is provided in the tawny documentation2 .
The basic structure of our ontology is shown in Figure 1.
• Which karyotypes have abnormalities in a given chromosome?
• Which karyotypes increase the copy number of a given band?
• Which karyotypes affects a given band in any way?
Currently, these are hard to answer computationally because of the
complexity of the ISCN strings, as well as intrinsic complexity of
the biology. Our karyotype ontology aims to address the former, and
contain the latter.
3
OUR METHODOLOGY
For the karyotype ontology, we have a very specific requirement
which is to enable machine interpretation of the knowledge that
is currently represented in ISCN strings. One of the implications
of this is that our knowledge capture is extremely contained;
essentially all the knowledge we require is present in the ISCN2009
specification1 . Our task is to formalise and represent this.
Our initial experiments with a realist ontology showed a number
of difficulties; the distinction between a chromosome (as a piece
of DNA and protein), the experimental artefact (following staining)
and the visualisation of the experimental artefact are all different
Fig. 1. The abstract structure of the karyotype ontology.
While ISCN2009 contains most of the knowledge we need, it is
informally represented; it sometimes lacks the clarity and contains
contradictions; trivially, for instance, Group G is described as
having satellites (page 7), while Chromosome Y (a member of
Group G) is described as having no satellites. A similar confusion
lies over the instance/type distinction. So, while for an individual
1
During the course of the work described here, ISCN2013 was released (in
2012!). We have not updated for this version yet.
2
2
https://github.com/phillord/tawny-owl
• Insertion: Band insertion between chromosomes. Specialised
with DirectInsertion or InverseInsertion.
organism or cell line all (normal) Chromosome 21s have (visible)
satellites, an individual chromosome, in an individual cell may
not3 . Similarly small supernumerary marker chromosomes (sSMCs)
cannot be fully represented in ISCN2009 (Liehr, 2009). This means
that an ontological representation of the ISCN2009 specification
cannot therefore be a completely faithful representation.
• Inversion: Band inversion, both paracentric and pericentric.
• Quadruplication: Band quadruplication.
• Translocation: Band translocation between chromosomes.
• Triplication: Band triplication.
4
REPRESENTING ABNORMALITIES
Our initial experiments attempt to represent karyotypes as a rich
partonomy, using the concepts described in the previous section. For
example, the normal male karyotype 46,XY would be described
as having 46 chromosomes of the appropriate types. One problem
with this approach is that the definition of any karyotype is relatively
large; while tawny enables the generation of this form of concept, it
cannot reduce the complexity for reasoners; so this form of ontology
is not likely to scale well. Further, a simple partonomy is not a
rich enough representation, as it cannot represent simple inversions
which contain all the same parts, but not necessarily in the right
order. While it is possible to represent order in OWL (Drummond
et al., 2006), this would add more complexity and scaling issues.
Further, there are some strong edge cases that are not complex,
but impossible to represent as a partonomy; for example, the
karyotype 45,X,-Y describes the chromosomes of a cell line,
isolated from a normal male, which has lost its Y chromosome. This
is a different karyotype but is partonomically indistinguishable from
45,X, a phenotypically female individual with Turners syndrome.
Therefore we represented karyotypes using events: a karyotype is
described as a set of changes from a base or normal karyotype.
For example, 45,X,-Y is described as a 46,XY male, with a
deletion event of Y, while 45,X female is described as a 46,XN,
with a deletion event of a sex chromosome (see Listing1): in this
case N represents an unknown autosome. OWL can represent partial
knowledge straight-forwardly.
(defclass k45_X
:label "The 45,X karyotype"
:subclass ISCNExampleKaryotype
(owlsome derivedFrom k46_XN)
(deletion 1 HumanSexChromosome))
Listing 1. A basic class definition
In total there are 13 events that represent the following key
concepts:
• Addition: Chromosome gain or band addition.
• Deletion: Chromosome loss or band deletion, including
terminal deletion with break (:) and interstitial deletion with
breakage and reunion (::).
• Duplication: Band duplication. Specialised with
DirectDuplication or InverseDuplication.
• Fisson: Centric fission.
3
As well as variation at a genetic level in even a clonal population, the
cytogenetic definition of satellite is a staining region. So even for two
genetically identical chromosomes, one may show a satellite, and another
may not.
These concepts are supported by a number of properties created
for the purpose, such as isBandOf.
The simplest representation in our ontology shares some
limitations with the ISCN “short system” – a usable subset of ISCN,
which is more generally used. For example, with both triplication
or quadruplication events, we do not represent the orientation of
all the repeats. It would be possible to differentiate these as two
direct duplications, or one direct and one inverted duplication
(for triplication); however, again it is useful to represent partial
information; for many existing ISCN strings which use the short
system, this knowledge is not available.
5
DEFINING SEX
One interesting outcome of both our representation and normal
custom within cytology is that the definition of the sex of
a karyotype is quite different from what might be expected.
Intuitively, male would be defined as a karyotype with a Y
chromosome5 , while female would be defined as a karyotype
without. However, this intuitive definition is not correct. For
example, the previously described 45X,-Y has no chromosome
Y and yet would generally be considered to be a male karyotype,
since the organism from which the cell line originated was male.
Our definitions of male and female, therefore, consider the
“history” of the karyotype. Female is defined as derived from
46,XX (Listing2), Male from 46,XY (Listing3). This definition
also copes with Turner’s syndrome which is not defined as
either male or female, nor describe sex for haploid karyotypes:
these can contain a Y chromosome (or not), but sex is not
meaningful for these karyotypes. Karyotypes which are definitional
for syndromes such as Turner’s or being male, are categorised under
NamedKaryotype.
(defclass FemaleKaryotype
:equivalent
(owlor k46_XX
(owlsome derivedFrom k46_XX)))
Listing 2. Definition of Female Karyotype
(defclass MaleKaryotype
:equivalent
(owlor k46_XY
(owlsome derivedFrom k46_XY)))
Listing 3. Definition of Male Karyotype
While this is ontologically correct, it does remove some
inferencing power that might reasonably be expected; the karyotype
45,X is effectively stated to be female, as it is not formally possible
5
We ignore Y chromosome translocations for simplicity
3
Jennifer D. Warrender and Phillip Lord7
to determine the sex from the components of a karyotype; for future
work, we may be able to address this, by describing the origin of the
karyotype (45,X,-Y is only valid as the karyotype of a cell line).
6
ASSESSMENT
As well as providing a specification, we are fortunate that ISCN2009
provides many examples; we are using these examples as an initial
evaluation for our ontology, to determine whether the ontology is
expressive enough to represent these exemplar karyotypes.
We wished this to be related to the ISCN string as, in most cases,
there is no other more humanly readable name. In order to represent
a karyotype in tawny a name is needed which is “safe” both as a
URL and in Clojure, the language used to implement tawny, and,
pragmatically, in Manchester syntax also6 .
• All karyotypes start with a “k” — Clojure symbols cannot start
with numbers
• Replaced ; character with — comment in Clojure
• Replaced ( and ) characters with ! — list delimiter in Clojure
• Replaced , character with — separator in Manchester syntax
Currently, we have represented 71 karyotypes in our karyotype
ontology. During this process, we have also discovered two
difficulties with the existing ISCN2009 specification; in both cases
a simple and intuitive correction is possible. These are:
• The lack of a band Xq12 in figures showing chromosome
bands (page 31); the figures are the only list of all chromosome
bands in ISCN2009.
• The absence of a band Yq11.2 in the 300 band resolution
(11.21, 11.22, 11.23 do exist on page 31) while this band
is used in several exemplars (for example on page 78). This
band does exist in ISCN2005 – the previous specification.
Taken together, these 71 karyotypes use all of the distinctions
necessary to answer questions given in Section 2.
7
DISCUSSION
The development of a karyotype ontology is potentially valuable
for cytogenetics, as the current ISCN specification is not
computationally amenable, reducing the value of collections of
karyotypes as they are hard to query, check and maintain. The
work described here presents an initial step towards this goal.
The process of producing this ontology is already of use; we
have discovered some errors or inconsistencies within ISCN which
prevent its direct interpretation computationally; we expect to find
more as we continue.
We have found the use of an ontology to be an appropriate
mechanism; the knowledge that needs to be represented is complex,
and overlapping. Cytogenetic data also requires the representation
of partial knowledge, such as locations that are only known to a
given resolution. The open world assumption of OWL copes well
6
It is possible to dissociate these two with tawny, but that did not seem
useful in this case
4
with this situation. Cytogenetics databases are also relatively small
(100,000’s rather than millions or billions), sizes to which OWL
should scale.
Existing tools for ontology development are, however, rather
limited in their support for building this form of ontology; it is to
address this need that we have developed tawny. This has proven
to be highly useful; for example, the current karyotype ontology
consists of 1466 classes, of which 1293 are used to represent the
chromosomes and their bands at different resolutions. All of these
classes have been generated from simpler data structures in Clojure.
Additionally, the arbitrary expressiveness of tawny has allowed us
to add syntax specific for the karyotype; many definitions in our
ontology follow patterns. Using tawny these can be encoded as
Clojure functions, such as that shown in and used in Listing4.
(defn inversion
[n band1 band2]
(exactly n hasEvent
(owland Inversion
(owlsome hasBreakPoint
band1 band2))))
(e/inversion 1 h/Hu2Bandp21 h/Hu2Bandq31)
Listing 4. A function used to define inverse events
In addition to the convenience, this also aids significantly in
maintainability, as it is possible to change definitions for all classes
that use this function. In time, we expect to extend this work to
present an end-user syntax and parser, probably built directly using
Clojure. Through the use of tawny we aim to build an end-to-end
solution for the computational encoding of karyotypes.
ACKNOWLEDGEMENTS
This work was supported by Newcastle University.
REFERENCES
Drummond, N., Rector, A., Stevens, R., Moulton, G., Horridge, M., Wang, H. H.,
and Seidenberg, J. (2006). Putting OWL in order: Patterns for sequences in OWL.
Concrete, pages 1–10.
Egana Aranguren, M., Stevens, R., and Antezana, E. (2009). Transforming
the axiomisation of ontologies: The ontology pre-processor language. Nature
Precedings.
Jupp, S., Horridge, M., Iannone, L., Klein, J., Owen, S., Schanstra, J., Stevens, R.,
and Wolstencroft, K. (2010). Populous: A tool for populating ontology templates.
Proceedings of the 3rd International Workshop on Semantic Web Applications and
Tools for the Life Sciences BerlinGermany December 810 2010.
Liehr, T. (2009). Small supernumerary marker chromosomes (sSMCs): a spotlight
on some nomenclature problems. Journal of Histochemistry and Cytochemistry,
57(11), 991–993.
Lord, P. (2013). The Semantic Web takes Wing: Programming Ontologies with TawnyOWL. http://arxiv.org/abs/1303.0213.
Lord, P. and Stevens, R. (2010). Adding a little reality to building ontologies for
biology. PLoS One.
McNaught, A. D. (2006). The IUPAC International Chemical Identifier (InChI).
Chemistry International, 2006(September 18).
Mungall, C., Ruttenberg, A., and Osumi-Sutherland, D. (2010). Taking shortcuts with
OWL using safe macros. Nature Preceedings.
Shaffer, L., on Human Cytogenetic Nomenclature, I. S. C., Slovak, M., and
Campbell, L. (2009). ISCN 2009: An International System for Human Cytogenetic
Nomenclature (2009). Karger.
| 5cs.CE
|
1
Production Line Technique for Autonomous Vehicle Scheduling
Nasser Aloufi
Department of Computer Science
California State University, Dominguez Hills
Carson, CA USA
[email protected]
Abstract - This paper considers the problem of scheduling autonomous vehicles in intersections. A new system is proposed
that could be an additional choice to the recently introduced autonomous intersection management (AIM) model. The
proposed system is based on the production line technique, where the environment of the intersection, vehicles position,
speeds and turning are specified and determined in advance. The goal of the proposed system is to eliminate vehicle collision
and the waiting time inside the intersection. 3 different patterns of the vehicles’ flow toward the intersection have been
considered for the evaluation of the model. The system requires less waiting time –compared to the other models- in the
random case where the flow is unpredictable. The KNN algorithm is used to predict the right turn vehicle. The experimental
results show that there is no single chance of collision inside the intersection; however, the system requires more free space
in the traffic lane.
I. INTRODUCTION
number of vehicles on streets is increasing day after
Theanother.
The more vehicles we have in roads, the more
furious people get which resulting in breaking traffic rules
and delaying the vehicles behind. As vehicles are
expecting to be completely autonomous, a new solution
to traffic jams has sparked. The necessity of creating an
autonomous vehicles and intelligent transportation
systems is more than ever before. This intelligent system
can be done by making vehicles interact with each others
and updating their routes according to the traffic flow.
If we look at the history of autonomous vehicle we can
see that they went through 5 stages [1]: A) No
Automation: in this stage the human was responsible
about all the driving. No automation interaction was
involved. B) Driver Assistance: in this stage the system
was responsible about simple tasks such as steering and
acceleration/deceleration. C) Partial Automation: this
stage is a continuation of the previous one with some
enhancement. Features such as cruise control and lanecentering have been added to the vehicles. D) Conditional
Automation: the autonomous driving system performs all
aspect of tasks; however, the human should be standing
by, in case anything went wrong. E) High Automation: in
this stage, the vehicle operates autonomously to all
conditions in the domain. Any scenario in the system
domain, the vehicle should do it fully autonomously. F)
Fully Automation: this is the final stage where human
intervention is not needed. The system’s performance is
equal to human driver in all scenarios without any domain
constraints. It is expected for all vehicles to reach the fully
automation stage and work without any human control by
the year of 2035 [2]. The Institute of Electrical and
Electronics Engineers (IEEE) have claimed that driverless
vehicles will be the most viable form of intelligent
transportation. They estimate that up to 75% of all
vehicles around the world will be autonomous by the year
of 2040 [3].
In 2008, The Defense Advanced Research Projects
Agency (DARPA) held an event which ask teams to build
autonomous vehicles that have the ability to drive in
traffic, make maneuvers and park. The event was
described as the first time that autonomous vehicles make
interaction between manned and unmanned vehicles in a
real environment [4]. The VisLab in Italy that had made
many advanced driver assistance systems [5], [6] [7] [8],
and prototype vehicles including ARGO [9], TerraMax
[5], [10], [11], and BRAiVE [12], [13] made another
experiment related to the autonomous vehicle
development. The team that consists of Massimo
Bertozzi, Alberto Broggi, Alessandro Coati, and Rean
Isabella Fedriga made a long trip experiment for four
autonomous vehicles [14]. The trip took a place from Italy
to China crossing more than 15,000 km for a consecutive
three months. The result of their experiment told us three
major things: a- the autonomous vehicle faced troubles
while making maneuver; b- being one of the few vehicles
that follow the street rules might cause a long waiting time
(specially while applying FIFO); c- hard to combine
autonomous with un-autonomous vehicles in the same
environment. In [15], an intersection algorithm model
was proposed. The model is based on (MixedInteger
Linear Program) MILP controller. Their contributions
are: 1) an algorithm that predict vehicle arrivals. 2)
Applying mixed-integer linear program. 3) Develop a
simulation that based on the proposed MILP controller.
For a maneuver and changing lanes, a politely change lane
(PCL) [16] paper was proposed. The main goal of the PCL
2
is to provide safety and efficiency while maneuver.
In 2013, Mladenović and Abbas proposed a selforganizing control framework for driverless vehicles [27].
Their proposal is based on cooperative control framework
and intersections as agent system, where distributed
vehicle intelligence has been used to get the vehicle’s
velocity. The priority level system determines the vehicle
that pass the intersection first and adjust the speed for the
following vehicle in the queue so they don’t collide. In
2014, Yan, Wu, and Dridi [28] studied the complexity of
the sequencing problem. Their simulation was based the
AIM framework. They converted the autonomous
vehicles scheduling to a single machine scheduling
problem, then they proved that it is an NP-hard problem.
Dresner and Stone proposed a new intersection control
mechanism called Autonomous Intersection Management
(AIM) [29]. Their paper shows that by making a smart
intersection controlling system, vehicles’ flow would be
more efficient than the current situation (traffic signals
and stop signs). They suggested that the drivers and the
designed intersections should be treated as agents. The
system could have more than one agent leading us to have
a Multiagent Systems (MAS). The MAS includes all the
interacted elements in the environment such as drivers,
pedestrians, speeds and road signs. Whenever a vehicle
wants to reserve a place in the intersection, it sends a
request to the intersection manager, and then the
intersection manager takes an action by weather accepting
or rejecting the vehicle’s request. Figure1 below shows
how
the
AIM
system
works[29].
Figure1.How an AIM driver agent make a
reservation.
In this paper, we propose a new scheduling model
that based on the producion line technique where items
don’t collide and interfere with each other. Our proposed
method can be implemented for faster computation
[19][20] using GPUs as well [21][25][26]. Also, the
network data overhead can be reduced using compression
techniques [18][22][23], or using cloud computing
resources [24].
The rest of the paper is organized as the following:
In section 2 we made a simulation to show the cahnce of
collision and the amount of waiting time in the current
model. In section 3 we propose our model. Then we
provide the simulation and the result of our system.
II. PROBLEM DESCRIPTION
It seems a mission impossible to make all vehicles go
through an intersection safely. We need to take into
account the speed, timing, distance, the chance of
collision, and the appropriate response that we need to
make if there is a chance of collision. In AIM and similar
systems, the vehicle sends a request to the intersection
manager asking for a permission to go through the
intersection. The intersection manager must make one of
those decisions: 1) Accept vehicle’s request (when the
vehicle meets the requirements). 2) Reject vehicle’s
request (when the vehicle failed to meet all the
requirement). The rejected request has two options too: 1ask the vehicle to accelerate or decelerate its speed (in this
case the vehicle should send a new request to the
intersection manager. 2- ask the vehicle to stop due to
requirements failure. Figure2 below shows a general
overview of the current intersection scheduling model.
Figure2. The current system structure.
The current model works well, but still has some
problems that we have to overcome before applying it to
a real autonomous environment. Asking vehicles to send
requests few hundred meters before approaching the
intersection cause a critical time situation. There are some
condition where stopping the vehicle completely is a
must, and this happens whenever:
Vx(T,S,P) = Vz(T,S,P)
Where: T is the time of approaching a specific point in the
intersection ,S is the speed of the vehicle,P is the occupied
point in the intersection.
The more the intersection manager asks vehicles to stop,
the more traffic jam happens which make the system
inefficient especially in big cities. Figure3 shows one of
the conditions that cause traffic jam. Let’s make an
3
example and assume that we have two vehicles. Vehicle
A is going to the east side with a speed of 80mph, and
vehicle B going to the west side with the same speed. Both
have the same distance from the intersection, 600 ft. Both
had sent requests to the intersection manager asking to go
through the intersection. The intersection manager
calculated the characteristics and the conditions of the two
vehicles and came to conclusion that both vehicles will
approach the same point at the same time, specifically
after 5.11 seconds (converting 80 mph to foot -equal to
422400- then divided by 60 –equal to 7040- then divided
by 60 –equal to 117.33- foot per second). To avoid the
collision, the intersection manager has two options as we
stated before; either reject the requests and ask vehicle A
or B to change its speed; or simply ask one of them to stop
in order to avoid the collision.
vehicles. When we had 300 vehicles, the number of
collisions increased to be 4.3 for each vehicle. The same
thing happened for the waiting time. It was increasing as
the number of vehicles increase. The waiting time was 85
seconds per vehicle when we had 50 vehicles inside the
intersection. Then it increased to be 320 seconds when we
had 200. In the end when we had 300 vehicles, the waiting
time was 515 seconds per vehicle. Figure4 and Figure5
show the number of collisions and waiting times
respectively.
Figure4. Chances of collisions for every vehicle.
Figure3. A point of collision between vehicles A and B.
The arisen problem is, during rush hours when the
intersection gets hundreds or thousands of requests, the
chances of collisions increase, leading to more stopping
and waiting times.
We made a simulation of a similar environment to this
model in order to show the chance of collisions that each
vehicle might have. The simulated intersection has the
following characteristics:
I.
Two directions (one direction is going from north
to south and the other one is going from west to
east.).
II.
Each side of the intersection can occupy 722
vehicle in total.
III.
All vehicles have the same speed 100 mph.
After running the simulation for 100 times, the
experiment shows that, as the number of vehicles
increase, the number of expected collisions and waiting
times increase. When we had 50 vehicles in the
intersection, the number of expected collisions was 1
collision for each vehicle. Then it was 3 when we got 200
Figure5. The waiting time for every vehicle.
Increasing and decreasing speed is not an option when the
intersection get enormous requests because there are
vehicles in front and behind each vehicle, and whenever
we fix one collision, another one appears. To sum up, the
more vehicles approaching the intersection we have, the
more chances of collision we get, leading us to more
stopping time.
III. THE PROPOSED SYSTEM
In the old model, the scheduling process starts after
receiving the vehicle’s request (Figure2). In this paper we
flipped the current approach and created another model
where the scheduling should be sat up in the intersection
4
prior receiving any request. Our proposal is based on the
production line system where every position (container)
in the line is reserved for a specific item.
Figure 6 shows the architecture of the proposed system
where the intersection’s spots should be sat up as the first
step. We prepare the intersection by making containers
that based on the length of the vehicles. Once setting up
the vehicles position, entering timing, and the speed of the
lanes; then vehicles can send requests to the intersection
manager.
Figure7 shows a closer look of the design where:
• S1; the minimum accepted speed.
• S2; the maximum accepted speed.
• Spin; the times where the lane is open.
Figure6: The architecture of the proposed system.
consideration, and instead of applying one specific speed
value, we applied S1 and S2 which are the minimum and
maximum speed. Any vehicle speed that meets this range
should be accepted.
A) ARISEN PROBLEMS:
There are 2 problems that had arisen after the initial
design:
1. After applying S1 and S2, vehicles speed are vary,
leading vehicles to collide after certain distance.
Let’s assume we sat up the entering speed for lane A to
60MPH as a minimum speed (S1), and to 65MPH as a
maximum speed (S2). Two vehicles (V1 and V2) sent
requests to enter the intersection. V1 came first with a
61MPH speed, and V2 came after it with a 64MPH speed.
The problem is after a certain distance, V2 will approach
V1 and collide with it.
2. Entering an intersection with one speed range
while the following intersection requires a
different speed range.
Let’s assume we have two intersections, one after another.
The first intersection accept vehicles with any speed
between 60 and 65, and the second one accept vehicles
with 102.5 - 107.5 speed range. The problem now is how
to make those two intersections compatible so they get
connected together in one system. Figure9. Shows this
problem.
Figure9. Varying speed between two intersections.
3. Making a right turn.
The system must produce a container that makes a right
turn, otherwise vehicles will be forced to go in straight
direction.
B) PROBLEMS SOLVING
Figure7. Production line intersection.
The timer is used to switch between lanes. If lane A is
open at this moment, lane B should be closed. In [30] a
setpoint system for generating setpoints for the PID
controllers was proposed. The purpose of the system is to
make sure that the vehicle arrives at the exact expected
time with the exact expected speed. However, the system
doesn’t guarantee the arrival time of the vehicle to a
specific point, so it’s not ideal where an unexpected
latency cause a catastrophic result. We took that into
To solve the first problem we applied an average speed
value. This average speed should be applied to every
vehicle as soon as entering the intersection.
Avg = (S1+S2)/2
The goal of the average speed it to make sure that the
vehicle stays in its given container and doesn’t jump to
another one. Figure10 shows the problem before and after
applying the average speed.
5
Figure10. The intersection flow before and after
applying the average speed.
To solve the second problem, we created an increasing
and decreasing speed areas. The intersection manager ask
all the vehicles either to increase or decrease the speed
depending on the speed of the next intersection.
Figure11. The vehicle in the yellow triangle is
going to make a right turn when K=3.
IV. SIMULATION AND RESULT
S = Ep1 + (Tp1- Ep1); where: Ep is the exit speed of the
intersection. Tp is the entering speed of the following
intersection. Figure11 shows how the increasing speed
works.
Figure11. Areas of changing the speed.
Regarding the third problem which is making a right turn.
We made a classifier by using the key nearest neighbor
(KNN) technique. In the KNN the intersection produce
right turn containers based on whatever features we would
like to apply. Some of the features that we could apply for
the KNN: day, time, population of the city, event
happening in the area,…etc. For our KNN we used the
Euclidean distance function:
In our experiment we made a virtual intersection with the
following features:
1- Vehicles are coming from 4 different directions:
A1: from east to west.
A2: from west to east.
B1: from north to west.
B2: from west to north.
2- Vehicle arriving speed to the intersection is
ranged between (60-65) mph.
3- The length for every container is 26.2467 ft.
4- The running time for the intersection is 1 minute.
5- The system produce one container for every
opening lane.
6- Gates A1 and A2 should open together while
gates B1 and B2 closed.
7- Gates B1 and B2 should open together while
gates A1 and A2 closed.
8- Every lane has 60 containers in total.
Figure12 below shows an illustration of the virtual
intersection that we built for the experiment.
This classifier will be used to calculate the possibility of
making a right turn for the next container (vehicle).
Figure11 shows an example the KNN method.
Figure12. Experiment virtual intersection.
For expecting the right turn in the KNN we need to enter
initial testing data in the system. The following are the
features and initial data:
6
Day
Hour
1
3
4
3
4
2
5
1
2
9
10
8
8
10
20
19
4
7
Event in Class
the area
turn
0
+
0
+
0
+
0
+
0
+
1
1
1
1
-
of
Where:
1- Days are numbered from 1:Monday to 5:Friday.
2- Hour is a 24 format.
3- Event in the area is :
0; if there is no event in the area of the
intersection.
1; if there is no event..
4. “+” means the vehicle is going to make a right
turn. “-” means the vehicle is going in a straight
direction.
We also created 3 different patterns of the vehicles flow.
The goal of these three patters is to calculate the waiting
time if we got unordered flow. In addition we wanted to
calculate the required space for each pattern. The first
pattern represents the best case where the flow match the
opening and the closing times of the intersection. The
second pattern represents the worst case of the flow where
vehicles keep coming even if the intersection is closed.
The third and final pattern represent the random (normal)
case where the flow is unpredictable.
Each of the three patterns has the follow features:
I.
It has 60 spots.
II.
Accept only one request per spot.
III.
Runs for 1 minute.
For the patterns, we got different results based on the flow
toward the intersection. For all of the following results:
1 represents a request
0 represent no request
In the average case where the flow pattern is: [1, 0, 1, 0,
1, 0, 1, 0, 1, 0, 1, 0,…..]: the arriving times for vehicles in
one minute was: [0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22,
24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52,
54, 56, 58]. We don’t need to arrange them because they
are already compatible with the opening and closing times
for our lane. The waiting time in this case is always zero.
In the second case the flow is constant (worst case): [1, 1,
1, 1, 1, 1, 1, 1, 1, 1……]. The arriving times for this
pattern in one minute was: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25,
26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55,
56, 57, 58, 59]. Arranging them to match the intersection
opening and closing moments was a time consuming
since it took 164.84 seconds as an average waiting time
for each vehicle. Here is the order of arrivals after
arranging them: [0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22,
24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52,
54, 56, 58, 60, 62, 64, 66, 68, 70, 72, 74, 76, 78, 80, 82,
84, 86, 88, 90, 92, 94, 96, 98, 100, 102, 104, 106, 108,
110, 112, 114, 116, 118]. The worst part of this case is not
only the high waiting time, but also we needed to increase
the space by 100% to arrange the vehicles. The last pattern
where the flow is unpredictable (random), the system was
really promising since the waiting time was less than the
old models. In our model, the waiting time was 35
seconds compared to 101 seconds for the old model. For
testing, we created a random function that creates random
requests. The requests we got were : [1, 1, 0, 1, 1, 1, 0, 0,
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 1, 0, 1,
0, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 1, 0, 1, 1,
0, 0, 0, 1, 1, 1], and their arriving times were: [0, 1, 3, 4,
5, 9, 11, 13, 15, 17, 18, 21, 24, 26, 28, 30, 32, 33, 34, 36,
38, 41, 43, 48, 50, 52, 53, 57, 58, 59]. In this random
pattern it only took 6.79 seconds as an average waiting
time. Here is the arranged arrivals : [0, 0, 1, 0, 3, 0, 4, 0,
5, 0, 9, 0, 11, 0, 13, 0, 15, 0, 17, 0, 18, 0, 21, 0, 24, 0, 26,
0, 28, 0, 30, 0, 32, 0, 33, 0, 34, 0, 36, 0, 38, 0, 41, 0, 43, 0,
48, 0, 50, 0, 52, 0, 53, 0, 57, 0, 58, 0, 59]. To sum up the
patterns result: -The first pattern (matched pattern) has 0
waiting time and 0 additional space. The second pattern
(worst pattern) has increased by double in both timing and
spacing. The third pattern (random flow) result depends
on the number of requests: 0 if the number of requests
below the giving spots, otherwise it will be ((n-30)
/30*100) where n is the number of requests. However, it
was less than the previous model in all tested random
flow.
For predicting the right turn, whenever a new vehicle
comes to the lane, its (day, hour, event) will be calculated
by the KNN classifier to predict if it’s going to make a
right turn or not. The resulted data of the entered vehicle
will be added to the classifier so it helps to predict the next
generated spot. After running the simulation for one
minute we got the following results:
Lane A1 and A2 instances:
7
Day
Hour
3
1
2
5
4
1
2
2
3
2
4
1
2
4
4
4
3
2
3
1
4
1
1
1
2
4
2
3
2
4
3
5
11
10
1
23
13
18
14
8
6
21
3
15
22
22
8
21
0
20
12
3
7
23
2
6
0
16
1
11
B1 and B2 instances:
Day
Hour
1
4
5
4
4
2
2
4
3
5
2
5
5
3
8
12
2
3
12
1
13
20
18
16
1
3
9
11
Event in the
area
0
0
0
0
0
0
0
1
1
1
1
1
0
0
1
0
1
0
1
0
0
1
1
1
1
0
1
0
1
0
Event in the
area
0
0
1
1
0
0
0
1
1
0
0
0
0
1
4
2
2
5
3
3
2
5
3
4
3
3
5
2
2
5
2
9
9
19
11
18
16
15
11
17
16
22
23
12
11
19
1
0
0
1
1
1
1
1
1
1
1
1
0
1
0
1
Turns prediction for lane A1 and lane A2:
1
2
3
4
5
6
7
8
+
+
+
11 12 13 14 15 16 17 18
+
+
+
21 22 23 24 25 26 27 28
+
+
+
+
9
+
19
29
-
10
+
20
30
+
9
19
+
29
+
10
20
30
-
Figure14. KNN for lane A1 and A2.
Turns prediction for lane B1 and lane B2:
1
2
3
4
5
6
7
8
+
+
+
+
11 12 13 14 15 16 17 18
+
+
+
+
21 22 23 24 25 26 27 28
+
+
8
working on for the second version of the system.
VI. REFERENCES
Figure14. KNN for lane B1 and B2.
The following Table1 shows the first 5 vehicles for each
lane.
Vehicle
ID
123
206
301
413
106
224
306
405
115
200
324
422
109
213
311
418
105
222
313
418
Lane
A1
A2
B1
B2
A1
A2
B1
B2
A1
A2
B1
B2
A1
A2
B1
B2
A1
A2
B1
B2
Arriving
Time
0.0
0.0
1.0
1.0
2.0
2.0
3.0
3.0
4.0
4.0
5.0
5.0
6.0
6.0
7.0
7.0
8.0
8.0
9.0
9.0
Right
Turn
No
No
Yes
Yes
No
No
Yes
Yes
Yes
Yes
No
No
Yes
Yes
No
No
No
No
Yes
Yes
Exiting Time
17.179657557103365
17.179657557103365
18.179657557103365
18.179657557103365
19.179657557103365
19.179657557103365
20.179657557103365
20.179657557103365
21.179657557103365
21.179657557103365
22.179657557103365
22.179657557103365
23.179657557103365
23.179657557103365
24.179657557103365
24.179657557103365
25.179657557103365
25.179657557103365
26.179657557103365
26.179657557103365
Table1. Data of the first 5 vehicles that entered the intersection.
V. CONCLUSION
The proposed autonomous scheduling system introduced
in this paper is based on the manufacturing production
line system. The intersection environment is sat up in
advance before any vehicle arrives to the intersection. We
tested three different patterns (best-worst-random). In the
first pattern where the flow matches the intersection
requirements, the queue waiting time was zero. In the
worst case the waiting time was high and could be
presented as X!. In the third pattern the waiting time was
promising for a real environment since it got less waiting
time compared to the old model. Having said that, using
a huge amount of space is still an issue that we are
[1]: SAE On-Road Automated Vehicle Standards Committee.
"Taxonomy and definitions for terms related to on-road motor
vehicle automated driving systems." Society of Automotive
Engineers , International J3016_201609, October 2016 .
[2]:Bimbraw, Keshav. "Autonomous cars: Past, present and
future a review of the developments in the last century, the
present scenario and the expected future of autonomous
vehicle technology." Informatics in Control, Automation and
Robotics (ICINCO), 2015 12th International Conference on.
Vol. 1. IEEE, 2015.
[3]:“IEEE News Releases.” IEEE - News Releases,
www.ieee.org/about/news/2012/5september_2_2012.html.
Last visited: 11/1/2017.
[4]: DARPA. (2007) Urban challenge. [Online]. Avilable:
http://archive.darpa.mil/grandchallenge/ Last visited :
10/4/2017.
[5] Y.-L. Chen, V. Sundareswaran, C. Anderson, A. Broggi, P.
Grisleri, P. P. Porta, P. Zani, and J. Beck, “TerraMax: Team
oshkosh urban robot,” in The DARPA Urban Challenge,
Autonomous Vehicles in City Traffic, M. Buehler, K.
Iagnemma, and S. Singh, Eds. Berlin, Germany:
SpringerVerlag, 2009, pp. 595–622.
[6] P. Cerri, L. Gatti, L. Mazzei, F. Pigoni, and G. Ho, “Day
and night pedestrian detection using cascade adaboost
system,” in Proc. IEEE Intelligent Transportation Systems
2010, pp. 1843–1848.
[7] C. Caraffi, E. Cardarelli, P. Medici, P. P. Porta, and G.
Ghisio, “Real time road signs classification,” in Proc. IEEE
Int. Conf. Vehicular Electronics Safety, Sept. 2008, pp. 253–
258.
[8] M. Bertozzi, A. Broggi, and A. Fascioli, “Vislab and the
evolution of vision-based UGVs,” IEEE Computer, vol. 39,
no. 12, pp. 31–38, Dec. 2006.
[9] A. Broggi, M. Bertozzi, A. Fascioli, and G. Conte,
Automatic Vehicle Guidance: The Experience of the ARGO
Vehicle. Singapore: World Scientific, 1999.
[10] D. Braid, A. Broggi, and G. Schmiedel, “The TerraMax
autonomous vehicle,” J. Field Robot., vol. 23, no. 9, pp. 693–
708, Sept. 2006.
[11] Y.-L. Chen, V. Sundareswaran, C. Anderson, A. Broggi,
P. Grisleri, P. P. Porta, P. Zani, and J. Beck, “TerraMax: Team
oshkosh urban robot,” J. Field Robot., vol. 25, no. 10, pp.
841–860, Oct. 2008.
[12] L. Bombini, S. Cattani, P. Cerri, R. I. Fedriga, M. Felisa,
and P. P. Porta, “Test bed for unified perception & decision
architecture,” in Proc. 13th Int. Forum Advanced
Microsystems Automotive Applications, 2009, pp. 287–298.
[13] P. Grisleri and I. Fedriga, “The braive platform,” in Proc.
7th IFAC Symp. Intelligent Autonomous Vehicles, 2010, pp.
497–502.
[14]: Bertozzi, Massimo, et al. "A 13,000 km intercontinental
trip with driverless vehicles: The VIAC experiment." IEEE
Intelligent Transportation Systems Magazine 5.1 (2013): 2841.
9
[15]: Fayazi, S. Alireza, Ardalan Vahidi, and Andre Luckow.
"Optimal scheduling of autonomous vehicle arrivals at
intelligent intersections via MILP." American Control
Conference (ACC), 2017. IEEE, 2017.
[16]: Hu, Jiajun, et al. "Scheduling of connected autonomous
vehicles on highway lanes." Global Communications
Conference (GLOBECOM), 2012 IEEE. IEEE, 2012.
[17] A. Chatterjee, A. Aceves, R. Dungca, H. Flores, and K.
Giddens. Classification of wearable computing: A survey of
electronic assistive technology and future design. In Research
in Computational Intelligence and Communication Networks
(ICRCICN), pages 22–27. IEEE, 2016.
[18] A. Chatterjee, M. Levan, C. Lanham, M. Zerrudo, M.
Nelson, and S. Radhakrishnan. Exploiting topological
structures for graph compression based on quadtrees. In
Research in Computational Intelligence and Communication
Networks (ICRCICN), 2016 Second International Conference
on, pages 192–197. IEEE, 2016.
[19] A. Chatterjee, S. Radhakrishnan, and J. K. Antonio. Data
Structures and Algorithms for Counting Problems on Graphs
using GPU. International Journal of Networking and
Computing (IJNC), Vol. 3(2):264–288, 2013.
[20] A. Chatterjee, S. Radhakrishnan, and J. K. Antonio. On
Analyzing Large Graphs Using GPUs. In Parallel and
Distributed Processing Symposium Workshops & PhD Forum
(IPDPSW), 2013 IEEE 27th International, pages 751–760.
IEEE, 2013.
[21] A. Chatterjee, S. Radhakrishnan, and John K. Antonio.
Counting Problems on Graphs: GPU Storage and Parallel
Computing Techniques. In Parallel and Distributed Processing
Symposium Workshops & PhD Forum, 2012 IEEE 26th
International, pages 804–812. IEEE, 2012.
[22] M. Nelson, S. Radhakrishnan, A. Chatterjee, and C. N.
Sekharan. On compressing massive streaming graphs with
Quadtrees. In Big Data, 2015 IEEE International Conference
on, pages 2409–2417, 2015.
[23] M. Nelson, S. Radhakrishnan, A. Chatterjee and C. N.
Sekharan, "Queryable compression on streaming social
networks," 2017 IEEE International Conference on Big Data
(Big Data), Boston, MA, 2017, pp. 988-993.
[24] A. Chatterjee, M. Levan, C. Lanham and M. Zerrudo,
"Job scheduling in cloud datacenters using enhanced particle
swarm optimization," 2017 2nd International Conference for
Convergence in Technology (I2CT), Mumbai, 2017, pp. 895900.
[25] Khondker S. Hasan, Amlan Chatterjee, Sridhar
Radhakrishnan, and John K. Antonio, “Performance
Prediction Model and Analysis for Compute-Intensive Tasks
on GPUs”, Network and Parallel Computin,g 2014, pp. 612617
[26] A. Chatterjee, “Parallel Algorithms for Counting
Problems on Graphs Using Graphics Processing Units”, Ph.D.
Dissertation, University of Oklahoma, 2014
[27]: M. N. Mladenovic and M. M. Abbas, “Self-organizing
control framework for driverless vehicles,” in Proc. 16th Int.
IEEE Conf. on Intell. Transp. Syst., The Hague, The
Netherlands, Oct. 2013.
[28]: Yan, Fei, Jia Wu, and Mahjoub Dridi. "A scheduling
model and complexity proof for autonomous vehicle
sequencing problem at isolated intersections." Service
Operations and Logistics, and Informatics (SOLI), 2014 IEEE
International Conference on. IEEE, 2014.
[29] : Dresner, Kurt, and Peter Stone. "A multiagent approach
to autonomous intersection management." Journal of artificial
intelligence research 31 (2008): 591-656.
[30]: Au, Tsz-Chiu, Michael Quinlan, and Peter Stone.
"Setpoint scheduling for autonomous vehicle controllers."
Robotics and Automation (ICRA), 2012 IEEE International
Conference on. IEEE, 2012.
| 3cs.SY
|
Extremal attractors of Liouville copulas
Léo R. Belzile∗ and Johanna G. Nešlehová†
arXiv:1704.03377v2 [math.ST] 6 Jul 2017
This is a copyedited, author-produced version of an article accepted for publication following peer-review in the
Journal of Multivariate Analysis, an Elsevier publication, c 2017. This manuscript version is made available under
the CC-BY-NC-ND 4.0 license. Please cite as
Belzile, L. R and J. G. Nešlehová. Extremal attractors of Liouville copulas, Journal of Multivariate
Analysis (2017), 160C, pp. 68–92. doi:10.1016/j.jmva.2017.05.008
Abstract
Liouville copulas introduced in [31] are asymmetric generalizations of the ubiquitous Archimedean copula class.
They are the dependence structures of scale mixtures of Dirichlet distributions, also called Liouville distributions. In
this paper, the limiting extreme-value attractors of Liouville copulas and of their survival counterparts are derived.
The limiting max-stable models, termed here the scaled extremal Dirichlet, are new and encompass several existing
classes of multivariate max-stable distributions, including the logistic, negative logistic and extremal Dirichlet. As
shown herein, the stable tail dependence function and angular density of the scaled extremal Dirichlet model have a
tractable form, which in turn leads to a simple de Haan representation. The latter is used to design efficient algorithms
for unconditional simulation based on the work of [10] and to derive tractable formulas for maximum-likelihood
inference. The scaled extremal Dirichlet model is illustrated on river flow data of the river Isar in southern Germany.
1. Introduction
Copula models play an important role in the analysis of multivariate data and find applications in many areas, including
biostatistics, environmental sciences, finance, insurance, and risk management. The popularity of copulas is rooted
in the decomposition of Sklar [39], which is at the heart of flexible statistical models and various measures, concepts
and orderings of dependence between random variables. According to Sklar’s result, the distribution function of any
random vector X = (X1 , . . . , Xd ) with continuous univariate margins F1 , . . . , Fd satisfies, for any x1 , . . . , xd ∈ R,
Pr(X1 ≤ x1 , . . . , Xd ≤ xd ) = C{F1 (x1 ), . . . , Fd (xd )},
for a unique copula C, i.e., a distribution function on [0, 1]d whose univariate margins are standard uniform. Alternatively, Sklar’s decomposition also holds for survival functions, i.e., for any x1 , . . . , xd ∈ R,
Pr(X1 > x1 , . . . , Xd > xd ) = Ĉ{F̄1 (x1 ), . . . , F̄d (xd )},
where F̄1 , . . . , F̄d are the marginal survival functions and Ĉ is the survival copula of X, related to the copula of X as
follows. If U is a random vector distributed as the copula C of X, Ĉ is the distribution function of 1 − U.
In risk management applications, the extremal behavior of copulas is of particular interest, as it describes the dependence between extreme events and consequently the value of risk measures at high levels. Our purpose is to study
the extremal behavior of Liouville copulas. The latter are defined as the survival copulas of Liouville distributions
[14, 17, 38], i.e., distributions of random vectors of the form R Dα , where R is a strictly positive random variable
independent of the Dirichlet random vector Dα = (D1 , . . . , Dd ) with parameter vector α = (α1 , . . . , αd ). Liouville
copulas were proposed by McNeil and Nešlehová [31] in order to extend the widely used class of Archimedean
copulas and create dependence structures that are not necessarily exchangeable. The latter property means that for
∗ École
Polytechnique Fédérale de Lausanne, EPFL-SB-MATH-STAT, Station 8, CH-1015 Lausanne, Switzerland. [email protected]
of Mathematics and Statistics, McGill University, 805 rue Sherbrooke Ouest, Montréal, Québec, H3A 0B9, Canada.
[email protected]
† Department
2
L. R. Belzile and J. G. Nešlehová
any u1 , . . . , ud ∈ [0, 1] and any permutation π of the integers 1, . . . , d, C(u1 , . . . , ud ) = C(uπ(1) , . . . , uπ(d) ). When
α = 1d ≡ (1, . . . , 1), Dα = D1d is uniformly distributed on the unit simplex
Sd = {x ∈ [0, 1]d : x1 + · · · + xd = 1}.
(1)
In this special case, one recovers Archimedean copulas. Indeed, according to [30], the latter are the survival copulas
of random vectors R D1d , where R is a strictly positive random variable independent of D1d . When α , 1d , the survival
copula of RDα is not Archimedean anymore. It is also no longer exchangeable, unless α1 = · · · = αd .
In this article, we determine the extremal attractor of a Liouville copula and of its survival counterpart. As a byproduct, we also obtain the lower and upper tail dependence coefficients of Liouville copulas that quantify the strength
of dependence at extreme levels [25]. These results are complementary to [21], where the upper tail order functions
of a Liouville copula and its density are derived when α1 = · · · = αd , and to [19], where the extremal attractor of
RDα is derived when R is light-tailed. The extremal attractors of Liouville copulas are interesting in their own right.
Because non-exchangeability of Liouville copulas carries over to their extremal limits, the latter can be used to model
the dependence between extreme risks in the presence of causality relationships [15]. The limiting extreme-value
models can be embedded in a single family, termed here the scaled extremal Dirichlet, whose members are new, nonexchangeable generalizations of the logistic, negative logistic, and Coles–Tawn extremal Dirichlet models given in
[7]. We examine the scaled extremal Dirichlet model in detail and derive its de Haan spectral representation. The latter
is simple and leads to feasible stochastic simulation algorithms and tractable formulas for likelihood-based inference.
The article is organized as follows. The extremal behavior of the univariate margins of Liouville distributions is
first studied in Section 2. The extremal attractors of Liouville copulas and their survival counterparts are then derived
in Section 3. When α is integer-valued, the results of [27, 31] lead to closed-form expressions for the limiting stable
tail dependence functions, as shown in Section 4. Section 5 is devoted to a detailed study of the scaled extremal
Dirichlet model. In Section 6, the de Haan representation is derived and used for stochastic simulation. Estimation
is investigated in Section 7, where expressions for the censored likelihood and the gradient score are also given. An
illustrative data analysis of river flow of the river Isar is presented in Section 8, and the paper is concluded by a
discussion in Section 9. Lengthy proofs are relegated to the Appendices.
In what follows, vectors in Rd are denoted by boldface letters, x = (x1 , . . . , xd ); 0d and 1d refer to the vectors
(0, . . . 0) and (1, . . . , 1) in Rd , respectively. Binary operations such as x + y or a · x, xa are understood as componentwise operations. k · k stands for the `1 -norm, viz. kxk = |x1 | + · · · + |xd |, ⊥
⊥ for statistical independence. For any x, y ∈ R,
let x ∧ y = min(x, y) and x ∨ y = max(x, y). The Dirac delta function Ii j is 1 if i = j and zero otherwise. Finally, Rd+ is
the positive orthant [0, ∞)d and for any x ∈ R, x+ denotes the positive part of x, max(0, x).
2. Marginal extremal behavior
A Liouville random vector X = RDα is a scale mixture of a Dirichlet random vector Dα = (D1 , . . . , Dd ) with
parameters α = (α1 , . . . , αd ) > 0d . In what follows, R is referred to as the radial variable of X and ᾱ denotes the sum
of the Dirichlet parameters, viz. ᾱ = kαk = α1 + · · · + αd . Recall that Dα has the same distribution as Z/kZk, where
Zi ∼ Ga(αi , 1), i = 1, . . . , d are independent Gamma variables with scaling parameter 1. The margins of X are thus
scale mixtures of Beta distributions, i.e., for i = 1, . . . , d, Xi = RDi with Di ∼ Beta(αi , ᾱ − αi ).
As a first step towards the extremal behavior of Liouville copulas, this section is devoted to the extreme-value
properties of the univariate margins of the vectors X and 1/X, where X is a Liouville random vector with parameters
α and a strictly positive radial part R, i.e., such that Pr(R ≤ 0) = 0. To this end, recall that a univariate random variable
X with distribution function F is in the maximum domain of attraction of a non-degenerate distribution F0 , denoted
F ∈ M(F0 ) or X ∈ M(F0 ), if and only if there exist sequences of reals (an ) and (bn ) with an > 0, such that, for any
x ∈ R,
lim F n (an x + bn ) = F0 (x).
n→∞
By the Fisher–Tippett Theorem, F0 must be, up to location and scale, either the Fréchet (Φρ ), the Gumbel (Λ) or the
Weibull distribution (Ψρ ) with parameter ρ > 0. Further recall that a measurable function f : R+ → R+ is called
regularly varying with index ρ ∈ (−∞, ∞), denoted f ∈ Rρ , if for any x > 0, f (tx)/ f (t) → xρ as t → ∞. If ρ = 0, f is
called slowly varying. For more details and conditions for F ∈ M(F0 ), see, e.g., [12, 35].
Extremal attractors of Liouville copulas
3
Because the univariate margins of X are scale mixtures of Beta distributions, their extremal behavior, detailed in
Proposition 1, follows directly from Theorems 4.1, 4.4. and 4.5 in [20].
Proposition 1 Let X = RDα be a Liouville random vector with parameters α = (α1 , . . . , αd ) and a strictly positive
radial variable R, i.e., Pr(R ≤ 0) = 0. Then the following statements hold for any ρ > 0:
(a) R ∈ M(Φρ ) if and only if Xi ∈ M(Φρ ) for all i = 1, . . . , d.
(b) R ∈ M(Λ) if and only if Xi ∈ M(Λ) for all i = 1, . . . , d.
(c) R ∈ M(Ψρ ) if and only if Xi ∈ M(Ψρ+ᾱ−αi ) for all i = 1, . . . , d.
Proposition 1 implies that the univariate margins of X are all in the domain of attraction of the same distribution if the
latter is Gumbel or Fréchet. This is not the case when R is in the Weibull domain of attraction. Note also that there are
cases not covered by Proposition 1, in which the univariate margins Xi are in the Weibull domain while R is not in the
domain of attraction of any extreme-value distribution. For example, when d = 2, α = (1, 1) and R = 1 almost surely,
the margins of X are standard uniform and hence in the maximum domain of attraction of Ψ1 ; see Example 3.3.15 in
[12]. At the same time, R is clearly neither in the Weibull, nor the Gumbel, nor the Fréchet domain of attraction.
In subsequent sections, we shall also need the extremal behavior of the univariate margins of 1/X. The proposition
below shows that the latter is determined by the properties of 1/R. In contrast to Proposition 1, however, the univariate
margins of 1/X are always in the Fréchet domain. The proof may be found in A.
Proposition 2 Let X = RDα be a Liouville random vector with parameters α = (α1 , . . . , αd ) and a strictly positive
radial variable R with Pr(R ≤ 0) = 0. The following statements hold for any i = 1, . . . , d.
(a) If 1/R ∈ M(Φρ ) for ρ ∈ (0, αi ], then 1/Xi ∈ M(Φρ ).
(b) If E(1/Rαi +ε ) < ∞ for some ε > 0, then 1/Xi ∈ M(Φαi ).
3. Extremal behavior of Liouville copulas
In this section, we will identify the extremal behavior of a Liouville random vector X = RDα and of the random
vector 1/X, assuming that Pr(R ≤ 0) = 0. As a by-product, we will obtain the extremal attractors of Liouville copulas
and their survival counterparts. To this end, recall that a random vector Y with joint distribution function H is in the
maximum domain of attraction of a non-degenerate distribution function H0 , in notation H ∈ M(H0 ) or Y ∈ M(H0 ),
iff there exist sequences of vectors (an ) in (0, ∞)d and (bn ) in Rd such that for all x ∈ Rd ,
lim H n (an x + bn ) = H0 (x).
n→∞
When the univariate margins F1 , . . . , Fd of H are continuous, H ∈ M(H0 ) holds if and only if Fi ∈ M(F0i ) for all
i = 1, . . . , d, where F01 , . . . , F0d are the univariate margins of H0 , and further if the unique copula C of H is in the
domain of attraction of the unique copula C0 of H0 , denoted C ∈ M(C0 ), i.e., iff for all u ∈ [0, 1]d ,
lim C n (u1/n ) = C0 (u).
n→∞
In particular, the univariate margins of the max-stable distribution H0 must each follow a generalized extreme-value
distribution, and C0 must be an extreme-value copula. This means that for all u ∈ [0, 1]d ,
C0 (u) = exp[−`{− log(u1 ), . . . , − log(ud )}],
(2)
where ` : Rd+ → [0, ∞) is a stable tail dependence function, linked to the so-called exponent measure ν viz.
ν{[0d , x)c } = `(1/x), see, e.g., [35]. The latter Rcan be characterized through an angular (or spectral) probability
measure σd on Sd given in Eq. (1) which satisfies S wi dσd (w) = 1/d for all i = 1, . . . , d. For all x ∈ Rd+ , one has
d
`(x) = d
Z
Sd
max(w1 x1 , . . . , wd xd ) dσd (w).
(3)
4
L. R. Belzile and J. G. Nešlehová
Because ` is homogeneous of order 1, i.e., for any c > 0 and x ∈ Rd+ , `(cx) = c`(x), C0 can also be expressed via the
Pickands dependence function A : Sd → [0, ∞) related to ` through `(x) = kxkA(x/kxk). Then at any u ∈ [0, 1]d ,
)#
"
(
log(ud )
log(u1 )
,...,
.
C0 (u) = exp log(u1 · · · ud )A
log(u1 · · · ud )
log(u1 · · · ud )
When d = 2, it is more common to define the Pickands dependence function A : [0, 1] → [0, 1] through `(x1 , x2 ) =
(x1 + x2 )A{x2 /(x1 + x2 )} so that, for all u1 , u2 ∈ [0, 1],
"
(
)#
log(u2 )
C0 (u1 , u2 ) = exp log(u1 u2 )A
.
(4)
log(u1 u2 )
Now consider a Liouville vector X = RDα with a strictly positive radial variable. Theorem 1 specifies when
X ∈ M(H0 ) and identifies H0 . While part (a) follows from regular variation of X, parts (b) and (c) are special cases
of the results discussed in Section 2.2 in [19]. Details of the proof may be found in B.
Theorem 1 Let X = RDα , Dα = (D1 , . . . , Dd ), α > 0d , and Pr(R ≤ 0) = 0. Then the following statements hold.
(a) If R ∈ M(Φρ ) for some ρ > 0, then X ∈ M(H0 ), where H0 is a multivariate extreme-value distribution with
univariate margins F0i = Φρ , i = 1, . . . , d, and a stable tail dependence function given, for all x ∈ Rd+ , by
ρ
Γ(αd )xd Dρd
Γ(ᾱ + ρ)
Γ(α1 )x1 D1
E max
,...,
`(x) =
.
Γ(ᾱ)
Γ(α1 + ρ)
Γ(αd + ρ)
(b) If R ∈ M(Λ), then X ∈ M(H0 ), where for all x ∈ Rd , H0 (x) =
Qd
i=1
Λ(xi ).
(c) If R ∈ M(Ψρ ) for some ρ > 0, then X ∈ M(H0 ), where for all x ∈ Rd , H0 (x) =
Qd
i=1
Ψρ+ᾱ−αi (xi ).
The next result, also proved in B, specifies the conditions under which 1/X ∈ M(H0 ) and gives the form of the limiting
extreme-value distribution H0 .
Theorem 2 Let X = R Dα , Dα = (D1 , . . . , Dd ), α > 0d , and assume that Pr(R ≤ 0) = 0. Let αM = max(α1 , . . . , αd ).
The following cases can be distinguished:
P
(a) If 1/R ∈ M(Φρ ) for ρ ∈ (0, αM ], set I1 = {i : αi ≤ ρ}, I2 = {i : αi > ρ} and ᾱ2 = i∈I2 αi . Then
1/X ∈ M(H0 ), where the univariate margins of H0 are F0i = Φρ∧αi , i = 1, . . . , d, and the stable tail dependence
function is given, for all x ∈ Rd+ , by
−ρ
X
e−ρ
Γ(ᾱ2 − ρ)
Γ(ᾱ − ρ)
Γ(αi )xi Di
X
Γ(αi )xi D
i
,
E max
xi +
E max
xi +
`(x) =
=
i∈I2 Γ(αi − ρ)
i∈I2 Γ(αi − ρ)
Γ(ᾱ)
Γ(ᾱ2 )
i∈I1
i∈I1
ei , i ∈ I2 ) is a Dirichlet random vector with parameters (αi , i ∈ I2 ) if |I2 | > 1 and D
ei ≡ 1 if I2 = {i}.
where (D
Q
(b) If E 1/Rβ < ∞ for β > αM , then 1/X ∈ M(H0 ), where for all x ∈ Rd , H0 (x) = di=1 Φαi (xi ).
Remark 1 Note that in the case of asymptotic independence between the components of X (Theorem 1 (b–c)) or
1/X (Theorem 2 (b)), dependence between component-wise maxima of finitely many vectors may still be present.
Refinements of asymptotic independence are then needed, but these considerations surpass the scope of this paper.
One option would be to consider triangular arrays as in [23]; extremes of arrays of Liouville vectors can be obtained
as a special case of extremes of arrays of weighted Dirichlet distributions developed in [18]. Another avenue worth
exploring might be the limits of scaled sample clouds, as in [2] and [32].
The stable tail dependence functions appearing in Theorems 1 and 2 will be investigated in greater detail in the
subsequent sections. Before proceeding, we introduce the following terminology, emphasizing that they can in fact be
embedded in one and the same parametric class.
Extremal attractors of Liouville copulas
5
Definition 1 For any α > 0 and ρ ∈ (−α, ∞), let c(α, ρ) = Γ(α + ρ)/Γ(α) denote the rising factorial. For d ≥ 2
and α1 , . . . , αd > 0 and let (D1 , . . . , Dd ) denote a Dirichlet random vector with parameters α = (α1 , . . . , αd ) and set
ᾱ = α1 + · · · + αd . For any − min(α1 , . . . , αd ) < ρ < ∞, the scaled extremal Dirichlet stable tail dependence function
with parameters ρ and α is given, for all x ∈ Rd+ , by
ρ
xd Dρd
x 1 D1
,
.
.
.
,
`D (x; ρ, α) = c(ᾱ, ρ)E max
,
c(α1 , ρ)
c(αd , ρ)
(5)
when ρ , 0 and by max(x1 , . . . , xd ) when ρ = 0. For any ρ > 0, the positive scaled extremal Dirichlet stable tail
dependence function `pD with parameters ρ and α is given, for all x ∈ Rd+ , by `pD (x; ρ, α) = `D (x; ρ, α), while for any
0 < ρ < min(α1 , . . . , αd ), the negative scaled extremal Dirichlet stable tail dependence function `nD is given, for all
x ∈ Rd+ , by `nD (x; ρ, α) = `D (x; −ρ, α).
Remark 2 As will be seen in Section 5, distinguishing between the positive and negative scaled extremal Dirichlet
models makes the discussion of their properties slightly easier because the sign of ρ impacts the shape of the corresponding angular measure. When ρ → 0, `D (x; ρ, α) becomes max(x1 , . . . , xd ), the stable tail dependence function
corresponding to comonotonicity, while when ρ → ∞, `D (x; ρ, α) becomes x1 + · · · + xd , the stable tail dependence
function corresponding to independence. Note also that ρ ∈ (−∞, ∞) can be allowed, with the convention that all
variables whose indices i are such that ρ ≤ −αi are independent, i.e., `nD is then of the form given in Theorem 2 (a).
From Theorems 1 and 2, we can now easily deduce the extremal behavior of Liouville copulas and their survival
counterparts. To this end, recall that a Liouville copula C is defined as the survival copula of a Liouville random
vector X = RDα with Pr(R ≤ 0) = 0. The following corollary follows directly from Theorem 2 upon noting that C is
also the unique copula of 1/X.
Corollary 1 Let C be the unique survival copula of a Liouville random vector X = RDα with Pr(R ≤ 0) = 0. Let
αM = max(α1 , . . . , αd ) and set I1 = {i : αi ≤ ρ}, I2 = {i : αi > ρ}. Then the following statements hold.
(a) If 1/R ∈ M(Φρ ) for ρ ∈ (0, αM ] and |I2 | > 1, then C ∈ M(C0 ), where C0 is an extreme-value copula of the form
(2) whose stable tail dependence function is given, for all x ∈ Rd+ , by
X
xi + `nD (x{2} ; ρ, α{2} ),
`(x) =
i∈I1
where and x{2} = (xi , i ∈ I2 ), α{2} = (αi , i ∈ I2 ). If 1/R ∈ M(Φρ ) for ρ ∈ (0, αM ] and |I2 | ≤ 1, then C ∈ M(Π),
where Π is the independence copula given, for all u ∈ [0, 1]d , by Π(u) = u1 · · · ud .
(b) If E 1/Rβ < ∞ for β > αM , then C ∈ M(Π).
Remark 3 Observe that Corollary 1 (a) in particular implies that when d = 2, α1 < α2 and 1/R ∈ M(Φρ ) for
α1 ≤ ρ < α2 , C ∈ M(Π). Also note that when α1 = · · · = αd ≡ α and 1/R ∈ M(Φρ ) for ρ ∈ (0, α), the result in
Corollary 1 (a) can be derived from formula (5) in Proposition 3 in [21] by relating the tail order function to the stable
tail dependence function when the tail order equals 1.
The survival counterpart Ĉ of a Liouville copula C is given as the distribution function of 1 − U, where U is a
random vector distributed as C. As C is the unique survival copula of X, Ĉ is the unique copula of X. The following
result thus follows directly from Theorem 1.
Corollary 2 Let Ĉ be the unique copula of a Liouville random vector X = R Dα with Pr(R ≤ 0) = 0. Then the
following statements hold.
(a) If R ∈ M(Φρ ) for ρ > 0, then Ĉ ∈ M(C0 ), where C0 is an extreme-value copula of the form (2) with the positive
scaled extremal Dirichlet stable tail dependence function given, for all x ∈ Rd+ , by `pD (x; ρ, α).
(b) If R ∈ M(Λ) or R ∈ M(Ψρ ) with ρ > 0, then Ĉ ∈ M(Π), where Π is the independence copula.
6
L. R. Belzile and J. G. Nešlehová
4. The case of integer-valued Dirichlet parameters
When α is integer-valued, Liouville distributions are particularly tractable because their survival function is explicit.
In this section, we will use this fact to derive closed-form expressions for the positive and negative scaled extremal
Dirichlet stable tail dependence functions. To this end, first recall the notion of the Williamson transform. The latter
is related to Weyl’s fractional integral transform and was used to characterize d-monotone functions in [43]; it was
adapted to non-negative random variables in [30].
Definition 2 Let X be a non-negative random variable with distribution function F, and let k ≥ 1 be an arbitrary
integer. The Williamson k-transform of X is given, for all x > 0, by
Wk F(x) =
∞
Z
x
x k−1
x k−1
dF(r) = E 1 −
.
1−
r
X +
For any k ≥ 1, the distribution of a positive random variable X is uniquely determined by its Williamson k-transform,
the formula for the inverse transform being explicit [30, 43]. If ψ = Wk F, then, for all x > 0,
F(x) = W−1
k ψ(x) = 1 −
k−2
X
(−1) j x j ψ( j) (x)
j!
j=0
−
(−1)k−1 xk−1 ψ+(k−1) (x)
,
(k − 1)!
where for j = 1, . . . , k − 2, ψ( j) is the jth derivative of ψ and ψ(k−1)
is the right-hand derivative of ψ(k−2) . These deriva+
tives exist because a Williamson k-transform ψ is necessarily k-monotone [43]. This means that ψ is differentiable up
to order k − 2 on (0, ∞) with derivatives satisfying (−1) j ψ( j) ≥ 0 for j = 0, . . . , k − 2 and such that (−1)k−2 ψ(k−2) is
non-increasing and convex on (0, ∞). Moreover, ψ(x) → 0 as x → ∞ and if F(0) = 0, ψ(x) → 1 and x → 0.
Now let C be a Liouville copula corresponding to a Liouville random vector X = RDα with integer-valued
parameters α = (α1 , . . . , αd ) and a strictly positive radial part R, i.e., Pr(R ≤ 0) = 0. Let ψ be the Williamson
ᾱ-transform of R and set Iα = {0, . . . , α1 − 1} × · · · × {0, . . . , αd − 1}. By Theorem 2 in [31], one then has, for all x ∈ Rd+ ,
Pr(X > x) = H̄(x) =
X
ψ( j1 +···+ jd ) (x1 + · · · + xd ) Y ji
xi .
j1 ! · · · jd !
i=1
d
(−1) j1 +···+ jd
( j1 ,..., jd )∈Iα
(6)
In particular, the margins of X have survival functions satisfying, for all x > 0 and i = 1, . . . , d,
Pr(Xi > x) = H̄i (x) =
α
i −1
X
j=0
(−1) j x j ψ( j) (x)
= 1 − W−1
αi ψ(x).
j!
(7)
By Sklar’s Theorem for survival functions, the Liouville copula C is given, for all u ∈ [0, 1]d , by
C(u) = H̄{H̄1−1 (u1 ), . . . , H̄d−1 (ud )}.
Although this formula is not explicit, it is clear from Equations (6) and (7) that C depends on the distribution of X
only through the Williamson ᾱ-transform ψ of R and the Dirichlet parameters α. For this reason, we shall denote
the Liouville copula in this section by Cψ,α and refer to ψ as its generator, reiterating that ψ must be an ᾱ-monotone
function satisfying ψ(1) = 0 and ψ(x) → 0 as x → ∞. When α = 1d , Cψ,1 is the Archimedean copula with generator
ψ, given, for all u ∈ [0, 1]d by Cψ,1 (u) = ψ{ψ−1 (u1 ) + · · · + ψ−1 (ud )}. Because the relationship between ψ and R is
one-to-one [30, Proposition 3.1], we will refer to R as the radial distribution corresponding to ψ.
Now suppose that 1/R ∈ M(Φρ ) with ρ ∈ (0, 1). By Theorem 2 in [27], this condition is equivalent to 1 − ψ(1/·) ∈
R−ρ . It further follows from Corollary 1 (a) that Cψ,α ∈ M(C0 ) where C0 is an extreme-value copula with the negative
scaled extremal Dirichlet stable tail dependence function `nD (·; ρ, α). This is because ρ < 1 ≤ min(α1 , . . . , αd ) so that
I1 = ∅ in Corollary 1 (a). Eq. (6) and the results of [27] can now be used to derive the following explicit expression
for `nD , as detailed in C.
7
Extremal attractors of Liouville copulas
Proposition 3 Let Cψ,α be a Liouville copula with integer-valued parameters α = (α1 , . . . , αd ) and generator ψ. If
1 − ψ(1/·) ∈ R−ρ for some ρ ∈ (0, 1), then Cψ,α ∈ M(C0 ), where C0 is an extreme-value copula with scaled negative
extremal Dirichlet stable tail dependence function `nD as given in Definition 1. Furthermore, for all x ∈ Rd+ ,
n
o1/ρ ji
ρ
d (
)
xi
d
1/ρ
X Γ( j1 + · · · + jd − ρ) Y
X
xj
1
c(αi ,−ρ)
.
1 − ρ
`nD (x; ρ, α) = Γ(1 − ρ)
c(α j , −ρ)
Γ(1 − ρ)
Γ( ji + 1) Pd n xk o1/ρ
( j1 ,..., jd )∈Iα
i=1
j=1
k=1 c(α ,−ρ)
( j1 ,..., jd ),(0,...,0)
k
When α = 1d , the index set Iα reduces to the singleton {0}, and the expression for `nD given in Proposition 3 simplifies,
for all x ∈ Rd+ , to the stable tail dependence function of the Gumbel–Hougaard copula, viz.
`nD (x; ρ, 1d ) = x11/ρ + · · · + xd1/ρ ρ .
The Liouville copula Cψ,1d , which is the Archimedean copula with generator ψ, is thus indeed in the domain of
attraction of the Gumbel–Hougaard copula with parameter 1/ρ, as shown, e.g., in [6, 27].
Remark 4 When α = 1d and 1 − ψ(1/·) ∈ R−1 , it is shown in Proposition 2 of [27] that Cψ,1 is in the domain of
attraction of the independence copula. However, when α is integer-valued but such that max(α1 , . . . , αd ) > 1, regular
variation of 1 − ψ(1/·) does not suffice to characterize those cases in Corollary 1 that are not covered by Proposition 3.
This is because by Theorem 2 of [27], 1/R ∈ M(Φρ ) for ρ ≥ 1, 1/R ∈ M(Λ) and 1/R ∈ M(Ψρ ) for ρ > 0 all imply
that 1 − ψ(1/·) ∈ R−1 . At the same time, by Corollary 1, Cψ,α ∈ M(Π) clearly does not hold in all these cases.
Next, let Ĉψ,α be the survival copula of a Liouville copula Cψ,α , i.e., the distribution function of 1 − U, where U is a
random vector with distribution function Cψ,α . The results of [27] can again be used to restate the conditions under
which Ĉψ,α ∈ M(C0 ) in terms of ψ and to give an explicit expression for the stable tail dependence function of C0 .
Proposition 4 Let Ĉψ,α be the survival copula of a Liouville copula Cψ,α with integer-valued parameters α and a
generator ψ. Then the following statements hold.
(a) If ψ ∈ R−ρ for some ρ > 0, then Ĉψ,α ∈ M(C0 ), where C0 has a positive scaled extremal Dirichlet stable tail
dependence function `pD as given in Definition 1. The latter can be expressed, for all x ∈ Rd+ , as
−1/ρ −ρ
d
k
X
X
X
x
Γ(1
+
ρ)
i
j
(−1)k+1
`pD (x; ρ, α) =
×
Γ(ρ) k=1 1≤i <···<i ≤d
j=1 c(αi j , ρ)
1
k
n
o−1/ρ jm
xim
k
Y
Γ( j1 + · · · + jk + ρ)
c(αim , ρ)
.
−1/ρ
Pk
xi j
j
!
·
·
·
j
!
1
k
( j1 ,..., jk )∈I(αi ,...,αi )
m=1
1
k
X
j=1
c(αi j , ρ)
(b) If ψ ∈ M(Λ) or ψ ∈ M(Ψρ ) for some ρ > 0, C̄ψ,α ∈ M(Π), where Π is the independence copula.
When α = 1d , the expression for `pD in part (a) of Proposition 4 simplifies, for all x ∈ Rd+ , to
X
X −1/ρ −ρ
pD
|A|+1
` (x; ρ, 1d ) =
(−1)
x
,
i
A⊆{1,...,d}, A,∅
i∈A
which is the stable tail dependence function of the Galambos copula [24]. When ψ ∈ R−ρ for some ρ > 0, Ĉψ,1d is thus
indeed in the domain of attraction of the Galambos copula, as shown, e.g., in [27].
5. Properties of the scaled extremal Dirichlet models
In this section, the scaled extremal Dirichlet model with stable tail dependence function given in Definition 1 is
investigated in greater detail. In Section 5.1 we derive formulas for the so-called angular density and relate the
positive and negative scaled extremal Dirichlet models to classical classes of stable tail dependence functions. In
Section 5.2 we focus on the bivariate case and derive explicit expressions for the stable tail dependence functions and,
as a by-product, obtain formulas for the tail dependence coefficients of Liouville copulas.
8
L. R. Belzile and J. G. Nešlehová
scaled Dirichlet, ρ < 0
2.0
1.5
0.0
0.0
0.5
1.0
Angular density
2.0
1.0
Angular density
3.0
2.5
3.0
scaled Dirichlet, ρ > 0
0.0
0.2
0.4
0.6
w
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
w
Figure 1: Angular density of the scaled extremal Dirichlet model. Left panel: ρ = 4/5 and α = (2, 1/2) (black full),
ρ = 1/4 and α = (1/10, 1/10) (red dashed), ρ = 1/4 and α = (1/2, 1/2) (blue dotted). Right panel: ρ = −1/4 and
α = (2, 1/2) (black full), α = (2/5, 2/5) (red dashed) and α = (1/2, 1/2) (blue dotted).
5.1. Angular density
The first property worth noting is that the positive and negative scaled extremal Dirichlet models are closed under
marginalization. Indeed, letting xi → 0 for some arbitrary 1 ≤ i ≤ d, we can easily derive from Lemma 2 that for
any α > 0d , ρ > 0, and any x ∈ Rd+ , `pD (x; ρ, α) → `pD (x−i ; ρ, α−i ) as xi → 0, where for any y ∈ Rd , y−i denotes the
vector (y1 , . . . , yi−1 , yi+1 , . . . , yd ). Similarly, for any 1 ≤ i ≤ d, α > 0d , 0 < ρ < min(α1 , . . . , αd ), and any x ∈ Rd+ ,
`nD (x; ρ, α) → `nD (x−i ; ρ, α−i ) as xi → 0.
Since none of the scaled extremal Dirichlet models places mass on the vertices or facets of the simplex Sd when
ρ , 0, the density of the angular measure σd completely characterizes the stable tail dependence function and hence
also the associated extreme-value copula. This so-called angular density of the scaled extremal Dirichlet models is
given below and derived in D.
Proposition 5 Let d ≥ 2 and set α1 , . . . , αd > 0 and ᾱ = α1 + · · · + αd . For any ρ > − min(α1 , . . . , αd ), let also
c(α, ρ) = (c(α1 , ρ), . . . , c(αd , ρ)), where c(α, ρ) is as in Definition 1. Then for any − min(α1 , . . . , αd ) < ρ < ∞, ρ , 0,
the angular density of the scaled extremal Dirichlet model with parameters ρ > 0 and α is given, for all w ∈ Sd , by
d
−ρ−ᾱ d
Y
X n
o1/ρ
Γ(
ᾱ
+
ρ)
hD (w; ρ, α) =
c(α
,
ρ)w
{c(αi , ρ)}αi /ρ wiαi /ρ−1 .
Q
j
j
d
d−1
d|ρ|
i=1 Γ(αi ) j=1
i=1
The angular density of the positive scaled extremal Dirichlet model with parameters ρ > 0 and α is given, for all
w ∈ Sd , by hpD (w; ρ, α) = hD (w; ρ, α), while the angular density of the negative scaled extremal Dirichlet model with
parameters 0 < ρ < min(α1 , . . . , αd ) and α is given, for all w ∈ Rd+ , by hnD (w; ρ, α) = hD (w; −ρ, α).
From Proposition 5, it is easily seen that when α = 1d , the angular density hpD reduces, for any ρ > 0 and w ∈ Sd ,
to the angular density of the symmetric negative logistic model; see, e.g., Section 4.2 in [7]. In general, the angular
density hpD is not symmetric unless α = α1d .
The positive scaled Dirichlet model can thus be viewed as a new asymmetric generalization of the negative logistic
model which does not place any mass on the vertices or facets of Sd , unless at independence or comonotonicity, i.e.,
when ρ → ∞ and ρ → 0, respectively. Furthermore, hpD can also be interpreted as a generalization of the Coles–
Tawn extremal Dirichlet model. Indeed, hpD (x; 1, α) is precisely the angular density of the latter model given, e.g., in
Equation (3.6) in [7]. Similarly, the negative scaled extremal Dirichlet model is a new asymmetric generalization of
Gumbel’s logistic model [16]. Indeed, when α = 1d , hnD simplifies to the logistic angular density, given, e.g., on p.
381 in [7].
9
Extremal attractors of Liouville copulas
scaled Dirichlet,
α = (1, 1/2, 1/5), ρ = 1/5
scaled Dirichlet,
α = (5/4, 2, 1), ρ = −2/5
w3 = 1
w3 = 1
w1 = 1
w1 = 1
w2 = 1
w1 = 1
w2 = 1
scaled Dirichlet,
α = (1/5, 1/5, 1/5), ρ = 1/5
scaled Dirichlet,
α = (5/4, 5/4, 5/4), ρ = −2/5
w3 = 1
w3 = 1
w2 = 1
w1 = 1
w2 = 1
Figure 2: Angular density of the scaled extremal Dirichlet model with α = (1, 1/2, 1/5), ρ = 1/5 (top left) and
α = (1/5, 1/5, 1/5), ρ = 1/5 (bottom left), α = (5/4, 2, 1), ρ = −2/5 (top right) and α = (5/4, 5/4, 5/4), ρ = −2/5
(bottom right). The colors correspond to log density values and range from red (high density) to blue (low density).
Figures 1 and 2 illustrate the various shapes of hpD and hnD that obtain through various choices of α and ρ. The
asymmetry when α , α1d is clearly apparent. For the same value of ρ, the shapes of the angular density can be quite
different depending on α. In view of the aforementioned closure of both the positive and negative scaled extremal
Dirichlet models under marginalization, this means that these models are able to capture strong dependence in some
pairs of variables (represented by a mode close to 1/2 of the angular density) and at the same time weak dependence
in others pairs (represented by a bathtub shape).
5.2. The bivariate case
When d = 2, the stable tail dependence functions of the positive and negative scaled extremal Dirichlet models have
a closed-form expression in terms of the incomplete beta function given, for any t ∈ (0, 1) and α1 , α2 > 0, by
B(t; α1 , α2 ) =
Z
0
t
xα1 −1 (1 − x)α2 −1 dx.
When t = 1, this integral is the beta function, viz. B(α1 , α2 ) = Γ(α1 )Γ(α2 )/Γ(α1 + α2 ). A direct calculation yields the
corresponding Pickands dependence function, for any t ∈ [0, 1], ApD (t; ρ, α1 , α2 ) = `pD (1 − t, t; ρ, α1 , α2 ), i.e.,
10
L. R. Belzile and J. G. Nešlehová
scaled Dirichlet, ρ < 0
1.0
1.0
0.9
0.9
0.8
0.8
A(t)
A(t)
scaled Dirichlet, ρ > 0
0.7
0.7
0.6
0.6
0.5
0.5
0.0
0.2
0.4
0.6
0.8
1.0
0.0
t
0.2
0.4
0.6
0.8
1.0
t
Figure 3: Pickands dependence function of the scaled extremal Dirichlet model. Left panel: ρ = 4/5 and α = (2, 1/2)
(black full), ρ = 1/4 and α = (1/10, 1/10) (red dashed), ρ = 1/4 and α = (1/2, 1/2) (blue dotted). Right panel:
ρ = −1/4 and α = (2, 1/2) (black full), α = (2/5, 2/5) (red dashed) and α = (1/2, 1/2) (blue dotted).
ApD (t; ρ, α1 , α2 ) =
"
#
(1 − t)
{c(α2 , ρ)(1 − t)}1/ρ
B
;
α
,
α
+
ρ
2
1
B(α2 , α1 + ρ) {c(α2 , ρ)(1 − t)}1/ρ + {c(α1 , ρ)t}1/ρ
"
#
t
{c(α1 , ρ)t}1/ρ
B
+
; α1 , α2 + ρ .
B(α1 , α2 + ρ) {c(α2 , ρ)(1 − t)}1/ρ + {c(α1 , ρ)t}1/ρ
When α1 = α2 = 1, ApD becomes the Pickands dependence function of the Galambos copula, viz. ApD (t; ρ, 1, 1) =
1 − t−1/ρ + (1 − t)−1/ρ −ρ , as expected given that the positive scaled extremal Dirichlet model becomes the symmetric
negative logistic model in this case.
Similarly, for any t ∈ [0, 1], the Pickands dependence function AnD (t; ρ, α1 , α2 ) = `nD (1 − t, t; ρ, α1 , α2 ) equals
AnD (t; ρ, α1 , α2 ) =
"
#
(1 − t)
{(1 − t)c(α2 , −ρ)}1/ρ
B
;
α
−
ρ,
α
1
2
B(α1 − ρ, α2 ) {tc(α1 , −ρ)}1/ρ + {(1 − t)c(α2 , −ρ)}1/ρ
#
"
t
{tc(α1 , −ρ)}1/ρ
; α2 − ρ, α1 .
+
B
B(α2 − ρ, α1 ) {tc(α1 , −ρ)}1/ρ + {(1 − t)c(α2 , −ρ)}1/ρ
When α1 = α2 = 1, AnD simplifies to the stable tail dependence function of the Gumbel extreme-value copula, viz.
AnD (t; ρ, 1, 1) = t1/ρ + (1 − t)1/ρ ρ . This again confirms that the negative scaled extremal Dirichlet model becomes
the symmetric logistic model when α1 = α2 = 1. The Pickands dependence functions ApD and AnD are illustrated in
Figure 3, for the same choices of parameters and the corresponding angular density shown in Figure 1.
The above formulas for ApD and AnD now easily lead to expressions for their upper tail dependence coefficients.
Recall that for an arbitrary bivariate copula C, the lower and upper tail dependence coefficients of [25] are given by
C(u, u)
,
u→0
u
λ` (C) = lim
C(u, u) − 1
Ĉ(u, u)
= lim
,
u→1
u→0
u−1
u
λu (C) = 2 − lim
where Ĉ is the survival copula of C, provided these limits exist. When C is bivariate extreme-value with Pickands
dependence function A, it follows easily from (4) that λ` (C) = 0 and λu (C) = 2 − 2A(1/2).
Extremal attractors of Liouville copulas
11
pD
Now suppose that Cρ,α is a bivariate extreme-value copula with positive scaled extremal Dirichlet Pickands dependence function ApD and parameters ρ > 0 and α1 , α2 > 0. Then
pD
λu (Cρ,α )
(
)
c(α2 , ρ)1/ρ
1
B
; α2 , α1 + ρ
=2−
B(α2 , α1 + ρ)
c(α2 , ρ)1/ρ + c(α1 , ρ)1/ρ
(
)
1
c(α1 , ρ)1/ρ
−
B
;
α
,
α
+
ρ
. (8)
1
2
B(α1 , α2 + ρ)
c(α2 , ρ)1/ρ + c(α1 , ρ)1/ρ
nD
Similarly, if Cρ,α
is a bivariate extreme-value copula with negative scaled extremal Dirichlet Pickands dependence
nD
function A and parameters α1 , α2 > 0 and 0 < ρ < min(α1 , α2 ), then
nD
λu (Cρ,α
)=2−
(
)
1
c(α2 , −ρ)1/ρ
B
;
α
−
ρ,
α
1
2
B(α1 − ρ, α2 )
c(α1 , −ρ)1/ρ + c(α2 , −ρ)1/ρ
)
(
1
c(α1 , −ρ)1/ρ
;
α
−
ρ,
α
+
B
2
1 . (9)
B(α2 − ρ, α1 )
c(α1 , −ρ)1/ρ + c(α2 , −ρ)1/ρ
In the symmetric case α1 = α2 ≡ α, Expressions (8) and (9) simplify to
pD
λu (Cρ,α )
!
1
2
=2−
B ; α, α + ρ ,
B(α, α + ρ) 2
nD
λu (Cρ,α
)
!
2
1
=2−
B ; α − ρ, α .
B(α − ρ, α) 2
Formulas (8) and (9) lead directly to expressions for the tail dependence coefficients of Liouville copulas. This is
because if C ∈ M(C0 ), where C0 is an extreme-value copula with Pickands tail dependence function A0 , λu (C) =
2 − 2A0 (1/2) [29, Proposition 7.51]. Similarly, if Ĉ ∈ M(C0∗ ), where C0∗ is an extreme-value copula with Pickands
tail dependence function A∗0 , λ` (C) = 2 − 2A∗0 (1/2). The following corollary is thus an immediate consequence of
Corollaries 1 and 2.
Corollary 3 Suppose that C is the survival copula of a Liouville random vector RDα with parameters α > 0 and a
radial part R such that Pr(R ≤ 0) = 0. Then the following statements hold.
pD
(a) If R ∈ M(Φρ ) for some ρ > 0, λ` (C) = λu (Cρ,α ) is given by Eq. (8).
(b) If R ∈ M(Λ) or R ∈ M(Ψρ ) for some ρ > 0, λ` (C) = 0.
nD
(c) If 1/R ∈ M(Φρ ) for some 0 < ρ < α1 ∧ α2 , λu (C) = λu (Cρ,α
) is given by Eq. (9).
(d) If 1/R ∈ M(Φρ ) for ρ > α1 ∧ α2 or if E(1/Rβ ) < ∞ for β > α1 ∨ α2 , λu (C) = 0.
The role of the parameters α and ρ is best explained if we consider the reparametrization ∆α = |α1 − α2 | and
Σα = α1 + α2 . As is the case for the Dirichlet distribution, the level of dependence is higher for large values of Σα .
Furthermore, λu is monotonically decreasing in ρ. Higher levels of extremal asymmetry, as measured by departures
from the diagonal on the copula scale, are governed by both Σα and ∆α . The larger Σα , the lower the asymmetry.
Likewise, the larger ∆α , the larger the asymmetry. Contrary to the case of extremal dependence, the behavior in ρ is
not monotone. For the negative scaled extremal Dirichlet model, asymmetry is maximal when ρ ≈ α1 ∧ α2 . When
Σα is small, smaller values of ρ induce larger asymmetry, but this is not the case for larger values of Σα where the
asymmetry profile is convex with a global maximum attained for larger values of ρ.
6. de Haan representation and simulation algorithms
Random samples from the scaled extremal Dirichlet model can be drawn efficiently using the algorithms recently
developed in [10]. We first derive the so-called de Haan representation in Section 6.1 and adapt the algorithms from
[10] to the present setting in Section 6.2.
12
L. R. Belzile and J. G. Nešlehová
6.1. de Haan representation
First, introduce the following family of univariate distributions, which we term the scaled Gamma family and denote
by sGa(a, b, c). It has three parameters a, c > 0 and b , 0 and a density given, for all x > 0, by
( b )
|b| −bc bc−1
x
f (x; a, b, c) =
a x
exp −
.
(10)
Γ(c)
a
d
Observe that when Z ∼ Ga(c, 1) is a Gamma variable with shape parameter c > 0 and scaling parameter 1, Y = aZ 1/b is
scaled Gamma sGa(a, b, c). Consequently, E (Y) = aΓ(c + 1/b)/Γ(c) < ∞ provided that b < −1/c. The scaled Gamma
family includes several well-known distributions as special cases, notably the Gamma when b = 1, the Weibull when
c = 1 and b > 0, the inverse Gamma when b = −1, and the Fréchet when c = 1 and b < 0. When b > 0, the scaled
Gamma is the generalized Gamma distribution of [41], albeit in a different parametrization.
Now consider the parameters α = (α1 , . . . , αd ) with α > 0d and ρ > − min(α1 , . . . , αd ), ρ , 0. Let V be a random
vector with independent scaled Gamma margins Vi ∼ sGa{1/c(αi , ρ), 1/ρ, αi }, where for α > 0, c(α, ρ) = Γ(α+ρ)/Γ(α)
as in Definition 1. If Z is a random vector with independent Gamma margins Zi ∼ Ga(αi , 1) then for all i = 1, . . . , d,
d
Vi = Ziρ /c(αi , ρ). Furthermore, recall that kZk ∼ Ga(ᾱ, 1) is independent of Z/kZk, which has the same distribution
as the Dirichlet vector Dα = (D1 , . . . , Dd ). One thus has, for all x ∈ Rd+ ,
)#
)#
"
(
"
(
xi Dρi
xi Ziρ
ρ
= E (kZk ) E max
= `D (x; ρ, α),
(11)
E max (xi Vi ) = E max
1≤i≤d c(αi , ρ)
1≤i≤d
1≤i≤d c(αi , ρ)
where `D is as in Definition 1, given that E (kZkρ ) = c(ᾱ, ρ).
When ρ = 1, the positive scaled Dirichlet extremal model becomes the Coles–Tawn Dirichlet extremal model,
Vi ∼ Ga(αi , 1) and Eq. (11) reduces to the representation derived in [37]. When α = 1d , `D becomes the stable tail
dependence function of the negative logistic model, Vi is Weibull and Eq. (11) is the representation in Appendix A.2.4
of [10]. Similarly, when ρ < 0 and α = 1d , the negative scaled Dirichlet extremal model becomes the logistic model,
Vi is Fréchet and Eq. (11) is the representation in Appendix A.2.4 of [10]. The requirement that ρ > − min(α1 , . . . , αd )
ensures that the expectation of Vi is finite for all i ∈ {1, . . . , d}.
Eq. (11) implies that the max-stable random vector Y with unit Fréchet margins and extreme-value copula with
stable tail dependence function `D (·; ρ, α) admits the de Haan [9] spectral representation
d
Y = max ζk V k ,
k∈N
(12)
−2
where Z = {ζk }∞
dζ and V k is an i.i.d. sequence of rank=1 is a Poisson point process on (0, ∞) with intensity ζ
dom vectors independent of Z. Furthermore, the univariate margins of V k are independent and such that Vk j ∼
sGa{1/c(α j , ρ), 1/ρ, α j } for j = 1, . . . , d with E (V k ) = 1d for all k ∈ N.
6.2. Unconditional simulation
The de Haan representation (12) offers, among other things, an easy route to unconditional simulation of max-stable
random vectors that follow the scaled Dirichlet extremal model, as laid out in [10] in the more general context of maxstable processes. To see how this work applies in the present setting, fix an arbitrary j0 ∈ {1, . . . , d} and recall that the
j0 th extremal function φ+j0 is given, almost surely, as ζk V k such that Y j0 = ζk Vk j0 . From Eq. (12) and Proposition 1 in
d
[10] it then directly follows that φ+j0 /Y j0 = (W j0 1 /W j0 j0 , . . . , W j0 d /W j0 j0 ), where W j0 = (W j0 1 , . . . , W j0 d ) is a random
vector with density given, for all x ∈ Rd+ , by
d
Y
h
i
h
i
|1/ρ|
|1/ρ|
α j /ρ
α /ρ−1
c(α j0 , ρ)α j0 /ρ x j0 0 exp −{c(α j0 , ρ)x j0 }1/ρ ×
c(α j , ρ)α j /ρ x j j
exp −{c(α j , ρ)x j }1/ρ .
Γ(α j0 )
Γ(α j )
j=1, j, j
0
This means that the components of W j0 are independent and such that W j0 j ∼ sGa{1/c(α j , ρ), 1/ρ, α j } when j , j0
and W j0 j0 ∼ sGa{1/c(α j0 , ρ), 1/ρ, α j0 + ρ}. In other words, W j0 j0 ∼ Z ρj0 /c(α j0 , ρ) where Z j0 ∼ Ga(α j0 + ρ, 1), while for
all j , j0 , W j0 j = Z ρj /c(α j , ρ) where Z j ∼ Ga(α j , 1).
d
Extremal attractors of Liouville copulas
13
The exact distribution of φ+j0 /Y j0 given above now allows for an easy adaptation of the algorithms in [10]. To
draw an observation from the extreme-value copula with the scaled Dirichlet stable tail dependence function `D with
parameters α > 0d and ρ > − min(α1 , . . . , αd ), ρ , 0, one can follow Algorithms 1 and 2 below. The first procedure
corresponds to Algorithm 1 in [10] and relies on [36]; the second is an adaptation of Algorithm 2 in [10].
Algorithm 1 Exact simulations from the extreme-value copula based on spectral densities.
1: Simulate E ∼ Exp(1) .
2: Set Y = 0.
3: while 1/E > min(Y1 , . . . , Yd ) do
4:
Simulate J from the uniform distribution on {1, . . . , d}.
5:
Simulate independent Z j ∼ Ga(α j , 1) for j ∈ {1, . . . , d} \ J and Z J ∼ Ga(α J + ρ, 1).
6:
Set W j ← Z ρj /c(α j , ρ), j = 1, . . . , d.
7:
Set S ← W/kWk.
8:
Update Y ← max{Y, dS/E}.
9:
Simulate E ∗ ∼ Exp(1) and update E ← E + E ∗ .
10: return U = exp(−1/Y).
Algorithm 2 Exact simulations based on sequential sampling of the extremal functions.
Simulate Z1 ∼ Ga(α1 + ρ, 1) and Z j ∼ Ga(α j , 1), j = 2, . . . , d.
Compute W where W j ← Z ρj /c(α j , ρ), j = 1, . . . , d.
Simulate E1 ∼ Exp(1).
Set Y ← W/(W1 E1 ).
for k = 2, . . . , d do
Simulate Ek ∼ Exp(1).
while 1/Ek > Yk do
Simulate independent Zk ∼ Ga(αk + ρ, 1) and Z j ∼ Ga(α j , 1), j = 1, . . . , d, j , k .
Set W = (W1 , . . . , Wd ) where W j ← Z ρj /c(α j , ρ), j = 1, . . . , d.
if Wi /(Wk Ek ) < Yi for all i = 1, . . . , k − 1 then
Update Y ← max{Y, W/(Wk Ek )}.
12:
Simulate E ∗ ∼ Exp(1) and update Ek ← Ek + E ∗ .
13: return U = exp(−1/Y).
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
Note that S obtained in Step 7 of Algorithm 1 has the angular distribution σd of `D ; see Theorem 1 in [10].
Similar algorithms for drawing samples from the angular distribution of the extremal logistic and Dirichlet models
were obtained in [3]. Algorithm 2 requires a lower number of simulations and is more efficient on average, cf. [10].
Both algorithms are easily implemented using the function rmev in the mev package within the R Project for Statistical
Computing [34], which returns samples of max-stable scaled extremal Dirichlet vectors with unit Fréchet margins,
i.e., Y in Algorithms 1 and 2.
7. Estimation
The scaled extremal Dirichlet model can be used to model dependence between extreme events. To this end, several
schemes can be envisaged. For example, one can consider the block-maxima approach, given that max-stable distributions are the most natural for such data. Another option is peaks-over-threshold models. Yet another alternative,
used in [13] for the Brown–Resnick model, is to approximate the conditional distribution of a random vector with unit
Fréchet margins given that the j0 th component exceeds a large threshold by the distribution of φ+j0 /Y j0 discussed in
Section 6.2.
Here, we focus on the multivariate tail model of [28]; see also Section 16.4 in [29]. To this end, let X1 , . . . , Xn be a
random sample from some unknown multivariate distribution H with continuous univariate margins which is assumed
to be in the maximum domain of attraction of a multivariate extreme-value distribution H0 . To model the tail of H, its
margins F j , j = 1, . . . , d can first be approximated using the univariate peaks-over-threshold method. For all x above
14
L. R. Belzile and J. G. Nešlehová
some high threshold u j , one then has
F j (x) ≈ F̃ j (x; η j , ξ j ) = 1 − ν j 1 + ξ j
(x − u j )
ηj
!−1/ξ j
,
(13)
+
where ν j = 1 − F j (u j ), and η j > 0 and ξ j are the parameters of the generalized Pareto distribution. Furthermore, for
w sufficiently close to 1d , the copula of H can be approximated by the extreme-value copula C0 of H0 , so that, for
x ≥ u, H(x) ≈ H̃(x) = C0 {F̃1 (x1 ), . . . , F̃d (xd )}. The parameters of this multivariate tail model, i.e., the parameters
θ of the stable tail dependence function `0 of C0 as well as the marginal parameters ν, η and ξ can be estimated
using likelihood methods; this allows, e.g., for Bayesian inference, generalized additive modeling of the parameters
and model selection based on likelihood-ratio tests. For a comprehensive review of likelihood inference methods for
extremes, see, e.g., [22].
The multivariate tail model can be fitted in low-dimensions using the censored likelihood L(X; ν, η, ξ, θ) =
Qn
i=1 Li (X i ; ν, η, ξ, θ), where for i = 1, . . . , n,
Li (Xi ; u, ν, η, ξ, θ) =
∂mi H̃(y1 , . . . , yd )
∂y j1 · · · ∂y jmi
=
y=max(Xi ,u)
∂mi exp{−`0 (1/y)}
∂y j1 · · · ∂y jmi
mi
Y
J jk (Xi jk )
(14)
y=t{max(Xi ,u)} k=1
In this expression, the indices j1 , . . . , jmi are those of the components of Xi exceeding the thresholds u and for x ≥ u,
t(x) = (t1 (x1 ), . . . , td (xd )), where for j = 1, . . . , d,
(x j − u j )
νj
1 + ξj
J j (x j ) =
ηj
ηj
1
,
t j (x j ) = −
log{F̃ j (x j ; η j , ξ j )}
!−1/ξ j −1
1
.
[log{F̃ j (x j ; η j , ξ j )}]2 F̃ j (x j ; η j , ξ j )
(15)
The censored likelihood L(X; ν, η, ξ, θ) can be maximized either over all parameters at once, or the marginal parameters ν, η and ξ can be estimated from each univariate margin separately, so that only the estimate of θ is obtained
through maximizing L. When d is large, one can also maximize the likelihood in [40] that uses the tail approximation
H̄(x) ≈ 1 − `(1/x). In either case, `0 and the higher-order partial derivatives of `0 (1/x) need to be computed.
When `0 is the scaled extremal Dirichlet stable tail dependence function `D (·; ρ, α) given in Definition 1 with
parameters α > 0 and ρ > − min(α1 , . . . , αd ), ρ , 0, its expression is not explicit. However, `D can be calculated numerically using adaptive numerical cubature algorithms for integrals of functions defined on the simplex, as
implemented in, e.g., the R package SimplicialCubature. Given the representation in Eq. (5), `D is also easily
approximated using Monte Carlo methods. Instead of employing Eq. (5) directly and sampling from the Dirichlet
vector Dα , one can use the more efficient importance sampling estimator
h
i
−1 ρ
B max
X
1≤ j≤d {c(α j , ρ)u j } Di j
1
`bD (1/u, ρ, α) =
,
1 Pd
−1 ρ
B i=1
j=1 c(α j , ρ) Di j
d
P
where Di ∼ d−1 dj=1 Dir(α + I j ρ1d ) is sampled from a Dirichlet mixture.
The partial derivatives of `D can be calculated using the following result, shown in E.
Proposition 6 Let `D be the scaled extremal Dirichlet stable tail dependence function with parameters α > 0d and
− min(α1 , . . . , αd ) < ρ < ∞, ρ , 0. Then, for any x ∈ Rd+ ,
d
−ρ−ᾱ d
Y
X n
o1/ρ
∂d `D (1/x)
Γ(ᾱ + ρ)
D
= −dh (x; ρ, α) = −
c(α j , ρ)x j
{c(αi , ρ)}αi /ρ xiαi /ρ−1 ,
Qd
d−1
∂x1 · · · ∂xd
|ρ|
Γ(α
)
i
i=1
j=1
i=1
where hD is as given in Proposition 5. Furthermore, for all k = 1, . . . , d − 1 and x ∈ Rd+ ,
∂k `D (1/x)
=−
∂x1 · · · ∂xk
Z
∞
tk
0
k
Y
i=1
f xi t;
!Y
!
d
1
1
1
1
, , αi
F xi t;
, , αi dt,
c(αi , ρ) ρ
c(αi , ρ) ρ
i=k+1
(16)
15
Extremal attractors of Liouville copulas
where f (; a, b, c) and F(; a, b, c) denote, respectively the density and distribution function of the Rscaled Gamma
x
distribution with parameters a, c > 0 and b , 0 given in Eq. (10). Furthermore, if γ(c, x) = 0 tc−1 e−t dt denotes the lower incomplete gamma function, then for x > 0, F(x; a, b, c) = γ{c, (x/a)b }/Γ(c) when b > 0 while
F(x; a, b, c) = 1 − γ{c, (x/a)b }/Γ(c) when b < 0.
Other estimating equations could be used to circumvent the calculation of `D (1/x) and its partial derivatives. An
interesting alternative to likelihoods in the context of proper scoring functions is proposed in [8]. Specifically, the
authors advocate the use of the gradient score, adapted by them for the peaks-over-threshold framework,
(
)
d
X
∂2 log h(x) 1 ∂ log h(x) 2
∂wi (x) ∂ log h(x)
2
+
+ wi (x)
δw (x) =
2wi (x)
∂xi
∂xi
2
∂xi
∂xi2
i=1
for a differentiable weighting function w(x), unit Fréchet observations x and density h(x) that would correspond in
the setting of the scaled extremal Dirichlet to dhD (x; ρ, α). Explicit expressions for the derivatives of log dhD may be
P
found in E. The parameter estimates are obtained as the solution to argmaxθ∈Θ ni=1 δw (xi )IR(xi /u)>1 , where θ = (ρ, α)
is the vector of parameters of the model and R is a differentiable risk functional, usually the ` p norm for some p ∈ N.
Although the gradient score is not asymptotically most efficient, weighting functions can be designed to reproduce
approximate censoring, lending the method robustness and tractability.
8. Data illustration
In this section, we illustrate the use of the scaled extremal Dirichlet model on a trivariate sample of daily river
flow data of the river Isar in southern Germany; this dataset is a subset of the one analyzed in [1]. All the code
can be downloaded from https://github.com/lbelzile/ealc. For this analysis, we selected data measured at
Lenggries (upstream), Pupplinger Au (in the middle) and Munich (downstream). To ensure stationarity of the series
and given that the most extreme events occur during the summer, we restricted our attention to the measurements
for the months of June, July and August. Since the sites are measuring the flow of the same river, dependence at
extreme levels is likely to be present, as is indeed apparent from Figure 4. Directionality of the river may further lead
to asymmetry in the asymptotic dependence structure, suggesting that the scaled extremal Dirichlet model may be
well suited for these data. Furthermore, given that other well-known models like the extremal Dirichlet, logistic and
negative logistic are nested within this family, their adequacy can be assessed through likelihood ratio tests.
To remove dependence between extremes over time, we decluster each series and retain only the cluster maxima
based on three-day runs. Rounding of the measurements has no impact on parameter estimates and is henceforth
neglected. The multivariate tail model outlined in Section 7 is next fitted to the cluster maxima. The thresholds
u = (u1 , u2 , u3 ) were selected to be the 92% quantiles using the parameter stability plot of [42] (not shown here).
Next, set θ = (η, ξ, α, ρ), where η and ξ are the marginal parameters of the generalized Pareto distribution in Eq. (13)
and ρ and α are the parameters of the scaled Dirichlet model. To estimate θ, the trivariate censored likelihood (14)
could be used. To avoid numerical integration and because of the relative robustness to misspecification, we employed
the pairwise composite log-likelihood lC of [28] instead; the loss of efficiency in this trivariate example is likely small.
Specifically, we maximized
lC (θ) =
n X
d−1 X
d h
X
i
log g{t j (xi j ), tk (xik ); θ, t j (u j ), tk (uk )} + I xi j >u j log J j (xi j ) + I xik >uk log Jk (xik ) ,
i=1 j=1 k= j+1
where
exp{−`(1/u j , 1/uk )},
−∂`(1/y j , 1/uk )/∂y j exp{−`(1/y j , 1/uk )},
g(y j , yk ; θ, u j , uk ) =
−∂`(1/u j , 1/yk )/∂yk exp{−`(1/u j , 1/yk )},
on
o
hn
i
∂`(1/y j , 1/yk )/∂y j ∂`(1/y j , 1/yk )/∂yk − dhD (y j , yk ) exp{−`(1/y j , 1/yk )},
where ` = `D and for all j = 1, . . . , d, t j and J j are as in Eq. (15).
yj
yj
yj
yj
≤ u j , yk
> u j , yk
≤ u j , yk
> u j , yk
≤ uk
≤ uk
> uk
> uk
16
L. R. Belzile and J. G. Nešlehová
Isar daily river flow
800
1958-11-01 / 2014-01-18
Muenchen
800
600
600
400
400
200
200
400 Pupplinger Au
400
300
300
200
200
100
100
400
400
Lenggries
300
300
200
200
100
100
1958
1962
1964
1966
1968
1970
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014
Figure 4: Daily river flow of the Isar river at three sites
Uncertainty assessment can be done in the same way as for general estimating equations. Specifically, let g(θ)
denote an unbiased estimating function and define the variability matrix J, the sensitivity matrix H and the Godambe
information matrix G as
!
!
∂g(θ) ∂g(θ) >
∂2 g(θ)
J=E
,
H = −E
,
G = HJ−1 H.
(17)
∂θ ∂θ
∂θ∂θ>
The maximum composite likelihood estimator is strongly consistent and asymptotically normal, centered at the true
parameter θ with covariance matrix given by the inverse Godambe matrix G−1 .
Using the pairwise composite log-likelihood lC , we fitted the scaled extremal Dirichlet model as well as the
logistic and negative logistic models that correspond to the negative and positive scaled extremal Dirichlet models,
respectively, and the parameter restriction α = 1d . The estimates of the marginal generalized Pareto parameters η and
ξ are given in Table 1. As the estimates were obtained by maximizing `C , their values depend on the fitted model; the
line labeled “Marginal" corresponds to fitting the generalized Pareto distribution to threshold exceedances of each one
of the three series separately. The marginal Q-Q plots displayed in Figure 5 indicate a good fit of the model as well.
Scaled Dirichlet
Neg. logistic
Logistic
ext. Dirichlet
Marginal
η1
η2
η3
ξ1
ξ2
ξ3
123.2 (7.5)
117.1 (6.8)
117.3 (6.8)
114.4 (6.8)
129.1 (14.5)
84.4 (5)
86.2 (5.1)
86.6 (5.1)
84.3 (4.9)
95.1 (10.6)
68.1 (4.2)
70 (4.3)
70.4 (4.3)
68.2 (4.1)
76 (8.7)
0.05 (0.04)
0.08 (0.04)
0.08 (0.04)
0.12 (0.04)
-0.01 (0.08)
-0.03 (0.04)
-0.05 (0.04)
-0.05 (0.04)
-0.02 (0.04)
-0.15 (0.08)
0.02 (0.04)
0 (0.04)
0 (0.04)
0.04 (0.04)
-0.08 (0.08)
Table 1: Generalized Pareto parameter estimates and standard errors (in parenthesis) for the trivariate river example
for four different models.
17
Extremal attractors of Liouville copulas
600
700
500
400
300
500
100
200
100
200
400
200
300
400
Sample quantiles
Sample quantiles
600
1200
1000
800
600
Sample quantiles
Q-Q plot, Lenggries
700
Q-Q plot, Pupplinger Au
1400
Q-Q plot, Muenchen
200 300 400 500 600 700 800 900
200
Theoretical quantiles
300
400
500
100
200
Theoretical quantiles
300
400
Theoretical quantiles
Figure 5: Marginal Q-Q plots for the three sites based on pairwise composite likelihood estimates for the scale and
shape parameters obtained from the scaled Dirichlet model, retaining marginal exceedances of the 92% quantiles. The
pointwise confidence intervals were obtained from the transformed Beta quantiles of the order statistics.
Scaled Dirichlet
Neg. logistic
Logistic
ext. Dirichlet
Gradient score
α1
α2
α3
ρ
0.76 (0.3)
1
1
3.34 (0.52)
1
1.65 (0.82)
1
1
10.2 (2.84)
2.72
2.03 (1.15)
1
1
12.78 (3.93)
2.66
−0.32 (0.1)
0.36 (0.02)
0.28 (0.01)
1
−0.39
Table 2: Dependence parameters estimates and standard errors (in parenthesis) for the trivariate river example.
w1 = 1
Negative logistic
Logistic
scaled
Dirichlet
w3 = 1
w3 = 1
w3 = 1
w2 = 1
w1 = 1
w2 = 1
w1 = 1
w2 = 1
Figure 6: Angular density plots for the three models, the negative logistic (left), logistic (middle) and scaled Dirichlet
(right). The colours correspond to log density values and range from red (high density) to blue (low density).
The estimates of the dependence parameters α and ρ are given in Table 2. The last line displays the maximum gradient score estimates were obtained from the raw data, i.e., ignoring the clustering, after transforming the observations
to the standard Fréchet scale using the probability integral transform. We retained only the 10% largest values based
on the ` p norm with p = 20; this risk functional is essentially a differentiable approximation of `∞ . We selected the
weight function w(x, u) = x[1 − exp{−(kxk p /u − 1)}] based on [8] to reproduce approximate censoring. The estimates
are similar to the composite maximum likelihood estimators, though not efficient.
18
L. R. Belzile and J. G. Nešlehová
The angular densities of the fitted logistic, negative logistic and scaled extremal Dirichlet models are displayed
in Figure 6. The right panel of this figure shows asymmetry caused by a few extreme events that only happened
downstream. Whether this asymmetry is significant can be assessed through composite likelihood ratio tests; recall
that the logistic model, the negative logistic model and the extremal Dirichlet model of [7] are all nested within
the scaled extremal Dirichlet model. To this end, consider a partition of θ = (ψ, λ) into a q dimensional parameter of
interest ψ and a 3d +1−q dimensional nuisance parameter λ, and the corresponding partitions of the matrices H, J and
G. Let b
θC = (b
ψC , b
λC ) denote the maximum composite likelihood parameter estimates and b
θ0 = (ψ0 , b
λ0 ) the restricted
parameter estimates under the null hypothesis that the simpler model is adequate. The asymptotic distribution of
P
the composite likelihood ratio test statistic 2{log lC (θbC ) − log lC (θb0 )} is equal to qi=1 ci Zi where Zi are independent
−1
χ21 variables and ci are the eigenvalues of the q × q matrix (Hψψ − Hψλ H−1
λλ Hλψ )Gψψ ; see [26]. We estimated the
−1
inverse Godambe information matrix, G , by the empirical covariance of B nonparametric bootstrap replicates. The
sensitivity matrix H was obtained from the Hessian matrix at the maximum composite likelihood estimate and the
variability matrix J from Eq. (17). Since the Coles–Tawn extremal Dirichlet, negative logistic and logistic models are
nested within the scaled Dirichlet family, we test for a restriction to these simpler models; the respective approximate
P-values were 0.003, 0.74 and 0.78. These values suggest that while the Coles–Tawn extremal Dirichlet model is
clearly not suitable, there is not sufficient evidence to discard the logistic and negative logistic models. The effects of
possible model misspecification are also visible for the Coles–Tawn extremal Dirichlet model, as the parameter values
of α1 , α2 and α3 are very large (viz. Table 2) and this induces negative bias in the shape parameter estimates, as can
be seen from Table 1.
9. Discussion
In this article, we have identified extremal attractors of copulas and survival copulas of Liouville random vectors
RDα , where Dα has a Dirichlet distribution on the unit simplex with parameters α, and R is a strictly positive random
variable independent of Dα . The limiting stable tail dependence functions can be embedded in a single family,
which can capture asymmetry and provides a valid model in dimension d. The latter is novel and termed here the
scaled extremal Dirichlet; it includes the well-known logistic, negative logistic as well as the Coles–Tawn extremal
Dirichlet models as special cases. In particular, therefore, this paper is first to provide an example of a random vector
attracted to the Coles–Tawn extremal Dirichlet model, which was derived by enforcing moment constraints on a
simplex distribution rather than as the limiting distribution of a random vector.
A scaled extremal Dirichlet stable tail dependence function `D has d + 1 parameters, ρ and α. The parameter
vector α is inherited from Dα and induces asymmetry in `D . The parameter ρ comes from the regular variation of
R at zero and infinity, respectively; this is reminiscent of the extremal attractors of elliptical distributions [33]. The
magnitude of ρ has impact on the strength of dependence while its sign changes the overall shape of `D . Having
d + 1 parameters, the scaled extremal Dirichlet model may not be sufficiently rich to account for spatial dependence,
unlike the Hüsler–Reiss or the extremal Student-t models, which have one parameter for each pair of variables and
are thus easily combined with distances. Also, it is less flexible than Dirichlet mixtures [4], which are however hard
to estimate in high dimensions and require sophisticated machinery. To achieve greater flexibility, the scaled extremal
Dirichlet model could perhaps be extended by working with more general scale mixtures, such as of the weighted
Dirichlet distributions considered, e.g., in [18].
Nonetheless, the scaled extremal Dirichlet model may naturally find applications whenever asymmetric extremal
dependence is suspected; the latter may be caused, e.g., by causal relationships between the variables [15]. The
stochastic structure of the scaled extremal Dirichlet model has several major advantages, that make the model easy
to interpret, estimate and simulate from. Its angular density has a simple form; in contrast to the asymmetric generalizations of the logistic and negative logistic models, this model does not place any mass on the vertices and
lower-dimensional facets of the unit simplex. Another plus is the tractability of the de Haan representation and of
the extremal functions, both expressible in terms of independent scaled Gamma variables; this allows for feasible
inference and stochastic simulation. While the scaled extremal Dirichlet stable tail dependence function `D does not
have a closed form in general, closed-form algebraic expressions exist when α is integer-valued and in the bivariate
case. Model selection for well-known families of extreme-value distributions can be performed through likelihood
ratio tests. Another potentially useful feature is that ρ ∈ (−∞, ∞) can be allowed, with the convention that all variables
whose indices i are such that −ρ ≤ −αi are independent.
Extremal attractors of Liouville copulas
19
Acknowledgment
Funding in partial support of this work was provided by the Natural Sciences and Engineering Research Council
(RGPIN-2015-06801, CGSD3-459751-2014), the Canadian Statistical Sciences Institute, and the Fonds de recherche
du Québec – Nature et technologies (2015–PR–183236). We thank the acting Editor-in-Chief, Richard Lockhart, the
Associate Editor and two anonymous referees for their valuable suggestions.
References
[1] P. Asadi, A. C. Davison, and S. Engelke. Extremes on river networks. Ann. Appl. Stat., 9(4):2023–2050, 2015.
[2] G. Balkema and N. Nolde. Asymptotic independence for unimodal densities. Adv. in Appl. Probab., 42(2):411–
432, 2010.
[3] M.-O. Boldi. A note on the representation of parametric models for multivariate extremes. Extremes, 12(3):211–
218, 2009.
[4] M.-O. Boldi and A. C. Davison. A mixture model for multivariate extremes. J. Roy. Stat. Soc. B Met., 69(2):217–
229, 2007.
[5] L. Breiman. On some limit theorems similar to the arc-sin law. Theory Probab. Appl., 10(2):323–331, 1965.
[6] A. Charpentier and J. Segers. Tails of multivariate Archimedean copulas. J. Multivariate Anal., 100:1521–1537,
2009.
[7] S. G. Coles and J. A. Tawn. Modelling extreme multivariate events. J. Roy. Stat. Soc. B Met., 53(2):377–392,
1991.
[8] R. de Fondeville and A. C. Davison. High-dimensional peaks-over-threshold inference for the Brown-Resnick
process. ArXiv e-prints, 2016.
[9] L. de Haan. A spectral representation for max-stable processes. Ann. Probab., 12(4):1194–1204, 1984.
[10] C. Dombry, S. Engelke, and M. Oesting. Exact simulation of max-stable processes. Biometrika, 103(2):303–317,
2016.
[11] P. Embrechts and C. M. Goldie. On closure and factorization properties of subexponential and related distributions. J. Austral. Math. Soc. Ser. A, 29(2):243–256, 1980.
[12] P. Embrechts, C. Klüppelberg, and T. Mikosch. Modelling Extremal Events for Insurance and Finance. Springer,
New York, 1997.
[13] S. Engelke, A. Malinowski, Z. Kabluchko, and M. Schlather. Estimation of Hüsler–Reiss distributions and
Brown–Resnick processes. J. Roy. Stat. Soc. B, 77(1):239–265, 2015.
[14] K.-T. Fang, S. Kotz, and K. W. Ng. Symmetric Multivariate and Related Distributions. Chapman & Hall,
London, 1990.
[15] C. Genest and J. G. Nešlehová. Assessing and modeling asymmetry in bivariate continuous data. In P. Jaworski,
F. Durante, and W. K. Härdle, editors, Copulae in Mathematical and Quantitative Finance, pages 91–114.
Springer, Berlin, 2013.
[16] É. J. Gumbel. Distributions des valeurs extrêmes en plusieurs dimensions. Publ. Inst. Statist. Univ. Paris,
9:171–173, 1960.
[17] R. D. Gupta and D. S. P. Richards. Multivariate Liouville distributions. J. Multivariate Anal., 23(2):233–256,
1987.
[18] E. Hashorva. Extremes of weighted dirichlet arrays. Extremes, 11(4):393–420, 2008.
[19] E. Hashorva. Extremes of aggregated dirichlet risks. J. Multivariate Anal., 133:334 – 345, 2015.
[20] E. Hashorva and A. G. Pakes. Tail asymptotics under beta random scaling. J. Math. Anal. Appl., 372(2):496–514,
2010.
[21] L. Hua. A note on upper tail behaviour of Liouville copulas. Risks, 4(4):40, 2016.
[22] R. Huser, A. C. Davison, and M. G. Genton. Likelihood estimators for multivariate extremes. Extremes,
19(1):79–103, 2016.
[23] J. Hüsler and R.-D. Reiss. Maxima of normal random vectors: between independence and complete dependence.
Statist. Probab. Lett., 7(4):283–286, 1989.
[24] H. Joe. Families of min-stable multivariate exponential and multivariate extreme value distributions. Statist.
Probab. Lett., 9:75–81, 1990.
[25] H. Joe. Multivariate dependence measures and data analysis. Comput. Statist. Data Anal., 16:279–297, 1993.
20
L. R. Belzile and J. G. Nešlehová
[26] J. T. Kent. Robust properties of likelihood ratio test. Biometrika, 69(1):19–27, 1982.
[27] M. Larsson and J. Nešlehová. Extremal behavior of Archimedean copulas. Adv. Appl. Probab., 43:195–216,
2011.
[28] A. W. Ledford and J. A. Tawn. Statistics for near independence in multivariate extreme values. Biometrika,
83(1):169–187, 1996.
[29] A. J. McNeil, R. Frey, and P. Embrechts. Quantitative risk management. Princeton Series in Finance. Princeton
University Press, Princeton, NJ, revised edition, 2015.
[30] A. J. McNeil and J. Nešlehová. Multivariate Archimedean copulas, d-monotone functions and `1 -norm symmetric distributions. Ann. Statist., 37:3059–3097, 2009.
[31] A. J. McNeil and J. Nešlehová. From Archimedean to Liouville copulas. J. Multivariate Anal., 101:1772–1790,
2010.
[32] N. Nolde. The effect of aggregation on extremes from asymptotically independent light-tailed risks. Extremes,
17(4):615–631, 2014.
[33] T. Opitz. Extremal processes: Elliptical domain of attraction and a spectral representation. J. Multivariate Anal.,
122:409 – 413, 2013.
[34] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2016.
[35] S. I. Resnick. Extreme values, regular variation, and point processes. Springer-Verlag, Berlin; New York, 1987.
[36] M. Schlather. Models for stationary max-stable random fields. Extremes, 5(1):33–44, 2002.
[37] J. Segers. Max-stable models for multivariate extremes. REVSTAT, 10(1):61–82, 2012.
[38] B. D. Sivazlian. On a multivariate extension of the gamma and beta distributions. SIAM J. Appl. Math.,
41(2):205–209, 1981.
[39] A. Sklar. Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Statist. Univ. Paris, 8:229–231,
1959.
[40] R. L. Smith, J. A. Tawn, and S. G. Coles. Markov chain models for threshold exceedances. Biometrika,
84(2):249–268, 1997.
[41] E. W. Stacy. A generalization of the gamma distribution. Ann. Math. Stat., 33(3):1187–1192, 1962.
[42] J. L. Wadsworth. Exploiting structure of maximum likelihood estimators for extreme value threshold selection.
Technometrics, 58(1):116–126, 2016.
[43] R. E. Williamson. Multiply monotone functions and their Laplace transforms. Duke Math. J., 23:189–207, 1956.
21
Extremal attractors of Liouville copulas
Appendix A. : Proofs from Section 2
Proof of Proposition 2. To prove parts (a) and (b), recall that 1/Xi is distributed as 1/(RDi ), where Di ∼ Beta(αi , ᾱ−αi )
is independent of R. Furthermore, it is easy to show that 1/Di ∈ M(Φαi ), which implies that E(1/Dβi ) < ∞ for any
β < αi . The extremal behavior of 1/Xi will thus be determined by the extremal behavior of either 1/R or 1/Di ,
depending on which one has a heavier tail. Indeed, Breiman’s Lemma [5] implies that 1/Xi ∈ M(Φρ ) if 1/R ∈ M(Φρ )
for some ρ < αi and that 1/Xi ∈ M(Φαi ) if E(1/Rαi +ε ) < ∞ for some ε > 0. Finally, the fact that 1/Xi ∈ M(Φαi ) when
1/R ∈ M(Φαi ) follows directly from the Corollary to Theorem 3 in [11].
The following lemma is a side result of Proposition 2, which is needed in the subsequent proofs.
Lemma 1 Suppose that X = RDα . If 1/R ∈ M(Φαi ) for some i ∈ {1, . . . , d}, then
lim
x→∞
Pr (1/R > x)
= 0.
Pr (1/Xi > x)
Proof of Lemma 1.
Because 1/R ∈ M(Φαi ), Pr(1/R > x) is regularly varying with index −αi . In particular, for any b ∈ (0, 1),
Pr(1/R > xb)/ Pr(1/R > x) → b−αi as x → ∞. An application of Fatou’s lemma thus gives
Z 1
Z 1 −1
Pr(1/R > xb)
b (1 − b)ᾱ−αi −1
Pr(1/Xi > x)
= lim inf
fDi (b) db ≥
db = ∞
lim inf
x→∞
x→∞ Pr(1/R > x)
B(αi , ᾱ − αi )
0 Pr(1/R > x)
0
and hence the result.
Appendix B. : Proofs from Section 3
First recall the following property of the Dirichlet distribution, which is easily shown using the transformation formula
for Lebesgue densities.
Lemma 2 Let Dα be a Dirichlet random vector with parameters α. Then for any 2 ≤ k ≤ d and any collection of
distinct indices 1 ≤ i1 < · · · < ik ≤ d,
(Di1 , . . . , Dik ) = Bi1 ,...,ik × D(αi1 ,...,αik ) ,
d
where Bi1 ,...,ik ∼ Beta(αi1 + · · · + αik , ᾱ − (αi1 + · · · + αik )) is independent of the k-variate Dirichlet vector D(αi1 ,...,αik )
with parameters (αi1 , . . . , αik ).
d
d
Proof of Theorem 1. In order to prove part (a), recall that kXk = R is independent of X/kXk = Dα . Because
R ∈ M(Φρ ), there exists a sequence (bn ) of constants in (0, ∞) such that, for any Borel set B ⊆ Sd and any r > 0,
!
X
lim n Pr kXk > bn r,
∈ B = lim n Pr(R > bn r) Pr(Dα ∈ B) = r−ρ Pr( Dα ∈ B).
n→∞
n→∞
kXk
By Corollary 5.18 in [35], X ∈ M(H0 ) where for all x ∈ Rd+ ,
ρ
Dρd
D1
H0 (x) = exp −E
max xρ , . . . , xρ
.
1
d
Let B(·, ·) denote the Beta function. The univariate margins of H0 are given, for all i = 1, . . . , d and x > 0, by
)
(
)
(
n
o
B(ρ + αi , ᾱ − αi )
Γ(αi + ρ)Γ(ᾱ)
F0i (x) = exp −x−ρ E(Dρi ) = exp −x−ρ
= exp −x−ρ
B(αi , ᾱ − αi )
Γ(ᾱ + ρ)Γ(αi )
for i ∈ {1, . . . , d}. The copula of H0 then satisfies, for all u ∈ [0, 1]d ,
C0 (u) =
−1
−1
H0 {F01
(u1 ), . . . , F0d
(ud )}
"
(
)#!
(− log ui )Γ(αi )Dρi
Γ(ᾱ + ρ)
= exp −
E max
.
1≤i≤d
Γ(ᾱ)
Γ(αi + ρ)
22
L. R. Belzile and J. G. Nešlehová
By Eq. (2), the stable tail dependence function of C0 thus indeed equals, for all x ∈ Rd+ ,
ρ
xd Γ(αd )Dρd
Γ(ᾱ + ρ)
x1 Γ(α1 )D1
E max
,...,
`(x) =
.
Γ(ᾱ)
Γ(α1 + ρ)
Γ(αd + ρ)
The part (b) follows directly from Proposition 2.2 in [19] upon setting p = 1 and taking, for i = 1, . . . , d and
j = 1, . . . , d, λi j = 1 whenever i = j and λi j = 0 otherwise.
To prove part (c), recall first that from Proposition 1, for i = 1, . . . , d, Xi ∈ M(Ψρ+ᾱ−αi ) and hence there exist
sequences (ani ) ∈ (0, ∞), (bni ) ∈ R, such that for all x ∈ R,
lim n Pr(Xi > ani x + bni ) = − log{Ψρ+ᾱ−αi (x)}.
n→∞
Next, observe that as in the proof of Proposition 5.27 in [35], X ∈ M(H0 ) follows if for all 1 ≤ i < j ≤ d and xk such
that Ψρ+ᾱ−αk (xk ) > 0 for k = i, j,
lim n Pr(Xi > ani xi + bni , X j > an j x j + bn j ) = 0.
n→∞
(B.1)
To prove that (B.1) indeed holds, it suffices to assume that d = 2. This is because for arbitrary indices 1 ≤ i <
d
d
j ≤ d, Lemma 2 implies that (Xi , X j ) = R∗ (B, 1 − B), where B ∼ Beta(αi , α j ), R∗ = RY is independent of B and
∗
Y ∼ Beta(αi + α j , ᾱ − αi − α j ) is independent of B and R. Because Pr(R ≤ 0) = 0, Theorem 4.5 in [20] implies that
R∗ ∈ M(Ψρ+ᾱ−αi −α j ) when R ∈ M(Ψρ ) for some ρ > 0. Thus suppose that d = 2 and write (D1 , D2 ) ≡ (B, 1 − B),
where B ∼ Beta(α1 , α2 ). Fix arbitrary x1 , x2 ∈ R are such that Ψρ+ᾱ−αi (xi ) > 0 for i = 1, 2. Then because for any
a, c > 0 and b ∈ (0, 1), max{a/b, c/(1 − b)} ≥ a + c, one has
(
!)
an1 x1 + bn1 an2 x2 + bn2
0 ≤ Pr(X1 > an1 x1 + bn1 , X2 > an2 x2 + bn2 ) = Pr R > max
,
B
1−B
≤ Pr(R > an1 x1 + bn1 + an2 x2 + bn2 ).
In order to prove Eq. (B.1), it thus suffices to show that
lim n Pr(R > an1 x1 + bn1 + an2 x2 + bn2 ) = 0.
n→∞
(B.2)
This however follows immediately from the fact that if R ∈ M(Ψρ ) for some ρ > 0, the upper end-point r of R,
viz. r = sup{x : Pr(R ≤ x) < 1}, is finite. Because for i = 1, 2, r is also the upper endpoint of Xi , ani xi + bni → r
as n → ∞. This means that there exists n0 ∈ N so that for all n ≥ n0 , an1 x1 + bn1 + an2 x2 + bn2 > r and Pr(R >
an1 x1 + bn1 + an2 x2 + bn2 ) = 0. This proves Eq. (B.1) and hence also Theorem 1 (c). Note that alternatively, part (c)
could be proved using Theorem 2.1 in [19] similarly to the proof of Proposition 2.2 therein.
The proof of Theorem 2 requires the following technical lemma.
Lemma 3 Suppose that Dα = (D1 , . . . , Dd ) is a Dirichlet random vector with parameters α. Further let R be a
positive random variable independent of Dα such that Pr(R ≤ 0) = 0, and let X = R Dα . Then for any 1 ≤ i < j ≤ d
and any xi , x j ∈ (0, ∞),
!
1
1
lim n Pr
> ani xi ,
> an j x j = 0
n→∞
Xi
Xj
if either:
(i) 1/R ∈ M(Φρ ) with ρ ∈ [αi ∧ α j , α1 ∨ α2 ], and for k = i, j, (ank ) is a sequence of positive constants such that
n Pr(1/Xk > ank xk ) → xk−(αk ∧ρ) as n → ∞;
(ii) E(1/Rβ ) < ∞ for some β > αi ∨ α j and for k = i, j, (ank ) is a sequence of positive constants such that
n Pr(1/Xk > ank xk ) → xk−αk as n → ∞.
23
Extremal attractors of Liouville copulas
d
Proof of Lemma 3. Observe first that when d > 2, Lemma 2 implies that (Xi , X j ) = R∗ (B, 1 − B), where R∗ ⊥
⊥ B,
B ∼ Beta(αi , α j ) and R∗ = RY, with Y ⊥
⊥ R and Y ∼ Beta(αi + α j , ᾱ − αi − α j ). Now note that 1/Y ∈ M(Φαi +α j ). Thus
if 1/R ∈ M(Φρ ) for ρ ∈ [αi ∧ α j , α1 ∨ α2 ], ρ < αi + α j and Breiman’s Lemma implies that 1/R∗ ∈ M(Φρ ). Further, if
E(1/Rβ ) < ∞ for some β ∈ (αi ∨ α j , αi + α j ), E{1/(R∗ )β } < ∞ given that E(1/Y β ) < ∞. We can thus assume without
loss of generality that d = 2 and α1 ≤ α2 ; we shall also write (D1 , D2 ) ≡ (B, 1 − B), where B ∼ Beta(α1 , α2 ).
To prove part (i), note first that the existence of the sequences (ank ), k = i, j, follows from Proposition 2, by which
1/Xk ∈ M(Φρ∧αk ) for k = i, j, and the Poisson approximation [12, Proposition 3.1.1]. Next, observe that for any
constants a, c > 0 and any b ∈ (0, 1),
ac
≤ ab ∨ c(1 − b) < a ∨ c.
(B.3)
a+c
Indeed, when b < c/(a + c), ab ∨ c(1 − b) = c(1 − b) and c(1 − b) ∈ (ac/(a + c), c), while when b ≥ c/(a + c),
ab ∨ c(1 − b) = ab and ab ∈ [ac/(a + c), a). To show the claim in part (i), distinguish the cases below:
d
Case I. α1 = α2 . Here, ρ = α1 = α2 and X1 = X2 , so that an1 /an2 → 1 by the Convergence to Types Theorem [35].
By Eq. (B.3),
#
!
"
!
1
an1 an2 x1 x2
1
1
1
> max{an1 x1 B, an2 x2 (1 − B)} ≤ n Pr
>
> an1 x1 ,
> an2 x2 = n Pr
. (B.4)
0 ≤ n Pr
X1
X2
R
R an1 x1 + an2 x2
Because (x1 x2 )/{(an1 /an2 )x1 + x2 } → (x1 x2 )/(x1 + x2 ) as n → ∞,
!
!−α1
1
an1 an2 x1 x2
x1 x2
lim n Pr
>
=
.
n→∞
X1 an1 x1 + an2 x2
x1 + x2
Furthermore, by Lemma 1, given that (an1 an2 x1 x2 )/(an1 x1 + an2 x2 ) → ∞ as n → ∞,
Pr 1/R > aan1n1xa1n2+axn21 xx22
= 0,
lim
n→∞ Pr 1/X > an1 an2 x1 x2
1
an1 x1 +an2 x2
so that the right-hand side in Eq. (B.4) tends to 0 as n → ∞, and this implies the claim.
Case II. α1 < α2 and ρ = α2 . Then for i = 1, 2, there exists a slowly varying function Li such that ani = n1/αi Li (n).
Hence an2 /an1 → 0 and (x1 x2 )/{x1 + x2 (an2 /an1 )} → x2 as n → ∞. Consequently,
!
1
an1 an2 x1 x2
lim n Pr
>
= x2−α2 .
n→∞
X2 an1 x1 + an2 x2
Moreover, by Lemma 1, given that (an1 an2 x1 x2 )/(an1 x1 + an2 x2 ) → ∞ as n → ∞,
Pr 1/R > aan1n1xa1n2+axn21 xx22
= 0,
lim
n→∞ Pr 1/X > an1 an2 x1 x2
2
an1 x1 +an2 x2
so that again the right-hand side in Eq. (B.4) tends to 0 as n → ∞.
Case III. α1 < α2 and ρ ∈ [α1 , α2 ). In this case, 1/X1 ∈ M(Φα1 ) and 1/X2 ∈ M(Φρ ). Therefore, either directly
when ρ > α1 or by Lemma 1, one can easily deduce that
lim
x→∞
Pr(1/R > x)
= 0.
Pr(1/X1 > x)
(B.5)
At the same time, Breiman’s Lemma [5] implies that
lim
x→∞
Pr(1/R > x)
1
B(α1 , α2 )
=
=
.
ρ
Pr(1/X2 > x) E{1/(1 − B) } B(α1 , α2 − ρ)
(B.6)
24
L. R. Belzile and J. G. Nešlehová
Hence, for any b ∈ (0, 1), the limit of n Pr {1/R > an2 x2 (1 − b)} as n → ∞ equals
lim n Pr {1/X2 > an2 x2 (1 − b)}
n→∞
B(α1 , α2 )
Pr {1/R > an2 x2 (1 − b)}
= {x2 (1 − b)}−ρ
Pr {1/X2 > an2 x2 (1 − b)}
B(α1 , α2 − ρ)
so that
lim
n→∞
Z
0
1
bα1 −1 (1 − b)α2 −1
db = n Pr(1/X2 > an2 x2 ) = x2−ρ
n Pr{1/R > an2 x2 (1 − b)}
B(α1 , α2 )
Z 1 α1 −1
Z 1
b
(1 − b)α2 −ρ−1
bα1 −1 (1 − b)α2 −1
= x2−ρ
db =
lim n Pr{1/R > an2 x2 (1 − b)}
db. (B.7)
B(α1 , α2 − ρ)
B(α1 , α2 )
0
0 n→∞
Given that for any b ∈ (0, 1), Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)} ≤ Pr{1/R > an2 x2 (1 − b)},
Z 1
bα1 −1 (1 − b)α2 −1
lim inf n Pr{1/R > an2 x2 (1 − b)} − Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)}
db
n→∞
B(α1 , α2 )
0
Z 1
bα1 −1 (1 − b)α2 −1
n Pr{1/R > an2 x2 (1 − b)} − Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)}
≤ lim inf
db
n→∞
B(α1 , α2 )
0
by Fatou’s Lemma. Because of Eq. (B.7), this inequality simplifies to
Z 1
bα1 −1 (1 − b)α2 −1
db
x2−ρ −
lim sup n Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)}
B(α1 , α2 )
n→∞
0
Z 1
bα1 −1 (1 − b)α2 −1
−ρ
≤ x2 − lim sup
n Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)}
db
B(α1 , α2 )
n→∞
0
and hence
0 ≤ lim sup n Pr(1/X1 > an1 x1 , 1/X2 > an2 x2 )
n→∞
Z
≤
0
1
bα1 −1 (1 − b)α2 −1
db.
lim sup n Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)}
B(α1 , α2 )
n→∞
To show the desired claim, it thus suffices to show that for arbitrary b ∈ (0, 1),
lim n Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)} = 0.
(B.8)
n→∞
To this end, fix b ∈ (0, 1) and observe that an1 /an2 → ∞. Indeed, if ρ > α1 , this follows directly from the fact
that an1 = n1/α1 L1 (n) and an2 = n1/ρ L2 (n) for some slowly varying functions L1 , L2 . When ρ = α1 , suppose that
lim inf n→∞ an1 /an2 were finite. Then there exists a subsequence ank 1 /ank 2 such that ank 1 /ank 2 → a as k → ∞ for some
a ∈ [0, ∞). Hence, for a fixed ε > 0 and all k ≥ k0 , ank 1 /ank 2 ≤ a + ε. Using the latter observation and Eq. (B.6),
lim nk Pr(1/R > ank 1 ) ≥ lim nk Pr{1/R > ank 2 (a + ε)}
k→∞
k→∞
= lim nk Pr{1/X2 > ank 2 (a + ε)}
k→∞
Pr{1/R > ank 2 (a + ε)}
B(α1 , α2 )
= (a + ε)−ρ
> 0.
Pr{1/X2 > ank 2 (a + ε)}
B(α1 , α2 − ρ)
At the same time, by Eq. (B.5),
lim nk Pr(1/R > ank 1 ) = lim nk Pr(1/X1 > ank 1 )
k→∞
k→∞
Pr(1/R > ank 1 )
=0
Pr(1/X1 > ank 1 )
and hence a contradiction. Therefore, lim inf n→∞ an1 /an2 = ∞ and hence an1 /an2 → ∞ as n → ∞. Because an1 b >
an2 (1 − b) if and only if b > an2 /(an1 + an2 ) and an2 /(an1 + an2 ) → 0 as n → ∞, there exists n0 such that for all n ≥ n0 ,
n Pr{1/R > an1 x1 b, 1/R > an2 x2 (1 − b)} = n Pr(1/R > an1 x1 b) = n Pr(1/X1 > an1 x1 b)
Pr(1/R > an1 x1 b)
.
Pr(1/X1 > an1 x1 b)
25
Extremal attractors of Liouville copulas
The last expression tends to 0 as n → ∞ by Eq. (B.5) and hence Eq. (B.8) indeed holds.
To prove part (ii), first recall that by Proposition 2 (b), 1/Xi ∈ M(Φαi ), i = 1, 2, and hence the scaling sequences
(an1 ) and (an2 ) indeed exist. Recall that for i = 1, 2, ani = n1/αi Li (n) for some slowly varying function Li . As in
the proof of part (i), n Pr(1/X1 > an1 x1 , 1/X2 > an2 x2 ) can be bounded above by the right-hand side in Eq. (B.4).
Markov’s inequality further implies that for β ∈ (α2 , α1 + α2 ) such that E(1/Rβ ) < ∞,
)β
!
(an1 x1 + an2 x2 )β E 1/Rβ (
x1
an1 an2 x1 x2
x2
1
β
=
>
≤ nE 1/R
+
n Pr
R an1 x1 + an2 x2
(an1 an2 x1 x2 )β
(x1 x2 )β n1/α2 −1/β L2 (n) n1/α1 −1/β L1 (n)
The right-most expression tends to 0 as n → ∞ because for any i = 1, 2 and ρ > 0, nρ Li (n) → ∞.
Proof of Theorem 2. First note that a positive random vector Y is in the maximum domain of attraction of a multivariate
extreme-value distribution H0 with Fréchet margins if and only if there exist sequences of positive constants (ani ) ∈
(0, ∞), i = 1, . . . , d, so that, for all y ∈ Rd+ ,
lim n{1 − Pr(Y1 ≤ an1 y1 , . . . , Yd ≤ and yd )} =
d
X
X
lim n
n→∞
n→∞
k=1 1≤i1 <···<ik ≤d
= − log H0 (y).
(−1)k+1 Pr(Yi1 > ani1 yi1 , . . . , Yik > anik yik )
(B.9)
This multivariate version of the Poisson approximation holds by the same argument as in the univariate case [12,
Proposition 3.1.1].
To prove part (a), suppose that 1/R ∈ M(Φρ ) for some ρ ∈ (0, α M ]. By Proposition 2, one then has that for
any i ∈ I1 , 1/(RDi ) ∈ M(Φαi ). For any i ∈ I1 , let (ani ) be a sequence of positive constants such that, for all
x > 0, n Pr{1/(RDi ) > ani x} → x−αi as n → ∞; such a sequence exists by the univariate Poisson approximation [12,
Proposition 3.1.1]. The same result also guarantees the existence of a sequence (an ) of positive constants such that,
for all x > 0, n Pr(1/R > an x) → x−ρ as n → ∞. Now set, for any i ∈ I2 ,
Γ(αi − ρ)Γ(ᾱ − αi )
Γ(ᾱ)
Γ(ᾱ)/Γ(ᾱ − ρ)
×
=
,
bi = E D−ρ
=
i
Γ(ᾱ − ρ)
Γ(αi )Γ(ᾱ − αi ) Γ(αi )/Γ(αi − ρ)
(B.10)
and define, for any i ∈ I2 and n ∈ N, ani = b1/ρ
i an . As detailed in the proof of Proposition 2 (a), Breiman’s Lemma
then implies that, for all i ∈ I2 and x > 0,
(
)
(
) Pr n 1 > a (b1/ρ x)o
n i
1
1
RD
−ρ
n i
o = x−ρ b−1
lim n Pr
> ani x = lim n Pr
> an (b1/ρ
i bi = x ,
i x)
1
n→∞
n→∞
RDi
R
Pr R > an (b1/ρ
x)
i
given that for all i ∈ I2 , Di ∼ Beta(αi , ᾱ − αi ).
Next, fix an arbitrary x ∈ (0, ∞)d , k ∈ {2, . . . , d} and indices 1 ≤ i1 < · · · < ik ≤ d. To calculate the limit of
n Pr(1/(RDi1 ) > ani1 xi1 , . . . , 1/(RDik ) > anik xik ), two cases must be distinguished:
Case I. {i1 , . . . , ik } ∩ I1 , ∅. In this case, suppose, without loss of generality, that i1 ∈ I1 . Then
!
!
1
1
1
1
0 ≤ n Pr
> ani1 xi1 , . . . ,
> anik xik ≤ n Pr
> ani1 xi1 ,
> ani2 xi2 .
RDi1
RDik
RDi1
RDi2
Now either i2 ∈ I1 , in which case ρ ≥ αi1 ∨ αi2 , or i2 ∈ I2 , so that αi1 ≤ ρ < αi2 . Either way, Lemma 3 implies that
!
1
1
lim n Pr
> ani1 xi1 ,
> ani2 xi2 = 0
n→∞
RDi1
RDi2
and consequently n Pr{1/(RDi1 ) > ani1 xi1 , . . . , 1/(RDik ) > anik xik } → 0 as n → ∞.
26
L. R. Belzile and J. G. Nešlehová
Case II. {i1 , . . . , ik } ∩ I1 = ∅. In this case, let Zi1 ,...,ik = max(xi1 (bi1 )1/ρ Di1 , . . . , xik (bik )1/ρ Dik ) and observe that for
any ε > 0 such that ρ + ε < min(α1 , . . . , αd ),
1
1
−ρ−ε −(ρ+ε)/ρ
E ρ+ε < ∞.
E ρ+ε ≤ xi1 bi1
Zi1 ,...,ik
Di1
Therefore, by Breiman’s Lemma,
lim n Pr
n→∞
!
!
1
1
1
> ani1 xi1 , . . . ,
> anik xik = lim n Pr
> an = E Zi−ρ
,...,i
1
k
n→∞
RDi1
RDik
RZi1 ,...,ik
"(
)−ρ #
1/ρ
= E max xi j bi j Di j
= E min
−ρ
xi j Di j
.
1≤ j≤k
bi j
1≤ j≤k
Putting the above calculations together, one then has, for any x ∈ Rd+ ,
1
1
lim n 1 − Pr
≤ an1 x1 , . . . ,
≤ and xd
n→∞
RD1
RDd
(
!)
=
X
xi−αi +
|I2 |
X
X
k=1 {i1 ,...,ik }⊆I2
i1 <···<ik
i∈I1
−ρ
x
D
i
i
j
j
(−1)k+1 E min
1≤ j≤k
bi j
Furthermore, one can readily establish by induction that for any t ∈ Rd ,
|I2 |
X
X
(−1)k+1 min(ti1 , . . . , tik ) = max(ti ).
i∈I2
k=1 {i1 ,...,ik }⊆I2
i1 <···<ik
Hence, for any x ∈ Rd+ ,
(
lim n 1 − Pr
n→∞
1
1
≤ an1 x1 , . . . ,
≤ and xd
RD1
RDd
!)
=
X
i∈I1
)#
(
"
(xi Di )−ρ
.
xi−αi + E max
i∈I2
bi
By the multivariate Poisson approximation (B.9), 1/X ∈ M(H0 ), where for all x ∈ Rd+ ,
"
)#
−ρ
X
(x
D
)
i
i
−α
.
H0 (x) = exp −
xi i − E max
i∈I2
bi
i∈I1
The univariate margins of H0 are given, for all i ∈ I1 , by F0i (x) = x−αi and for all i ∈ I2 , F0i (x) = exp(−x−ρ ). By
Sklar’s Theorem, the unique copula of H0 is given, for all u ∈ [0, 1]d , by (2), where for all x ∈ Rd+ ,
−ρ
X
xi Di
.
`(x) =
xi + E
max
i∈I2 bi
i∈I1
The first expression for ` follows immediately from Eq. (B.10). The second expression is readily verified using
Lemma 2, given the fact that if B ∼ Beta(ᾱ2 , ᾱ − ᾱ2 ), E(B−ρ ) = Γ(ᾱ2 − ρ)Γ(ᾱ)/Γ(ᾱ − ρ)Γ(ᾱ2 ).
To prove part (b), recall that by Proposition 2 (b), 1/Xi ∈ M(Φαi ). Hence, there exist sequences of positive
constants (ani ), i = 1, . . . , d, such that for all i = 1, . . . , d and all x > 0, n Pr(1/(RDi ) > ani x) → x−αi as n → ∞. By
Lemma 2 (ii), it also follows that for arbitrary x ∈ (0, ∞)d , k ∈ {2, . . . , d} and indices 1 ≤ i1 < · · · < ik ≤ d,
!
!
1
1
1
1
0 ≤ lim n Pr
> ani1 xi1 , . . . ,
> anik xik ≤ lim n Pr
> ani1 xi1 ,
> ani2 xi2 = 0.
n→∞
n→∞
RDi1
RDik
RDi1
RDi2
Thus, by Eq. (B.9), 1/X is in the domain of attraction of the multivariate extreme-value distribution given, for all
x ∈ Rd+ , by H0 (x) = exp(−x1−α1 − · · · − xd−αd ), as was to be showed.
27
Extremal attractors of Liouville copulas
Appendix C. : Proofs from Section 4
Proof of Proposition 3. In view of Corollary 1 and Theorem 2 in [27], it only remains to derive the explicit expression
for `nD . Because 1−ψ(1/·) ∈ R−ρ , there exists a slowly varying function L such that for all x > 0, 1−ψ(1/x) = x−ρ L(x).
Given that the distribution function ψ(1/·) is in the domain of attraction of Φρ , the Poisson approximation implies that
there exists a sequence (an ) of positive constants such that, for all x > 0,
lim n 1 − ψ{1/(an x)} = lim n(an x)−ρ L(an x) = x−ρ .
(C.1)
n→∞
n→∞
Furthermore, by Equation (A6) in the proof of Theorem 2 (a) in [27], one has, for any j = 1, . . . , ᾱ − 2,
(−1) j x− j ψ( j) (1/x)
= 1,
x→∞
κ j x−ρ L(x)
(C.2)
lim
where κ j = ρΓ( j − ρ)/Γ(1 − ρ). Now for all i = 1, . . . , d, Eq. (7) yields, for any x > 0,
!)
!
(
α
i −1
j
− j ( j)
X
1
(−1)
(a
x)
ψ
(1/a
x)
1
n
n
1
−
.
= n(an x)−ρ L(an x)
n Pr
> an x = n 1 − H̄i
−ρ L(a x)
Xi
an x
j!(a
x)
n
n
j=1
Given that an → ∞ as n → ∞, the last expression converges by Equations (C.1) and (C.2) as n → ∞ to
α
α
i −1
i −1
X
X
κ
Γ( j − ρ)
c(αi , −ρ)
j
−ρ
1
−
x−ρ
=
x
1
−
ρ
= x−ρ
.
j!
Γ(
j
+
1)Γ(1
−
ρ)
Γ(1 − ρ)
j=1
j=1
The Poisson approximation thus implies that, as n → ∞, for all i = 1, . . . , d and x > 0,
!
(
)
1
n
−ρ c(αi , −ρ)
H̄i
→ exp −x
.
an x
Γ(1 − ρ)
(C.3)
For any x ∈ (0, ∞)d , let 1/(an x) = {1/(an x1 ), . . . , 1/(an xd )} and denote by x̄H the harmonic mean of x, viz. x̄H =
d/(1/x1 + · · · + 1/xd ). From Eq. (6) one then has
− j1 −···− jd
(
!)
!
d
j1 +···+ jd an x̄H
( j1 +···+ jd )
d
j
i
X
Y
−ρ
(−1)
ψ
1
an x̄H
an x̄H
x̄
d
an x̄H
H
1
−
n 1 − H̄
=n
L
−ρ
a
x̄
a
x̄
n H
n H
an x
d
d
dxi
j
!
·
·
·
j
!
L
(
j
,...,
j
)∈I
i=1
1
d
1
d
α
d
d
( j ,..., j ),0
1
d
d
By Eq. (C.2), the right most expression in the curly brackets converges, as n → ∞, to
1−ρ
X
( j1 ,..., jd )∈Iα
( j1 ,..., jd ),0d
d
Γ( j1 + · · · + jd − ρ) Y x̄H
Γ(1 − ρ) j1 ! · · · jd ! i=1 dxi
! ji
=1−ρ
X
( j1 ,..., jd )∈Iα
( j1 ,..., jd ),0d
d
Γ( j1 + · · · + jd − ρ) Y
1
1/xi
Γ(1 − ρ)
Γ(
j
+
1)
1/x
+
· · · + 1/xd
i
1
i=1
Furthermore, Eq. (C.1) implies that, as n → ∞,
!ρ
a x̄ −ρ a x̄
1
1
n H
n H
n
L
→
+ ··· +
.
d
d
x1
xd
Consequently, as n → ∞, n{1 − H̄(1/an x)} → − log H0 (x), where
j
!ρ
i
d
X Γ( j1 + · · · + jd − ρ) Y
1/xi
1
1
1
− log H0 (x) =
+ ··· +
1
−
ρ
.
P
d
1
x1
xd
Γ(1 − ρ)
Γ( ji + 1)
j=1 x j
i=1
(
j
,...,
j
)∈I
1
d
α
( j ,..., j ),0
1
d
d
! ji
.
28
L. R. Belzile and J. G. Nešlehová
By Eq. (B.9), 1/X ∈ M(H0 ). From Eq. (C.3), the univariate margins of H0 are scaled Fréchet, and Sklar’s theorem
implies that the unique copula of H0 is of the form (2) with stable tail dependence function as in Proposition 3.
Proof of Proposition 4. In view of Corollary 2 and Theorem 1 in [27], it only remains to compute the expression for
`pD given in part (a). Suppose that ψ ∈ R−ρ for some ρ > 0. This means that there exists a slowly varying function
such that for all x > 0, ψ(x) = x−ρ L(x). Because ψ is itself a survival function, ψ ∈ M(Φρ ) and by the univariate
Poisson approximation, there exists a sequence (an ) of strictly positive constants such that, for all x > 0,
lim nψ(an x) = x−ρ .
(C.4)
n→∞
Furthermore, by Equation (A1) in the proof of Theorem 1 (a) in [27], one has, for any j = 1, . . . , ᾱ − 1,
(−1) j x j ψ( j) (x)
= c( j, ρ).
x→∞
ψ(x)
(C.5)
lim
Now let X be the Dirichlet random vector with parameters α and radial part R whose Williamson ᾱ-transform is ψ.
Denote the distribution function of X by H and its univariate margins by Fi , i = 1, . . . , d. Then for all i = 1, . . . , d,
Equations (C.4) and (C.5) imply that
lim nF̄i (an x) = lim n
n→∞
n→∞
α
i −1
X
j=0
α −1
i
X
Γ( j + ρ)
x−ρ c(αi , ρ)
(−1) j (an x) j ψ( j) (an x)
= x−ρ
=
j!
Γ(ρ)Γ( j + 1)
Γ(ρ + 1)
j=0
(C.6)
and hence, by the Poisson approximation, Fin (x) → exp{−x−ρ c(αi , ρ)/Γ(ρ + 1)} as n → ∞.
Next, for arbitrary k = 1, . . . , d and 1 ≤ i1 < · · · < ik ≤ d, let I(αi1 ,...,αik ) = {0, . . . , αi1 − 1} × · · · × {0, . . . , αik − 1}.
For any x ∈ (0, ∞)d , Equations (6), (C.4) and (C.5) imply that
n→∞
n→∞
(−1) j1 +···+ jk
( j1 ,..., jk )∈I(αi
1
,...,αi )
k
= (xi1 + · · · + xik )−ρ
X
(−1)
k+1
k=1 1≤i1 <···<ik ≤d
k
X
( j1 ,..., jk )∈I(αi
Therefore, for any x ∈ (0, ∞)d ,
d
X
lim n
n→∞
ψ( j1 +···+ jk ) {an (xi1 + · · · + xik )} Y
(an xim ) jm
j1 ! · · · jk !
m=1
k
X
lim n Pr(Xi1 > xi1 , . . . , Xik > xik ) = lim n
,...,αi )
1
k
x im
Γ( j1 + · · · + jk + ρ) Y
Γ(ρ) j1 ! · · · jk ! m=1 xi1 + · · · + xik
! jm
.
Pr(Xi1 > an xi1 , . . . , Xik > an xik )
= − log H0 (x),
where
− log H0 (x) =
d
X
X
(−1)k+1 (xi1 + · · · + xik )−ρ
k=1 1≤i1 <···<ik ≤d
k
X
( j1 ,..., jk )∈I(αi
,...,αi )
1
k
xim
Γ( j1 + · · · + jk + ρ) Y
Γ(ρ) j1 ! · · · jk ! m=1 xi1 + · · · + xik
! jm
.
By Eq. (B.9), X ∈ M(H0 ). As argued above, the univariate margins of H0 are given, for all i = 1, . . . , d and x > 0, by
exp{−x−ρ c(αi , ρ)/Γ(ρ + 1)}. Sklar’s theorem thus implies that the unique copula of H0 is of the form (2) with stable
tail dependence function indeed as given by the expression in part (a).
Appendix D. : Proofs from Section 5
Proof of Proposition 5. First, we show that for any ρ > − min(α1 , . . . , αd ), ρ , 0,
d
−ρ−ᾱ d
)#
"
(
Z
Y
X
xi Dρi
Γ(ᾱ)
1/ρ
E max
=
max(xi ti ) {c(αi , ρ)ti }
{c(αi , ρ)}αi /ρ (ti )αi /ρ−1 dt.
Qd
d−1
1≤i≤d c(αi , ρ)
|ρ|
Γ(α
)
S
i
d
i=1
i=1
i=1
(D.1)
29
Extremal attractors of Liouville copulas
d
Indeed, using the fact that (D1 , . . . , Dd ) = Z/kZk, where Zi ∼ Ga(αi , 1), i = 1, . . . , d are independent,
)# Z
)
(
"
(
d
Y
e−zi zαi i −1
xi zρi
xi Dρi
=
(z1 + · · · + zd )−ρ
max
dz.
E max
1≤i≤d c(αi , ρ)
Γ(αi )
Rd+ 1≤i≤d c(αi , ρ)
i=1
P
P
Make a change of variable ti = {zρi /c(αi , ρ)}/ dj=1 zρj /c(α j , ρ) for i = 1, . . . , d − 1 and w = dj=1 zρj /c(α j , ρ). For ease of
P
1/ρ
notation, set also td = 1 − d−1
and the absolute value of the Jacobian is
i=1 ti . Then, for i = 1, . . . , d, zi = {c(αi , ρ)ti w}
d
1 d/ρ−1 Y
c(αi , ρ)1/ρ ti1/ρ−1 .
w
|ρ|d
i=1
|J| =
Therefore,
xi Dρi
E max
1≤i≤d c(αi , ρ)
"
(
)#
−ρ d
d
Y
X
{c(αi , ρ)ti }1/ρ
c(αi , ρ)αi /ρ tiαi /ρ−1
=
max
(x
t
)
Q
i
i
d
d
1≤i≤d
|ρ|
i=1 Γ(αi ) Sd
i=1
i=1
Z ∞
1/ρ Pd
1/ρ
×
wᾱ/ρ−1 e−w i=1 {c(αi ,ρ)ti } dw dt.
1
Z
0
Eq. (D.1) now follows from the fact that
Z
∞
ᾱ/ρ−1 −w1/ρ
w
e
1/ρ
i=1 {c(αi ,ρ)ti }
Pd
0
d
−ᾱ
X
1/ρ
dw = |ρ| Γ(ᾱ) {c(αi , ρ)ti } .
i=1
The expression for hD now follows directly from Eqs. (3) and (D.1), while the formulas for hpD and hnD obtain upon
setting ρ = ρ and ρ = −ρ, respectively.
Appendix E. : Proofs from Section 7
Proof of Proposition 6. For k = 1, . . . , d, the formula for the kth order mixed partial derivatives of `D (1/x) can
be established from Eq. (12). Indeed, if V denotes a random vector with independent scaled Gamma components
Vi ∼ sGa{1/c(αi , ρ), 1/ρ, αi }, then the point process representation Eq. (12) implies that, for all x ∈ Rd+ ,
(
)
Z ∞
Z ∞
d
Y
Vi
1
1
D
1 −
` (1/x) =
Pr
F xi t;
(18)
> xi for at least one i ∈ {1, . . . , d} dt =
, , αi dt.
t
c(αi , ρ) ρ
0
0
i=1
For any k = 1, . . . , d, the expression on the right-hand side of Eq. (18) can be differentiated with respect to x1 , . . . , xk
under the integral sign. This gives the formulas for ∂`D (1/x)/∂x1 . . . ∂xk . When k = d, Eq. (18) implies that
∂d `D (1/x)
=−
∂x1 · · · ∂xd
(
d
Y
t
f xi t;
)
1
1
, , αi dt
c(αi , ρ) ρ
0
i=1
Z
d
d
X
1 Y c(α j , ρ){c(α j , ρ)x j }α j /ρ−1 ∞ α j /ρ
1/ρ
1/ρ
= d
t
exp −t
{c(α j , ρ)x j } dt
Γ(α j )
ρ j=1
0
j=1
=
Z
∞
d
Γ(ᾱ + ρ)
ρd−1
d
Y
c(α j , ρ){c(α j , ρ)x j }α j /ρ−1
iᾱ+ρ
d
1/ρ
j=1
j=1 {c(α j , ρ)x j }
hP
Γ(α j )
,
P
where the last equality follows upon making the change of variable u = dj=1 {c(α j , ρ)x j }1/ρ t1/ρ . Alternatively, Theorem 1 in [7] implies that that the dth order mixed partial derivative of ` D (1/x) equals −dkxk−d−1 hD (x/kxk; ρ, α), which
indeed simplifies to −dhD (x; ρ, α) given that hD (x/kxk; ρ, α) = kxkd+1 hD (x; ρ, α).
Finally, the formulas for F(x; a, b, c) follow immediately from the fact that the scaled Gamma distribution is also
the distribution of the random variable aZ 1/b , where Z is Gamma with shape c and unit scaling.
30
L. R. Belzile and J. G. Nešlehová
Derivation of the gradient score. Straightforward calculations show that
!
(ᾱ + ρ)c(αi , ρ)1/ρ xi1/ρ−1
∂ log dhD (x)
1
αi
= − Pd
−1
+
∂xi
ρ
xi
ρ j=1 {c(α j , ρ)x j }1/ρ
!
!
c(αk , ρ)1/ρ xk1/ρ−1
(ᾱ + ρ)c(αi , ρ)1/ρ xi1/ρ−1 1
αi
Iik
Iik
∂2 log dhD (x)
−1 2.
= − Pd
− Pd
− 1
−
1/ρ
1/ρ
∂xi ∂xk
ρ
xi ρ j=1 {c(α j , ρ)x j }
ρ
xi
ρ j=1 {c(α j , ρ)x j }
| 10math.ST
|
A Unification and Generalization of Exact
Distributed First Order Methods
arXiv:1709.01317v2 [cs.IT] 25 Dec 2017
Dušan Jakovetić
Abstract—Recently, there has been significant progress in
the development of distributed first order methods. (At least)
two different types of methods, designed from very different
perspectives, have been proposed that achieve both exact and
linear convergence when a constant step size is used – a favorable
feature that was not achievable by most prior methods. In this
paper, we unify, generalize, and improve convergence speed of
these exact distributed first order methods. We first carry out
a novel unifying analysis that sheds light on how the different
existing methods compare. The analysis reveals that a major
difference between the methods is on how a past dual gradient of
an associated augmented Lagrangian dual function is weighted.
We then capitalize on the insights from the analysis to derive
a novel method – with a tuned past gradient weighting –
that improves upon the existing methods. We establish for the
proposed generalized method global R-linear convergence rate
under strongly convex costs with Lipschitz continuous gradients.
Index Terms—Distributed optimization, Consensus optimization, Exact distributed first order methods, R-linear convergence
rate.
I. I NTRODUCTION
Context and motivation. Distributed optimization methods
for solving convex optimization problems over networks,
e.g., [1]–[8], have gained a significant renewed and growing
interest over the last decade, motivated by various applications,
ranging from inference problems in sensor networks, e.g., [9],
[10], to distributed learning, e.g., [11], to distributed control
problems, e.g., [12].
Recently, there have been significant advances in the context
of distributed first order methods. A distinctive feature of
the novel methods, proposed and analyzed in [13]–[26], is
that, unlike most of prior distributed first order algorithms,
e.g., [1], [27], [5], [7], they converge to the exact solution
even when a constant (non-diminishing) step size is used.
This property allows the methods to achieve global linear
convergence rates, when the nodes’ local costs are strongly
convex and have Lipschitz continuous gradients. Among the
exact distributed first order methods, (at least) two different
types of methods have been proposed, each designed through
a very different methodology. The method in [13], dubbed
Extra, see also [25], [23], modifies the update of the standard
distributed gradient method, e.g., [1], by introducing two
different sets of weighting coefficients (weight matrices) for
any two consecutive iterations of the algorithm, as opposed
to a single weight matrix with the standard method [1].
The second type of methods, [14], [19], [18], [20], [21],
replaces the nodes’ local gradient with a “tracked” value
D. Jakovetić is with the Department of Mathematics and Informatics,
Faculty of Sciences, University of Novi Sad, Novi Sad, Serbia. The research is
supported by Ministry of Education, Science and Technological Development,
Republic of Serbia, grant no. 174030. Author’s e-mail: [email protected].
of the network-wide average of the nodes’ local gradients.
The method in [14] has been proved to be convergent under
time-varying networks, e.g., [19], node-varying step sizes,
e.g.,[18], and Nesterov acceleration [22], while such studies
have not been provided to date for the method in [13]. Both
references [13] and [14] establish global linear convergence
rates of the exact distributed first order methods that they
study. Further, reference [28] shows that Extra is equivalent
to a primal-dual gradient-like method (see, e.g., [29], [30],
[31], for primal-dual (sub)gradient methods) applied on the
augmented Lagrangian dual problem of a reformulation of the
original problem of interest. Reference [19] demonstrates that
the method therein, equivalent to the method in [14], can be
put in the form of Extra, with a specific choice of the two
Extra’s weight matrices, and it also provides a primal-dual
interpretation of the exact method therein. (A more detailed
review of works [13]–[26] is provided further ahead.)
Contributions. The main contributions of this paper are
to unify, generalize, and improve convergence speed of exact
distributed first order methods. First, with both the methods
in [13] and [14], we provide an augmented Lagrangian-based
analysis. that gives for both methods a characterization of the
(joint) evolution of primal and dual errors along iterations.
This characterization reveals the effect of the difference between the two methods on their performance and sheds light
on understanding of how the two methods compare under
various problem model scenarios. We further provide a generalized method, parameterized with an additional, easy-to-tune
network-wide weight matrix, that determines the weighting
of the term that corresponds to a past dual gradient of the
associated augmented Lagrangian function. The generalized
method subsumes the two known methods in [13] and [14]
upon setting a weighting parameter matrix to appropriate
specific values. We describe how to set the weighting parameter matrix to improve upon both existing methods. The
proposed parameter matrix tuning is based on an insight from
our analysis on how the negative effect of the primal error
on the dual error can be reduced (compensated) through the
introduced parameter matrix.
The paper carries out convergence rate analysis of the
proposed generalized method when nodes’ local costs are
strongly convex, have Lipschitz continuous gradient, and the
underlying network is static. With the proposed generalized
method, we establish a global R-linear convergence rate,
when the algorithm’s step size is appropriately set. Numerical
examples confirm the insights from our analysis on the mutual
comparison of the methods in [13] and [14], as well as the
improvements of the proposed generalized method over the
two existing ones.
2
Brief literature review. Distributed computation and optimization has been studied for a long time, e.g., [32]. More
recently, reference [1] proposes a distributed first order (subgradient) method for unconstrained problems, allowing for
possibly non-differentiable, convex nodes’ local costs and carries out its convergence and convergence rate analysis for deterministically time varying networks. A distributed projected
subgradient method for constrained problems and possibly
non-differentiable local costs has been proposed and analyzed
in [27]. References [7], [17], [13], [14], [15], [16], [18], [19],
[20], [21], [22] study unconstrained distributed optimization
under more structured local costs. In [7], distributed gradient
methods with an acceleration based on the Nesterov (centralized) gradient method [33] have been proposed and analyzed
under differentiable local costs with Lipschitz continuous and
bounded gradients. The methods in [1], [27], [7], [34] converge
to the exact solution only when a diminishing step-size is
used; when a constant step size is used they converge to a
solution neighborhood. References [35], [36] propose different
types of primal-dual methods, prove their convergence to
the exact solution for a wide class of problems assuming
diminishing step-sizes, and are not concerned with establishing
the methods’ convergence rates. The authors of [37] use
insights from control theory to propose a gradient-like algorithm for which they prove exact convergence under certain
conditions, while the paper is not concerned with analyzing the
method’s convergence rate. Reference [17] proposes several
variants of distributed AL methods that converge to the exact
solution under a constant step size and establishes their linear
convergence rates for twice continuously differentiable costs
with bounded Hessian.
References [13], [14], [15], [16], [18], [19], [20], [21], [22]
develop and/or analyze different variants of exact distributed
first order methods under various assumptions on the nodes’
local costs, algorithm step sizes, and the underlying network.
The papers [13], [14] propose two different exact distributed
first order methods and analyze their convergence rates, as
already discussed above. References [15], [16] develop exact
methods based on diffusion algorithms (see, e.g., [3]) and
through a primal-dual type method on an associated AL
function. Under twice differentiable local costs, each node’s
cost having Lipschitz continuous gradient, and at least one
node’s cost being strongly convex, the papers show linear
convergence rates for the methods therein, allowing for different step sizes across nodes and for a wider range of step
sizes and admissible weight matrices with respect to [13]. The
papers [18], [19], [20], [21] consider strongly convex local
costs with Lipschitz continuous gradients. Under this setting,
reference [18] establishes global linear convergence of the
method studied therein when nodes utilize uncoordinated step
sizes. The paper [19] establishes global linear convergence for
time-varying networks. The authors of [21] prove global linear
rates under both time varying networks and uncoordinated step
sizes, while the issue of uncoordinated step-sizes is previously
considered in [20]. Finally, the paper [22] proposes an accelerated exact distributed first order method based on the Nesterov
acceleration [33] and establishes its convergence rates for local
costs with Lipschitz continuous gradients, both in the presence
and in absence of the strong convexity assumption. Under
strongly convex local costs that have Lipschitz continuous
gradients, the authors of [26] develop optimal distributed
methods, where the optimality is in terms of the number of
oracle calls of a therein appropriately defined oracle. However,
their method is of a different type than [18], [19], [20], [21],
[22] and is different from the method proposed here. Namely,
the method in [26] requires evaluation of Fenchel conjugates of
the nodes’ local costs at each iteration, and hence it in general
has a much larger computational cost per iteration than the
methods in [18], [19], [20], [21], [22] and the method proposed
in this paper. To the best of our knowledge, communication
and computation optimality of distributed first order methods
that do not involve Fenchel conjugates has not been studied
to date.
Paper organization. The next paragraph introduces notation. Section II explains the model that we assume and reviews
existing distributed first order methods. Section III presents the
proposed algorithm and relates it with the existing methods.
Section IV establishes global R-linear convergence rate of the
proposed method, while Section V provides simulation examples. Finally, we conclude in Section VI. Certain auxiliary
proofs are provided in the Appendix.
Notation. We denote by: R the set of real numbers; Rd
the d-dimensional Euclidean real coordinate space; Aij the
entry in the i-th row and j-th column of a matrix A; A>
the transpose of a matrix A; ⊗ the Kronecker product of
matrices; I, 0, and 1, respectively, the identity matrix, the
zero matrix, and the column vector with unit entries; J the
N × N matrix J := (1/N )11> ; A 0 (A 0) means
that the symmetric matrix A is positive definite (respectively,
positive semi-definite); k·k = k·k2 the Euclidean (respectively,
spectral) norm of its vector (respectively, matrix) argument;
λi (·) the i-th largest eigenvalue; Diag (a) the diagonal matrix
with the diagonal equal to the vector a; | · | the cardinality
of a set; ∇h(w) and ∇2 h(w), respectively, the gradient and
Hessian evaluated at w of a function h : Rd → R, d ≥ 1.
Finally, for two positive sequences ηn and χn , we have:
ηn = O(χn ) if lim supn→∞ χηnn < ∞; and ηn = Ω(χn ) if
lim inf n→∞ χηnn > 0.
II. M ODEL AND PRELIMINARIES
Subsection II-A describes the optimization and network
models that we assume. Subsection II-B reviews the standard
(inexact) distributed gradient method in [1], as well as the
exact methods in [13] and [14], and it reviews the known
equivalence (see [28]) of the method in [13] and a primaldual gradient-like method.
A. Optimization and network models
We consider distributed optimization where N nodes in a
connected network solve the following problem:
minx∈Rd f (x) :=
N
X
fi (x).
(1)
i=1
Here, fi : Rd → R is a convex function known only by node i.
Throughout the paper, we impose the following assumption on
the fi ’s.
3
Assumption 1 Each function fi : Rd → R, i = 1, ..., N , is
strongly convex with strong convexity parameter µ, and it
has Lipschitz continuous gradient with Lipschitz constant L,
where L ≥ µ > 0. That is, for all i = 1, ..., N , there holds:
fi (y) ≥ fi (x) + ∇fi (x)> (y − x)
µ
+ kx − yk2 , x, y ∈ Rd
2
k∇fi (x) − ∇fi (y)k ≤ L kx − yk, x, y ∈ Rd .
For a specific result (precisely, Lemma 3 ahead), we additionally assume the following.
Assumption 2 Each function fi : Rd → R, i = 1, ..., N , is
twice continuously differentiable.
Under Assumptions 1 and 2, there holds for every x ∈ Rd
that:
µ I ∇2 fi (x) L I.
Further, under Assumption 1, problem (1) is solvable and has
the unique solution x? ∈ Rd .
Nodes i = 1, ..., N constitute an undirected network G =
(V, E), where V is the set of nodes and E is the set of edges.
The presence of edge {i, j} ∈ E means that the nodes i and
j can directly exchange messages through a communication
link. Further, let Ωi be the set of all neighbors of a node i
(including i).
Assumption 3 The network G = (V, E) is connected, undirected and simple (no self-loops nor multiple links).
We associate with network G a N × N symmetric, (doubly)
stochastic weight matrix W . Further, we let: Wij = Wji > 0,
if {i, j} ∈ E,
/ E, i 6= j; and
P i 6= j; Wij = Wji = 0, if {i, j} ∈
Wii = 1 − j6=i Wij > 0, for all i = 1, ..., N . Denote by λi ,
i = 1, 2, ..., N , the eigenvalues of W , ordered in a descending
order; it can be shown that they obey 1 = λ1 > λ2 ≥ ... ≥
λN > −1.
Throughout the paper, we consider several iterative distributed methods to solve (1). An algorithm’s iterations are
(k)
indexed by k = 0, 1, 2, ... Further, we denote by xi ∈ Rd the
estimate of the solution to (1) available to node i at iteration k.
To avoid notational clutter, we will keep the same notation
(k)
xi across different methods, while it is clear from context
which method
is in question. For compact
notation, we use
(k)
(k)
(k)
>
∈ RN d to denote
x(k) = (x1 )> , (x2 )> , ..., (xN )>
the vector that stacks the solution estimates by all nodes at
iteration k. We use analogous notation for certain auxiliary
(k)
sequences that the algorithms maintain (e.g., see ahead si
and s(k) with (14).) Again, to simplify notation, with all
methods to be considered we will assume equal initialization
(0)
(0)
(0)
across all nodes, i.e., we let x1 = x2 = ... = xN ; e.g.,
(0)
nodes can set xi = 0, for all i.
For future reference, we introduce the following quantities.
We let the (N d) × (N d) matrix W = W ⊗ I, where I is the
d × d identity matrix. That is, W is a N × N block matrix
with d × d blocks, such that the block at the (i, j)-th position
equals Wij I.1 Note that the (i, j)-th and (j, i)-th block of W
equal to zero if {i, j} ∈
/ E, i 6= j. In other words, we say that
W respects the sparsity pattern of the underlying graph G.
Further, recall that J = N1 11> is the ideal consensus matrix,2
f = W −J = (W −J)⊗I the
and let J = J ⊗I. Denote by W
matrix that describes how far is W from the ideal matrix J .
f = max{λ2 , −λN } =: σ ∈ [0, 1).
It can be shown that kWk
•
?
Next, let x = 1 ⊗ x , where we recall that 1 is an all-ones
vector (here of size N × 1), and x? is the (d × 1) solution to
(1). In other words, x• concatenates N repetitions of x? on
top of each other; the i-th repetition corresponds to node i in
the network. Our goal is that, with a distributed method, we
(k)
have x(k) → x• , or, equivalently, xi → x? , for all nodes
i = 1, ..., N . Further, we define function F : RN d → R, by
F (x) := F (x1 , ..., xN ) =
N
X
fi (xi ).
(2)
i=1
Note that, under Assumption 1, F is strongly convex with
strong convexity parameter µ, and it has Lipschitz continuous
gradient with Lipschitz constant L.
B. Review of distributed first order methods
We review three existing distributed first order methods
that are of relevance to our studies – the standard distributed
gradient method in [1], the Extra method in [13], and the
method in [14].
Standard distributed gradient method in [1]. The method
(k)
updates the solution estimate xi at each node i by weightaveraging node i’s solution estimate with the estimates of its
immediate neighbors j ∈ Ωi \ {i}, and then by taking a step
in the negative local gradient’s direction. This corresponds to
the following update rule at arbitrary node i:
X
(k+1)
(k)
(k)
xi
=
Wij xj − α ∇fi (xi ), k = 0, 1, ... (3)
j∈Ωi
In matrix format, the network-wide update is as follows:3
x(k+1) = W x(k) − α ∇F (x(k) ), k = 0, 1, ...
(4)
Here, constant α > 0 is the algorithm’s step size. A drawback
of algorithm (4) is that (when constant step size α is used)
it does not converge to the exact solution x• := 1 ⊗ x? ,
but only to a point in a O(α) solution neighborhood; see,
e.g., [39]. The algorithm can be made convergent to x• by
1 We will frequently work with Kronecker products of types A ⊗ B, a ⊗ B,
and a ⊗ b, where A is an N × N matrix, a – N × 1 vector, B – d × d matrix,
and b – d×1 vector. In other words, the Kronecker products throughout always
appear with the first argument of dimension either N × N or N × 1, and the
second argument either d × d or d × 1. As the capital letters denote matrices
and the lower case letters denote vectors, the dimensions of the Kronecker
products arguments will be clear.
2 The consensus algorithm, e.g., [38], computes the global average of nodes’
local quantities through a linear system iteration with a stochastic system
matrix W . When W = J, consensus converges in a single iteration, hence
we name J the ideal consensus matrix. Clearly, J is not realizable over a
generic graph as it is not sparse, but it is usually desirable to have W as
close as possible to J in an appropriate sense, given the constraints on the
network sparsity.
3 To save space, we will present all subsequent algorithms in compact forms
only. From a compact form, it is straightforward to recover local update forms
at any node i.
4
taking an appropriately set diminishing step size (e.g., square
summable, non-summable), but the convergence rate of the
resulting method is sublinear.
Extra [13]. The method modifies the update rule (1) and
achieves exact (global linear) convergence with a fixed step
size α. Extra works as follows. It uses the following update
rule:4
x(1)
x
(k+1)
= W x(0) − α ∇F (x(0) )
(5)
= 2 W x(k) − α ∇F (x(k) ) − W x(k−1)
(6)
+ α ∇F (x
(k−1)
), k = 1, 2, ...
Reference [28] demonstrates that algorithm (5)–(6) is a
primal-dual gradient-like method; see, e.g., [29], [30], [31], for
primal-dual (sub)gradient methods. Denote by L := I − W,
and consider the following constrained problem:
minimize F (x)
1 1/2
L x = 0.
α
subject to
(7)
It can be shown, e.g., [28], that L1/2 x = 0, if and only if x1 =
x2 = ... = xN , where xi ∈ Rd is the i-th consecutive block
of x = ( (x1 )> , ..., (xN )> )> . Therefore, (7) is equivalent to
(1). Introduce the augmented Lagrangian function A : RN d ×
RN d → R (with the penalty parameter equal to α) associated
with (7):
1 >
1
x Lx,
(8)
A(x, µ) = F (x) + µ> L1/2 x +
α
2α
and the corresponding dual problem:
maximizeµ∈RN d
inf A(x, µ).
x∈RN d
(9)
Consider the following primal-dual method to solve (9):5
x(k+1)
(k+1)
µ
=
=
x(k) − α ∇x A(x(k) , µ(k) )
(k)
µ
+ α ∇µ A(x
(k+1)
(0)
(k)
,µ
(10)
=
= x(k) − α
1
L x(k) + ∇F (x(k) ) + u(k) (12)
α
1
L x(k+1) , k = 0, 1, ...
(13)
α
It turns out that (12)–(13) with a proper initialization is
equivalent to (5)–(6).7
u(k+1)
= u(k) +
Lemma 1 ([28]) The sequence of iterates {x(k) } generated by
(0)
(0)
algorithm (5)–(6), with initialization xi = xj , for all i, j,
is the same as the sequence of iterates generated by (12)–(13),
(0)
with the same initialization of xi , i = 1, ..., N , and with u(k)
initialized to zero.
Under Assumptions 1 and 3 an appropriately chosen step size
α, one has that x(k) → x• , and u(k) → −∇F (x• ), at an
R-linear rate [13].
The exact method in [14]. The authors of [14], see
also [18], [19], [20], [21], [22], consider a distributed first
order method that, besides solution estimate x(k) ∈ RN d , also
maintains an auxiliary variable s(k) ∈ RN d . Here, at each node
(k)
i, quantity si ∈ Rd serves to approximate the network-wide
PN
(k)
gradient average N1 i=1 ∇fi (xi ); then, the gradient con(k)
tribution −α∇F (x ) with the standard distributed gradient
method (4) is replaced with s(k) . More precisely, the update
rule for k = 0, 1, ... is as follows:
x(k+1)
=
W x(k) − α s(k)
(k+1)
=
W s(k) + ∇F (x(k+1) ) − ∇F (x(k) ), (15)
s
(14)
with s(0) = ∇F (x(0) ). The method also achieves a global
R-linear convergence under appropriately chosen step-size α,
when Assumptions 1 and 3 are in force.
Nd
III. T HE PROPOSED METHOD
Subsection III-A presents the method that we propose and
explains how to recover the existing methods in [13], [14] by
a particular setting of a proposed method’s parameter matrix.
Subsection III-B gives further insights into the proposed
method and explains how to tune the parameter matrix. Finally,
Subsection III-C provides a primal-dual interpretation of the
proposed method and the methods in [13], [14].
µ(k) + L1/2 x(k+1) , k = 0, 1, ...
Introducing the new variable u(k) :=
4 The
x(k+1)
), k = 0, 1, ...(11)
with step size α > 0, arbitrary x ∈ R , and µ(0) = 0.
Here, ∇x and ∇µ denote the partial derivatives with respect
to x and µ, respectively. Evaluating the partial derivatives, we
arrive at the following method:
1
L x(k)
x(k+1) = x(k) − α
α
L1/2 (k)
+ ∇F (x(k) ) +
µ
α
µ(k+1)
at the following method:6
1 1/2
αL
µ(k) , one arrives
Extra method utilizes two different weight matrices in general. When
these matrices are tuned as it is suggested in [13], Extra can be represented
as in (5)–(6). The formulas here and in [13] appear slightly different due to
f = 1 (I + W ) in [13] corresponds to W here. It is
a different notation: W
2
required in [13] for the analysis of (5)–(6) that (in our notation) (2W − I)
be doubly stochastic and that W be, besides being doubly stochastic, also
positive definite. The latter assumptions are not imposed here when analyzing
the proposed method (16)–(17) (Section IV).
5 More precisely, under appropriate conditions, with (10) one has that
x(k) , µ(k) converges to a saddle point of function A(x, µ), which then
implies that x(k) solves (7) and µ(k) solves (9).
A. The proposed method and its relation with existing algorithms
We now describe the generalized exact first order method
that we propose. The method subsumes the known methods [13] and [14] upon a specific choice of the tuning
parameters, as explained below. The algorithm maintains over
iterations k the primal variable (solution estimate) x(k) ∈ RN d
6 See also [17] for similar methods with multiple primal updates per each
dual update.
7 Reference [28] shows the equivalence under a more general initialization;
we keep the one as in Lemma 1 for simplicity.
5
>
(k)
(k)
(ex )> , (eu )> , for k = 0, 1, ..., satisfies the following
recursion:
"
#
" (k) #
(k+1)
ex
W − α Hk
−α I
ex
(k+1) = (W − I) (H − B) W − J
(k) . (18)
k
eu
eu
and the dual variable u(k) ∈ RN d , initialized with the zero
vector. The update rule is for k = 0, 1, ... given as follows:
x(k+1) = W x(k) − α ∇F (x(k) ) + u(k)
(16)
u(k+1) = u(k) − L ∇F (x(k) ) + u(k) − B x(k) .(17)
Here, quantities W, L, and α are the same as before. We
note that one can use two different (doubly stochastic) weight
matrices W1 and W2 in (16) and (17) above; more precisely,
one can replace W with W1 in (16) and L with I − W2 in
(17). However, throughout we keep a single weight matrix W
for simplicity of the analysis and presentation. Quantity B
is a (N d) × (N d) symmetric matrix that respects the blocksparsity pattern of the underlying graph G and satisfies the
property that, for any y ∈ Rd , there exists some c ∈ R, such
that B (1 ⊗ y) = c (1 ⊗ y). Specifically, we will consider
the following choices (that clearly obey the latter conditions):
1) B = b I, where b ≥ 0 is a scalar parameter; and 2)
B = b0 W, for b0 ≥ 0. These choices are easy to implement
and incur no additional communication overhead while lead
to efficient algorithms (see also Section V). That is, with
the above choices of matrix B, the proposed method (16)–
(17) requires communicating 2 vectors of size d per node,
per iteration. We note that, for a generic choice of matrix B,
method (16)–(17) incurs one additional communication (of a
d-dimensional vector) per node, per iteration, i.e., it requires
communicating 3 vectors of size d per node, per iteration.
The next lemma, proved in the Appendix, explains how
to recover the existing exact distributed first order methods
from (16)–(17).8
Lemma 3 expresses the primal dual error e(k) as a recursion,
guided by a time varying matrix:
W − α Hk
−α I
Mk :=
,
(19)
(W − I) (Hk − B) W − J
Lemma 2 Consider algorithm (16)–(17). Then, the following
holds.
(a) Algorithm (16)–(17) with B = 0 is equivalent to
method (14), proposed in [14].
(b) Algorithm (16)–(17) with B = α1 W is equivalent to
method (5)–(6), proposed in [13].
B. Further insights into the proposed method and parameter
tuning
We give further intuition and insights into the proposed
method, the existing algorithms in [14] and [13], and we also
describe how to improve upon existing methods by appropri(k)
(k)
ately setting matrix B. Denote by ex := x(k) −x• and eu :=
u(k) + ∇F (x• ) the primal and dual errors, respectively. Also,
>
(k)
(k)
let the primal-dual error vector e(k) := (ex )> , (eu )> ,
R1
and Hk := t=0 ∇2 F x• + t (x(k) − x• ) dt. We have the
following Lemma.9
Lemma 3 Let Assumptions 1–3 hold, and consider algorithm (16)–(17). Then, the primal-dual error vector e(k) :=
8 The equivalence claimed in Lemma 2 is in the sense that the methods
generate the same sequence of iterates {x(k) } under the same initialization
x(0) , the appropriate initialization of methods’ auxiliary variables, and the
same step size α.
9 In Section IV, we show that, under Assumptions 1 and 3, e(k) converges
to zero R-linearly.
that we refer to as the error dynamics matrix.10 Equation (19)
shows the partitioning of the error dynamics matrix as a
2 × 2 block matrix with blocks of size (N d) × (N d). The
(1, 1)-th block (W − α Hk ) describes how the current primal
error affects the next primal error, the (1, 2)-th block (−α I)
describes how the current dual error affects the next primal
error, and so on. Specifically, with the methods in [14]
and [13], the blocks on the positions (1, 1), (1, 2), and (2, 2)
are of the same structure for both methods, and are of the
same structure as in (19). For clarity of explanation, assume
that the fi ’s are strongly convex quadratic functions, so that
Hk = H = ∇2 F (x) = const, for any x ∈ RN d , with
µ I H L I, and of course the same H appearing in
the error dynamics matrices for each of the three methods.
Then, the blocks (1, 1), (1, 2), and (2, 2) match completely
for the three methods. However, the error dynamics matrices
for different methods differ in the (2, 1)-th block. Specifically,
with [14], the (2, 1)-th block of the corresponding error
dynamics matrix equals (W − I) H. On the
other hand, with
[13] the block equals (W − I) H − α1 W . Intuitively, one
may expect that if [Mk ]2,1 is smaller (as measured by an
appropriate matrix norm), then the algorithm’s convergence
is likely to be faster.11 This provides an intuition on the
comparison between [14] and [13]. Namely, if H is small
relative to (H − α1 W) (for instance, when α is very small
and so α1 W has a very large norm), then we expect that the
method in [14] is faster than the method in [13], and vice
versa. This intuition is confirmed in Section V by numerical
examples.
We can go one step further and seek to tune matrix B such
that [Mk ]2,1 = (W − I)(H − B) is smallest in an appropriate
sense. We consider separately the cases B = b I and B =
b0 W. For the former, a possible worst case-type approach is as
follows: choose parameter b that solves the following problem:
min b≥0 sup kH − b Ik ,
(20)
H∈H
where H is the set of all (N d) × (N d) symmetric block
diagonal matrices H with arbitrary d × d diagonal blocks that
in addition obey the following condition: µ I H L I. It is
easy to show (see the Appendix for details) that the solution
µ+L
to (20) is b? = µ+L
2 . Note that the choice B = 2 I in (16)–
(17) does not match either [14] or [13] method and thus
10 In Section IV, we show that, under appropriate choice of parameters α
and B, the primal-dual error e(k) converges to zero R-linearly.
11 This intuition is corroborated more formally in Theorem 4 and Remark 4.
6
represents a novel algorithm. For the latter choice B = b0 W,
the analogous problem:
min b0 ≥0 sup kH − b0 Wk
(21)
H∈H
is more challenging. A sub-optimal choice for λN > 0, as
L+µ
shown in the Appendix, is b0 = 1+λ
, where we recall that λN
N
is the smallest eigenvalue of matrix W . Extensive simulations
(see Also Section V) show that the simple choice b0 = L
works well in practice.
Note that, ideally, one would like to minimize with respect
to b the following quantity: supH: µ IHL I kMk k, i.e., one
wants to take into account the full matrix Mk . This problem
is challenging in general and is hence replaced here by
method (20) (or, similarly, (21)), i.e., by considering the
(2, 1)-th block of Mk only. Note that a smaller norm of
the (2, 1)-th block of Mk might not necessarily imply a
smaller norm of the full matrix Mk . However, extensive
numerical experiments on quadratic and logistic losses (see
also Section 5) demonstrate that tuning method (20) yields
fast algorithms while at the same time is very cheap.
C. Primal-dual interpretations
We now provide a primal-dual interpretation of the proposed
method. The interpretation builds on construction (8)–(11)
from [28]. It is worth noting that [19] provides a primal-dual
interpretation of the method therein, equivalent to the method
in [14]. The construction in [19] starts from a different problem
reformulation than (8) and utilizes a different quadratic penalty
term for the Lagrangian function.
To start, we write the methods (5)–(6) and (14)–(15) in
another equivalent form (see the Appendix as to why this
equivalence also holds.) Namely, (5)–(6) can be equivalently
represented as follows (this is essentially a re-write of (12)–
(13)) but is useful to present it here):
x(k+1) = W x(k) − α ∇F (x(k) ) + u(k)
(22)
1
L x(k+1) , k = 0, 1, ...
(23)
α
Similarly, (14) can be, for k = 0, 1, ..., equivalently represented as:
x(k+1) = W x(k) − α ∇F (x(k) ) + u(k)
(24)
u(k+1)
u(k) +
=
1
1
L x(k+1) − W L x(k) . (25)
α
α
We can re-interpret (24) as a primal-dual gradient-like method
for solving (9). Namely, due to the fact that matrices W and
L commute, it is easy to see that (24)–(25) corresponds to the
following method:
u(k+1)
x(k+1)
(k+1)
µ
u(k) +
=
=
=
x(k) − α ∇x A(x(k) , µ(k) )
(k)
µ
+ α ∇µ A(x
(k+1)
(k)
,µ
(26)
)
(27)
−α W ∇µ A(x(k) , µ(k) ).
Hence, (24) is a primal-dual gradient-like method that modifies the dual update step to also incorporate the (weighted)
previous dual gradient term. Clearly, the proposed generalized
method (16)–(17) also incorporates the (weighted) previous
dual gradient term. It is shown in the Appendix that, assuming
that matrices B and L commute (which is the case for the two
specific choices B = b I and B = b0 W considered here), we
have that (16)–(17) is equivalent to the following primal-dual
method:
x(k+1)
(k+1)
µ
= x(k) − α ∇x A(x(k) , µ(k) )
(k)
= µ
(k+1)
+ α ∇µ A(x
,µ
−α (W − α B) ∇µ A(x
(28)
(k)
(k)
)
(29)
(k)
,µ
).
IV. C ONVERGENCE RATE ANALYSIS
This Section presents the results on global R-linear convergence of the proposed method (16)–(17), and provides
the needed intermediate results and proofs. The Section is
organized as follows. First, Subsection IV-A states the result
(Theorem 4) on global R-linear convergence of the proposed
method (16)–(17) and discusses the implications of the result.
Subsection IV-B sets up the analysis and gives preliminary
Lemmas. Finally, Subsection IV-C proves a series of intermediate Lemmas, followed by the proof of Theorem 4.
A. Global R-linear convergence rate: Statement of the result
We show that, under appropriate choice of step size α, the
proposed method (16)–(17) converges to the exact solution at
a global R-linear rate. Recall quantity σ = max{λ2 , −λN } ∈
[0, 1), with λi the i-th largest eigenvalue of W .
Theorem 4 Consider algorithm (16)–(17) with B = b I,
and nlet Assumptions o1 and 3 hold. Further, let α <
1/2
µ (1−σ)2 µ
0
2
2
.
min (1−σ)
19 L2 , 192 L0 L , where L = L + b − 2 b µ
Then, the sequence of iterates x(k) generated by algo•
?
rithm (16)–(17) converges to
x = 1⊗x R-linearly, i.e., there
(k)
•
k
holds kx − x k = O r , with r ∈ (0, 1). The convergence
factor r is at most max{1 − α2µ , 1+σ
2 } + , where > 0 is
arbitrarily small.
Several Remarks on Theorem 4 are now in order.
Remark 1 When B = b I is replaced with a generic symmetric
matrix B that respects the sparsity pattern of graph G and
satisfies the property that, for any y ∈ Rd , there exists some
c ∈ R, such that B (1 ⊗ y) = c (1 ⊗ y), Theorem 4 continues
to hold with L0 replaced with constant (L + kBk); see the
Appendix for the proof.
Remark 2 The maximal admissible step size for which Theorem 4 guarantees R-linear convergence can be very small
for poorly conditioned problems (L/µ large) and/or weekly
connected networks (σ close to one). However, extensive
simulations show that the proposed method converges (Rlinearly) in practice with large step sizes, e.g., α = 31L . Similar
theoretical and practical admissible values for step size α are
reported in earlier works for the methods studied therein [14],
[13].
7
Remark 3 Theorem 4 generalizes existing R-linear convergence rate results of exact distributed first order methods, e.g.,
[14], to a wider class of algorithms. Recall that for b = 0
we recover the method in [14]. Note that for b = 0 we
have L0 = L. Abstracting universal constants, the convergence
factor obtained in [14] (see Theorem 1 and Lemma 2 therein)
and the one obtained here are the same and equal
(1 − σ)2 µ2
.
(30)
1−Ω
L2
This bound is obtained here by setting the maximal
step
size
(1−σ)2 µ
α permitted by Theorem 4, which is α = Ω
.
L2
Remark 4 Adapting the results in [18] to our setting by letting
the nodes’ step-sizes and their strong convexity parameters be
equal, the results from [18] yield the convergence factor:
(1 − σ)2 µ3/2
√
.
1−Ω
N L3/2
Comparing this with bound (30) obtained here, we can see that
our bound exhibits better scaling with respect to the number
of nodes N ; on the other hand, the bounds in [18], [19] exhibit
better scaling in terms of condition number (L/µ).
Remark 5 It is interesting to observe what happens in Theorem 4 when b = µ – the case that corresponds to a novel
1/2
exact method. Then, one has L0 = L2 − µ2
, which is
arbitrarily small when L becomes close to µ. Hence, for wellconditioned problems (L close to µ), the maximal
admissible
step size in Theorem 4 becomes α = Ω (1−σ)µ
, and the
2
L 2
µ
corresponding convergence factor is 1 − Ω (1−σ)
. Hence,
L2
according to the upper bounds derived here and in [14], the
proposed method with
b =2 µ
improvesconvergence
factor
(1−σ) µ2
(1−σ) µ2
over [14] from 1−Ω
to 1−Ω
for wellL2
L2
conditioned problems (L sufficiently close to µ). Simulations
confirm large gains in convergence speed of the new method
over [14] (see Section V).
B. Setting up analysis and preliminary Lemmas
The proof of Theorem 4 is based on the small gain Theorem [40]. It is worth noting that analytical approaches using
small gain theorem have been developed in [18] and [19] prior
to our work and hold for a more general network and stepsize models than what we study here. Namely, reference [19]
carries out analysis under undirected and directed time varying
graphs, while reference [18] allows for node-dependent stepsizes. However, the analysis here is different. First, it is carried
out for a different algorithm in general (the proposed method
(16)–(17) that involves matrix B); as such, it reveals the novel
insight that the negative coupling from the primal to the dual
error can be partially compensated by tuning matrix B; see
also Remark 5. Second, when the results here are specialized
to the method in [18], [19], the obtained convergence factors
are different. (See Remark 4 for details.)
Besides the small gain Theorem, this Subsection also reviews some known results on the convergence of inexact
(centralized) gradient methods that will be used subsequently.
We start by reviewing the setting of the small gain theorem [40]. Denote by a := a(0) , a(1) , ..., a(k) , ... an infinite
sequence of vectors, a(k) ∈ Rp , k = 0, 1, ... For a fixed δ ∈
(0, 1), define the following quantities:
1 (k)
ka
k
(31)
kakδ,K :=
max
k=0,...,K
δk
1 (k)
kakδ := sup
ka
k
.
(32)
δk
k≥0
0
Clearly, we have that kakδ,K ≤ kakδ,K ≤ kakδ , for
K 0 ≥ K ≥ 0. It is also clear that, if kakδ is finite, then the sequence a(k) converges to zero R-linearly. Indeed, provided that
kakδ ≤ Ca < ∞, there holds: ka(k) k ≤ Ca δ k , k = 0, 1, ...
Hence, if for some δ ∈ (0, 1) we have that kakδ is finite,
then the sequence {a(k) } converges to zero R-linearly with
convergence factor at most δ. We state the small gain theorem
for two infinite sequences a and b as this suffices for our
analysis; for the more general Theorem involving an arbitrary
(finite) number of infinite sequences, see, e.g., [40], [18].
Theorem 5 Consider two infinite sequences a = a(0) , a(1) , ...,
and b = b(0) , b(1) , ..., with a(k) , b(k) ∈ Rp , k = 0, 1, ...
Suppose that for some δ ∈ (0, 1), and for all K = 0, 1, ...,
there holds:
kakδ,K
kbk
δ,K
≤ γ1 kbkδ,K + ω1
≤ γ2 kak
δ,K
+ ω2 ,
(33)
(34)
where γ1 , γ2 ≥ 0 and γ1 γ2 < 1. Then, there holds:
kakδ ≤
1
(ω2 γ1 + ω1 ) .
1 − γ1 γ2
(35)
We will frequently use the following simple Lemma.
Lemma 6 Consider two infinite sequences a = a(0) , a(1) , ...,
and b = b(0) , b(1) , ..., with a(k) , b(k) ∈ Rp . Suppose that, for
all k = 0, 1, ..., there holds:
ka(k+1) k ≤ c1 ka(k) k + c2 kb(k) k,
(36)
where ci ≥ 0, i = 1, 2. Then, for all K = 0, 1, ..., for any
δ ∈ (0, 1), we have:
c1
c2
kakδ,K ≤
kakδ,K +
kbkδ,K + ka(0) k.
(37)
δ
δ
1
Proof: Divide inequality (36) by δk+1
, δ ∈ (0, 1). The
resulting inequality implies, for all K = 1, 2, ... that:
c1
1 (k)
(k+1)
ka
k
ka
k
≤
max
k=0,...,K−1
δ k+1
δ k=0,...,K−1 δ k
c2
c1
1 (k)
c2
max
kb k = kakδ,K−1 + kbkδ,K−1
δ k=0,...,K−1 δ k
δ
δ
c1
c2
≤ kakδ,K + kbkδ,K .
(38)
δ
δ
max
+
1
Note that (38) implies that, for all K = 0, 1, ...
1
(k+1)
kakδ,K =
max
ka
k
k=−1,0,...,K−1 δ k+1
c1
c2
≤
kakδ,K + kbkδ,K + ka(0) k,
δ
δ
which is precisely what we wanted to show.
8
The use of the Lemma will be to bound kakδ,K by kbkδ,K .
Namely, whenever c3 := cδ1 < 1, (36) implies the following
bound:
c2 /δ
1
kakδ,K ≤
kbkδ,K +
ka(0) k.
(39)
1 − c3
1 − c3
We will also need the following Lemma from [41] on the
convergence of inexact (centralized) gradient methods.
Lemma 7 Consider unconstrained minimization of function
φ : Rp → R, where φ is assumed to be strongly convex
with strong convexity parameter m, and it also has Lipschitz
continuous gradient with Lipschitz constant M , M ≥ m > 0.
Consider the following inexact gradient method with step size
1
:
γ≤M
y (k+1) = y (k) − γ ∇φ(y (k) ) + (k) , k = 0, 1, ...,
with arbitrary initialization y (0) ∈ Rp and (k) ∈ Rp . Then,
for all k = 0, 1, ..., there holds:
ky (k+1) − y ? k ≤ (1 − γ m) ky (k) − y ? k + γ k(k) k,
Proof: Consider (17). Multiplying the equality from the
left by N1 (1 ⊗ I)> , using (1 ⊗ I)> L = (1> ⊗ I)((I − W ) ⊗ I)
= (1> (I − W )) ⊗ I = 0, we obtain that:
u(k+1) = u(k) , k = 0, 1, ...
(41)
Recall that u(0) = 0, by assumption. Thus, the result.
Lemma 9 Consider algorithm
(16)–(17), with α ≤ 1/L. Then,
for any δ ∈ 1 − α2µ , 1 , there holds:
4L
2
kex kδ,K ≤ √
ke(0) k.
ke
xkδ,K +
αµ x
Nµ
(42)
Proof: Consider (16). Multiplying the equation from the
left by N1 ( 1 ⊗ I )> , using the fact that N1 ( 1 ⊗ I )> W =
1
1
>
> (k)
= u(k) = 0,
N ( 1 ⊗ I ) , and the fact that N ( 1 ⊗ I ) u
for all k (by Lemma 8), we obtain:
x(k+1) = x(k) −
(40)
N
N
α X
α X
(k)
∇fi (xi ) = x(k) −
∇fi (x(k) )
N i=1
N i=1
N
α X
(k)
∇fi (xi ) − ∇fi (x(k) )
N i=1
α
∇f (x(k) ) + (k) , where
= x(k) −
N
N
X
(k)
∇fi (xi ) − ∇fi (x(k) ) .
=
−
where y ? = arg miny∈Rp φ(y).
Quantity (k) in (40) is an inexactness measure that says how
far is the employed search direction from the exact gradient
at the iterate y (k) .
(k)
i=1
(k)
C. Intermediate Lemmas and proof of Theorem 4
We now carry out convergence proof of Theorem 4 through
a sequence of intermediate Lemmas. We first split theprimal
(k)
error as follows: ex = x(k) − x• = x(k) − 1 ⊗ x(k) +1⊗
(k)
(k)
?
(k)
e(k) = x(k) −1⊗x(k)
x − x =: x
e +1⊗ex . Quantity x
(k)
says how mutually different are the solution estimates xi ’s at
(k)
different nodes; quantity ex says how far is the global average
PN (k)
1
(k)
x = N i=1 xi from the solution x? . We decompose the
(k)
dual error eu = u(k) + ∇F (x• ) in the following way:
e(k)
e(k) + 1 ⊗ e(k)
u .
u =u
P
(k)
(k)
(k)
N
Here, eu = N1 (1 ⊗ I)> eu = N1 i=1 [eu ]i , and u
e(k) =
(k)
(k)
(k)
(k)
(I − J )eu = eu − 1 ⊗ eu . Note that eu = u(k) +
PN
PN (k)
1
(k)
?
, where u(k) = N1 i=1 ui .
i=1 ∇fi (x ) = u
N
We will be interested in deriving bounds on kex kδ,K =
(k)
maxk=0,...,K δ1k kex k, for some δ ∈ (0, 1), and on the
analogous quantities that correspond to other primal and dual
errors that we defined above. Specifically, the proof path is as
(k)
follows. First, Lemma 8 shows that eu = 0, for all k, which
simplifies further analysis. Our goal is to apply Theorem 5
e,
with the following identification of infinite sequences: a → x
e . Then, Lemmas 9-11 are devoted to deriving
and b → u
a bound like in (33), while Lemma 12 devises a bound
that corresponds to (34). As shown below, this sequence of
Lemmas will be sufficient to complete the proof of Theorem 4.
(k)
Lemma 8 With algorithm (16)–(17), there holds: eu
u(k) = 0, for all k = 0, 1, ...
=
Now, applying Lemma 7, with ex = x(k) − x? , we obtain:
α (k)
ke(k+1)
k ≤ (1 − α µ) ke(k)
k k.
(43)
x
x k+
N
We next upper bound
α (k)
k k
N
≤
α
N
k(k) k as follows:
N
α X
(k)
∇fi (xi ) − ∇fi (x(k) )
N i=1
≤
N
α X
(k)
L xi − x(k)
N i=1
(44)
αL
≤√
x
e(k) .
(45)
N
Inequality (44) is by the Lipschitz continuity of the ∇fi ’s,
while (45) is by noting that
N
X
(k)
kxi
i=1
− x(k) k =
N
X
(k)
ke
xi k ≤
√
N ke
x(k) k.
i=1
Substituting the last bound in (43), we obtain:
α L (k)
ke(k+1)
k ≤ (1 − α µ) ke(k)
ke
x k.
x
x k+ √
N
Now, applying Lemma 6, we obtain:
(46)
1
αL 1
(1 − α µ) kex kδ,K + √
ke
xkδ,K + ke(0)
x k.
δ
Nδ
(47)
From (47), using α ≤ 1/L, one can verify that, for all δ ≥
1 − α2µ , there holds:
α µ
kex kδ,K ≤
1−
kex kδ,K
2
kex kδ,K ≤
9
αL
2 √ ke
xkδ,K + ke(0)
x k.
N
The last bound yields the desired result.
Lemma 10 Let α <
holds:
ke
xk
δ,K
≤
1−σ
3L ,
and δ ∈ 1 −
αµ
2 ,1
(48)
. Then, there
√
12 α L N
kex kδ,K
5 1−σ
12 α
2
+
ke
ukδ,K +
ke
x(0) k.
5 1−σ
1−σ
Proof: Consider (16). Subtracting x• = 1 ⊗ x? from both
sides of the equation, and noting that W x• = (W ⊗I)(1⊗x? )
= (W 1) ⊗ (I x? ) = 1 ⊗ x? = x• , we obtain:
(k)
(k)
(k)
.
(49)
e(k+1)
=
W
e
−
α
∇F
(x
)
+
u
x
x
Next, note that
∇F (x(k) ) + u(k) = ∇F (x(k) ) − ∇F (x• )
+ ∇F (x• ) + u(k)
(50)
= ∇F (x(k) ) − ∇F (x• ) + u
e(k) .
(51)
(k)
e(k) (by
In (51), we use that ∇F (x• ) + u(k) = eu = u
Lemma 8). Substituting (51) into (49), we get:
(k)
e(k+1)
= W e(k)
) − ∇F (x• ) − α u
e(k) (52)
x
x − α ∇F (x
We next multiply (52) from the left by I − J = (I − J) ⊗ I,
(k)
e(k) . Noting that
and we use that [ (I − J) ⊗ I ]ex = x
[ (I − J) ⊗ I ] W
=
[ (I − J) ⊗ I ] [ W ⊗ I ]
Next, note that for δ ∈ 1 − α2µ , 1 , and α <
holds:
1
σ+1
(σ + α L) <
.
δ
2
2
5 (1−σ)µ
12 α L
Next, note that, for α < 96
L2 , we have that (1−σ) µ <
1/2. Substituting the latter bound in (59) and manipulating the
terms, the desired result follows.
1+σ
Lemma 12 Let δ
∈
, 1 , and recall L0
=
2
1/2
2
2
L + b − 2bµ
. Then, the following holds:
fx
W
e(k) − α (I − J ) ∇F (x(k) )
−
∇F (x• )) − α u
e(k) .
(k)
•
(53)
(k)
ex
(k)
(k)
+1⊗ex ,
Next, use the decomposition x −x =
=x
e
and Lipschitz continuity of ∇F , to note that:
√
k∇F (x(k) ) − ∇F (x• )k ≤ L ke
x(k) k + L N ke(k)
x k.
(54)
Using the latter bound and taking the 2-norm in (53), while
using its sub-additive and sub-multiplicative properties, we
get:
√
u(k) k.
ke
x(k+1) k ≤ (σ + α L) ke
x(k) k + α L N ke(k)
x k + α ke
(55)
Now, similarly to Lemma 6, it is easy to see that the last
equation implies:
1
ke
xk
≤ (σ + α L) ke
xkδ,K
δ
√
1
α
+ α L N kex kδ,K + ke
ukδ,K + ke
x(0) k. (56)
δ
δ
δ,K
≤
40 L0 L
2
ke
xkδ,K +
ke
u(0) k
µ (1 − σ)
1−σ
√
16 L0 N
kex (0) k.
+
1 − σ αµ
Proof: Consider (17). Adding ∇F (x• ) to both sides of
(k)
e(k) ,
the equality, and recalling that eu = u(k) + ∇F (x• ) = u
we obtain:
=
(57)
Proof: We combine Lemmas 9 and 10, to obtain:
√
48 α L2
24
NL
δ,K
δ,K
ke
xk
≤
ke
xk
+
ke(0) k
5 (1 − σ) µ
5 (1 − σ) µ x
12 α
2
+
ke
ukδ,K +
ke
x(0) k.
(59)
5 1−σ
1−σ
= [ (I − J)W ] ⊗ I
f ⊗ I = W,
f
=W
x
e(k+1)
there
1
Also, as α < 1−σ
3L < 3L , we have that 1/δ ≤ 6/5. Substituting
the last two bounds in (56), we obtain:
√
1+σ
6
ke
xkδ,K + α L N kex kδ,K
ke
xkδ,K ≤
2
5
6α
δ,K
+
ke
uk
+ ke
x(0) k.
5
Rearranging terms in the last equality, the desired result
follows.
αµ
5 (1−σ)µ
Lemma 11 Let α < 96
L2 , and δ ∈ 1 − 2 , 1 . Then,
there holds:
4
24 α
ke
ukδ,K +
ke
x(0) k (58)
ke
xkδ,K ≤
5 1−
σ
1
−
σ
√
48 L N
+
ke(0) k.
5 µ (1 − σ) x
kukδ,K
and (I − J )e
u(k) = u
e(k) , we get:
1−σ
3L ,
u
e(k+1)
= u
e(k) − L ( ∇F (x(k) ) − ∇F (x• ) + u(k)
+ ∇F (x• ) − b (x(k) − x• ) )
(60)
= Wu
e(k) − L ( ∇F (x(k) ) − ∇F (x• ) )
+ b L (x(k) − x• )
fu
= W
e(k) − L ( ∇F (x(k) ) − ∇F (x• ) )
+ b L (x(k) − x• ).
(61)
Here, (60) uses the fact that L x• = 0, while (61) holds
fu
because W
e(k) = (W − J ) u
e(k) = W u
e(k) , as J u
e(k) =
J (I − J )eu (k) = (J − J )eu (k) = 0. We nextupper bound
the term ∇F (x(k) ) − ∇F (x• ) − b (x(k) − x• ) as follows:
k∇F (x(k) ) − ∇F (x• ) − b (x(k) − x• )k2
= k∇F (x(k) ) − ∇F (x• )k2 + b2 kx(k) − x• k2
>
−2 b ∇F (x(k) ) − ∇F (x• )
x(k) − x•
10
≤ L2 kx(k) − x• k2 + b2 kx(k) − x• k2
max{ 1+σ
2 , 1−
αµ
2 }, 1
, as desired.
− 2 b µ kx(k) − x• k2 ,
where the last inequality holds by the Lipschitz continuity of
∇F , and by the strong monotonicity of ∇F :
V. S IMULATIONS
This Section provides a simulation example on learning a
linear
classifier via minimization of the `2 -regularized logistic
(∇F (x) − ∇F (y)) (x − y) ≥ µ kx−yk2 , for all x, y ∈ Rd . loss. Simulations confirm the insights gained through the
theoretical analysis and demonstrate that the proposed generHence, we have that:
alized method improves convergence speed over the existing
k∇F (x(k) ) − ∇F (x• ) − b (x(k) − x• )k ≤ L0 kx(k) − x• k.
methods [13], [14].
The simulation setup is as follows. We consider distributed
Next, taking the norm in (61), using kLk ≤ 2, exploiting its
learning
of a linear classifier via the `2 -regularized logistic
sub-additive and sub-multiplicative properties, and using the
loss,
e.g.,
[11]. Each node i has J data samples {aij , bij }Jj=1 .
Lipschitz continuity of ∇F , we obtain:
Here, aij ∈ Rd−1 , d ≥ 2, is a feature vector, and bij ∈
(k+1)
(k)
0
(k)
ke
u
k ≤ σ ke
u k + 2 L kex k.
(62) {−1, +1} is its class label. The goal is to learn a vector x =
>
d−1
(x>
, and xP
0 ∈ R, d ≥ 1, such that the total
(k)
(k)
1 , x0 ) , x1 ∈ R
Decomposing ex = x
e(k) + 1 ⊗ ex , and applying Lemma 6,
N
`2 -regularized surrogate loss i=1 fi (x) is minimized, where,
we obtain:
for i = 1, ..., N , we have:
2 L0
σ
δ,K
δ,K
δ,K
1
ke
uk
+
ke
xk
ke
uk
≤
+ Rkxk2 .
fi (x) = ln 1 + exp −bij ( a>
δ √
δ
ij x1 + x0 )
2
2 L0 N
kex kδ,K + ke
+
u(0) k.
Here, R is a positive regularization parameter. We can take the
δ
strong convexity constant
µ = R, while a Lipschitz constant
Pas
N PJ
Next, applying Lemma 9, we get:
1
>
L
can
be
taken
as
k
i=1
j=1 cij cij k + R, where cij =
4N
0
>
>
σ
2L
4L
(bij aij , bij ) .
ke
ukδ,K ≤
ke
ukδ,K +
1+
ke
xkδ,K
With all experiments, we test the algorithms on a connected
δ
δ
µ
√
network
with N = 30 nodes and 123 links, generated as
0
4 NL
ke(0)
u(0) k.
+
a realization of the random
geometric graph model with
x k + ke
p
αµδ
communication radius ln(N )/N .
From now on, assume that δ ∈ 1+σ
We generate data and set the algorithm parameters as
2 , 1 . Then, it is easy to
.
Also,
note that 1δ ≤ 2. Thus, follows. Each node i has J = 2 data points whose dimension
see that there holds: σδ < 1+σ
2
we have:
is d − 1 = 5. The aij ’s are generated independently over
0
i and j; each entry of aij is drawn independently from the
1
+
σ
20
L
L
ke
ukδ,K ≤
ke
ukδ,K +
ke
xkδ,K
standard normal distribution. We generate the “true” vector
2
µ
√
x? = ((x?1 )> , x?0 )> by drawing its entries independently from
0
8L N
(0) δ,K
(0)
kex k
+ ke
u k.
+
standard normal distribution. The class
labels are generated
αµ
as bij = sign (x?1 )> aij + x?0 + ij , where ij ’s are drawn
After rearranging expressions, the desired result follows.
independently from normal distribution with zero mean and
We are now ready to prove Theorem 4.
variance 0.4. We set the regularization parameter as R = 0.03.
(0)
With all algorithms, we initialize xi to zero for all
Proof of Theorem 4: We apply Theorem 5 with the
e and b → u
e , by utilizing Lemma 11 i = 1, ..., N . The auxiliary variables for each algorithm
identification a → x
and Lemma 12. Assume the algorithm parameters obey the are initialized as described in Subsection II-B. Further, the
1
weight matrix is as follows: Wij = 2 (max{deg(i),deg(j)}+1)
,
conditions of Lemmas 11 and 12, namely that:
for
i
=
6
j,
{i,
j}
∈
E;
W
=
0,
for
i
=
6
j,
{i,
j}
∈
/
E;
and
ij
5 µ (1 − σ)
1+σ
αµ
P
α<
and δ ∈ max{
, 1−
}, 1 . Wii = 1 − j6=i Wij , for i = 1, ..., N . Here, deg(i) is the
2
96
L
2
2
number of neighbors of node i (excluding i).
Then, by the two Lemmas, the product of gains equals:
As an error metric, we use the following quantity:
6α 40 L0 L
?
PN kx(k)
γ1 γ2 = 1−σ
1
i −x k
µ (1−σ) We need that γ1 γ2 < 1 in order for (35)
, x? 6= 0, that we refer to as the relative
i=1
N
kx? k
µ (1−σ)2 µ(1−σ)
?
to hold. Therefore, when α < min{ 192 L0 L , 19 L2 }, we have error. Quantity x is obtained numerically beforehand by a
that ke
xkδ ≤ C < ∞, for a constant C ∈ (0, ∞). Note also centralized Nesterov gradient method [33].
that, by Lemma 9, we have:
We compare four methods: the method in [14], that we
refer
to here as “harnessing”; the Extra method in [13]; the
4L
2
kex kδ ≤ √
ke
xkδ +
ke(0)
(63) proposed method (16)–(17) with B = L+µ I – that we refer to
x k.
2
αµ
Nµ
as the modified “harnessing”; and the proposed method (16)–
(k)
(k)
As ex = x
e(k) +1⊗ex , we have: kex kδ ≤ 1 + √4NLµ ke
xkδ (17) with B = L W – that we refer to as the modified Extra.
Figure 1 plots the relative error versus number of iterations k
(0)
(0)
+ α2µ kex k = 1 + √4NLµ C+ α4µ kex k=: C 0 < +∞. for the four methods, for different values of step sizes: the
(k)
Therefore, kex k ≤ C 0 δ k , for all k, for any δ ∈ top Figure: α = 1/(3L); middle: α = 1/(9L); and bottom:
>
11
α = 1/(15L). First, on the top Figure, we can see that the
proposed modifications yield improvements in the convergence
speed over the respective original methods in [14] and [13].
While the improvement is not very large for [13], it is quite
significant for the method in [14].
As the step size decreases (the middle and bottom Figures),
we can see that the gain of the proposed method is reduced,
and the four methods tend to behave mutually very similarly.
Next, while for the large step size (the top Figure) Extra [13]
performs better than the method in [14], for the small step
size (bottom Figure) the performance of the two methods
is reversed, as predicted by our theoretical considerations.
(Though the difference between the methods is quite small
for the small step size.)
Figure 2 repeats the experiment for a 100-node, 561-link,
connected network (generated also as
p an instance of a random
geometric graph model with radius ln(N )/N ), and for stepsizes α = 1/(6L) (top Figure); α = 1/(18L) (middle); and
α = 1/(54L) (bottom). We can see that a similar behavior of
the four methods can be observed here, as well.
VI. C ONCLUSION
We considered exact first order methods for distributed optimization problems where N nodes collaboratively minimize
the aggregate sum of their local convex costs. Specifically, we
unified, generalized, and improved convergence speed of the
existing methods, e.g., [13], [14]. While it was known that
the method in [13] is equivalent to a primal-dual gradient-like
method, we show here that this is true also with [14], where the
corresponding primal-dual update rule incorporates a weighted
past dual gradient term, in addition to the current dual gradient term. We then generalize the method by proposing an
optimized, easy-to-tune weighting of the past dual gradient
term, and show both theoretically and by simulation that the
modification yields significant improvements in convergence
speed. We establish for the proposed exact method global
R-linear convergence rate, assuming strongly convex local
costs with Lipschitz continuous gradients and static networks.
Possible future work directions include extensions to uncoordinated step-sizes across nodes and time varying, random,
and directed networks. Specifically, with directed networks,
it would be interesting to extend the methods to work with
singly stochastic matrices, i.e., to relax the requirement on
the double stochasticity, as doubly stochastic methods become
undesirable when dealing with time-varying or directed graph
topologies.
.
A PPENDIX
A. Proof of Lemma 2
We first prove part (b), i.e., we set B = α1 W. We must
show that (16)–(17) is equivalent to (5)–(6). It suffices to show
that (16)–(17) is equivalent to (12)–(13), due to Lemma 1.
First, note that, due to identity L = I − W, we have that (16)
and (12) are the same. Next, from (16), we have that:
1
1
∇F (x(k) ) + u(k) = − x(k+1) + W x(k) .
α
α
Fig. 1: Relative error versus number of iterations k, for three
different values of step-size α: Top: α = 31L ; middle: α = 91L ;
and bottom: α = 151L .
12
Substituting this into (17), and using B = α1 W, we recover (13). Hence, the equivalence of (16)–(17) and (5)–(6)
with B = α1 W. Next, we prove part (a). That is, we consider
(14) and (16)–(17) with B = 0. The key is to note that s(k) ,
k = 0, 1, ..., can be written as: s(k) = u(k) +∇F (x(k) ), where
u(k) is defined by the following recursion:
u(k+1)
=
W u(k) + (W − I) ∇F (x(k) )
=
u(k) − L (u(k) + ∇F (x(k) )),
(64)
for k = 0, 1, ..., and u(0) = 0. Substituting (64) into (16),
yields the desired equivalence.
B. Proof of Lemma 3
Consider (16). Subtracting x• from both sides of the equality, noting that W x• = x• , and adding and subtracting
∇F (x• ) to the term in the parenthesis on the right hand side
of the equality, we obtain:
e(k+1)
x
=
(k)
W e(k)
) − ∇F (x• ) + ∇F (x• ) + u(k)
x − α ∇F (x
=
(k)
(k)
W e(k)
x − α Hk ex − α eu ,
(65)
by the
definition of Hk , Hk =
R 1where2 (65)• follows
(k)
•
∇
F
x
+
t
(x
−
x
)
dt, and by the definition of
t=0
(k)
eu . Next, consider (17). Using identity L = I − W, the
equality can be equivalently written as:
u(k+1) = W u(k) − L ∇F (x(k) ) − B x(k) .
(66)
Next, add ∇F (x• ) to both sides of the equality, and express
the quantity on the right hand side as ∇F (x• ) = W ∇F (x• )+
L ∇F (x• ). We obtain:
e(k+1)
= W (u(k) + ∇F (x• )) − L ∇F (x(k) )
u
−
∇F (x• )) + L B x(k)
(k)
(k)
= W e(k)
.
u − L Hk ex + L B x
(67)
Next, by the property B x• = c x• , for some c ∈ R, it follows
that: L B x• = c L x• = c [ (I − W ) ⊗ I ] [1 ⊗ x? ] = c [ (I −
W )1 ] ⊗ [ I ⊗ x? ] = 0. Hence, we can subtract L B x• = 0
from the right hand side of (67) to obtain the following:
e(k+1)
u
(k)
(k)
= W e(k)
u − L Hk ex + L B ex
(k)
= W e(k)
u + L (B − Hk ) ex .
(68)
Finally, note from (68) that:
J e(k+1)
= J e(k)
u
u , k = 0, 1, ...,
because J L = [ J ⊗ I ] [ (I − W ) ⊗ I ] = [ J(I − W ) ] ⊗ I =
(0)
[J − J] ⊗ I = 0. Note that J eu = J (0 + ∇F (x• )) =
P
(k)
N
1
?
i=1 ∇fi (x ) ) = 0. Thus, we conclude that J eu = 0,
N 1(
for all k. Applying the latter fact to (68), we obtain:
(k)
e(k+1)
= (W − J ) e(k)
u
u + L (B − Hk ) ex .
(69)
The relations (65) and (69) yield the claim of the Lemma.
Fig. 2: Relative error versus number of iterations k, for three
different values of step-size α: Top: α = 61L ; middle: α =
1
1
18 L ; and bottom: α = 54 L .
C. Derivation of the solution to (20) and of an approximate
solution to (21)
We first consider (20). Note that, for any H ∈ H,
we have that: kb I − Hk = maxi=1,...,N d |b − hi | =
13
max {|b − hN d |, |h1 − b|} ≤ max {|b − µ|, |L − b|} . Here,
hi denotes the i-th largest eigenvalue of H. In the last
inequality above, we used the fact that µ I H L I, for all
H ∈ H. Therefore, we have that:
max kb I − Hk = max{|b − µ|, |L − b|}.
H∈H
(70)
The maximum in (70) is attained, e.g., for H =
Diag(L, µ, ..., µ), where µ is repeated (N d − 1) times. The
quantity (70) is clearly minimized over b ≥ 0 at b = b? =
L+µ
2 .
Now, consider (21), and assume that λN > 0. Note that
the maximal eigenvalue of W = W ⊗ I equals one, and the
minimal eigenvalue of W equals λN . We have:
kb0 W − Hk ≤ max {|b0 − µ|, |L − b0 λN |} .
(71)
We choose a sub-optimal b0 that minimizes the upper bound
in (71) on the desired function supH∈H kb0 W − Hk. It is easy
L+µ
.
to see that the corresponding value is b0 = 1+λ
N
D. Proof of equivalence of (16)–(17) and (28)–(29)
Consider (16). From the equation,
we have: ∇F (x(k) ) +
1
(k+1)
(k)
u = −α x
−Wx
. Substituting the latter relation
in (17), and using the fact that matrices L and W commute,
as well as that L and B commute, (17) leads to (25). Now,
consider (28). Proceeding by the same steps as in deriving
(12) from (10), it is straightforward to verify that (28) leads
to (24). Thus, the equivalence between (16)–(17) and (28)–
(29). The respective equivalence for the method in [14] follows
by setting B = 0 in (28)–(29).
(k)
E. Proof of Theorem 4 for generic matrices B
We show here that, when B = b I is replaced with a
generic symmetric matrix B that respects the sparsity pattern
of graph G and obeys that for any y ∈ Rd , there exists some
c ∈ R, such that B (1 ⊗ y) = c (1 ⊗ y), Theorem 4 continues
to hold with L0 replaced with constant (L + kBk). Namely,
it is easy to see that Lemma 8 continues to hold unchanged.
Further, Lemmas 9–11 pertain to the update equation (16) that
does not depend on B, and hence they also hold unchanged.
The only modifications occur with Lemma 12. Namely, (61)
fu
becomes: u
e(k+1) = W
e(k) − L ( ∇F (x(k) ) − ∇F (x• ) )
(k)
•
+L B (x − x ). Using Lipschitz continuity of ∇F and
the fact that kLk ≤ 2, the latter equality implies: ke
u(k+1) k
≤ σ ke
u(k) k + 2 L kx(k) − x• k +2kBk kx(k) − x• k. The proof
of the modified Lemma 12 then proceeds in the same way
as the remaining part of the proof of Lemma 12. Finally, the
proof of the modified Theorem 4 then also proceeds in the
same way as the proof of Theorem 4.
R EFERENCES
[1] A. Nedic and A. Ozdaglar, “Distributed subgradient methods for multiagent optimization,” IEEE Transactions on Automatic Control, vol. 54,
no. 1, pp. 48–61, January 2009.
[2] C. Lopes and A. H. Sayed, “Adaptive estimation algorithms over
distributed networks,” in 21st IEICE Signal Processing Symposium,
Kyoto, Japan, Nov. 2006.
[3] F. Cattivelli and A. H. Sayed, “Diffusion LMS strategies for distributed
estimation,” IEEE Trans. Sig. Process., vol. 58, no. 3, pp. 1035–1048,
March 2010.
[4] A. H. Sayed, S.-Y. Tu, J. Chen, X. Zhao, and Z. Towfic, “Diffusion
strategies for adaptation and learning over networks,” IEEE Sig. Process.
Mag., vol. 30, no. 3, pp. 155–171, May 2013.
[5] J. Duchi, A. Agarwal, and M. Wainwright, “Dual averaging for distributed optimization: Convergence and network scaling,” IEEE Trans.
Aut. Contr., vol. 57, no. 3, pp. 592–606, March 2012.
[6] I. Lobel and A. Ozdaglar, “Convergence analysis of distributed subgradient methods over random networks,” in 46th Annual Allerton
Conference onCommunication, Control, and Computing, Monticello,
Illinois, September 2008, pp. 353 – 360.
[7] D. Jakovetic, J. Xavier, and J. M. F. Moura, “Fast distributed gradient
methods,” IEEE Trans. Autom. Contr., vol. 59, no. 5, pp. 1131–1146,
May 2014.
[8] ——, “Convergence rates of distributed Nesterov-like gradient methods
on random networks,” IEEE Transactions on Signal Processing, vol. 62,
no. 4, pp. 868–882, February 2014.
[9] S. Kar, J. M. F. Moura, and K. Ramanan, “Distributed parameter estimation in sensor networks: Nonlinear observation models and imperfect
communication,” IEEE Transactions on Information Theory, vol. 58,
no. 6, pp. 3575–3605, June 2012.
[10] M. Rabbat and R. Nowak, “Distributed optimization in sensor networks,”
in IPSN 2004, 3rd International Symposium on Information Processing
in Sensor Networks, Berkeley, California, USA, April 2004, pp. 20 –
27.
[11] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction method
of multipliers,” Foundations and Trends in Machine Learning, Michael
Jordan, Editor in Chief, vol. 3, no. 1, pp. 1–122, 2011.
[12] F. Bullo, J. Cortes, and S. Martinez, Distributed control of robotic
networks: A mathematical approach to motion coordination algorithms.
Princeton University Press, 209.
[13] W. Shi, Q. Ling, G. Wu, and W. Yin, “EXTRA: An exact first-order
algorithm for decentralized consensus optimization,” SIAM J. Optim.,
vol. 25, no. 2, pp. 944–966, 2015.
[14] G. Qu and N. Li, “Harnessing smoothness to accelerate distributed
optimization,” to appear in IEEE Transactions on Control of Network
Systems, 2017, DOI: 10.1109/TCNS.2017.2698261.
[15] K. Yuan, B. Ying, X. Zhao, and A. H. Sayed, “Exact diffusion for
distributed optimization and learning — Part I: Algorithm development,”
2017, arxiv preprint, arXiv:1702.05122.
[16] ——, “Exact diffusion for distributed optimization and learning — Part
II: Convergence analysis,” 2017, arxiv preprint, arXiv:1702.05142.
[17] D. Jakovetic, J. Xavier, and J. M. F. Moura, “Linear convergence rate of
a class of distributed augmented Lagrangian algorithms,” IEEE Trans.
Autom. Contr., vol. 60, no. 4, pp. 922–936, April 2015.
[18] A. Nedic, A. Olshevsky, W. Shi, and C. A. Uribe, “Geometrically convergent distributed optimization with uncoordinated step-sizes,” 2016,
arXiv preprint arXiv:1609.05877.
[19] A. Nedic, A. Olshevsky, and W. Shi, “Achieving geometric convergence
for distributed optimization over time-varying graphs,” 2016, arXiv
preprint arXiv:1607.03218.
[20] J. Xu, S. Zhu, Y. Soh, and L. Xie, “Augmented distributed gradient
methods for multi-agent optimization under uncoordinated constant
stepsizes,” in 54th IEEE Conference on Decision and Control (CDC),
2015, pp. 2055–2060.
[21] Q. Lu and H. Li, “Geometrical convergence rate for distributed optimization with time-varying directed graphs and uncoordinated step-sizes,”
2016, arXiv preprint arXiv:1611.00990.
[22] G. Qu and N. Li, “Accelerated distributed Nesterov gradient descent,”
2017, arxiv preprint arXiv:1705.07176.
[23] C. Xi and U. A. Khan, “Dextra: A fast algorithm for optimization over
directed graphs,” IEEE Transactions on Automatic Control, 2017, to
appear, DOI: 10.1109/TAC.2017.2672698.
[24] J. Zeng and W. Yin, “Extrapush for convex smooth decentralized
optimization over directed networks,” Journal of Computational Mathematics, vol. 35, no. 4, pp. 381–394, 2017.
[25] W. Shi, Q. Ling, G. Wu, and W. Yin, “A proximal gradient algorithm
for decentralized composite optimization,” Journal of Machine Learning
Research, vol. 63, no. 22, pp. 6013–6023, 2015.
[26] K. Scaman, F. Bach, S. Bubeck, Y. T. Lee, and L. Massoulie, “Optimal
algorithms for smooth and strongly convex distributed optimization in
networks,” 2017, arxiv preprint, arXiv:1702.08704.
[27] A. Nedic, A. Ozdaglar, and A. Parrilo, “Constrained consensus and
optimization in multi-agent networks,” IEEE Transactions on Automatic
Control, vol. 55, no. 4, pp. 922–938, April 2010.
14
[28] A. Mokhtari and A. Ribeiro, “DSA: Decentralized double stochastic
averaging gradient algorithm,” Journal of Machine Learning Research,
vol. 17, pp. 1–35, 2016.
[29] A. Nedic and A. Ozdaglar, “Subgradient methods for saddle point
problems,” Journal of Optimization Theory and Applications, vol. 145,
no. 1, pp. 205–228, July 2009.
[30] M. Kallio and C. H. Rosa, “Large-scale convex optimization via saddlepoint computation,” Oper. Res., pp. 93–101, 1999.
[31] H. Uzawa, “Iterative methods in concave programming,” 1958, in Arrow,
K., Hurwicz, L., Uzawa, H. (eds.) Studies in Linear and Nonlinear
Programming, pp. 154-165. Stanford University Press, Stanford.
[32] J. Tsitsiklis, D. Bertsekas, and M. Athans, “Distributed asynchronous
deterministic and stochastic gradient optimization algorithms,” IEEE
Trans. Autom. Contr., vol. 31, no. 9, pp. 803–812, Sep. 1986.
[33] Y. E. Nesterov, “A method for solving the convex programming problem
with convergence rate O(1/k2 ),” Dokl. Akad. Nauk SSSR, vol. 269, pp.
543–547, 1983, (in Russian).
[34] D. Jakovetic, D. Bajovic, N. Krejic, and N. K. Jerinkic, “Distributed
gradient methods with variable number of working nodes,” IEEE Transactions on Signal Processing, vol. 64, no. 15, pp. 4080–4095, August
2016.
[35] M. Zhu and S. Martinez, “On distributed convex optimization under
inequality and equality constraints,” IEEE Transactions on Automatic
Control, vol. 57, no. 1, pp. 151–164, Jan. 2012.
[36] T.-H. Chang, A. Nedic, and A. Scaglione, “Distributed constrained optimization by consensus-based primal-dual perturbation method,” IEEE
Transactions on Automatic Control, vol. 59, no. 6, pp. 1524–1538, June
2014.
[37] J. Wang and N. Elia, “Control approach to distributed optimization,”
in 48th Annual Allerton Conference onCommunication, Control, and
Computing, Monticello, IL, Oct. 2010.
[38] A. Dimakis, S. Kar, J. M. F. Moura, M. Rabbat, and A. Scaglione,
“Gossip algorithms for distributed signal processing,” Proceedings of
the IEEE, vol. 98, no. 11, pp. 1847–1864, 2010.
[39] K. Yuan, Q. Ling, and W. Yin, “On the convergence of decentralized
gradient descent,” SIAM J. Optim., vol. 26, no. 3, pp. 1835–1854, 2016.
[40] C. Desoer and M. Vidyasagar, Feedback Systems: Input-Output Properties. SIAM, 2009.
[41] M. Schmidt, N. L. Roux, and F. Bach, “Convergence rates of inexact
proximal-gradient methods for convex optimization,” in Advances in
Neural Information Processing Systems 24, 2011, pp. 1458–1466.
| 7cs.IT
|
arXiv:1710.10898v1 [cs.CV] 30 Oct 2017
Learning to solve inverse problems
using Wasserstein loss
Jonas Adler1,2 , Axel Ringh1 , Ozan Öktem1 , Johan Karlsson1
1
Department of Mathematics, KTH Royal Institute of Technology
2
Elekta Instrument AB
{ jonasadl, aringh, ozan } @kth.se, [email protected]
Abstract
We propose using the Wasserstein loss for training in inverse problems. In particular,
we consider a learned primal-dual reconstruction scheme for ill-posed inverse
problems using the Wasserstein distance as loss function in the learning. This is
motivated by miss-alignments in training data, which when using standard mean
squared error loss could severely degrade reconstruction quality. We prove that
training with the Wasserstein loss gives a reconstruction operator that correctly
compensates for miss-alignments in certain cases, whereas training with the mean
squared error gives a smeared reconstruction. Moreover, we demonstrate these
effects by training a reconstruction algorithm using both mean squared error and
optimal transport loss for a problem in computerized tomography.
1
Introduction
In inverse problems the goal is to determine model parameters from indirect noisy observations.
Example of such problems arise in many different fields in science and engineering, e.g., in X-ray
computed tomography (CT) [27], electron tomography [28], and magnetic resonance imaging [8].
Machine learning has recently also been applied in this area, especially in imaging applications.
Using supervised machine-learning to solve inverse problems in imaging requires training data where
ground truth images are paired with corresponding noisy indirect observations. The learning provides
a mapping that associates observations to corresponding images. However, in several applications
there are difficulties in obtaining the ground truth, e.g., in many cases it may have undergone a
distortion. For example, a recent study showed that MRI images may be distorted by up to 4 mm
due to, e.g., inhomogeneities in the main magnetic field [36]. If these images are used for training,
the learned MRI reconstruction will suffer in quality. Similar geometric inaccuracies arise in several
other imaging modalities, such as Cone Beam CT and full waveform inversion in seismic imaging.
This work seeks to provide a scheme for learning a reconstruction scheme for an ill-posed inverse
problem with a Wasserstein loss by leveraging upon recent advances in efficient solutions of optimal
transport [10, 22] and learned iterative schemes for inverse problems [3]. The proposed method
is demonstrated on a computed tomography example, where we show a significant improvement
compared to training the same network using mean squared error loss. In particular, using the
Wasserstein loss instead of standard mean squared error gives a result that is more robust against
potential miss-alignment in training data.
2
2.1
Background
Inverse problems
In inverse problems the goal is to reconstruct an estimate of the signal ftrue ∈ X from noisy indirect
measurements (data) g ∈ Y assuming
g = T (ftrue ) + δg.
(1)
In the above X and Y are referred to as the reconstruction and data space, respectively. Both are
typically Hilbert or Banach spaces. Moreover T : X → Y denotes the forward operator, which
models how a given signal gives rise to data in absence of noise. Finally, δg ∈ Y is the noise
component of data. Many inverse problems of interest are ill-posed, meaning that there is no uniques
solution to (1) and hence there is no inverse to T . Typically reconstructions of ftrue are sensitive to
the data and small errors gets amplified. One way to mitigate these effects is to use regularization
[12].
Variational regularization In variational regularization one formulates the reconstruction problem as an optimization problem. To this end, one introduces a data discrepancy functional
f 7→ L(T (f ), g), where L : Y × Y → R, that quantifies the miss-fit in data space, and a regularization functional S : X → R that encodes a priori information about ftrue by penalizing undesirable
solutions. For a given g ∈ Y , this gives an optimization problem of the form
min L(T (f ), g) + λS(f ).
f ∈X
(2)
Here, λ acts as a trade-off parameter between the data discrepancy and regularization functional. In
many cases L is taken to be the negative data log-likelihood, e.g., L(T (f ), g) = k T (f ) − gk22 in
the case of additive white Gaussian noise. Moreover, a typical choice for regularization functional
is total variation (TV) regularization, S(f ) = k∇f k1 [33]. These regularizers typically give rise
to large scale and non-differentiable optimization problems, which requires advanced optimization
algorithms.
Learning for inverse problems In many applications, and so also in inverse problems, data driven
approaches have shown dramatic improvements over the state-of-the-art [24]. Using supervised
learning to solve an inverse problem amounts to finding a parametrized operator T †Θ : Y → X where
the parameters Θ are selected so that
g = T (ftrue ) + δg =⇒ T †Θ (g) ≈ ftrue .
For inverse problems in image processing, such as denoising and deblurring, we have Y = X and
it is possible to apply a wide range of widely studied machine learning techniques, such as fully
convolutional deep neural networks with various architectures, including fully convolutional networks
[18] and denoising auto-encoders [38].
However, in more complicated inverse problems as in tomography, the data and reconstruction spaces
are very different, e.g., their dimension after discretization may differ. For this reason, learning a
mapping from Y to X becomes nontrivial, and classical architectures that map, e.g., images to images
using convolutional networks cannot be applied as-is. One solution is to use fully-connected layers as
in [30] for very small scale tomographic reconstruction problems. A major disadvantage with such a
fully learned approach is that the parameters space has to be very high dimensional in order to be
able to learn both the prior and the data model, which often renders it infeasible due to training time
and lack of training data.
A more successful approach is to first apply some crude reconstruction operator T † : Y → X and then
use machine learning to post process the result. This separates the learning from the complications
of mapping between spaces since the operator T † can be applied off-line, prior to training. Such
an approach has been demonstrated for tomographic reconstruction in [31, 37]. Its drawback for ill
posed inverse problems is that information is typically lost by using T † , and this information cannot
be recovered by post processing.
Finally, another approach is to incorporate the forward operator T and its adjoint T ∗ into the neural
network. In these learned iterative schemes, classical neural networks are interlaced with applications
2
of the forward and backward operator, thus allowing for the learned reconstruction operator to work
directly from data without having to learn the data model. For example, in [39] an alternating direction
method of multipliers (ADMM)-like scheme for Fourier inversion is learned and [32] consider solving
inverse problems typically arising in image restoration by a learned gradient-descent scheme. In [4]
this later approach is shown to be applicable to large scale tomographic inversion. Finally, in [3] they
apply learning in both spaces X and Y , yielding a Learned Primal-Dual scheme, and show that it
outperforms learned post-processing for reconstruction of medical CT images.
Loss functions for learning Once the Θ parametrization of T †Θ is set, the parameters are typically
chosen by minimization of some loss functional L. Without doubt, the most common loss function is
the mean squared error, also called L2 loss, given by
h
i
L(Θ) = Ef,g k T †Θ (g) − fk22 .
(3)
It has however been noted that it is sub-optimal for imaging, and a range of other loss functions have
been investigated. These include the classical `p norms and the structural similarity index (SSIM)
[40], as well as more complex losses such as perceptual losses [20] and adversarial networks [26].
Recently, optimal mass transport has also been considered as loss function for classification [14] and
generative models [5]. In this work we consider using optimal transport for training a reconstruction
scheme for ill-posed inverse problems.
2.2
Optimal mass transport and Sinkhorn iterations
In optimal mass transport the aim is to transform one distribution into another by moving the mass in
a way that minimizes the cost of the movement. For an introduction and overview of the topic, see,
e.g., the monograph [35]. Lately, the area has attracted a lot of research [10, 11, 9] with applications
to, e.g., signal processing [17, 15, 19, 13] and inverse problems [7, 22].
The optimal mass transport problem can be formulated as follows: let Ω ⊂ Rd be a compact set, and
let µ0 and µ1 be two measures, defined on Ω, with the same total mass. Given a cost c : Ω × Ω → R+
that describes the cost for transporting a unit mass from one point to another, find a (mass preserving)
transference plan M that is as cheap as possible. Here, the transference plan characterizes how to
move the mass of µ0 in order to deform it into µ1 . Letting the transference plan be a nonnegative
measure dM on the space Ω × Ω yields a linear programming problem in the space of measures:
Z
T (µ0 , µ1 ) =
min
c(x0 , x1 )dM (x0 , x1 )
(4)
dM ≥0
(x0 ,x1 )∈Ω×Ω
Z
subject to µ0 (x0 )dx0 =
dM (x0 , x1 ),
x ∈Ω
Z 1
µ1 (x1 )dx1 =
dM (x0 , x1 ).
x0 ∈Ω
Although this formulation is only defined for measures µ0 and µ1 with the same total mass, it can also
be extended to handle measures with unbalanced masses [15, 9]. Moreover, under suitable conditions
one can define the Wasserstein metrics Wp using T , by taking c(x0 , x1 ) = d(x0 , x1 )p for p ≥ 1
and where d is a metric on Ω, and Wp (µ0 , µ1 ) := T (µ0 , µ1 )1/p [35, Definition 6.1]. As the name
indicates, Wp is a metric on the set of nonnegative measures on Ω with fixed mass [35, Theorem 6.9],
and T is weak∗ continuous on this set. One important property is that T (and thus also Wp ) does not
only compare objects point by point, as standard Lp metrics, but instead quantifies how the mass is
moved. This makes optimal transport natural for quantifying uncertainty and modelling deformations
[19, 21].
One way to solve the optimal transport problem in applications is to discretize Ω and solve the
corresponding finite-dimensional linear programming problem. In this setting the two measures are
represented by point masses on the discretization grid, i.e., by two vectors µ0 , µ1 ∈ Rn+ where the
element [µk ]i corresponds to the mass in the point x(i) ∈ Ω for i = 1, . . . , n and k = 0, 1. Moreover,
a transference plan is represented by a matrix M ∈ Rn×n
where the value mij := [M ]ij denotes
+
the amount of mass transported from point x(i) to x(j) . The associated cost of a transference plan
Pn
is i,j=1 cij mij = trace(C T M ), where [C]ij = cij = c(x(i) , x(j) ) is the transportation cost from
3
x(i) to x(j) , and by discretizing the constraints we get that M is a feasible transference plan from µ0
to µ1 if the row sums of M is µ0 and the column sums of M is µ1 . The discrete version of (4) thus
takes the form
T (µ0 , µ1 ) =
min
M ≥0
trace(C T M )
subject to µ0 = M 1n
(5)
µ1 = M T 1n ,
where M ≥ 0 denotes element-wise non-negativity of the matrix. However, even though (5) is a
linear programming problem it is in many cases computationally infeasible due to the vast number of
variables. Since M ∈ Rn×n
the number of variables is n2 , and thus if one seek to solve the optimal
+
transport problem between two 512 × 512 images this results in more than 6 · 1010 variables.
One approach for addressingP
this problem was proposed by Cuturi [10] that introduces an entropic
n
regularizing term D(M ) = i,j=1 (mij log(mij ) − mij + 1) for approximating the transference
plan, so the resulting perturbed optimal transport problem reads as
min
M ≥0
trace(C T M ) + εD(M )
subject to µ0 = M 1n
(6)
µ1 = M T 1n .
One can show that an optimal solution to (6) is of the form
M = diag(u)Kdiag(v),
(7)
Rn+
where K = exp(−C/ε) (point-wise exponential) is known, and u, v ∈
are unknown. This shows
that the solution is parameterized by only 2n variables. Moreover, the two vectors can be computed
iteratively by so called Sinkhorn iterations, i.e., alternatingly compute u and v that matches µ0 and
µ1 respectively. This is summarizied in Algorithm 1 where denotes elementwise multiplication
and ./ elementwise division. The procedure has been shown to have a linear convergence rate, see
[10] and references therein.
Moreover, when the underlying cost c(x0 , x1 ) is translation invariant the discretized cost matrix C,
and thus also the transformation K, gets a Toeplitz-block-Toeplitz structure. This structure can be
used in order to compute Kv and K T u efficiently using the fast Fourier transform in O(n log n),
instead of naive matrix-vector multiplication in O(n2 ) [22]. This is crucial for applications in imaging
since for images of size 512×512 pixels one would have to explicitly store and multiply with matrices
of size 262144 × 262144 .
Algorithm 1 Sinkhorn iterations for computing entropy-regularized optimal transport [10]
1:
2:
3:
4:
5:
6:
3
Input C, ε, µ0 , µ1
initialize v0 > 0 and K = exp(−C/ε)
for i = 1, . . . , N do
ui ← µ0 ./(Kvi−1 )
vi ← µ1 ./(K T ui )
return uTN (K C)vN
Learning a reconstruction operator using Wasserstein loss
In this work we propose to use entropy regularized optimal transport (6) to train a reconstruction
operator, i.e., to select the parameters as
h
i
Θ∗ ∈ arg min Ef,g T (T †Θ (g), f) .
(8)
Θ
This should give better results when data g is not aligned with the ground truth f . To see this,
consider the case when f is a point mass. In that case training the network with the L2 loss (3)
will (in the ideal case) result in a perfect reconstruction composed with a convolution that “smears”
the reconstruction over the area of possible miss-alignment. On the other hand since optimal mass
transport does not only compare objects point-wise, the network will (in the ideal case) learn a perfect
reconstruction combined with a movement of the object to the corresponding barycenter (centroid) of
the miss-alignment. These statements are made more precise in the following propositions. Formal
definitions and proofs are deferred to the appendix.
4
Proposition 1. Let g ∈ L2 (Rn ), let τ be a Rn -valued random variable with probability measure
dP (t), and let gτ (x) := g(x − τ ). Then there exists a function f ∈ L2 (Rn ) that minimizes
Eτ [kf − gτ k22 ], and this f has the form
Z
f (x) = (dP ∗ g)(x) :=
g(x − t)dP (t).
Rn
Proposition 2. Let δ(x) be the Dirac delta function on Rn , let τ be a Rn -valued random variable
with probability measure dP (t), and let δτ (x) := δ(x − τ ). Then
Z Z
Z
Eτ [T (δτ , µ)] =
c(t, x)dP (t) dµ(x) for µ ≥ 0 and
dµ(x) = 1
n
Rn
Rn
|R
{z
}
:=F (x)
and Eτ [T (δτ , µ)] = ∞ otherwise. Furthermore, finding a µ that minimizes Eτ [T (δτ , µ)] is equivalent
to finding the global minimizers to F (x). In particular, if (i) the probability measure dP is symmetric
around its mean, (ii) the underlying cost c is of the form c(t, x) = d(x − t), where d is convex and
symmetric, and (iii) dP and d are such that
Z
Z
d(x − t)dP (t)
and
∂d(x − t)dP (t)
Rn
Rn
n
both exist and are finite for all x ∈ R , then µ(x) = δ(x − E[τ ]) is an optimal solution. Furthermore,
if d is also strictly convex, then this is the unique minimizer.
To illustrate Propositions 1 and 2 we consider the following example.
Example 3. Let τ be uniformly distributed on [−1, 1], and let c(x0 , x1 ) = (x0 − x1 )2 . This gives
Z
1 1
1
F (x) =
(x − t)2 dt = + x2 ,
2 −1
3
which has minimum x = 0, and hence the (unique) minimizer to Eτ [T (δτ , µ)] is µ(x) = δ(x). For
the L2 case with the uniform distribution, the minimizer of Eτ [kf − gτ k22 ] is the smoothed function
g ∗ 12 χ[−1,1] .
The most common choice of distance c is to use the squared norm c(x0 , x1 ) = kx0 − x1 k2 , as in
the previous example. In this case the result of Proposition 2 can be strengthened, as shown in the
following example.
Example 4. Let τ be a Rn -valued random variable with probability measure dP (t) with finite first
and second moments, and let c(x0 , x1 ) = kx0 − x1 k2 . This gives
Z
F (x) =
(x − t)2 dP (t) = x2 − 2x E[τ ] + E[τ 2 ],
Rn
which has a unique global minimum in x = E[τ ] and hence µ(x) = δ(x − E[τ ]).
4
Implementation and evaluation
We use the recently proposed learned primal-dual structure in [3] for learning a reconstruction
operator T †Θ for solving the inverse problem in (1). In this algorithm, a sequence of small blocks work
alternatingly in the data (dual) space Y and the reconstruction (primal) space X and are connected
using the forward operator T and its adjoint T ∗ . The algorithm works with any differentiable operator
T , but we state the version for linear operators for simplicity in algorithm 2.
Algorithm 2 Learned Primal-Dual reconstruction algorithm
1: Initialize f0 ∈ X Nprimal , h0 ∈ U Ndual
2: for i = 1, . . . , I do
(2)
3:
hi ← ΓΘd hi−1 , T (fi−1 ), g
i
(1)
fi ← ΛΘpi fi−1 , T ∗ (hi )
(1)
†
5: T Θ (g) := fI
4:
5
7
32
32
5
6
32
32
5
apply
apply
copy
3x3 conv + PReLU
3x3 conv
Figure 1: Network architecture used to solve the inverse problem. Dual and primal iterates are in
blue and red boxes, respectively. Several arrows pointing to the same box indicates concatenation.
The initial values f0 , h0 enter from the left, while the data g is supplied to the dual iterates.
→
−
(a) Phantom
→
−
(b) Translated phantom
(c) Data
Figure 2: Example of data generation process used for training and validation, where 2a shows
an example phantom, 2b is the phantom with a random translation and 2c is the data (sinogram)
corresponding to 2b with additive white noise on top. The pair (gi , fi ) = (2c, 2a) is what is used in
the training.
The method was implemented using ODL [2], ASTRA [34], and TensorFlow [1]. We used the
reference implementation1 with default parameters, i.e., the number of blocks in the primal and dual
space was I = 10, and the number of primal and dual variables was set to Nprimal = Ndual = 5.
Moreover, the blocks used a residual structure and had three layers of 3×3 convolutions with 32 filters.
PReLU nonlinearities were used. Thus, this corresponds to a residual CNN with convolutional depth
of 10 · 2 · 3 = 60, as shown in graphical format in fig. 1. We used zero initial values, f0 = h0 = 0.
We compare a learned reconstruction operator of this form when trained using L2 loss (3) and using
optimal transport loss (8). Moreover, the evaluation is done on a problem similar to the evaluation
problem in [3, 4], i.e., on a problem in computed tomography. More specifically, training is done on
data that consists of randomly generated circles on a domain of 512 × 512 pixel, and the forward
operator T is the ray transform [27]. What makes this an ill-posed problem is that the data acquisition
is done from only 30 views with 727 parallel lines. Moreover, the source of noise is two-fold in
this set-up: (i) the pairs (gi , fi ) of data sets and phantoms are not aligned, meaning that the data is
computed from a phantom with a random change in position. This random change is independent
for the different circles, and for each circle it is a shift which is uniformly distributed over [−40, 40]
pixels, both in up-down and left-right direction. (ii) on the data computed from the shifted phantom,
5% additive Gaussian noise was added. For an example, see fig. 2.
The optimal mass transport distance computed with Sinkhorn iterations was used as loss function,
where we used the transport cost
4
4
c(x1 , x2 ) = 1 − e−kx1 −x2 k /80 .
1
https://github.com/adler-j/learned_primal_dual
6
(a) Phantom
(b) Translated phantom (c) Mean squared error loss (d) Optimal transport loss
Figure 3: In 3a we show the validation phantom, which was generated from the same training set
but not used in training, in 3b the translated phantom from which the validation data was computed,
in 3c a reconstruction with neural network trained using mean squared error loss (3), and in 3d a
reconstruction with neural network trained using optimal mass transport loss (8).
This was chosen since it heavily penalizes large movements, while not diverging to infinity which
causes numerical instabilities. Moreover, c(x1 , x2 )1/4 is in fact a metric on R2 (see lemma 6 in the
appendix) and thus W4 (µ0 , µ1 ) := T (µ0 , µ1 )1/4 gives rise to a Wasserstein metric on the space of
images, where T (µ0 , µ1 ) is the optimal mass transport distance with the transport cost c(x1 , x2 ).
Since this cost is translation invariant, the matrix-vector multiplications Ku and K T v can be done
with fast Fourier transform, as mentioned in section 2.2, and this was implemented in Tensorflow. We
used 10 Sinkhorn iterations with entropy regularization ε = 10−3 , to approximate the optimal mass
transport. Automatic differentiation was used to back-propagate the result during training.
Since the optimal mass transport function (6) is only finite for marginals µ0 and µ1 with the
same total mass, in the training we normalize the output of the reconstruction T †Θ (g) with
mass(f )/mass(T †Θ (g)). This makes T †Θ (g) invariant with respect to the total mass, which is
undesirable. To compensate for this, a small penalization on the error in total mass was added to the
loss function.
The training also followed [3] closely. In particular, we used 2 · 104 batches of size 1, using the ADAM
optimizer [23] with default values except for β2 = 0.99. The learning rate (step length) used was
cosine annealing [25] with initial step length 10−3 . Moreover, in order to improve training stability
we performed gradient norm clipping [29] with norms limited to 1. The convolution parameters
were initialized using Xavier initialization [16], and all biases were initialized to zero. The training
took approximately 3 hours using a single Titan X GPU. The source code used to replicate these
experiments are available online 2
Results are presented in fig. 3. As can be seen, the reconstruction using L2 loss “smears” the reconstruction to an extent where the shape is impossible to recover. On the other hand, the reconstruction
using the Wasserstein loss retains the over-all global shape of the object, although relative and exact
positions of the circles are not recovered.
5
Conclusions and future work
In this work we have considered using Wasserstein loss to train a neural network for solving ill-posed
inverse problems in imaging where data is not aligned with the ground truth. We give a theoretical
motivation for why this should give better results compared to standard mean squared error loss, and
demonstrate it on a problem in computed tomography. In the future, we hope that this method can be
applied to other inverse problems and to other problems in imaging such as segmentation.
2
https://github.com/adler-j/wasserstein_inverse_problems
7
Appendix: Deferred definition and proofs
Proof of Proposition 1. To show that f (x) = (dP ∗ g)(x) ∈ L2 (Rn ) minimizes Eτ [kf − gτ k22 ] we
expand the expression and use Fubini’s theorem to get
Z Z
Z Z
2
2
2
Eτ [kf − gτ k2 ] =
|f (x) − gt (x)| dx dP (t) =
(f (x) − gt (x)) dP (t) dx.
Rn
Rn
Rn
Rn
R
dP (t)dt = 1, this can be written as
2
Z
Z
Eτ [kf − gτ k22 ] =
f (x) −
gt (x)dP (t) dx + c,
Rearranging terms and using that
Rn
Rn
Rn
where c is a constant. Using this it follows that the minimizing f is of the form
Z
f (x) =
gt (x)dP (t).
Rn
To see that f ∈ L2 (Rn ) we note that, by using Fubini’s theorem, we have
2
Z Z
Z Z Z
2
kf k2 =
gt (x)dP (t) dx =
gs (x)gt (x)dx dP (s)dP (t)
Rn
Rn
Rn Rn
Rn
Z Z Z
1
≤
gs (x)2 + gt (x)2 dx dP (s)dP (t) = kgk22 < ∞
2
Rn Rn
Rn
where the first inequality is the arithmetic-geometric mean inequality. This completes the proof.
Definition 5. Let h : Rn → R. A subgradient to h in a point y is a vector v so that
∀ x ∈ Rn .
h(x) ≥ h(y) + hv, x − yi,
The set of all subgradients in a point y is called the subdifferential of h at y, and is denoted by
n
R∂h(y). This is a set-valued operator, and for any measure dν on R we define (dν ∗ ∂h)(y) :=
∂h(t)dν(y
−
t)
to
be
the
set-valued
operator
Rn
Z
y 7→
v(t)dν(y − t) ∈ Rn | v(t) ∈ ∂h(t) .
Rn
Proof of Proposition 2. We consider finding the marginal µ that minimize Eτ [T (δτ , µ)]. Without
loss of generality we assume that τ is zero-mean, since otherwise we simply consider τ − E[τ ] which
is a zero-mean random variable. First we note that T (δt , µ) is only finite for nonegative measures µ
with total mass 1, and hence Eτ [T (δτ , µ)] is only finite for such measures. Second, for such a µ we
have
Z
T (δt , µ) =
c(t, x)dµ(x),
Rn
since one needs to transport all mass in µ into the point t where δt has its mass. Using this and
expanding the expression for the expectation gives that
Z
Z Z
Eτ [T (δτ , µ)] =
T (δt , µ)dP (t) =
c(t, x)dP (t) dµ(x),
Rn
Rn
Rn
where we have used Fubini’s theorem in the last step. This completes the first half of the statement.
To prove the second half of the statement, note that the optimal µ have support only in the global
minimas of the function
Z
Z
F (x) :=
d(x − t)dP (t) =
d(t)dP (x − t),
Rn
Rn
which by assumption exists and is finite. Now, since d is convex we have that
d(x) ≥ d(y) + h∂d(y), x − yi,
8
for all x, y ∈ Rn ,
(9)
and convolving this inequality with dP gives the inequality
Z
F (x) ≥ F (y) + h
∂d(y − t)dP (t), x − yi,
for all x, y ∈ Rn ,
(10)
Rn
where all terms exist and are bounded by assumption. This shows that
Z
∂d(y − t)dP (t) ⊂ ∂F (y).
Rn
Now, since d is symmetric we have that ∂d is anti-symmetric, i.e., that ∂d(x) = −∂d(−x), since
d(x) = d(−x) ≥ d(−y) + h∂d(−y), −x + yi = d(y) + h−∂d(−y), x − yi.
Therefore
Z
∂F (0) ⊃
∂d(−t)dP (t) 3 0,
Rn
where the last inclusion follows since dP is symmetric and ∂d is anti-symmetric. Now, since
0 ∈ ∂F (0) we have that x = 0 is a global minimizer to F (x) [6, Theorem 16.2], and thus one
optimal solution to the problem is µ(x) = δ(x). Now, if d is strictly convex, the inequality (9) is
strict for x 6= y, and thus (10) is also strict, which shows that the optimal solution is unique.
Lemma 6. Let k · k be a norm on Rm . Then
d(x1 , x2 ) = 1 − e−kx1 −x2 k
n
n1
.
is a metric on Rm for n ≥ 1.
Proof. It is easily seen that d(x1 , x2 ) is symmetric, nonnegative, and equal to zero if only if x1 = x2 .
Thus we only need to verify that the triangle inequality holds. To this end we note that if
n 1
n 1
n 1
1 − e−(a+b) n ≤ 1 − e−a n + 1 − e−b n , for all a, b ≥ 0,
(11)
for all n ≥ 1, then by taking a = kx1 − x2 k, b = kx2 − x3 k, and using the triangle inequality for
the norm k · k we have that
n 1
n 1
d(x1 , x3 ) = 1 − e−kx1 −x3 k n ≤ 1 − e−(kx1 −x2 k+kx2 −x3 k) n
n 1
n 1
≤ 1 − e−kx1 −x2 k n + 1 − e−kx2 −x3 k n = d(x1 , x2 ) + d(x2 , x3 ).
Therefore we will show that (11) holds for all n ≥ 1, and to do so we will
(i) show that if a function g : R+ → R+ fulfills g(0) = 0, g(x)0 ≥ 0, g 00 (x) ≤ 0 for all
x ∈ R+ , then g(x1 + x2 ) ≤ g(x1 ) + g(x2 ),
n
1
(ii) show that for x ≥ 0 the map x 7→ (1 − e−x ) n fulfills the assumptions in (i) for any n ≥ 1.
To show (i) we note that
x1 +x2
Z x1
Z x1 +x2
g 0 (t)dt =
g 0 (t)dt +
g 0 (t)dt
0
0
x1
Z x1
Z x2
0
0
≤
g (t)dt +
g (t)dt = g(x1 ) + g(x2 ),
Z
g(x1 + x2 ) =
0
0
where the inequality uses that g 0 (t) ≥ g 0 (x + t) for any x, t ≥ 0 since g 00 (x) ≤ 0 for all x ≥ 0.
n
1
To show (ii), let g(x) := (1 − e−x ) n and observe that g(0) = 0. Differentiating g twice gives
n
g 0 (x) =
n
1
e−x (1 − e−x ) n xn−1
1 − e−xn
=:hn (xn )
z
}|
{
1
xn n
n
xn
xn
−xn n
n−2
(1
−
e
)
x
ne
x
−
x
+
e
−
ne
+
n
−
1
g 00 (x) = −
.
(exn − 1)2
For x ≥ 0 we see that g 0 (x) ≥ 0 for all n ≥ 1. Moreover, for x ≥ 0 we see that g 00 (x) ≤ 0 for all
x ≥ 0 and for all n ≥ 1 if and only if hn (xn ) ≥ 0. With the change of variable xn = y, we thus
want to show that hn (y) ≥ 0 for all y ≥ 0 and all n ≥ 1. To see this we note that hn (0) = 0 and that
h0n (y) = ney y + ey − 1 ≥ 0 for all y ≥ 0 and n ≥ 1.
This shows (ii), and hence completes the proof.
9
References
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. Corrado, A. Davis, J. Dean, M. Devin,
et al. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint
arXiv:1603.04467, 2016.
[2] J. Adler, H. Kohr, and O. Öktem. ODL - a Python framework for rapid prototyping in inverse problems.
2017. In preparation, KTH Royal Institute of Technology. Code and documentation available online:
https://github.com/odlgroup/odl.
[3] J. Adler and O. Öktem. Learned primal-dual reconstruction. arXiv preprint arXiv:1707.06474, 2017.
[4] J. Adler and O. Öktem. Solving ill-posed inverse problems using iterative deep neural networks. Inverse
Problems, 2017.
[5] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN. arXiv preprint arXiv:1701.07875, 2017.
[6] H.H. Bauschke and P.L. Combettes. Convex analysis and monotone operator theory in Hilbert spaces.
Springer, New York, 2011.
[7] J.-D. Benamou, G. Carlier, M. Cuturi, L. Nenna, and G. Peyré. Iterative Bregman projections for regularized
transportation problems. SIAM Journal on Scientific Computing, 37(2):A1111–A1138, 2015.
[8] R.W. Brown, Y.-C. N. Cheng, E.M. Haacke, M.R. Thompson, and R. Venkatesan. Magnetic Resonance
Imaging: Physical Principles and Sequence Design. John Wiley & Sons Ltd, 2014.
[9] L. Chizat, G. Peyré, B. Schmitzer, and F.-X. Vialard. Unbalanced optimal transport: geometry and
Kantorovich formulation. arXiv preprint arXiv:1508.05216, 2015.
[10] M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural
Information Processing Systems, pages 2292–2300, 2013.
[11] M. Cuturi and G. Peyré. A smoothed dual approach for variational Wasserstein problems. SIAM Journal
on Imaging Sciences, pages 320–343, 2016.
[12] H.W. Engl, M. Hanke, and A. Neubauer. Regularization of inverse problems. Kluwer Academic Publisher,
2000.
[13] B. Engquist and B.D. Froese. Application of the Wasserstein metric to seismic signals. Communications
in Mathematical Sciences, 12(5), 2014.
[14] C. Frogner, C. Zhang, H. Mobahi, M. Araya, and T.A. Poggio. Learning with a Wasserstein loss. In
Advances in Neural Information Processing Systems, pages 2053–2061, 2015.
[15] T.T. Georgiou, J. Karlsson, and M.S. Takyar. Metrics for power spectra: an axiomatic approach. IEEE
Transactions on Signal Processing, 57(3):859–867, 2009.
[16] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In
Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pages
249–256, 2010.
[17] S. Haker, L. Zhu, A. Tannenbaum, and S. Angenent. Optimal mass transport for registration and warping.
International Journal of computer vision, 60(3):225–240, 2004.
[18] V. Jain and S. Seung. Natural image denoising with convolutional networks. In D. Koller, D. Schuurmans,
Y. Bengio, and L. Bottou, editors, Advances in Neural Information Processing Systems 21, pages 769–776.
Curran Associates, Inc., 2009.
[19] X. Jiang, Z.-Q. Luo, and T.T. Georgiou. Geometric methods for spectral analysis. IEEE Transactions on
Signal Processing, 60(3):1064–1074, 2012.
[20] J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual Losses for Real-Time Style Transfer and Super-Resolution,
pages 694–711. Springer International Publishing, Cham, 2016.
[21] J. Karlsson and T.T. Georgiou. Uncertainty bounds for spectral estimation. IEEE Transactions on Automatic
Control, 58(7):1659–1673, 2013.
[22] J. Karlsson and A. Ringh. Generalized Sinkhorn iterations for regularizing inverse problems using optimal
mass transport. arXiv preprint arXiv:1612.02273, 2016.
[23] D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
[24] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521:436—444, 2015.
[25] I. Loshchilov and F. Hutter.
arXiv:1608.03983, 2016.
SGDR: stochastic gradient descent with restarts.
arXiv preprint
[26] M. Mardani, E. Gong, J.Y. Cheng, S. Vasanawala, G. Zaharchuk, M.T. Alley, N. Thakur, S. Han, W.J.
Dally, J.M. Pauly, and L. Xing. Deep generative adversarial networks for compressed sensing automates
MRI. CoRR, abs/1706.00051, 2017.
10
[27] F. Natterer and F. Wübbeling. Mathematical methods in image reconstruction. SIAM, 2001.
[28] O. Öktem. Mathematics of electron tomography. In O. Scherzer, editor, Handbook of Mathematical
Methods in Imaging, pages 937–1031. Springer, New York, NY, 2015.
[29] R. Pascanu, T. Mikolov, and Y. Bengio. Understanding the exploding gradient problem. CoRR,
abs/1211.5063, 2012.
[30] P. Paschalis, N. D. Giokaris, A. Karabarbounis, G. K. Loudos, D. Maintas, C. N. Papanicolas,
V. Spanoudaki, Ch. Tsoumpas, and E. Stiliaris. Tomographic image reconstruction using artificial neural
networks. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment, 527(1–2):211–215, 2004.
[31] D. M. Pelt and K. J. Batenburg. Fast tomographic reconstruction from limited data using artificial neural
networks. IEEE Transactions on Image Processing, 22(12):5238–5251, 2013.
[32] P. Putzky and M. Welling. Recurrent inference machines for solving inverse problems. Submitted to ICLR
2017, Toulon, France, April 24–26, 2017. Report available from https://openreview.net/pdf?id=HkSOlP9lg,
2017.
[33] L.I. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based noise removal algorithms. Physica D:
Nonlinear Phenomena, 60(1):259–268, 1992.
[34] W. van Aarle, W.J. Palenstijn, J. Cant, E. Janssens, F. Bleichrodt, A. Dabravolski, J. De Beenhouwer, K.J.
Batenburg, and J. Sijbers. Fast and flexible X-ray tomography using the ASTRA toolbox. Optics express,
24(22):25129–25147, 2016.
[35] C. Villani. Optimal transport: old and new, volume 338. Springer Science & Business Media, 2008.
[36] A. Walker, G. Liney, P. Metcalfe, and L. Holloway. MRI distortion: considerations for MRI based
radiotherapy treatment planning. Australasian Physical & Engineering Sciences in Medicine, 37(1):103–
113, Mar 2014.
[37] T. Würfl, F. C. Ghesu, V. Christlein, and A. Maier. Deep learning computed tomography. In S. Ourselin,
L. Joskowicz, M. Sabuncu, G. Unal, and W. Wells, editors, MICCAI 2016: Medical Image Computing
and Computer-Assisted Intervention – MICCAI 2016, volume 9902 of Lecture Notes in Computer Science,
pages 432–440. Springer-Verlag, 2016.
[38] J. Xie, L. Xu, and E. Chen. Image denoising and inpainting with deep neural networks. In F. Pereira,
C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing
Systems 25, pages 341–349. Curran Associates, Inc., 2012.
[39] Y. Yang, J. Sun, H. Li, and Z. Xu. Deep ADMM-Net for compressive sensing MRI. In D. D.
Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29, pages 10–18. Curran Associates, 2016. Report available from
http://papers.nips.cc/paper/6406-deep-admm-net-for-compressive-sensing-mri.pdf.
[40] H. Zhao, O. Gallo, I. Frosio, and J. Kautz. Loss functions for image restoration with neural networks.
IEEE Transactions on Computational Imaging, 3(1):47–57, March 2017.
11
| 1cs.CV
|
A short characterization of relative entropy
arXiv:1712.04903v1 [cs.IT] 13 Dec 2017
Tom Leinster∗
Abstract
We prove characterization theorems for relative entropy (also known as
Kullback–Leibler divergence), q-logarithmic entropy (also known as Tsallis entropy), and q-logarithmic relative entropy. All three have been characterized axiomatically before, but we show that earlier proofs can be
simplified considerably, at the same time relaxing some of the hypotheses.
1
Introduction
The Shannon entropy of a finite probability distribution p = (p1 , . . . , pn ),
X
H(p) =
i : pi >0
pi log
1
,
pi
is such an important quantity that many authors have sought short lists of
properties that determine H uniquely. Many such characterization theorems
have been found, beginning with one in Shannon’s seminal paper of 1948 ([17],
Theorem 2). For instance, Faddeev [6] proved that up to a constant factor,
H is uniquely characterized by symmetry, continuity, and a certain recursivity
property.
Accompanying Shannon entropy is the concept of relative entropy, defined
as follows. Given probability distributions p and r on n elements, the entropy
of p relative to r is
X
pi
∈ [0, ∞].
pi log
H(p k r) =
ri
i : p >0
i
Relative entropy goes by a multitude of names: Kullback–Leibler divergence,
directed divergence, discrimination information, relative information, information gain, and so on. In information theory, it measures the wastage when a
language whose n letters have frequencies p = (p1 , . . . , pn ) is encoded using
a system optimized for a different language with frequencies r, instead of the
system optimized for the original language. There are other interpretations in
other fields, as the plethora of names suggests.
Axiomatic characterizations of relative entropy have also been sought and
found. One such theorem is implicit in work of Kannappan and Ng [12]. It
states that up to a constant factor, relative entropy is uniquely determined
∗ School of Mathematics, University of Edinburgh, UK; [email protected]. Key words:
relative entropy, Kullback–Leibler divergence, q-logarithm, q-logarithmic entropy, Tsallis entropy, Tsallis relative entropy
1
by measurability in each of p and r separately, invariance under permutations
of {1, . . . , n}, the vanishing property H(p k p) = 0, and a certain recursivity
equation. (Remark 2.7 gives further details.) Their proof was a tour de force of
functional equations, involving the solution of the functional equation
x
y
= h(y) + (1 − y)k
(1)
f (x) + (1 − x)g
1−x
1−y
in four unknown functions, as well as the four-variable functional equation
u
x
v
y
F (x, y) + (1 − x)F
= F (u, v) + (1 − u)F
.
,
,
1−x 1−y
1−u 1−v
We give a much simpler proof, at the same time weakening the measurability
hypothesis. Our proof involves neither of these equations. Instead, it borrows heavily from a categorical characterization of relative entropy by Baez and
Fritz [1]. This is the first main result, Theorem 2.1.
Shannon entropy is just one member (albeit a special one) of a one-parameter
family of entropies (Sq )q∈R , first investigated by Havrda and Charvát [10] and
often misattributed to Tsallis (Remark 3.2(ii)). These entropies Sq , and the
accompanying relative entropies, are defined as follows.
For q ∈ R, the q-logarithm is the function lnq : (0, ∞) → R given by
Z x
t−q dt.
lnq (x) =
1
The q-logarithmic entropy and q-logarithmic relative entropy are defined
by
X
1
pi lnq ,
Sq (p) =
p
i
i : pi >0
X
ri
Sq (p k r) = −
pi lnq ,
p
i
i : p >0
i
for probability distributions p and r on n elements. When q = 1, these reduce
to the ordinary Shannon entropy and relative entropy.
There are several existing theorems characterizing the q-logarithmic entropy
for a given q 6= 1. Up until now, the simplest appears to have been the 1970
result of Daróczy [5]. We simplify further, weakening the hypotheses and shortening the proof to just a few lines (Theorem 3.1).
Finally, we use a similar and equally short argument to characterize the
q-logarithmic relative entropies Sq (− k −) (Theorem 4.1).
It is remarkable that when q 6= 1, the characterizations of q-logarithmic
entropy and q-logarithmic relative entropy need no regularity conditions whatsoever (not even measurability), in contrast to the theorems for q = 1.
The remaining three sections of this paper establish our three theorems in
turn, characterizing first relative entropy, then q-logarithmic entropy, then qlogarithmic relative entropy.
2
Relative entropy
For n ≥ 1, write
n
o
X
∆n = p = (p1 , . . . , pn ) ∈ Rn : pi ≥ 0,
pi = 1
2
for the set of probability distributions on {1, . . . , n}, and write
An = (p, r) ∈ ∆n × ∆n : pi = 0 whenever ri = 0 .
Evidently, (p, r) ∈ An if and only if the relative entropy
H(p k r) =
X
pi log
i : pi >0
pi
ri
is finite. (Viewing p and r as measures on {1, . . . , n}, we have (p, r) ∈ An just
when p is absolutely continuous with respect to r.)
We will characterize the sequence of functions
H(− k −) : An → R n≥1
uniquely up to a constant factor. It is easy to check that this sequence has the
following four properties, as does any scalar multiple cH(− k −) (for c ∈ R).
Measurability in the second argument For each n ≥ 1 and p ∈ ∆n , the
function
{r ∈ ∆n : (p, r) ∈ An } →
R
r
7→ H(p k r)
is Lebesgue measurable.
Symmetry For each n ≥ 1, (p, r) ∈ An and permutation σ of {1, . . . , n},
H(p k r) = H(pσ k rσ),
(2)
where pσ = (pσ(1) , . . . , pσ(n) ).
Vanishing H(p k p) = 0 for all n ≥ 1 and p ∈ ∆n .
Chain rule To state this, we need some notation. Given n, k1 , . . . , kn ≥ 1 and
w ∈ ∆n , p1 ∈ ∆k1 , . . . , pn ∈ ∆kn , and writing pi = (pi1 , . . . , piki ), define
w ◦ (p1 , . . . , pn ) = (w1 p11 , . . . , w1 p1k1 , . . . , wn pn1 , . . . , wn pnkn ) ∈ ∆k1 +···+kn .
The chain rule for relative entropy is that
n
X
e i ) (3)
e ◦ (e
e n ) = H(w k w)
e +
wi H(pi k p
H w ◦ (p1 , . . . , pn ) w
p1 , . . . , p
i=1
e ∈ An and (pi , p
e i ) ∈ Aki . (Under these hypotheses, the
whenever (w, w)
pair of distributions on the left-hand side belongs to Ak1 +···+kn .)
Theorem 2.1 Let I(− k −) : An → R n≥1 be a sequence of functions. The
following are equivalent:
i. I(− k −) satisfies the four properties above: measurability in the second
argument, symmetry, vanishing, and the chain rule;
ii. I(− k −) = cH(− k −) for some c ∈ R.
3
We have just noted that (ii) implies (i). We now embark on the proof of the
converse. For the rest of this section, let I(− k −) : An → R n≥1 be a sequence
of functions satisfying the four conditions. Define a function L : (0, 1] → R by
L(α) = I (1, 0) (α, 1 − α) .
The idea is that if I(− k −) = H(− k −) then L = − log. We will show that in
any case, L is a scalar multiple of log.
Lemma 2.2 Let (p, r) ∈ An with pk+1 = · · · = pn = 0, where 1 ≤ k ≤ n. Then
r1 + · · · + rk > 0 and
I(p k r) = L(r1 + · · · + rk ) + I(p′ k r′ ),
where
p′ = (p1 , . . . , pk ),
r′ =
(r1 , . . . , rk )
.
r1 + · · · + rk
Proof The case k = n reduces to the statement that L(1) = 0, which follows
from the vanishing property. Suppose, then, that k < n.
Since p is a probability distribution with pi = 0 for all i > k, there is some
i ≤ k such that pi > 0, and then ri > 0 as (p, r) ∈ An . Hence r1 + · · · + rk > 0.
Let r′′ ∈ ∆n−k be the normalization of (rk+1 , . . . , rn ) if rk+1 + · · · + rn > 0, or
choose r′′ arbitrarily in ∆n−k otherwise (which is possible since k < n). Then
I(p k r) = I (1, 0) ◦ (p′ , r′′ ) (r1 + · · · + rk , rk+1 + · · · + rn ) ◦ (r′ , r′′ ) .
The result now follows from the chain rule.
Lemma 2.3 L(αβ) = L(α) + L(β) for all α, β ∈ (0, 1].
Proof We evaluate the real number
x := I (1, 0, 0) (αβ, α(1 − β), 1 − α)
in two ways. By Lemma 2.2 with k = 1 and the vanishing property,
x = L(αβ) + I (1) (1) = L(αβ),
where (1) is the unique element of ∆1 . But also, by Lemma 2.2 with k = 2,
x = L(α) + I (1, 0) (β, 1 − β) = L(α) + L(β).
Comparing the two expressions for x gives the result.
Lemma 2.4 There is some c ∈ R such that L(α) = −c log α for all α ∈ (0, 1].
Proof Define f : [0, ∞) → R by f (t) = L(e−t ). By Lemma 2.3, f (t + u) =
f (t) + f (u) for all t, u ∈ [0, ∞). Also, f is measurable, since L is. It is wellknown [2] that these conditions force f (t) = ct for some constant c, giving
L(α) = −c log α.
Our next lemma is an adaptation of the most ingenious part of Baez and
Fritz’s argument ([1], Lemma 4.2).
4
Lemma 2.5 Let (p, r) ∈ An with pi > 0 for all i ∈ {1, . . . , n}. Then I(p k r) =
cH(p k r).
Proof The hypotheses imply that ri > 0 for all i. We can therefore choose
some α ∈ (0, 1] such that ri − αpi ≥ 0 for all i. We will compute the number
x := I (p1 , . . . , pn , 0, . . . , 0) (αp1 , . . . , αpn , r1 − αp1 , . . . , rn − αpn )
| {z }
n
in two ways. First, by Lemma 2.2, Lemma 2.4, and the vanishing property,
x = L(α) + I(p k p) = −c log α.
Second, by symmetry, the chain rule, and Lemma 2.4,
x = I (p1 , 0, . . . , pn , 0) (αp1 , r1 − αp1 , . . . , pn , rn − αpn )
= I p ◦ (1, 0), . . . , (1, 0) r ◦ α pr11 , 1 − α pr11 , . . . , α prnn , 1 − α prnn
= I(p k r) +
n
X
i=1
pi L α prii
= I(p k r) − c log α − cH(p k r).
Comparing the two expressions for x gives the result.
We have now proved that I(p k r) = cH(p k r) when p has full support. It
only remains to prove it for arbitrary p.
Proof of Theorem 2.1 Let (p, r) ∈ An . By symmetry, we can assume that
p1 , . . . , pk > 0 and pk+1 = · · · = pn = 0 for some k ∈ {1, . . . , n}. Writing
R = r1 + · · · + rk , we have R > 0 since (p, r) ∈ An , and t
I(p k r) = L(R) + I (p1 , . . . , pk ) R1 (r1 , . . . , rk )
by Lemma 2.2. Hence by Lemmas 2.4 and 2.5,
I(p k r) = −c log R + cH (p1 , . . . , pk )
.
1
R (r1 , . . . , rk )
But by the same argument applied to cH in place of I (or by direct calculation),
we also have
cH(p k r) = −c log R + cH (p1 , . . . , pk ) R1 (r1 , . . . , rk ) .
The result follows.
Remarks 2.6
i. The vanishing P
axiom cannot be dropped from Theorem 2.1. Indeed, the quantity i : pi >0 pi log r1i satisfies the other three
axioms but not vanishing.
ii. In the literature on information functions, the chain rule is often replaced
by one of two superficially simpler rules. The first is the special case
k1 = 2, k2 = · · · = kn = 1, which is
H (pw1 , (1 − p)w1 , w2 , . . . , wn ) (e
pw
e1 , (1 − pe)w
e1 , w
e2 , . . . , w
en )
e + w1 H (p, 1 − p) (e
=H w w
p, 1 − pe) (4)
5
e ∈ An , ((p, 1 − p), (e
((w, w)
p, 1 − pe)) ∈ A2 ). This is known as recursivity.
The second is the special case n = 2 of the chain rule, which is
we
ep ⊕ (1 − w)e
er
e ) + (1 − w)H(r k e
= H (w, 1 − w) (w,
e 1 − w)
e + wH(p k p
r), (5)
H wp ⊕ (1 − w)r
where
wp ⊕ (1 − w)r = (wp1 , . . . , wpk , (1 − w)r1 , . . . , (1 − w)rℓ )
e ) ∈ Ak , (r, e
and ((w, 1 − w), (w,
e 1 − w))
e ∈ A2 , (p, p
r) ∈ Aℓ . However,
straightforward inductions (similar to those in Feinstein [7], p. 5–6) show
that in the presence of the symmetry axiom, either one of the special
cases (4) or (5) is equivalent to the full chain rule (3). Which to use is,
therefore, simply a matter of taste.
Remark 2.7 Here we compare Theorem 2.1 with some earlier characterizations
of relative entropy. One of the first such theorems was that of Hobson [11], who
used stronger hypotheses for the same conclusion. In common with Theorem 2.1,
he assumed symmetry, vanishing, and the chain rule (in the equivalent form (5)).
But he also assumed continuity in both variables (instead of measurability in
one) and a monotonicity hypothesis unlike anything in Theorem 2.1.
In 1973, Kannappan and Ng [12] proved a result very close to Theorem 2.1.
They did not state that result in [12], but the closing remarks in another paper
by the same authors [13] and the approach of a contemporaneous paper by
Kannappan and Rathie [14] strongly suggest the intent. The result was stated
explicitly by Csiszár ([4], Section 2.1), who attributed it to Kannappan and Ng.
There are some slight differences of hypotheses between Kannappan and
Ng’s theorem and Theorem 2.1. They assumed measurability in both variables,
whereas we only assumed measurability in the second. (In fact, all we used
was that I((1, 0) k −) is measurable.) On the other hand, they only needed the
vanishing condition for (1/2, 1/2), whereas we needed it for all p. They used
the chain rule in the equivalent form (4). And as indicated in the Introduction,
the proofs are entirely different.
3
q-logarithmic entropy
Let q ∈ R. The definition of q-logarithm in the Introduction gives, explicitly,
lnq (x) =
1
(x1−q − 1)
1−q
for x ∈ (0, ∞) and q 6= 1, while ln1 is the natural logarithm log. Hence,
explicitly, the q-logarithmic entropy is given by
!
X q
1
p −1
Sq (p) =
1 − q i : p >0 i
i
for p ∈ ∆n and q 6= 1, while S1 is the Shannon entropy H. We have lnq (x) →
log(x) as q → 1, hence also Sq (p) → H(p) as q → 1.
6
Fix q ∈ R. The q-logarithmic entropy satisfies a chain rule
X
wiq Sq (pi )
Sq w ◦ (p1 , . . . , pn ) = Sq (w) +
(6)
i : wi >0
(w ∈ ∆n , p1 ∈ ∆k1 , . . . , pn ∈ ∆kn ), as is easily checked. In particular, this
holds when p1 = · · · = pn = p, say. For w ∈ ∆n and p ∈ ∆k , write
w ⊗ p = w ◦ (p, . . . , p)
= (w1 p1 , . . . , w1 pk , . . . , wn p1 , . . . , wn pk ) ∈ ∆nk .
In this case, the q-chain rule (6) gives a q-multiplicativity property:
!
X
q
wi Sq (p)
Sq (w ⊗ p) = Sq (w) +
(7)
i : wi >0
(n, k ≥ 1, w ∈ ∆n , p ∈ ∆k ).
Note also that Sq is symmetric in its arguments:
Sq (p) = Sq (pσ)
(8)
for all p ∈ ∆n and permutations σ of {1, . . . , n}.
The left-hand side of equation (7) is symmetric in w and p, but the righthand side is not obviously so. This is the key to our second theorem.
Theorem 3.1 Let 1 6= q ∈ R and let (I : ∆n → R)n≥1 be a sequence of functions. The following are equivalent:
i. I has the q-multiplicativity property (7) and the symmetry property (8)
(both with I in place of Sq );
ii. I = cSq for some c ∈ R.
Proof By the observations just made, (ii) implies (i). Now assume (i). By
symmetry, I(w ⊗ p) = I(p ⊗ w), so
!
!
X q
X
q
pi I(w),
wi I(p) = I(p) +
I(w) +
i : pi >0
i : wi >0
or equivalently
X
wiq
!
− 1 I(p) =
i : wi >0
X
i : pi >0
pqi
!
− 1 I(w),
for all w ∈ ∆n and p ∈ ∆k . Take w = (1/2, 1/2): then for all p ∈ ∆k ,
!
X q
1−q
pi − 1 I(1/2, 1/2).
2
− 1 I(p) =
i : pi >0
Since q 6= 1, we can define c =
1−q
21−q −1
· I(1/2, 1/2), and then I = cSq .
7
Remarks 3.2
i. The q-logarithms were used in Hardy, Littlewood and
Pólya’s classic book on inequalities, first published in 1934 ([9], proof
of Theorem 84). They have been an explicit object of study since at least
a 1964 paper of Box and Cox in statistics ([3], Section 3). The name
‘q-logarithm’ appears to have been introduced by Umarov, Tsallis and
Steinberg in 2008 [20], working in statistical mechanics.
ii. The q-logarithmic entropies have been discovered and rediscovered repeatedly. To my knowledge, they first appeared in a 1967 paper on information
and classification by Havrda and Charvát [10], who used a form adapted
to base 2 logarithms. They were rediscovered in 1970 by Daróczy [5]. The
base e version Sq seems to have first appeared in a 1982 article of Patil and
Taillie ([16], Section 3.2), where it was studied as an index of biodiversity.
In physics, meanwhile, the q-logarithmic entropies appeared in a 1971 article of Lindhard and Nielsen [15] (according to Csiszar [4], Section 2.4), and
in a 1978 survey by Wehrl ([21], p. 247). Finally, they were rediscovered
again in a 1988 paper on statistical physics by Tsallis [19].
Despite the twenty years of active life that the q-logarithmic entropies had
already enjoyed, it is after Tsallis that they are most commonly named.
The term ‘q-logarithmic entropy’ is new, but has the benefits of being
descriptive and of not perpetuating a misattribution.
iii. As in Remark 2.6(ii) or Feinstein [7] (p. 5–6), a simple inductive argument
shows that the q-chain rule of equation (6) follows from the special case
(9)
Sq pw1 , (1 − p)w1 , w2 , . . . , wn = Sq (w) + w1q Sq (p, 1 − p)
(p ∈ [0, 1], n ≥ 1, w ∈ ∆n ).
iv. A characterization of the q-logarithmic entropies similar to Theorem 3.1
was published by Daróczy in 1970 [5]. He assumed the full q-chain rule
for I(w ◦ (p1 , . . . , pn )) (in the equivalent form (9)), rather than just the
special case of I(w ⊗ p) that we used. However, where we assumed that
I : ∆n → R is symmetric for all n ≥ 2, Daróczy only assumed it for n = 3.
The two proofs are very different; the main step in Daróczy’s was the
solution of the functional equation (1) in the case f = g = h = k.
Other characterizations of Sq have been proved, using stronger hypotheses
than Theorem 3.1 to obtain the same conclusion (such as the theorem in
Section 2 of [18], and Theorem V.2 of [8]).
4
q-logarithmic relative entropy
For q 6= 1, the q-logarithmic relative entropy Sq : An → R, defined in the Introduction, is given explicitly by
!
X q 1−q
1
p r
−1 ,
Sq (p k r) =
q − 1 i : p >0 i i
i
for (p, r) ∈ An and q 6= 1. In the case q = 1, it reduces to the ordinary
relative entropy H(p k r). As in that case, restricting the arguments to lie in
An guarantees that Sq (− k −) takes only finite values.
8
Our third and final theorem is a characterization of q-logarithmic relative
entropy, very similar to the characterization of q-logarithmic entropy itself.
We begin by noting two properties of q-logarithmic relative entropy. First,
there is an easily-checked chain rule:
X
ei)
e ◦ (e
e n ) = Sq (w k w)
e +
wiq w
ei1−q Sq (pi k p
p1 , . . . , p
Sq w ◦ (p1 , . . . , pn ) w
i : wi >0
e ∈ An , (pi , p
e i ) ∈ Aki ). This specializes to a q-multiplicativity formula
((w, w)
!
X
q 1−q
e)
e ⊗p
e ) = Sq (w k w)
e +
Sq (p k p
(10)
wi w
ei
Sq (w ⊗ p k w
i : wi >0
e ∈ An , (p, p
e ) ∈ Ak ). Second, q-logarithmic relative entropy has the
((w, w)
same symmetry property as ordinary relative entropy:
Sq (p k r) = Sq (pσ k rσ)
(11)
for all n ≥ 1, (p, r) ∈ An , and permutations σ of {1, . . . , n}.
Theorem 4.1 Let 1 6= q ∈ R and let I(− k −) : An → R n≥1 be a sequence of
functions. The following are equivalent:
i. I(− k −) has the q-multiplicativity property (10) and the symmetry property (11) (both with I in place of Sq );
ii. I(− k −) = cSq (− k −) for some c ∈ R.
Proof It is trivial that (ii) implies (i). Now assume (i). By symmetry,
e ⊗p
e ) = I(p ⊗ w k p
e ⊗ w)
e
I(w ⊗ p k w
e ∈ An , and (p, p
e ) ∈ Ak . So by q-multiplicativity,
for all n, k ≥ 1, (w, w)
!
!
X q 1−q
X
q 1−q
e
e ) = I(p k p
e) +
e +
I(w k w),
pi pei
I(p k p
wi w
ei
I(w k w)
i : pi >0
i : wi >0
or equivalently,
X
i : wi >0
wiq w
ei1−q
!
e ). =
− 1 I(p k p
e = (1/2, 1/2): then
Take w = (1, 0) and w
(2
q−1
e) =
− 1)I(p k p
X
i : pi >0
X
i : pi >0
pqi pe1−q
i
pqi pe1−q
i
!
e
− 1 I(w k w).
!
− 1 I (1, 0) (1/2, 1/2)
e ) ∈ Ak . But q 6= 1, so we can define
for all (p, p
c=
1−q
· I (1, 0) (1/2, 1/2) ,
−1
2q−1
and then I(− k −) = cSq (− k −).
Remarks 4.2 Other characterization theorems for q-logarithmic relative entropy have been proved. For example, Furuichi ([8], Section IV) obtained the
same conclusion, but also assumed continuity and essentially the full chain rule
(that is, an equivalent special case, as in Remarks 2.6(ii) and 3.2(iii)).
9
Acknowledgements I thank John Baez and Tobias Fritz for their comments.
This work was partially supported by a BBSRC FLIP award (BB/P004210/1).
References
[1] J. Baez and T. Fritz. A Bayesian characterization of relative entropy. Theory and
Applications of Categories, 29:421–456, 2014.
[2] S. Banach. Sur l’équation fonctionnelle f (x + y) = f (x) + f (y). Fundamenta
Mathematicae, 1:123–124, 1920.
[3] G. E. P. Box and D. R. Cox. An analysis of transformations. Journal of the Royal
Statistical Society. Series B (Methodological), 26:211–252, 1964.
[4] I. Csiszár. Axiomatic characterizations of information measures. Entropy, 10:261–
273, 2008.
[5] Z. Daróczy. Generalized information functions. Information and Control, 16:36–
51, 1970.
[6] D. K. Faddeev. On the concept of entropy of a finite probabilistic scheme (in
Russian). Uspekhi Matematicheskikh Nauk, 11:227–231, 1956.
[7] A. Feinstein. Foundations of Information Theory. McGraw–Hill, New York, 1958.
[8] S. Furuichi. Uniqueness theorems for Tsallis entropy and Tsallis relative entropy.
IEEE Transactions on Information Theory, 51:3638–3645, 2005.
[9] G. Hardy, J. E. Littlewood, and G. Pólya. Inequalities. Cambridge University
Press, Cambridge, 2nd edition, 1952.
[10] J. Havrda and F. Charvát. Quantification method of classification processes:
concept of structural α-entropy. Kybernetika, 3:30–35, 1967.
[11] A. Hobson. A new theorem of information theory. Journal of Statistical Physics,
1(3):383–391, 1969.
[12] P. Kannappan and C. T. Ng. Measurable solutions of functional equations related
to information theory. Proceedings of the American Mathematical Society, 38:303–
310, 1973.
[13] P. Kannappan and C. T. Ng. On functional equations connected with directed
divergence, inaccuracy and generalized directed divergence. Pacific Journal of
Mathematics, 54(1):157–167, 1974.
[14] P. Kannappan and P. N. Rathie. On a characterization of directed divergence.
Information and Control, 22:163–171, 1973.
[15] J. Lindhard and V. Nielsen. Studies in statistical dynamics. Matematisk-Fysiske
Meddelelser: Kongelige Danske Videnskabernes Selskab, 38:1–42, 1971.
[16] G. P. Patil and C. Taillie. Diversity as a concept and its measurement. Journal
of the American Statistical Association, 77(379):548–561, 1982.
[17] C. E. Shannon. A mathematical theory of communication. Bell System Technical
Journal, 27:379–423, 1948.
[18] H. Suyari. On the most concise set of axioms and the uniqueness theorem for
Tsallis entropy. Journal of Physics A: Mathematical and General, 35:10731–10738,
2002.
[19] C. Tsallis. Possible generalization of Boltzmann–Gibbs statistics. Journal of
Statistical Physics, 52:479–487, 1988.
[20] S. Umarov, C. Tsallis, and S. Steinberg. On a q-central limit theorem consistent
with nonextensive statistical mechanics. Milan Journal of Mathematics, 76:307–
328, 2008.
[21] A. Wehrl. General properties of entropy. Reviews of Modern Physics, 50(2):221–
260, 1978.
10
| 7cs.IT
|
Day-Ahead Solar Forecasting Based on Multi-level
Solar Measurements
Mohana Alanazi, Mohsen Mahoor, Amin Khodaei
Department of Electrical and Computer Engineering
University of Denver
Denver, USA
[email protected], [email protected], [email protected]
Abstract—The growing proliferation in solar deployment,
especially at distribution level, has made the case for power
system operators to develop more accurate solar forecasting
models. This paper proposes a solar photovoltaic (PV) generation
forecasting model based on multi-level solar measurements and
utilizing a nonlinear autoregressive with exogenous input
(NARX) model to improve the training and achieve better
forecasts. The proposed model consists of four stages of data
preparation, establishment of fitting model, model training, and
forecasting. The model is tested under different weather
conditions. Numerical simulations exhibit the acceptable
performance of the model when compared to forecasting results
obtained from two-level and single-level studies.
Keywords—Solar generation forecast, nonlinear autoregressive
with exogenous input (NARX).
NOMENCLATURE
Pactual
Pforecast
Actual solar generation
Forecasted solar generation
P actual
N
Ec
Ef
ES
Average actual solar generation
Number of sample
Calculated error at customer level
Calculated error at feeder level
Calculated error at substation level
I.
S
INTODUCTION
OLAR FORECASTING plays a key role in the planning,
control and operation of power systems. Although this
viable generation technology is making fast inroads in
electricity grids, solar forecasting is still facing various
challenges, due to the inherent variability and uncertainty in
solar photovoltaic (PV) generation [1], [2].
Numerous factors, including but not limited to the dropping
cost of solar technology, environmental concerns, and the state
and governmental incentives, have made the path for a rapid
growth of solar regeneration. More than 2 GW of solar PV
was installed only in the U.S. in the summer of 2016, which is
43% higher compared to the installed capacity in the same
timeframe in 2015, to achieve an accumulated capacity of 31.6
GW [3]. Accordingly, solar forecasting problem has attracted
more attention to properly incorporate solar generation into
power system planning, operation, and control.
Many research studies are carried out on solar forecasting
problem, and several approaches are suggested to improve
forecasting results [4]-[8]. In [9], a short-term one-hour-ahead
Global Horizontal Irradiance (GHI) forecasting framework is
developed using machine learning and pattern recognition
models. This model reduces normalized mean absolute error
and the normalized root mean square error compared to the
commonly-used persistence method by 16% and 25%,
respectively. In [10], an intelligent approach for wind and
solar forecasting is proposed based on linear predictive coding
and digital image processing. It is shown that the model can
outperform conventional methods and neural networks.
Ensemble methods are quite popular in statistics and machine
learning, as they reap the benefit of multiple predictors to
achieve not only an aggregated, but also a better and reliable
decision. A survey paper on using ensemble methods for wind
power forecasting and solar irradiance forecasting is proposed
in [11]. The paper concludes that the ensemble forecasting
methods in general outperform the non-ensemble ones. A
comprehensive review focusing on the state-of-the-art
methods applied to solar forecasting is conducted in [12]. A
variety of topics including the advantages of probabilistic
forecast methods over deterministic ones, and the current
computational approaches for renewable forecasting is
discussed in this paper.
Historical data are of great importance for solar forecasting.
By leveraging historical data, the solar PV generation can be
forecasted for various time horizons as discussed in [13]. This
study investigates least-square support vector machines,
artificial neural network (ANN), and hybrid statistical models
based on least square support vector machines with wavelet
decomposition. In addition, a variety of measures, including
the root mean square error, mean bias error and mean absolute
error, are employed to evaluate the performance of the
aforementioned methods. The hybrid method based on leastsquare support vector machines and wavelet decomposition
surpasses the other methods. A new two-stage approach for
online forecasting of solar PV generation is proposed in [14].
This approach leverages a clear sky model to achieve a
statistical normalization. Normalized power output is further
forecasted by using two adaptive linear time series models;
autoregressive and autoregressive with exogenous input.
Various types of ANN, including but not limited to recurrent
neural network, feed-forward neural network, and radial basis
function neural network are employed for solar forecasting.
Substation
II.
Feeder level
Customer level
Substation level
Fig. 1 Multi-level Solar PVs installed at different locations.
The ANNs not only can process complex and nonlinear
time series forecast problems, but also can learn and figure out
the relationship between the input and the target output. On
the basis of ANN, a statistical method for solar PV generation
forecasting is proposed in [15]. One of the lessons learned
from this paper is that neural networks can be well-trained to
enhance forecast accuracy. In [16], by levering stationary data
and employing post-processing steps, a feed-forward neural
network-based method for day ahead solar forecasting is
studied. A comprehensive review of solar forecasting by using
different ANNs is provided in [17].
Hybrid models are considered highly effective for solar
forecasting in a way that they reinforce capabilities of each
individual method. Hybrid models reap the benefits of two or
more forecasting methods with the objective of achieving a
better forecast result [18]-[21]. In [22], authors present a
hybrid model consisting of various forecasting methods for a
48-hour-ahead solar forecasting in North Portugal. This study
advocates that the hybrid model attains a significant
improvement compared to statistical models. Another hybrid
short-term model to forecast solar PV generation is studied in
[23]. This hybrid model is formed on the basis of both group
method of data handling and least-square support vector
machine, where the performance of the hybrid model
significantly outperforms the other two methods.
The existing literature in this research area lacks studies on
multi-level data measurements for day-ahead solar PV
generation forecasting. Leveraging the multi-level solar
measurements to provide a more accurate forecasting for the
solar PV generation is the primary objective of this paper. The
solar PV generation, which is measured at various locations
including customer, feeder and substation, is utilized for dayahead solar forecasting with the objective of enhancing the
forecast accuracy. These multi-level measurements could play
an instrumental role in enhancing solar forecasting in terms of
reaching lower error values. The proposed forecasting model,
which will be further discussed in details in this paper,
consists of four stages and takes advantage of multiple
datasets related to specific locations.
The rest of the paper is organized as follows: Section II
discusses outline and the architecture of the proposed
forecasting model. Numerical simulations are presented in
Section III. Discussions and conclusions drawn from the
studies are provided in Section IV.
FORECASTING MODEL OUTLINE AND ARCHITECTURE
Fig. 1 depicts the three levels of solar PV measurements:
customer, feeder, and substation. The proposed model aims to
outperform the forecast applied at each solar measurement
level. The forecast in each level is performed using a
nonlinear autoregressive neural network. The mean absolute
percent error MAPE is accordingly calculated as in (3) for
each level, and denoted as EC, EF, and ES for customer, feeder,
and substation, respectively. This model aims to reduce the
forecasting error to be less than the minimum of EC, EF, and
ES. Fig. 2 depicts the three datasets, which are processed under
different stages and explained in the following.
Hourly solar PV generation at customer
(C), feeder (F), and substation (S) levels
Forecast using NARNN
and calculate the errors
EC, EF , ES
Data Preparation
Preprocessed Data at customer, feeder,
and substation level
Fitting Model
Establish a fitting model using NARNN
at customer, feeder and substation level
Forecasting
Select the best fitting model with
2
maximum R
Train the model
based on the
previous
preprocessed data
Forecast the output using NARX
Calculate the new error (En)
NO
En< min (EC , EF , ES )
Yes
Data post-processing and final output
Fig. 2 The flowchart of the multi-level solar PV generation forecasting.
A. Data Preprocessing and Adjusment
The data used in the simulation represent the total solar PV
generation. The data preparation includes removing offset,
normalization,
removing
nighttime
values,
and
stationarization. More detail about data preprocessing can be
found in [16]. The data preparation is to ensure the quality of
dataset before it is inputted to the forecasting model. This step
includes the simulation of maximum power generated from
solar PV at clear sky conditions. This is achieved by
simulating the maximum solar PV generation at clear sky
conditions using the system advisory model (SAM) provided
by National Renewable Energy Laboratory (NREL) [24]. The
maximum solar irradiance along with different metrological
inputs in clear condition are fed to SAM in order to simulate
the maximum solar PV generation. Fig. 3 presents the
flowchart for data preparation.
Benchmark data
Clear sky solar
PV generation
Data
preprocessing
Stationarity
check
Fig. 3 The flowchart for data preparation.
B. Fitting Model
By using NARNN, the fitting model is created for each
level. In this respect, the NARNN model utilizes a large set of
historical data in order to train the model and then forecast the
output. It is applied to the three datasets, including customer,
feeder, and substation in order to establish the three fitting
models. The best fitting model among the three is selected
using the coefficient of determination R2. The coefficient of
determination examines the proportion variance of the
predicted fitting model. The coefficient of determination can
be expressed mathematically as in (1), where P(t)actual is the
average of the actual data over the number of sample. The R2
ranges from 0 to 1, where 0 represents that the fitting model is
not predictable, and 1 means that the NARNN is able to
predict the fitting without any error. So, the best selected
fitting model among the three is the one with maximum R2.
N
( P(t ) actual P(t ) forecast ) 2
2
t 1
R 1 N
2
t1( P(t ) actual P(t ) actual )
(1)
C. Forecasting
NARX is a time series model that predicts the output using
historical values y(t) as well as inputs x(t). The NARX model
is presented in (2), where d is the number of considered
historical values. The fitting model is fed as an input to NARX
along with the previously preprocessed data. The NARX is
trained and the output is forecasted. Fig. 4 depicts the
architecture of the NARX. The goal is to forecast a day-ahead
solar PV generation with a new error En, which is less than the
minimum of the three errors as shown in the flowchart in
terms of a condition.
(2)
y(t) = f (x(t -1),.., x(t - d), y(t -1),.., y(t - d))
D. Data Post-processing
The output from the forecasting model is post-processed by
denormalizing, adding nighttime values, and calculating the
final solar output as explained in detail in [16]. MAPE and the
root mean square error (RMSE) are calculated as in (3) and
(4), respectively.
MAPE
RMSE
1 N P(t ) actual P(t ) forecast
t 1
N
P(t ) actual
1 N
* ( P(t ) actual P(t ) forecast ) 2
N t 1
Hidden Layer
x(t-1)
…
x(t-d)
(3)
neuron
(4)
Output Layer
fn
fn
Σ
y(t-1)
…
y(t-d)
fn
y(t)
fn
Fig. 4 The architecture of the NARX
III.
NUMERICAL SIMULATIONS
The hourly solar PV generation of three levels including
customer (C), feeder (F), and substation (S), for a specific area
in Denver, Colorado are utilized to perform forecasting. The
data used in this model are available in [25]. The customer
level data are considered as the aggregated customers’ solar
PVs generation for a selected area. The feeder level data are
the aggregated solar PVs generation for each feeder, in which
four feeders are considered in this study. Finally, the
substation level data are the solar PV generation measured at
the substation level. In order to demonstrate the merits of the
proposed model, the following three cases with various
weather conditions are investigated:
Case 1: Forecast using NARNN for each level without data
processing.
Case 2: Forecast using NARX with three-level measurements
and data processing.
Case 3: Forecast using NARX with two-level measurements
and data processing.
Case 4: Forecast using NARX with single-level measurement
and data processing.
Case 1: In this case, by leveraging NARNN, day-ahead solar
PV generation is forecasted for all three levels, while ignoring
data processing. The calculated MAPE and RMSE for the
customer, feeder, and substation levels with different weather
conditions are listed in Table I. As highlighted in Table I, the
customer level forecast has achieved the minimum MAPE as
well as RMSE for the selected weather conditions. This case is
considered as a base case, in which the calculated values are
utilized in order to demonstrate the effectiveness of using the
three-level measurements for forecasting. The objective in the
following case is to apply the proposed model in order to get a
new error that is less than the minimum achieved under this
case.
8.29
23.12
7.34
38.59
6.73
Substation
70.77
10.41
43.65
10.13
53.03
10.37
Case 2: In this case, three-level measurements are
preprocessed in order to ensure the quality of the training data
fed to the forecasting model. This case includes three
forecasting stages: establishing the fitting model from each
measurement level using the NARNN, training the NARX
model using the previously preprocessed datasets, and
forecasting the solar PV generation using the three-level
measurements and the fitting model as input. The fitting
model with the minimum MAPE and the maximum R2 is
selected as input to NARX. Table II exhibits how well the
fitting model is established in terms of R2 and MAPE for the
three-level measurements under different weather conditions.
As highlighted in Table II, the fitting model established by
using the customer level measurement outperforms the ones
established by using the feeder and substation measurements.
In order to show the merit of using three-level measurements
for the same location, the three measurements along with the
best fitting model are fed as inputs to NARX to forecast the
solar PV generation. The forecast is simulated for the same
selected days in Case 1. Table III exhibits the MAPE and
RMSE for the selected days. The forecast errors in this case
are less than the minimum achieved in Case 1. Fig. 5, 6 and 7
depict the forecasted and actual solar PV generation for the
considered sunny, cloudy, and partly cloudy days,
respectively.
TABLE II
THE FITTING MODEL MAPE AND R2 FOR THE CONSIDERED LEVELS UNDER
DIFFERENT WEATHER CONDITIONS
Sunny
Cloudy
Partly Cloudy
Dataset
MAPE
MAPE
MAPE
2
2
level
R
R
R2
(%)
(%)
(%)
Customer
2.39
0.9987
2.29
0.9956
3.49
0.99
Feeder
3.78
0.996
2.80
0.9943
4.02
0.988
Substation
5.95
0.9873
4.98
0.9821
5.37
0.986
Case 3: In this case only two measurements at customer and
feeder levels are used for forecasting. The preprocessed data
along with the best selected fitting model are fed to NARX.
The forecasting performance of this case is shown in Table
III.. In sunny day, this case has reduced the MAPE compared
to Case 1 by 47%. In cloudy and partly cloudy weather
conditions, Case 3 has reduced the MAPE compared to Case 1
by 61% and 19%, respectively.
Case 4: To exhibit the effectiveness of the three-level
measurements, Case 2 is repeated, but only one measurement
(customer level) is included as an input to NARX. Similar to
Weather
Condition
Sunny
Cloudy
Partly
Cloudy
2000
PV power output (kW)
48.81
1500
TABLE III
THE MAPE FOR DIFFERENT CASE STUDIES
Minimum
MAPE
MAPE
MAPE
(Using
(Using
(Using
NARX and
NARX and
NARNN
three-level
two-level
without data
processed
processed
processing)
data)
data)
4.47
1.67
2.38
MAPE
(Using
NARX and
single-level
processed
data)
3.14
6.04
2.10
2.36
2.44
4.09
2.69
3.30
3.39
Forecasted
Actual
1000
500
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324
Time (hour)
Fig. 5 Actual and forecasted solar PV generation in a sunny day
PV power output (kW)
Feeder
the previous case, the best fitting model based on MAPE and
R2 is fed to NARX along with preprocessed customer level
measurement. Table III shows the forecast error using NARX
with single-level measurement comparing to NARX with
three-level measurements, two-level measurements, and the
minimum forecast error among the single-level measurement
using NARNN without data processing. A single-level
measurements considerably improve the results over NARNN
method, however achieve not as good solution as in two
previous cases with three- and two-level measurements.
700
600
500
400
300
200
100
0
Forecasted
Actual
1 2 3 4 5 6 7 8 9 101112131415161718192021222324
Time (hour)
Fig. 6 Actual and forecasted solar PV generation in a cloudy day
1200
PV power output (kW)
TABLE I
CASE 1: MAPE AND RMSE FOR THE CONSIDERED DATASETS UNDER
DIFFERENT WEATHER CONDITIONS
Sunny
Cloudy
Partly Cloudy
Dataset
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
level
(kW)
(%)
(kW)
(%)
(kW)
(%)
Customer
44.58
4.47
20.54
6.04
36.11
4.09
1000
Forecasted
Actual
800
600
400
200
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324
Time (hour)
Fig. 7 Actual and forecasted solar PV generation in a partly cloudy day
As shown in Table III, the model minimizes the forecast
error to outperform the minimum error reported at customer
level. The proposed model has reduced the error compared to
the minimum error in Case 1 by 63%, 65%, and 34% for
sunny, cloudy, and partly cloudy weather conditions,
respectively. Moreover, the merit of using three-level
measurements is shown by comparing the forecast error using
the proposed model with applying two-level measurements to
the model as in Case 3. The MAPE is reduced by 30%, 11%,
and 18% for sunny, cloudy, and partly cloudy weather
conditions, respectively. The three-level measurement also
outperforms Case 4 in which only single-level measurement
are included. The three-level measurements model has
reduced the MAPE by 47%, 14%, and 21% for sunny, cloudy,
and partly cloudy weather conditions, respectively. The
previous cases have shown that forecasting performance is
greatly impacted by the historical data used to train the model.
Multiple historical data for a specific location along with an
appropriate data processing will improve the training step and
minimize the forecasting error.
IV.
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
CONCLUSION
In this paper, a day-ahead solar PV generation forecast
model based on multi-level measurements was proposed. The
proposed model demonstrated an improvement in forecasting
accuracy by reducing the MAPE from 14% to 47% for various
weather conditions, compared to the case when only singlelevel measurements were included. It was further seen that the
data preprocessing was an important step to ensure the quality
of the data before it was used in the training process. The
numerical studies revealed that training the forecasting model
without data preprocessing might adversely impact the
forecasting accuracy. The proposed preprocessing model
could potentially reduce the MAPE by 34% to 65%. It was
further shown that the three-level measurements help achieve
a better forecasting accuracy compared to two-level
measurements. The proposed model can be further enhanced
by including multiple meteorological parameters such as cloud
cover, solar irradiance, and temperature along with three-level
measurements as inputs to NARX.
[14]
REFERENCES
[22]
[1]
[2]
[3]
[4]
[5]
H. Sangrody, et al, "Weather forecasting error in solar energy
forecasting," IET Renewable Power Generation, 2017.
S. Watetakarn and S. Premrudeepreechacharn, "Forecasting of solar
irradiance for solar power plants by artificial neural network,” Smart
Grid Technologies-Asia (ISGT ASIA), 2015 IEEE Innovative, 2015.
S. Kann, J. Baca, M. Shiao, C. Honeyman, A. Perea, and S. Rumery,
“US Solar Market Insight - Q3 2016 - Executive Summary,” GTM
Research,Wood Mackenzie Business and the Solar Energy Industries
Association.
S. Akhlaghi, H. Sangrody, M. Sarailoo, and M. Rezaeiahari, "Efficient
operation of residential solar panels with determination of the optimal
tilt angle and optimal intervals based on forecasting model," IET
Renewable Power Generation, 2017.
H. Sangrody, M. Sarailoo, A. Shokrollahzade, F. Hassanzadeh, and E.
Foruzan "On the Performance of Forecasting Models in the Presence of
Input Uncertainty," North American Power Symposium (NAPS),
Morgantown, West Virginia, USA, 2017.
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[23]
[24]
[25]
J. Remund, R. Perez, and E. Lorenz, “Comparison of solar radiation
forecasts for the USA,” in Proc. of the 23rd European PV Conference,
Valencia, Spain. 2008.
G. Reikard, S. E.Haupt, and T. Jensen, “Forecasting ground-level
irradiance over short horizons: Time series, meteorological, and timevarying parameter models,” Renewable Energy, vol. 112, pp.474-485,
Nov. 2017.
G. Reikard, “Predicting solar radiation at high resolutions: A
comparison of time series forecasts,” Solar Energy, vol. 83, no. 3,
pp.342-349, March 2009.
C. Feng, et al, “Short-term Global Horizontal Irradiance Forecasting
Based on Sky Imaging and Pattern Recognition,” IEEE General
Meeting, Chicago, IL, July 2017.
A. A. Moghaddam and A. Seifi, "Study of forecasting renewable
energies in smart grids using linear predictive filters and neural
networks," IET Renewable Power Generation, vol. 5, no.6, pp. 470480, Dec. 2011.
Y. Ren, P. Suganthan, and N. Srikanth, "Ensemble methods for wind
and solar power forecasting-A state-of-the-art review," Renewable and
Sustainable Energy Reviews, vol. 50, pp. 82-91, 2015.
R. Banos, F. Manzano-Agugliaro, F. Montoya, C. Gil, A. Alcayde, and
J. Gómez, "Optimization methods applied to renewable and sustainable
energy: A review," Renewable and Sustainable Energy Reviews, vol. 15,
no.4, pp. 1753-1766, 2011.
M. G. De Giorgi, P. M. Congedo, M. Malvoni, and D. Laforgia, "Error
analysis of hybrid photovoltaic power forecasting models: A case study
of mediterranean climate," Energy Conversion and Management, vol.
100, pp. 117-130, 2015.
P. Bacher, H.Madsen, and H. A. Nielsen, "Online short-term solar
power forecasting," Solar Energy, vol. 83, no.10, pp. 1772-1783, 2009.
C. Chen, S. Duan, T. Cai, and B. Liu, "Online 24-h solar power
forecasting based on weather type classification using artificial neural
network," Solar Energy, vol. 85, no.11, pp. 2856-2870, 2011.
M. Alanazi and A. Khodaei, "Day-ahead Solar Forecasting Using Time
Series Stationarization and Feed-Forward Neural Network," North
American Power Symposium (NAPS), 2016, Denver, CO, USA, 2016.
A. Mellit and S. A. Kalogirou, "Artificial intelligence techniques for
photovoltaic applications: A review," Prog. Energy Combust. Sci., vol.
34, no. 5, pp. 574–632, Oct. 2008.
M. Alanazi, M. Mahoor and A. Khodaei, "Two-stage hybrid day-ahead
solar forecasting," North American Power Symposium (NAPS),
Morgantown, WV, USA, 2017.
J. Wu and C. K. Chan, "Prediction of hourly solar radiation using a
novel hybrid model of ARMA and TDNN," Solar Energy, vol. 85, no. 5,
pp. 808-817, May 20111.
M. Bouzerdoum, A. Mellit, and A. M. Pavan, "A hybrid model
(SARIMA–SVM) for short-term power forecasting of a small-scale
grid-connected photovoltaic plant," Solar Energy, vol. 98, pp. 226-235,
Dec. 2013.
R. Marquez, et al, "Hybrid solar forecasting method uses satellite
imaging and ground telemetry as inputs to ANNs," Solar Energy, vol.
92, pp. 176-188, June 2013.
J. M. Filipe, R. J. Bessa, J. Sumaili, R. Tome, and J. N. Sousa, “A
hybrid short-term solar power forecasting tool,” in Intelligent System
Application to Power Systems (ISAP), 2015 18th International
Conference on, 2015, pp. 1–6.
M. De Giorgi, M. Malvoni, and P. Congedo, "Comparison of strategies
for multi-step ahead photovoltaic power forecasting models based on
hybrid group method of data handling networks and least square support
vector machine," Energy, vol. 107, pp. 360-373, 2016.
“System Advisor Model (SAM) |.” [Online]. Available:
https://sam.nrel.gov/.
S. Pfenninger and I. Staffell, “Long-term patterns of European PV
output using 30 years of validated hourly reanalysis and satellite data,”
Energy, vol. 114, pp. 1251–1265, Nov. 2016.
| 2cs.AI
|
3D Trajectory Reconstruction of Dynamic Objects Using Planarity Constraints
Sebastian Bullinger1 , Christoph Bodensteiner1 , Michael Arens1 and Rainer Stiefelhagen2
1
arXiv:1711.06136v1 [cs.CV] 16 Nov 2017
2
Fraunhofer IOSB
Karslruhe Institute of Technology
{sebastian.bullinger,christoph.bodensteiner,michael.arens}@iosb.fraunhofer.de
[email protected]
Abstract
mented reality applications. There are different platforms
like drones or wearable systems where one wants to achieve
this task with a minimal number of devices in order to reduce weight or lower production costs. We propose an approach to reconstruct three-dimensional object motion trajectories using a single camera as sensor.
The reconstruction of object motion trajectories in monocular video data captured by moving cameras is a challenging
task, since in general it cannot be solely solved exploiting
image observations. Each observed object motion trajectory
is scale ambiguous. Additional constraints are required to
identify a motion trajectory consistent to background structures. [23, 13, 3] assume that the camera is mounted on
a driving vehicle, i.e. the camera has specific height and
a known pose. [16, 28, 17] solve the scale ambiguity by
making assumptions about object and camera motion trajectories. We follow Ozden’s principle of non-accidental
motion trajectories [16] and introduce a new object motion
constraint exploiting semantic segmentation and terrain geometry to compute consistent object motion trajectories.
In many scenarios objects cover only a minority of pixels in
video frames. This increases the difficulty of reconstructing object motion trajectories using image data. In such
cases current state-of-the-art Structure from Motion (SfM)
approaches treat moving object observations most likely as
outliers and reconstruct background structures instead. Previous works, e.g. [11, 12], tackle this problem by considering multiple video frames to determine moving parts in the
video. They apply motion segmentation or keypoint tracking to detect moving objects. These kind of approaches are
vulnerable to occlusion and require objects to move in order
to separate them from background structures.
Our method exploits recent results in instance-aware semantic segmentation and rigid Structure from Motion techniques. Thus, our approach extends naturally to stationary objects. In addition, we do not exploit specific camera
pose constraints like a fixed camera-ground-angle or a fixed
We present a method to reconstruct the threedimensional trajectory of a moving instance of a known
object category in monocular video data. We track the
two-dimensional shape of objects on pixel level exploiting
instance-aware semantic segmentation techniques and optical flow cues. We apply Structure from Motion techniques
to object and background images to determine for each
frame camera poses relative to object instances and background structures. By combining object and background
camera pose information, we restrict the object trajectory to
a one-parameter family of possible solutions. We compute a
ground representation by fusing background structures and
corresponding semantic segmentations. This allows us to
determine an object trajectory consistent to image observations and reconstructed environment model. Our method
is robust to occlusion and handles temporarily stationary
objects. We show qualitative results using drone imagery.
Due to the lack of suitable benchmark datasets we present a
new dataset to evaluate the quality of reconstructed threedimensional object trajectories. The video sequences contain vehicles in urban areas and are rendered using the
path-tracing render engine Cycles to achieve realistic results. We perform a quantitative evaluation of the presented
approach using this dataset. Our algorithm achieves an average reconstruction-to-ground-truth distance of 0.31 meter. The dataset will be publicly available on our website1 .
1. Introduction
1.1. Trajectory Reconstruction
The reconstruction of three-dimensional object motion
trajectories is important for autonomous systems and aug1 Project
page: URL
1
camera-ground-distance. We evaluate the presented object
motion trajectory reconstruction algorithm in UAV scenarios, where such constraints are not valid.
struction related research.
1.4. Paper Overview
The paper is organized as follows. Section 2 describes
the structure and the components of the proposed pipeline.
In section 2.1 we derive an expression for a one-parameter
family of possible object motion trajectories combining object and background reconstruction results. Section 2.2 describes a method to approximate the ground locally. In section 2.3 we describe a method to compute consistent object
motion trajectories. In section 4 we provide an qualitative
and quantitative evaluation of the presented algorithms using drone imagery and rendered video data. Section 5 concludes the paper.
1.2. Related Work
Semantic segmentation or scene parsing is the task of
providing semantic information at pixel-level. Early semantic segmentation approaches using ConvNets, e.g. Farabet
et al. [5], exploit patchwise training. Long et al. [21] applied Fully Convolutinal Networks for semantic segmentation, which are trained end-to-end. Recently, [4, 14, 9] proposed instance-aware semantic segmentation approaches.
The field of Structure from Motion (SfM) can be divided
into iterative and global approaches. Iterative or sequential
SfM methods [22, 27, 15, 24, 20] are more likely to find
reasonable solutions than global SfM approaches [15, 24].
However, the latter are less prone to drift.
The determination of the correct scale ratio between object and background reconstruction requires additional constraints. Ozden et al. [16] exploit the non-accidentalness
principle in the context of independently moving objects.
Yuan et al. [28] propose to reconstruct the 3D object trajectory by assuming that the object motion is perpendicular to the normal vector of the ground plane. Kundu et al.
[11] exploit motion segmentation with multibody VSLAM
to reconstruct the trajectory of moving cars. They use an
instantaneous constant velocity model in combination with
Bearing only Tracking to estimate consistent object scales.
Park et al. propose an approach in [17] to reconstruct the
trajectory of a single 3D point tracked over time by approximating the motion using a linear combination of trajectory
basis vectors. Previous works, like [16, 28, 11, 17] show
only qualitative results.
2. Object Motion Trajectory Reconstruction
The pipeline of our approach is shown in Fig. 1. The
input is an ordered image sequence. We track twodimensional object shapes on pixel level across video sequences following the approach presented in [2]. In contrast
to [2], we identify object shapes exploiting the instanceaware semantic segmentation method presented in [14] and
associate extracted object shapes of subsequent frames using the optical flow approach described in [10]. Without
the loss of generality, we describe motion trajectory reconstructions of single objects. We apply SfM [15, 20] to object and background images as shown in Fig. 1. Object
images denote images containing only color information of
single object instance. Similarly, background images show
only background structures. We combine object and background reconstructions to determine possible, visually identical, object motion trajectories. We compute a consistent
object motion trajectory exploiting constraints derived from
reconstructed terrain ground geometry.
1.3. Contribution
2.1. Object Trajectory Representation
The core contributions of this work are as follows. (1)
We present a new framework to reconstruct the threedimensional trajectory of moving instances of known object categories in monocular video data leveraging sateof-the-art semantic segmentation and structure from motion approaches. (2) We propose a novel method to compute object motion trajectories consistent to image observations and background structures. (3) In contrast to previous work, we quantitatively evaluate the reconstructed object motion trajectories. (4) We created a new object motion trajectory benchmark dataset due to the lack of publicly
available video data of moving objects with suitable ground
truth data. The dataset consists of photo-realistic rendered
videos of urban environments. It includes animated vehicles
as well as set of predefined camera and object motion trajectories. 3D vehicle and environmental models used for rendering serve as ground truth. (5) We will publish the dataset
and evaluation scripts to foster future object motion recon-
In order to estimate a consistent object motion trajectory
we apply SfM simultaneously to object and background images as shown in Fig. 1. We denote the corresponding SfM
(o)
results with sf m(o) and sf m(b) . Let oj ∈ P (o) and
(b)
bk ∈ P (b) denote the 3D points contained in sf m(o) or
(o)
sf m(b) , respectively. The superscripts o and b in oj and
(b)
bk describe the corresponding coordinate frame. The variables j and k are the indices of points in the object or the
background point cloud, respectively. We denote the reconstructed intrinsic and extrinsic parameters of each registered input image as virtual camera. Each virtual camera
in sf m(o) and sf m(b) corresponds to a certain frame from
which object and background images are extracted. We associate virtual cameras in sf m(o) with the corresponding
virtual cameras in sf m(b) and vice versa. In the following, we consider only camera pairs, whose virtual cameras
2
Input Frames
Semantic
Segmentation and
Object Tracking
Object Segmentations
Background
Segmentations
SfM
Ground Segmentations
SfM
Background
SfM Result
Object SfM Result
Object Trajectory Family
Trajectory Family
Computation
Background
SfM Result
Scale Estimation
and Trajectory
Computation
Ground Representation
Consistent Object
Motion Trajectory
Ground
Computation
Figure 1: Overview of the Trajectory Reconstruction Pipeline.
are contained in sf m(o) and sf m(b) . Because of missing
image registrations this may not be the case for all virtual
cameras.
We reconstruct the object motion trajectory by combining
information of corresponding virtual cameras. For any virtual camera pair of an image with index i the object SfM
result sf m(o) contains information of object point positions
(o)
(o)
oj relative to virtual cameras with camera centers ci and
(o)
(o)
rotations Ri . We express each object point oj
(i)
coordinates oj of camera i using equation (1)
(i)
(o)
oj = Ri
(o)
(o)
· (oj − ci ).
described according to equation (3).
(b)
(b)
(b)
T (b)
(i)
· oj .
T (b)
(o)
· Ri
(o)
(o)
· (oj − ci )
:= ci(b) + r · v(b)
j,i
(3)
with
(b)
T (b)
vj,i = Ri
in camera
(o)
· Ri
(o)
(o)
(b)
(b)
· (oj − ci ) = oj,i − ci .
(4)
Given the scale ratio r, we can recover the full object motion
trajectory computing equation (4) for each virtual camera
(b)
pair. We use oj,i of all cameras and object points as object
motion trajectory representation. The ambiguity mentioned
in section 1 is expressed by the unknown scale ratio r.
(1)
The background SfM result sf m(b) contains the camera
(b)
(b)
center ci and the corresponding rotation Ri , which provide pose information of the camera with respect to the reconstructed background. Note, that the camera coordinate
systems of virtual cameras in sf m(o) and sf m(b) are equiv(b)
(b)
alent. We use ci and Ri to transform object points to the
background coordinate system using equation (2)
oj,i = ci + Ri
(b)
oj,i = ci + r · Ri
2.2. Terrain Ground Approximation
Further camera or object motion constraints are required
to determine the scale ratio r introduced in equation (4).
In contrast to previous work [16, 28, 17, 13, 23, 3] we assume that the object category of interest moves on top of
the terrain. We exploit semantic segmentation techniques
to estimate an approximation of the ground surface of the
scene. We apply the ConvNet presented in [21] to determine ground categories like street or grass for all input images on pixel level. We consider only stable background
points, i.e. 3D points that are observed at least four times.
We determine for each 3D point a ground or non-ground
label by accumulating the semantic labels of corresponding
keypoint measurement pixel positions. This allows us to determine a subset of background points, which represent the
ground of the scene. We approximate the ground surface
locally using plane representations. For each frame i we
(2)
In general, the scale ratio of object and background reconstruction does not match due to the scale ambiguity of
SfM reconstructions [8]. We tackle this problem by treating the scale of the background as reference scale and by
introducing a scale ratio factor r to adjust the scale of object point coordinates. The overall transformation of object
(o)
points given in object coordinates oj to object points in
(b)
the background coordinate frame system oj,i of camera i is
3
Equation (8) allows us to determine the scale ratio r
between object and background reconstruction using the
extrinsic parameters of two cameras and corresponding
ground approximations.
use corresponding estimated camera parameters and object
point observations to determine a set of ground points Pi
close to the object. We build a kd-tree containing all ground
measurement positions of the current frame. For each object point observation, we determine the numb closest background measurements. In our experiments, we set numb to
50. Let cardi be the cardinality of Pi . While cardi is less
than numb , we add the next background observation of each
point measurement. This results in an equal distribution of
local ground points around the vehicle. We apply RANSAC
[6] to compute a local approximation of the ground surface
using Pi . Each plane is defined by a corresponding normal
vector ni and an arbitrary point pi lying on the plane.
2.3.2
The accuracy of the estimated scale ratio r in equation (8)
is subject to the condition of the parameters of the particular view pair. For instance, if the numerator or denominator is close to zero, small errors in the camera poses or
ground approximations may result in negative scale ratios.
In addition, wrongly estimated local plane normal vectors
may disturb camera-plane distances. We tackle these problems by combining two different view pair rankings. The
first ranking uses for each view pair the difference of the
camera-plane distances, i.e. the numerator in equation (8).
The second ranking reflects the quality of the local ground
approximation w.r.t. the object reconstruction. For a view
pair with well reconstructed local planes the variance of the
corresponding scale ratios is small. This allows for the determination of ill conditioned view pairs. The second ranking uses the scale ratio difference to order the view pairs.
We sort the view pairs by weighting both ranks equally. Let
vp denote the view pair with the lowest overall rank. The final scale ratio is determined by using a least squares method
w.r.t. all equations of vp.
2.3. Scale Estimation using Constant Distance Constraints
In section 2.3, we exploit priors of object motion to improve the robustness of the reconstructed object trajectory.
We assume that the object of interest moves on a locally
planar surface. In this case the distance of each object point
(b)
oj,i to the ground is constant for all cameras i. The reconstructed trajectory shows this property only for the true
scale ratio and non-degenerated camera motion. For example, a degenerate case occurs when the camera moves exactly parallel to a planar object motion. For a more detailed
discussion of degenerated camera motions see [16].
2.3.1
Scale Ratio Estimation using a Single View Pair
We use the term view to denote cameras and corresponding local ground planes. The signed distance of an object
(b)
point oj,i to the ground plane can be computed according
to equation (5)
(b)
dj,i = ni · (oj,i − pi ),
3. Virtual Object Motion Trajectory Dataset
To quantitatively evaluate the quality of the reconstructed object motion trajectory we require accurate object
and environment models as well as object and camera poses
at each time step. The simultaneous capturing of corresponding ground truth data with sufficient quality is difficult
to achieve. For example, one could capture the environment
geometry with LIDAR sensors and the camera / object pose
with an additional system. However, the registration and
synchronization of all these different modalities is a complex and cumbersome process. The result will contain noise
and other artifacts like drift. To tackle these issues we exploit virtual models. Previously published virtually generated and virtually augmented datasets, like [18, 19, 7, 25],
provide data for different application domains and do not include three-dimensional ground truth information. We build
a virtual world including an urban environment, animated
vehicles as well as predefined vehicle and camera motion
trajectories. This allows us to compute spatial and temporal
error free ground truth data. We exploit procedural generation of textures to avoid artificial repetitions. Thus, our
dataset is suitable for evaluating SfM algorithms.
(5)
where pi is a point on the local ground plane and ni is the
corresponding normal vector. If the object moves on top of
the approximated terrain ground the distance dj,i should be
independent of a specific camera i. We substitute dj,i with
dj in equation (5). This allows us to combine equation (5)
of the same point and different cameras.
(b)
(b)
ni · (oj,i − pi ) = ni0 · (oj,i0 − pi0 ).
(6)
Substituting equation (3) in equation (6) results in (7)
(b)
(b)
(b)
(b)
ni · (ci + r · vj,i − pi ) = ni0 · (ci0 + r · vj,i0 − pi0 ) (7)
Solving equation (7) for r yields equation (8)
(b)
r=
(b)
ni0 · (ci0 − pi0 ) − ni · (ci − pi )
(b)
(b)
(ni · vj,i − ni0 · vj,i0 )
.
Scale Ratio Estimation using View Pair Ranking
(8)
4
(a) Environment Model from (b) Environment Rendered from (c) Car Model with Motion Path
a Bird’s Eye Perspective (in a Bird’s Eye Perspective (Ren- in Street Scene (in Blender)
Blender)
dered)
(d) Rendered Street Scene
Figure 2: Example Scenes from the Virtual Object Trajectory Dataset.
(a) Input Frame 0
(b) Input Frame 100
(c) Input Frame 200
(d) Reconstructed
(Top View)
Trajectory
Figure 3: Car Trajectory Reconstruction using 250 frames with a resolution of 1920 x 1080 pixels captured by a DJI drone.
3.1. Virtual World
quences cover a high variety of object and camera poses.
The object trajectories reflect common vehicle motions include vehicle acceleration, different curve types and motion
on changing slopes. We use the path-tracing render engine
Cycles [1] to achieve photo realistic rendering results. We
observed that the removal of artificial path-tracing artifacts
using denoising is crucial to avoid degenerated SfM reconstructions.
The dataset includes 6D object and camera poses for each
frame as well as ground truth meshes of corresponding vehicle models. In contrast to measured ground truth data, virtual ground truth data is free of noise and shows no spatial
registration or temporal synchronization inaccuracies. The
dataset contains semantic segmentations of objects, ground
and background to separate the reconstruction task from
specific semantic segmentation and tracking approaches. In
addition to the virtual data, the dataset also includes the
computed reconstruction results. We will make our evaluation scripts publicly available to foster future analysis of
object trajectory estimation.
We used Blender [1] to create a virtual world consisting of a city surrounded by a countryside. We exploit procedural generation to compute textures of large surfaces,
like streets and sidewalks, to avoid degenerated Structure
from Motion results caused by artificial texture repetitions.
The virtual world includes different assets like trees, traffic lights, streetlights, phone booths, bus stops and benches.
We collected a set of publicly available vehicle assets to
populate the scenes. We used skeletal animation, also referred to as rigging, for vehicle animation. This includes
wheel rotation and steering w.r.t. the motion trajectory as
well as consistent vehicle placement on uneven ground surfaces. The animation of wheels is important to avoid unrealistic wheel point triangulations. We adjusted the scale of
vehicles and virtual environment using Blender’s unit system. This allows us to set the virtual space in relation to the
real world. The extent of the generated virtual world corresponds to one square kilometer. We exploit environment
mapping to achieve realistic illumination. With Blender’s
built-in tools, we defined a set of camera and object motion
trajectories. This allows us to determine the exact 3D pose
of cameras and vehicles at each time step.
4. Experiments and Evaluation
We show qualitative results and quantitative evaluations
using real drone footage and virtual generated drone imagery, respectively. Fig. 3 shows the reconstruction of a car
motion trajectory using images captured by a DJI drone.
Fig. 4 shows an example of the virtual object trajectory
dataset. Fig. 4(j) and Fig. 4(i) show the object point cloud
transformed into the virtual world coordinate frame system.
3.2. Trajectory Dataset
We use the previously created virtual world to build a
new object trajectory dataset1 . The dataset consists of 35
sequences capturing five vehicles in different urban scenes.
Fig. 2 shows some example images. The virtual video se5
(a) Ground Truth Object Path
(b) Object Model in Blender
(c) Rendered Scene
(d) Object Segmentation
(e) Background Segmentation
(f) Ground Segmentation
(g) Object Reconstruction
(h) Background Reconstruction
(i) Reconstructed Object Motion (j) Reconstructed Object Motion (k) Ground Truth Model at Dif- (l) Ground Truth Model at DifferTrajectory
Trajectory
ferent Frames Overlayed With ent Frames Overlayed With ReReconstruction
construction
Figure 4: Car trajectory reconstruction using 60 virtual frames with a resolution of 1920 x 1080 pixels.
truth coordinates. For each sequence we define the trajectory error as the average trajectory-point-mesh distance.
Fig. 5 shows for each sequence the trajectory error in meter. The average trajectory error per vehicle using the full
dataset is shown in table 1. Overall, we achieve a trajectory
error of 0.31 meter. The error of the object trajectory reconstructions reflects four types of computational inaccuracies:
deviations of camera poses w.r.t. object and background
point clouds, wrong triangulated object points as well as
scale ratio discrepancies. Fig. 6 compares the estimated
scale ratios of the proposed and the baseline method w.r.t.
the reference scale ratio. The reference scale ratio computation is described in section 4.4. The overall estimated scale
ratio deviation w.r.t. the reference scale per vehicle is shown
in table 1. The provided reference scale ratios are subject to
the registration described in section 4.3. Wrongly reconstructed background camera poses may influence the reference scale ratio. The van object reconstruction was only
partial successful on the sequences crossing, overtaking and
steep street. The SfM algorithm registered 19%, 60% and
98% of the images, respectively. The object reconstruction
of the smart model contained 74% of the crossing input object images. Here, we use the subset of registered images to
perform the evaluation. The camera and the object motion
in bumpy road simulate a sequence close to a degenerated
case, i.e. equation (8) is ill conditioned for all view pairs.
Fig. 4(k) and Fig. 4(l) show the overlay result of transformed points and the corresponding virtual ground truth
model. To segment the two-dimensional shape of objects of
interest we follow the approach presented in [2]. In contrast to [2], we used [14] and [10] to segment and track
visible objects, respectively. We considered the following SfM pipelines for object and background reconstructions: Colmap [20], OpenMVG [15], Theia [24] and VisualSfM [27]. Our object trajectory reconstruction pipeline
uses Colmap for object and OpenMVG for background reconstructions, since Colmap and OpenMVG created in our
experiments the most reliable object and background reconstructions.
4.1. Virtual Object Motion Trajectory Evaluation
We use the dataset presented in section 3 to quantitatively evaluate the proposed object motion trajectory reconstruction approach. The evaluation is based on object, background and ground segmentations included in the dataset.
This allows us to show results independent from the performance of specific instance segmentation and tracking approaches. We compare the proposed method with the baseline presented in section 4.2 using 35 sequences contained
in the dataset. We automatically register the reconstructed
object trajectory to the ground truth using the method described in section 4.3. We compute the shortest distance of
each object trajectory point to the object mesh in ground
6
Trajectory Error in meter
Lancer (Baseline)
Lancer (Ours)
2
Lincoln (B.)
Lincoln (O.)
Smart (B.)
Smart (O.)
Golf (B.)
Golf (O.)
Van (B.)
Van (O.)
1
0
Right Curves
Left Curves
Crossing
Overtaking
Bridge
Steep Street
Bumpy Road
Deviation w.r.t. Reference
Figure 5: Quantitative evaluation of the trajectory error in meter computed by the methods constant distance (ours) and
intersection (baseline). We evaluate seven different vehicle trajectories (right curves, left curves, crossing, overtaking, bridge,
steep street and bumpy road) and five different vehicle models (Lancer, Lincoln Navigator, Smart, Golf and Van). The
cropped values of the baseline are 6.3m, 4m and 4.8m.
Lancer (Baseline)
Lancer (Ours)
0.4
Lincoln (B.)
Lincoln (O.)
Smart (B.)
Smart (O.)
Golf (B.)
Golf (O.)
Van (B.)
Van (O.)
0.2
0
Right Curves
Left Curves
Crossing
Overtaking
Bridge
Steep Street
Bumpy Road
Figure 6: Quantitative evaluation of the scale ratio of the methods constant distance (ours) and intersection (baseline). The
provided values are the deviations w.r.t. reference scales.
4.2. Scale Estimation Baseline: Intersection Constraints
The intersection parameter ri corresponds to the point being
closest to the ground surface, i.e. a point at the bottom of the
object. Plugging ri in equation (3) for camera i places the
object point cloud on top of the ground surface. Thus, the
smallest ray-plane-intersection-parameter ri is a value close
to the real object-to-background-scale. Finally, we use the
median s of all image scale ratio factors as scale ratio factor
to reconstruct the trajectory as computed in equation (11)
The baseline is motivated by the fact, that some reconstructed points of the bottom of an object should lie in the
proximity of the ground surface of the environment. Consider for example the reconstructed 3D points corresponding to the wheels of a car. This approach works only if
at least one camera-point-ray of an arbitrary point in the
object point cloud intersects the ground surface. For each
(b)
camera we generate a set of vectors vj,i pointing from the
(b)
r = med({ri |i ∈ {1, . . . , nI }}),
where med denotes the median and nI the number of images. We do not consider invalid image scale ratios ri , i.e.
cameras which have no camera-object-point-rays intersecting the ground representation.
(b)
camera center ci towards each object point oj,i . For non(b)
orthogonal direction vectors vj,i and normal vectors ni
we compute the ray plane intersection parameter for each
camera-object-point-pair according to equation (9)
4.3. Registration of Background Reconstruction
and Virtual Environment
(b)
rj,i =
(pi − ci ) · ni
(b)
vj,i · ni
.
(9)
A common approach to register different coordinate systems is to exploit 3D-3D correspondences. To determine
points in the virtual environment corresponding to background reconstruction points one could create a set of rays
from each camera center to all visible reconstructed back-
We compute the smallest ray-plane-intersection parameter
for each image i.
ri = min({rj,i |j ∈ {1, . . . , |P (o) |}})
(11)
(10)
7
Scale Ratio
Est. Type
Intersection (Baseline)
Constant Distance (Ours)
Average Scale Ratio Deviation w.r.t. Reference
Lancer Lincoln Smart Golf
Van
0.05
0.07
0.01
0.08
0.13
0.04
0.04
0.04
0.06
0.08
Lancer
0.42
0.20
Average Trajectory Error
Lincoln Smart Golf
0.53
0.25
0.95
0.23
0.33
0.33
Van
1.68
0.47
Table 1: Summary of the conducted evaluation. The second column shows the deviation of the estimated scale ratio w.r.t to
the reference scale ratio per vehicle. The third column contains the average distances of the full dataset in meter. Overall, the
trajectory error of the baseline and our approach is 0.77m and 0.31m.
ground points. The corresponding environment points are
defined by the intersection of these rays with the mesh of
the virtual environment. Due to the complexity of our environment model this computation is in terms of memory and
computational effort quite expensive. Instead, we use the
algorithm presented in [26] to estimate a similarity transformation Ts between the cameras contained in the background reconstruction and the virtual cameras used to render the corresponding video sequence. This allows us to
perform 3D-3D-registrations of background reconstructions
and the virtual environment as well as to quantitatively evaluate the quality of the reconstructed object motion trajectory. We use the camera centers as input for [26] to compute an initial reconstruction-to-virtual-environment transformation. Depending on the shape of the camera trajectory
there may be multiple valid similarity transformations using camera center positions. In order to find the semantically correct solution we enhance the original point set
with camera pose information, i.e. we add points reflect(b)
T (b)
ing up vectors ui = Ri · (0, 1, 0)T and forward vectors
(b)
T (b)
f i = Ri · (0, 0, 1)T . For the reconstructed cameras, we
adjust the magnitude of these vectors using the scale computed during the initial similarity transformation. We add
(b)
(b)
the corresponding end points of up ci + m · ui as well
(b)
(b)
as viewing vectors ci + m · f i to the camera center point
set. Here, m denotes the corresponding magnitude.
plicitly contains information about the scale ratio r(bv) between background reconstruction and virtual environment.
To compute r(ov) we use corresponding pairs of object reconstruction and virtual cameras. We use the extrinsic parameters of the object reconstruction camera to transform
all 3D points in the object reconstruction into camera coordinates. Similarly, the object mesh with the pose of the
corresponding frame is transformed into the camera coordinates leveraging the extrinsic camera parameters of the
corresponding virtual camera. The ground truth pose and
shape of the object mesh is part of the dataset. In camera
coordinates we generate rays from the camera center (i.e.
(i)
the origin) to each 3D point oj in the object reconstruc-
4.4. Reference Scale Ratio Computation
ref
Equation (14) shows that r(ob)
depends on the quality of the
cameras in the background reconstruction and may slightly
differ from the true scale ratio.
(i)
tion. We determine the shortest intersection mj of each
ray with the object mesh in camera coordinates. This allows us to compute r(ov) according to equation (13)
(i)
r(ov) = med({med({
kmj k
(i)
koj k
|j ∈ {1, . . . , nJ }}
(13)
|i ∈ {1, . . . , nI }}).
ref
This allows us to compute the reference scale ratios r(ob)
using equation (14)
ref
r(ob)
= r(ov) · r(bv) −1 .
As explained in section 4.1 the presented average trajectory errors in Fig. 5 are subject to four different error
sources. To evaluate the quality of the scale ratio estimation between object and background reconstruction we provide corresponding reference scale ratios. The scale ratios
between object reconstruction, background reconstruction
and virtual environment are linked via the relation shown in
equation (12)
r(ov) = r(ob) · r(bv) ,
(12)
(14)
5. Conclusions
This paper presents a pipeline to reconstruct the threedimensional trajectory of moving objects using monocular
video data. We propose a novel constraint to estimate consistent object motion trajectories. In contrast to previous
approaches, the presented scale estimation constraint is robust to occlusion and extents naturally to stationary objects.
Due to the lack of 3D object motion trajectory benchmark
datasets with suitable ground truth data, we present a new
virtual dataset to quantitatively evaluate object motion trajectories. The dataset contains rendered videos of urban en-
where r(ov) and r(bv) are the scale ratios between object
and background reconstructions and virtual environment,
respectively. The scale ratios r(ob) in Fig. 6 express the
spatial relation of object and background reconstructions.
The similarity transformation Ts defined in section 4.3 im8
vironments and accurate ground truth data including semantic segmentations, object meshes as well as object and camera poses for each frame. The proposed algorithm achieves
an average reconstruction-to-ground-truth distance of 0.31
m evaluating 35 trajectories. In future work, we will analyze
breakdown points of the proposed pipeline in more detail.
This includes minimal object sizes, object occlusions and
degeneracy cases. In addition, we intend to integrate previously published scale estimation approaches. These will
serve together with the dataset1 as benchmark references for
future object motion trajectory reconstruction algorithms.
[14]
[15]
[16]
[17]
References
[1] Blender Online Community. Blender - a 3d modelling and
rendering package, 2016.
[2] S. Bullinger, C. Bodensteiner, and M. Arens. Instance flow
based online multiple object tracking. In IEEE International
Conference on Image Processing (ICIP), 2017.
[3] F. Chhaya, N. D. Reddy, S. Upadhyay, V. Chari, M. Z. Zia,
and K. M. Krishna. Monocular reconstruction of vehicles:
Combining SLAM with shape priors. In 2016 IEEE International Conference on Robotics and Automation, ICRA
2016, Stockholm, Sweden, May 16-21, 2016, pages 5758–
5765, 2016.
[4] J. Dai, K. He, and J. Sun. Instance-aware semantic segmentation via multi-task network cascades. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2016.
[5] C. Farabet, C. Couprie, L. Najman, and Y. LeCun. Learning
hierarchical features for scene labeling. IEEE Transactions
on Pattern Analysis and Machine Intelligence, August 2013.
[6] M. A. Fischler and R. C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM,
24(6):381–395, June 1981.
[7] A. Gaidon, Q. Wang, Y. Cabon, and E. Vig. Virtual worlds as
proxy for multi-object tracking analysis. In IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2016.
[8] R. I. Hartley and A. Zisserman. Multiple View Geometry
in Computer Vision. Cambridge University Press, ISBN:
0521540518, second edition, 2004.
[9] K. He, G. Gkioxari, P. Dollar, and R. Girshick. Mask r-cnn.
In The IEEE International Conference on Computer Vision
(ICCV), Oct 2017.
[10] E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, and
T. Brox. Flownet 2.0: Evolution of optical flow estimation
with deep networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
[11] A. Kundu, K. M. Krishna, and C. V. Jawahar. Realtime multibody visual slam with a smoothly moving monocular camera. In ICCV, 2011.
[12] K. Lebeda, S. Hadfield, and R. Bowden. 2D or not 2D:
Bridging the gap between tracking and structure from motion. In Proceedings, Asian Conference on Computer Vision
(ACCV), 2014.
[13] B. Lee, K. Daniilidis, and D. D. Lee. Online self-supervised
monocular visual odometry for ground vehicles. In 2015
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
9
IEEE International Conference on Robotics and Automation
(ICRA), pages 5232–5238, May 2015.
Y. Li, H. Qi, J. Dai, X. Ji, and Y. Wei. Fully convolutional
instance-aware semantic segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
July 2017.
P. Moulon, P. Monasse, R. Marlet, and Others. Openmvg. an
open multiple view geometry library., 2013.
K. E. Ozden, K. Cornelis, L. V. Eycken, and L. J. V. Gool.
Reconstructing 3d trajectories of independently moving objects using generic constraints. Computer Vision and Image
Understanding, 96(3):453–471, 2004.
H. S. Park, T. Shiratori, I. Matthews, and Y. Sheikh. 3d trajectory reconstruction under perspective projection. International Journal of Computer Vision, 115(2):115–135, 2015.
S. R. Richter, V. Vineet, S. Roth, and V. Koltun. Playing
for data: Ground truth from computer games. In B. Leibe,
J. Matas, N. Sebe, and M. Welling, editors, European Conference on Computer Vision (ECCV), volume 9906 of LNCS,
pages 102–118. Springer International Publishing, 2016.
G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M.
Lopez. The synthia dataset: A large collection of synthetic
images for semantic segmentation of urban scenes. In 2016
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3234–3243, June 2016.
J. L. Schönberger and J.-M. Frahm. Structure-from-motion
revisited. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
E. Shelhamer, J. Long, and T. Darrell. Fully convolutional
networks for semantic segmentation. IEEE Trans. Pattern
Anal. Mach. Intell., 39(4):640–651, 2017.
N. Snavely, S. M. Seitz, and R. Szeliski. Photo tourism:
Exploring photo collections in 3d. ACM Trans. Graph.,
25(3):835–846, July 2006.
S. Song, M. Chandraker, and C. C. Guest. High accuracy
monocular SFM and scale correction for autonomous driving. IEEE Trans. Pattern Anal. Mach. Intell., 38(4):730–743,
2016.
C. Sweeney. Theia Multiview Geometry Library: Tutorial &
Reference. University of California Santa Barbara., 2014.
A. Tsirikoglou, J. Kronander, M. Wrenninge, and J. Unger.
Procedural modeling and physically based rendering for synthetic data generation in automotive applications. CoRR,
abs/1710.06270, 2017.
S. Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern
Anal. Mach. Intell., 13(4):376–380, 1991.
C. Wu. Visualsfm: A visual structure from motion system,
2011.
C. Yuan and G. G. Medioni. 3d reconstruction of background
and objects moving on ground plane viewed from a moving camera. In 2006 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR 2006), 1722 June 2006, New York, NY, USA, pages 2261–2268, 2006.
| 1cs.CV
|
Simultaneous Lightwave Information and Power
Transfer (SLIPT) for Indoor IoT Applications
Panagiotis D. Diamantoulakis∗ and George K. Karagiannidis∗
arXiv:1706.09624v1 [cs.IT] 29 Jun 2017
∗ Department
of Electrical and Computer Engineering, Aristotle University of Thessaloniki, GR-54124 Thessaloniki, Greece
e-mails: {padiaman, geokarag}@auth.gr
Abstract—We present the concept of Simultaneous Lightwave
Information and Power Transfer (SLIPT) for indoor Internet-ofThings (IoT) applications. Specifically, we propose novel and fundamental SLIPT strategies, which can be implemented through
Visible Light or Infrared communication systems, equipped
with a simple solar panel-based receiver. These strategies are
performed at the transmitter or at the receiver, or at both sides,
named Adjusting transmission, Adjusting reception and Coordinated adjustment of transmission and reception, correspondingly.
Furthermore, we deal with the fundamental trade-off between
harvested energy and quality-of-service (QoS), by maximizing the
harvested energy, while achieving the required user’s QoS. To this
end, two optimization problems are formulated and optimally
solved. Computer simulations validate the optimum solutions
and reveal that the proposed strategies considerably increase the
harvested energy, compared to SLIPT with fixed policies.
I. I NTRODUCTION
The era of Internet-of-Things (IoT) opens up the opportunity
for a number of promising applications in smart buildings,
health monitoring, and predictive maintenance. In the context of wireless access to IoT devices, radio frequency (RF)
technology is the main enabler. Furthermore, the exponential
growth in the data traffic puts tremendous pressure on the
existing global telecommunication networks and the expectations from the fifth generation (5G) of wireless networks.
However, it is remarkable that most of the data consumption/generation, which are related to IoT applications, occurs
in indoor environments [1]. Motivated by this, optical wireless
communication (OWC), such as visible light communications
(VLC) or infrared (IR), have been recognized as promising
alternative/complimentary technologies to RF, in order to give
access to IoT devices in indoor applications [1]. The data rates
reported for indoor VLC/IR networking are much higher than
those achieved by WiFi, especially when client and server
are closely located. Apart from the very high data rates [2],
the advantages of OWC technologies include: i) increase of
available bandwidth, ii) easy bandwidth reuse, iii) increase of
energy efficiency and considerable energy savings, iv) no RF
contamination, and v) free from RF interference. Moreover,
nowadays, light emitting diodes (LEDs) and photodetectors
(PDs) tend to be considerably cheaper than their RF counterparts, while the cost-efficiency is further improved due to the
potential to use the existing lighting infrastructure [3]–[5].
Due to the strong dependence of the IoT on wireless
access, their applications are constrained by the finite battery capacity of the involved devices [6]. Therefore, energy
harvesting (EH), which refers to harnessing energy from the
environment or other sources and converting to electrical
energy, is a critical part of the operation and maintain of the
IoT devices. Energy harvesting is regarded as a disruptive
technological paradigm to prolong the lifetime of energyconstrained wireless networks, which apart from offering a
promising solution for energy-sustainability of wireless nodes,
it also reduces the operating expenses (OPEX) [6]. However,
the main disadvantage of traditional EH methods is that they
rely on natural resources, such as solar and wind, which are
uncontrollable. For this reason, harvesting energy from radio
frequency signals, which also transfer information, seems to be
an interesting alternative. In order to enable simultaneous wireless information and power transfer and increase efficiency
of the utilized resources, two strategies have been proposed
named power-splitting, which is based on the division of the
signals power into two streams, and time-splitting, according
to which, during a portion of time, the received signal is used
solely for energy harvesting, instead of decoding [7].
Although RF based wireless power transfer is a well investigated topic in the last five years, optical wireless power
transfer (OWPT) is a new topic and only a few works have
been reported so far in the open literature. In the pioneering
work of Fakidis et. al. [8], the visible and infra-red parts
of the electromagnetic (EM) spectrum was used for OWPT,
through laser or LEDs at the transmitter and solar cells at
the receiver side. Also, in [9] and [10] energy harvesting
was performed by using the existing lighting fixtures for
indoor IoT applications. Regarding the simultaneous optical
wireless information and power transfer, in [11] the sum rate
maximization problem has been optimized in a downlink VLC
system with simultaneous wireless information and power
transfer. However, in this paper the utilized energy harvesting
model does not correspond to that of the solar panel, where
only the direct current DC component of the modulated light
can be used for energy harvesting, in contrast to the alternating
current AC component, which carries the information. The
separation of the DC and AC components was efficiently
achieved by the self-powered solar panel receiver proposed
in [12], [13], where it was proved that the use of the solar
panel for communication purposes does not limit its energy
harvesting capabilities. Thus, the utilization of the powersplitting in the useful recent work [14], where the received
photocurrent is splitted in two parts with each of them to
include both a DC and a AC part, reduces the EH efficiency.
Moreover, in [14] an oversimplified energy harvesting model
was used, assuming that the harvested energy is linearly proportional to the received optical power, while an optimization
of the splitting technique was not presented. Furthermore, in
the significant research works [15], [16], a dual-hop hybrid
VLC/RF communication system is considered, in order to
extend the coverage. In these papers, besides detecting the
information over the VLC link, the relay is also able to harvest
energy from the first-hop VLC link, by extracting the DC
component of the received optical signal. This energy can
be used to re-transmit the data to a mobile terminal over the
second-hop RF link. Also, in [15] the proposed hybrid system
was optimized, in terms of data rate maximization, while in
[16] the packet loss probability was evaluated.
In this paper, we present for first time a framework for
simultaneous optical wireless information and power transfer,
called from now on as Simultaneous Lightwave Information
and Power Transfer (SLIPT), which can be efficiently used
for indoor IoT applications through VLC or IR systems. More
specifically, we propose novel and fundamental strategies in
order to increase the feasibility and efficiency of SLIPT,
when a solar panel-based receiver is used. These strategies
are performed at the transmitter or at the receiver, or at
both sides, named Adjusting transmission, Adjusting reception,
and Coordinated adjustment of transmission and reception.
Regarding adjusting transmission two policies are proposed:
i) Time-splitting (TS), according to which the time frame is
separated in two distinct phases, where in each of them the
main focus is either on communication or energy transfer
and, ii) Time-splitting with DC bias optimization, which is
a generalization of TS. In contrast to RF-based wireless
powered networks, where the TS strategy and adjustment of
the related parameters takes place at the receiver’s side, TS in
SLIPT refers to the adaptation of specific parameters of the
transmitted signal. Regarding adjusting reception, the Fieldof-view (FoV) adjustment policy is proposed, while according
to the coordinated adjustment of transmission and reception
strategy, we propose the simultaneous optimization of the
former policies at both transmitter and receiver, in order to
maximize the harvested energy, while achieving the required
Quality-of-Service (QoS) (e.g. data rate and signal-to-noise
plus interference ratio (SINR)). Finally, the resulting two
optimization problems are formulated and optimally solved.
II. S YSTEM
AND
C HANNEL MODEL
We consider the downlink transmission of an OWC system,
consisting of one LED and a single user. We also assume
that the user is equipped with the functionality of energy
harvesting. The VLC/IR transmitter/receiver design is shown
in Fig. 1, while the VLC/IR downlink communication is
depicted in Fig. 2.
A. Optical Wireless Transmission
Let m(t) denote the modulated electrical signal that corresponds to the bit stream from the information source. A DC
bias B is added to m(t) to ensure that the resulting signal
Fig. 1. SLIPT transceiver design
Fig. 2. VLC/IR downlink communication
is non-negative, before being used to modulate the optical
intensity of the LED and regulate the LED in the proper
operation mode. The transmitted optical signal from the LED
is [15]
Pt (t) = PLED [B + m(t)],
(1)
where PLED is the LED power. The electrical signal varies
around the DC bias B ∈ [IL , IH ] with peak amplitude A,
where IL is the minimum and IH is the maximum input bias
currents, correspondigly. In order to avoid clipping distortion
by the nonlinearity of the LED, by restraining the input
electrical signal to the LED within the linear region of the
LED operation, the following limitation is induced
A ≤ min(B − IL , IH − B),
(2)
where min(z, y) denotes the minimum between z and y.
B. Channel Model
The channel power gain is given by [17]–[19]
h=
Lr
R0 (ϕ)Ts (ψ)g(ψ) cos(ψ),
d2
(3)
where Lr is the physical area of the photo-detector, d is the
transmission distance from the LED to the illuminated surface
of the photo-detector, Ts (ψ) is the gain of the optical filter
and g(ψ) represents the gain of the optical concentrator, given
by [17], [19]
(
ρ2
, 0 ≤ ψ ≤ Ψfov ,
2 (Ψ
sin
fov )
(4)
g(ψ) =
0, ψ > Ψfov .
is due to the dedicated LED, while I2 is due to different light
sources, e.g. neighboring LEDs. Also, Voc is
IDC
Voc = Vt ln(1 +
),
(12)
I0
where Vt is the thermal voltage and I0 is the dark saturation
current of the PD. Moreover, f is the fill factor, defined as the
ratio of the maximum power from the solar cell to the product
of the open-circuit voltage Voc and Isc ,
with ρ and Ψfov being the refractive index and FOV, respectively. Also in (3), R0 (ϕ) is the Lambertian radiant intensity
of the LED, given by
III. SLIPT S TRATEGIES
In this section we propose fundamental SLIPT strategies
for use in VLC/IR communication systems. These strategies
are performed either at the transmitter or at the receiver, or at
both sides: Adjusting transmission, Adjusting reception, and
Coordinated adjustment of transmission and reception.
ξ+1
cosξ ϕ,
(5)
2π
where ϕ is the irradiance angle, ψ is the incidence angle, and
R0 (ϕ) =
ξ=−
1
,
log2 cos(Φ1/2 )
(6)
with Φ1/2 being the semi-angle at half luminance.
C. Received Electrical SINR
The electrical current ir (t) at the output of the PD can be
written as
ir = η(hPt (t) + Po ) + n(t) = IDC (t) + i(t) + n(t),
(7)
where η is the photo-detector responsivity in A/W, Po is
the received optical signal from other sources, e.g. other
neighboring LEDs, IDC is the DC component, i(t) is the
AC component, and n(t) is the additive white Gaussian noise
(AWGN), which is created from background shot noise and
thermal noise.
The AC component i(t) is composed of two terms, i.e.
i(t) = i1 (t) + i2 (t), where
i1 (t) = ηhPLED m(t)
(8)
is due to the dedicated LED, and i2 (t) is due to other
interfering sources. Thus, the received SINR can be written
as
(ηhPLED A)2
,
(9)
γ=
PI + σ 2
where σ 2 is the noise power and PI is the electrical power of
the received interference.
D. Energy Harvesting Model
As it has already been mentioned the photocurrent consistes
of both the DC and AC signals. In order to perform energy
harvesting, the DC component is blocked by a capacitor
and passes through the energy harvesting branch [12]. The
harvested energy is given by [20]
E = f IDC Voc ,
(10)
with f being the fill factor [20] and IDC = I1 + I2 being the
DC component of the output current, where
I1 = ηhn PLED B
(11)
A. Adjusting Transmission
Next, we introduced two policies for the adjusting transmission strategy, named Time-splitting and Time-splitting with
DC Bias Optimization.
1) Time-splitting: According to the Time-splitting policy
the received optical signal is used for a portion of time solely
for energy harvesting, instead of decoding. During this period
of time the LED transmits by using the maximum DC bias,
in order to maximize the harvested energy by the receiver.
Thus, assuming time frames of unitary duration, there are the
following two distinct phases during a time frame:
Phase 1: The AC component of the received signal is
used for information decoding and the DC component for
energy harvesting. Let A1 and B1 ∈ [IL , IH ] denote the
peak amplitude of m(t) and DC bias, respectively. During
Phase 1, the aim is to maximize the received SINR. Since
SINR is an increasing function with respect to A1 , then A1
takes its maximum value, which, considering (2) is given by
L
L
and similarly, B1 = IH +I
. The duration of
A1 = IH −I
2
2
this phase is denoted by 0 ≤ T ≤ 1, which can be optimized
according to the QoS requirements. For a specific value of T ,
the amount of harvested energy is given by
IH + IL
[1]
+ I2 )Vt
ET S = f T (ηhPLED
2
(13)
IH +IL
ηhPLED 2 + I2
).
× ln(1 +
I0
Phase 2: In the time period 1−T , the aim is to maximize the
harvested energy, which is an increasing function with respect
to B. Thus, during Phase 2 the transmitter eliminates the AC
part and maximizes the DC bias, i.e., A = 0 and B = IH ,
where A2 and B2 ∈ [IL , IH ] denote the values of A and B,
respectively. Thus, the amount of harvested energy during this
phase, is given by
[2]
ET S =f (1 − T )(ηhPLED IH + I2 )Vt ×
(14)
ηhPLED IH + I2
ln(1 +
).
I0
Considering both phases, the total harvested energy is given
by
[1]
[2]
ET S = ET S + ET S .
(15)
2) Time-Splitting with DC Bias Optimization: This policy is
a generalization of Time-splitting. During Phase 1, the DC bias
is optimized in order to increase the harvested energy, while
it simultaneously enables information transfer, i.e., A1 > 0.
In this case, the total harvested energy is given by
ETSBO =f T (ηhPLED B1 + I2 )Vt ×
ηhPLED B1 + I2
[2]
) + ET S ,
ln(1 +
I0
(16)
where B1 is the DC bias during Phase 1.
B. Adjusting Reception
We propose the Adjustment of the field of view (FOV) policy
for the adjusting reception strategy, in order to balance the
trade-off between harvested energy and SINR. Controlling
FOV is particularly important especially when, except for
the used VLC/IR LED, there are extra light sources in the
serving area [21], e.g. neighboring LEDs that serve other
users. For the practical and efficient implementation of this
policy, electrically controllable liquid crystal (LC) lenses is a
promising technology [22].
When the aim is to maximize the SINR, the FOV is tuned
up to receive the beam of the dedicated LED only (if possible),
in order to reduce the beam overlapping. This is achieved by
tuning the FOV to the narrowest setting, that allows reception
only from that LED. On the other hand, when the aim is to
achieve a balance between SINR and harvested energy, a wider
FOV setting could be selected.
For the sake of practicality, we assume that the VLC/IR re[1]
[M]
ceiver has discrete FOV settings, i.e. Ψfov ∈ {Ψfov , .., Ψfov }.
Also, note that except for h, both PI and I2 are also discrete
functions of Ψfov , i.e., PI = PI (Ψfov ) and I2 = I2 (Ψfov ).
IV. SLIPT O PTIMIZATION
SLIPT induces an interesting trade-off between harvested
energy and QoS. In this section, we aim to balance this tradeoff by maximizing the harvested energy, while achieving the
required user QoS. In the present work, we focus on the
coordinated adjustment of transmission and reception strategy,
which can be considered as a generalization of the other
SLIPT strategies. The following optimization problems can
be formulated, based on the two techniques presented in
subsection III-C.
Regarding the QoS, two different criteria are taken into
account, namely SINR and information rate. Note that these
two criteria are not equivalent to each other, when either
of the two techniques is used, due to the time-splitting.
More specifically, since only Phase 1 is used for information
transmission (the duration of which is T ), the lower bound of
the capacity is given by [23]
e
γ .
(17)
R = T log2 1 +
2π
A. Time-Splitting with Tunable FOV
The corresponding optimization problem can be expressed
as
max
T,Ψfov,1
s.t.
EhTS
C1
C2
C3
C4
: R ≥ Rth ,
: γ ≥ γth ,
: 0 ≤ T ≤ 1,
[1]
[M]
: Ψfov,1 ∈ {Ψfov , .., Ψfov },
(18)
where Rth and γth denote the information rate SINR and
threshold, respectively.
Theorem 1: The optimal value of T in (18) is given by
Rth
,
T∗ =
(19)
e(ηhPLED (IH −IL ))2
log2 1 + 8π(PI (Ψ∗ )+σ2 )
fov,1
C. Coordinated Transmission and Reception Adjustment
Considering (9), (15), and (16), it is revealed that both
SINR and harvested energy -apart from A1 , B1 and T also depend on the selection of Ψfov , despite the utilized
adjusting transmission technique. This dependence motivates
the coordinated transmission and reception adjustment, i.e. the
coordination between the strategy III-A1 or III-A2 and III-B,
which results in the following two policies, i.e.
•
•
Policy 1: Time-splitting with tunable FOV (III-A1 and
III-B)
Policy 2: Time-splitting with DC bias optimization and
tunable FOV (III-A2 and III-B)
Note that in both policies, during Phase 2, where the aim
is to maximize the harvested energy, the FOV setting that
[2]
maximizes ET S should be used. This is not necessarily the
widest setting, because although it increases the received
beams (if there are neighboring LEDs), it reduces g(ψ). On
the other hand, the preferable FOV setting during phase 1,
denoted by Ψfov,1 , cannot be straightforwardly determined,
since it also depends on the required QoS.
where (·)∗ denotes optimality.
Proof: The optimization problem (18) is a combinatorial
one. In order to find the optimal solution, all possible values
of Ψfov,1 have to be checked before selecting the value that
maximizes the harvested energy, EhTS , while satisfying the
constraints C1 , C2 , and C3 . For a specific specific value of
Ψfov,1 , if
L 2
(ηhPLED IH −I
)
2
< γth ,
(20)
PI (Ψfov,1 ) + σ 2
then the optimization problem is infeasible, since C2 is not
satisfied. Also, due to constraint C1 , the following limitation
is induced for T ,
Rth
.
(21)
T ≥
I −I
e(ηhPLED H 2 L )2
log2 1 + 2π(PI (Ψfov,1 )+σ2 )
Moreover, the harvested energy is decreasing with respect to
T . Thus, the optimal value of T is given by (19) and the proof
is completed.
Note that if T ∗ > 1, the optimization problem in (18) is
infeasible, due to C3 .
B. Time-Splitting with DC Bias Optimization and tunable FOV
The corresponding optimization problem can be formulated
as
Proof: Considering Proposition 1 and for a specific
value of Ψfov,1 the optimization problem in (22) can be reformulated as
EhTSBO
max
B1 ,A1 ,T
C1
C3
C4
C5
C7
s.t.
: R ≥ Rth , C2 : γ ≥ γth ,
: A1 ≤ min(B1 − IL , Ih − B1 ),
: 0 ≤ T ≤ 1,
: A1 ≥ 0, C6 : IL ≤ B1 ≤ IH ,
[1]
[M]
: Ψfov,1 ∈ {Ψfov , .., Ψfov }.
(22)
Proposition
1: The optimal value of B in (22) belongs in
L
the range IH +I
, IH .
2
Proof: The constraint C3 can be rewritten as
C3a : A1 ≤ B1 − IL , C3b : A1 ≤ IH − B1 .
The optimal value Ψfov,1 is calculated similarly to the
solution of (18). Regarding the rest optimization variables of
(22) they are optimized according to the following theorem:
Theorem 2: For a specific value of Ψfov,1 , the optimal value
of T is given by
T ∗ = argmax ẼhTSBO
(24)
K1 ≤T ≤K2
with ẼhTSBO being solely a function of T and given by (16),
by replacing A1 and B1 by
s
Rth
2π(PI (Ψfov,1 ) + σ 2 )(2 T − 1)
1
, (25)
A1 =
nhPLED
e
and
B1 = IH − A1 ,
(26)
respectively. Also,
Rth
log2 1 +
K2 = min
e(ηh1 PLED (IH −IL ))2
2
8π (PI (Ψ∗
fov,1 )+σ )
log2
Rth
1+
!
eγth , 1 .
2π
(29)
The optimization problem in (29) still cannot be easily
solved in its current form, since the objective function as well
as the constraints C1 and C2 are not concave. However, it
can be solved with low complexity by using the following
reformulation.
First, the inequalities in C1 and C3 are replaced by equalities. Then, A1 and B1 are given by (25) and (26), respectively.
By substituting T1 and B1 by (25) and (26), C1 , C3 , and C3
of (29) vanish, and the optimization problem is rewritten as
(23)
For a specific value of B1 , only one of the constraints C3a and
C3b is activated. Now, let assume that the optimal solution is
L
, for which all the constraints are satisfied. In
B1∗ < IH +I
2
L
this case, C3a is activated. However, by setting B1 = IH +I
2
the objective function is increased, while the constraints are
still satisfied. Thus, we can infer that that B1∗ is not optimal.
Consequently, Proposition 1 has been proved by contradiction.
K1 =
C1 : R ≥ Rth , C2 : γ ≥ γth ,
C3 : A1 + B1 ≤ IH , C4 : 0 ≤ T ≤ 1,
L
.
C5 : A1 ≥ 0, C6 : B1 ≥ IH +I
2
s.t.
EhTSBO
max
B1 ,A1 ,T,Ψfov,1
max
B1 ,A,T1 ∀n
s.t.
ẼhTSBO
C2 : T ≤
Rth
eγth ,
log2 (1+ 2π
)
C4 : 0 ≤ T ≤ 1,
C6 : T ≥
log2 1+
Rth
e(ηh1 PLED (IH −IL ))2
8π (PI +σ2 )
(30)
,
which is equivalent to (24), and, thus, the proof is completed.
V. S IMULATIONS
AND DISCUSSION
We assume the downlink VLC/IR system of Fig. 2, where
the user is located in a distance d = 1.5 m from the LED,
ψ = 0, and the transmitter plane is parallel to the receiver one,
i.e., ϕ = ψ. In the same room there are N other LEDs, which
simultaneously use the same frequency band. The distance
between each of them and from the dedicated LED is D = 1.5
m. We also assume f = 0.75, PLED = 20 W/A, Φ1/2 = 60
deg, σ 2 = 10−15 A2 , Lr = 0.04 m2 , η = 0.4 A/W, I0 = 10−9
A, IL = 0 A, IH = 12 mA [15], Ts = 1, ρ = 1.5, γth = 10
dB, and two settings for the FOV, i.e., Ψfov ∈ {30, 50} deg,
are considered.
Regarding the neighboring LEDs, we assume that the DC
bias and the peak amplitude are given by A′n = Bn′ = 6 mA,
∀n ∈ {1, ..., N }, while the rest parameters are equal to those
of the dedicated LED. Furthermore, the channel between them
and the user’s receiver, denoted by hn is modeled according
to (3), using the corresponding parameters. Thus, when the
widest FOV setting is selected, PI and I2 are given by
(27)
PI =
N
X
(ηhn PLED A′n )2
(31)
n=1
and
(28)
Finally, the optimal values of A1 and B1 are given by (25) and
(26), by replacing Ψfov,1 and T by Ψ∗fov,1 and T ∗ , respectively.
I2 =
N
X
ηhn PLED Bn′ ,
(32)
n=1
otherwise their values are zero.
The performance of both optimized policies of Section III-C
are compared for N = 1, while they are also presented against
energy increases with the increase of LEDs.
2.2
Policy 2
2.0
R EFERENCES
Policy 1
A =B =6 mA, T =1,
A =B =6 mA, T =1,
Harvested Energy (mW)
1.8
1.6
1
1
1
fov,1
1
1
1
fov,1
=30 deg
=50 deg
1.4
1.2
1.0
0.8
0.6
0.4
Infeasible for these values of
R
th
0.2
2
4
6
8
R
th
10
12
14
(bps/Hz)
Fig. 3. Harvested energy vs Rth for N = 1.
2.2
Policy 2
Policy 1
2.0
A =B =6 mA, T =1,
Harvested Energy (mW)
1.8
1
1
1
fov,1
=30 deg
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
2
4
6
8
10
12
14
16
18
20
N
Fig. 4. Harvested energy vs N for Rth = 7 bps/Hz.
the case of fixed A1 , B1 , T1 , and Ψfov,1 , which is considered
as the baseline policy. More specifically, in Fig. 3 the harvested
energy is plotted against the rate threshold. As it is observed,
both policies significantly outperform the baseline for both
values of Ψfov,1 . Regarding the baseline, the value Ψfov,1 = 50
deg reduces the harvested energy compared to Ψfov,1 = 30
deg, because g(ψ) decreases and thus, cancels the benefit
of receiving the beam of the neighboring LED. Also, the
baseline policy with Ψfov,1 = 50 deg is infeasible for medium
and high values of Rth , because the rate threshold cannot be
reached, due to the received interference. Interestingly, Policy
2 outperforms Policy 1, especially for the high region of
Rth , which is due to the extra degrees of freedom. Similar
conclusions can be obtained by Fig. 4, where the harvested
energy is plotted against the number of neighboring LEDS.
For this specific setup, the baseline with Ψfov,1 = 50 deg is not
feasible, and, thus, it is omitted. We notice here that for a small
number of neighboring LEDs, the harvested energy remains
constant, since the receiver prefers the smallest FOV setting.
However, as the number of neighboring LEDs increases, the
receiver prefers the widest FOV setting and the harvested
[1] M. Ayyash, H. Elgala, A. Khreishah, V. Jungnickel, T. Little, S. Shao,
M. Rahaim, D. Schulz, J. Hilt, and R. Freund, “Coexistence of WiFi
and LiFi toward 5G: Concepts, Opportunities, and Challenges,” IEEE
Commun. Mag., vol. 54, no. 2, pp. 64–71, Feb. 2016.
[2] M. S. A. Mossaad, S. Hranilovic, and L. Lampe, “Visible Light Communications Using OFDM and Multiple LEDs,” IEEE Trans. Commun.,
vol. 63, no. 11, pp. 4304–4313, Nov. 2015.
[3] M. Kavehrad, “Sustainable Energy-Efficient Wireless Applications Using Light,” IEEE Commun. Mag., vol. 48, no. 12, pp. 66–73, Dec. 2010.
[4] S. Arnon, Ed., Visible Light Communication. Cambridge University
Press, 2015.
[5] D. Bykhovsky and S. Arnon, “Multiple Access Resource Allocation in
Visible Light Communication Systems,” J. Lightw. Technol., vol. 32,
no. 8, pp. 1594–1600, Apr. 2014.
[6] S. Sudevalayam and P. Kulkarni, “Energy Harvesting Sensor Nodes:
Survey and Implications,” IEEE Commun. Surveys Tuts., vol. 13, no. 3,
pp. 443–461, Jul. 2010.
[7] S. Nikoletseas, Y. Yang, and A. Georgiadis, Eds., Wireless Power
Transfer Algorithms, Technologies and Applications in Ad Hoc Communication Networks. Springer, 2016.
[8] J. Fakidis, S. Videv, S. Kucera, H. Claussen, and H. Haas, “Indoor
Optical Wireless Power Transfer to Small Cells at Nighttime,” J. Lightw.
Technol., vol. 34, no. 13, pp. 3236–3258, Jul. 2016.
[9] C. Carvalho and N. Paulino, “On the Feasibility of Indoor Light
Energy Harvesting for Wireless Sensor Networks,” Procedia Technology,
vol. 17, pp. 343–350, 2014.
[10] A. Nasiri, S. A. Zabalawi, and G. Mandic, “Indoor Power Harvesting
using Photovoltaic Cells for Low-Power Applications,” IEEE Trans. Ind.
Electron., vol. 56, no. 11, pp. 4502–4509, Nov. 2009.
[11] Y. Li, N. Huang, J. Wang, Z. Yang, and W. Xu, “Sum Rate Maximization
for VLC Systems with Simultaneous Wireless Information and Power
Transfer,” IEEE Photon. Technol. Lett., vol. 29, no. 6, pp. 531–534,
Mar. 2017.
[12] Z. Wang, D. Tsonev, S. Videv, and H. Haas, “Towards Self-Powered
Solar Panel Receiver for Optical Wireless Communication,” in Proc.
IEEE International Conference on Communications (ICC), Jun. 2014,
pp. 3348–3353.
[13] ——, “On the Design of a Solar-Panel Receiver for Optical Wireless
Communications with Simultaneous Energy Harvesting,” IEEE J. Sel.
Areas Commun., vol. 33, no. 8, pp. 1612–1623, Aug. 2015.
[14] H. Sandalidis, A. Vavoulas, T. Tsiftsis, and N. Vaiopoulos, “Illumination,
Data Transmission and Energy Harvesting: The Threefold Advantage of
VLC,” 2017.
[15] T. Rakia, H. C. Yang, F. Gebali, and M. S. Alouini, “Optimal Design of
Dual-Hop VLC/RF Communication System with Energy Harvesting,”
IEEE Comm. Lett., vol. 20, no. 10, pp. 1979–1982, Oct. 2016.
[16] ——, “Dual-Hop VLC/RF Transmission System with Energy Harvesting
Relay under Delay Constraint,” in Proc. IEEE Globecom Workshops,
Dec. 2016, pp. 1–6.
[17] T. Komine and M. Nakagawa, “Fundamental Analysis for Visible-Light
Communication System Using LED Lights,” IEEE Trans. Consum.
Electron., vol. 50, no. 1, pp. 100–107, Feb. 2004.
[18] H. Ma, L. Lampe, and S. Hranilovic, “Coordinated Broadcasting for
Multiuser Indoor Visible Light Communication Systems,” IEEE Trans.
Commun., vol. 63, no. 9, pp. 3313–3324, Sep. 2015.
[19] J. M. Kahn and J. R. Barry, “Wireless Infrared Communications,” Proc.
IEEE, vol. 85, no. 2, pp. 265–298, Feb. 1997.
[20] C. Li, W. Jia, Q. Tao, and M. Sun, “Solar Cell Phone Charger Performance in Indoor Environment,” in Proc. IEEE 37th Annual Northeast
Bioengineering Conference (NEBEC), 2011, pp. 1–2.
[21] A. Al-Ghamdi and J. Elmirghani, “Performance and Field of View Optimization of an Optical Wireless Pyramidal Fly-Eye Diversity Receiver,”
Journal of optical communications, vol. 23, no. 6, pp. 215–222, 2002.
[22] Y. S. Tsou, Y. H. Lin, and A. C. Wei, “Concentrating Photovoltaic
System using a Liquid Crystal Lens,” IEEE Photon. Technol. Lett.,
vol. 24, no. 24, pp. 2239–2242, Dec. 2012.
[23] J. B. Wang, Q. S. Hu, J. Wang, M. Chen, and J. Y. Wang, “Tight Bounds
on Channel Capacity for Dimmable Visible Light Communications,” J.
Lightw. Technol., vol. 31, no. 23, pp. 3771–3779, Dec. 2013.
| 7cs.IT
|
Logical Methods in Computer Science
Vol. 13(2:1)2017, pp. 1–57
www.lmcs-online.org
Submitted
Published
Jan. 15, 2016
Apr. 7, 2017
DYNAMIC CHOREOGRAPHIES: THEORY AND IMPLEMENTATION
MILA DALLA PREDA a , MAURIZIO GABBRIELLI b , SAVERIO GIALLORENZO c , IVAN LANESE d ,
AND JACOPO MAURO e
a
b,c,d
e
Department of Computer Science, University of Verona
e-mail address: [email protected]
Department of Computer Science and Engineering, University of Bologna/INRIA
e-mail address: {gabbri, sgiallor, lanese}@cs.unibo.it
Department of Informatics, University of Oslo
e-mail address: [email protected]
Abstract. Programming distributed applications free from communication deadlocks and
race conditions is complex. Preserving these properties when applications are updated at
runtime is even harder. We present a choreographic approach for programming updatable,
distributed applications. We define a choreography language, called Dynamic InteractionOriented Choreography (DIOC), that allows the programmer to specify, from a global
viewpoint, which parts of the application can be updated. At runtime, these parts may
be replaced by new DIOC fragments from outside the application. DIOC programs are
compiled, generating code for each participant in a process-level language called Dynamic
Process-Oriented Choreographies (DPOC). We prove that DPOC distributed applications
generated from DIOC specifications are deadlock free and race free and that these properties
hold also after any runtime update. We instantiate the theoretical model above into a
programming framework called Adaptable Interaction-Oriented Choreographies in Jolie
(AIOCJ) that comprises an integrated development environment, a compiler from an
extension of DIOCs to distributed Jolie programs, and a runtime environment to support
their execution.
1. Introduction
Programming distributed applications is an error-prone activity. Participants send and
receive messages but, if the application is badly programmed, they may get stuck waiting
for messages that never arrive (communication deadlock), or they may receive messages in
an unexpected order, depending on the speed of the other participants and of the network
(races).
Key words and phrases: Choreographies, Adaptable Systems, Deadlock Freedom.
Supported by the EU EIT Digital project SMAll .
d
Supported by the GNCS group of INdAM via project Logica, Automi e Giochi per Sistemi Auto-adattivi.
e
Supported by the EU project FP7-644298 HyVar: Scalable Hybrid Variability for Distributed, Evolving
Software Systems.
c
l
LOGICAL METHODS
IN COMPUTER SCIENCE
©
DOI:10.23638/LMCS-13(2:1)2017
CC
M. Dalla Preda, M. Gabbrielli, S. Giallorenzo, I. Lanese, and J. Mauro
Creative Commons
2
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Recently, language-based approaches have been proposed to tackle the complexity of
programming concurrent and distributed applications. These approaches provide high-level
primitives that avoid by construction some of the risks of concurrent programming. Some
renowned examples are the ownership/borrowing mechanisms of Rust [44] and the separate
keyword of SCOOP [41]. In these settings most of the work needed to ensure a correct
behaviour is done by the language compiler and runtime support. The use of these languages
requires a conceptual shift from traditional ones, but reduces times and costs of development,
testing, and maintenance by avoiding some of the most common programming errors.
We propose an approach based on choreographic programming [12, 13, 30, 47, 34] following
a similar philosophy. A choreography is a global description of a distributed application
expressed as the composition of the expected interactions between its components. Based
on such an abstraction, we present high-level primitives that avoid some of the main errors
linked to programming communication-centred concurrent and distributed applications that
can be updated at run time.
For an informal introduction of our approach, let us consider a simple application where
a Buyer asks the price of some product to a Seller.
priceReq : Buyer(prod) → Seller(order);
2 order price@Seller = getPrice(order);
3 offer : Seller(order price) → Buyer(prod price)
1
A unique choreographic program describes the behaviour of multiple participants, here
the Buyer and the Seller. The first line specifies that the Buyer sends, along channel
priceReq, the name of the desired product prod to the Seller, which stores it in its local
variable order. At Line 2, the Seller computes the price of the product with function
getPrice. Finally, at Line 3 the Seller sends the computed price to the Buyer on channel
offer . The Buyer stores the value in its local variable prod price.
Choreographic languages focus on describing message-based interactions between distributed participants. As shown by the example, the distinctive feature of choreographies is
that communications are atomic entities, i.e., not split into send and receive actions. This
makes impossible to write programs with common errors of concurrent and distributed applications like deadlocks or race conditions. However, a choreographic description is not directly
executable. Indeed, choreographies describe atomic interactions from a global point of view
whilst executable programs are written in lower level languages (like Java or C) that support
only local behaviours and communication/synchronisation primitives such as message send
and message receive. Hence, to run a choreographic description, we have to compile it into a
set of lower level programs. The correct compilation of choreographic descriptions is one of
the main challenges of choreographic languages, yet the ease of development and the strong
guarantees of deadlock and race freedom make it a worth endeavour.
In this work, we take this challenge one step further: we consider updatable distributed
applications whose code can change dynamically, i.e., while the application is running. In
particular, the applications we consider can integrate external code at runtime. Such a
feature, tricky in a sequential setting and even more in a distributed one, has countless
uses: to deal with emergency requirements, to cope with rules and requirements depending
on contextual properties or to improve and specialise the application to user preferences.
We propose a general mechanism to structure application updates. Inside applications, we
delimit blocks of code, called scopes, that may be dynamically replaced by new blocks of
DYNAMIC CHOREOGRAPHIES
3
code, called updates. Updates are loosely related to scopes: it is not necessary to know the
details of the behaviour of updates when writing scopes, and updates may be written while
applications are running.
To illustrate our approach, let us consider the previous example and let us suppose that
we would like to introduce new commercial offers at runtime, e.g., to provide a discount
on the computed prices. Since we want to be able to change how prices are computed, we
enclose Lines 2–3 of the example within a scope, as shown below.
priceReq : Buyer(prod) → Seller(order);
scope @Seller{
3
order price@Seller = getPrice(order);
4
offer : Seller(order price) → Buyer(prod price)
5 }
1
2
In essence, a scope is a delimiter that defines which part of the application can be updated.
Each scope identifies a coordinator of the update, i.e., the participant entitled to ensure that
either none of the participants updates, or they all apply the same update. In the example
above, the coordinator of the update is the Seller (Line 2).
Since now the code includes a scope, we can introduce runtime updates. Let us assume
that the Seller issued a fidelity card to some frequent customers and (s)he wants to update
the application to let Buyers insert their fidelity card to get a discount. The update in
Figure 1 answers this business need. At runtime, if the Seller (which is the coordinator of
the update) applies the update in Figure 1, the code of the update replaces the scope. When
this new code executes, the Buyer sends his/her card id to the Seller. If the card id is
valid, the Seller issues a 10% discount on the price of the selected good, otherwise it reports
the standard price to the Buyer.
We remark that expressing the behaviour described above using lower level languages
that lack a dedicated support for distributed runtime updates (e.g., Java, C, etc.) is extremely
error prone. For instance, in the example above, let us consider the case in which the Buyer
is updated first and it starts the execution whilst the Seller has not been updated yet. The
Buyer waits for a message from the Seller on channel cardReq, whilst the Seller sends a
message on channel offer . Thus, the application is deadlocked. Furthermore, in our setting
the available updates may change at any time, posing an additional challenge. To avoid
1
2
3
4
5
6
7
8
9
cardReq : Seller(null) → Buyer( );
card id@Buyer = getInput();
cardRes : Buyer(card id) → Seller(buyer id);
if isValid(buyer id)@Seller {
order price@Seller = getPrice(order) ∗ 0.9
} else {
order price@Seller = getPrice(order)
};
offer : Seller(order price) → Buyer(prod price)
Figure 1: Fidelity Card Update.
4
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
errors, participants must select the same update and, since updates appear and disappear at
runtime, they must also be able to retrieve the same selected update.
Since at choreographic level updates are applied atomically to all the involved participants, these problems cannot arise if both the original application and the updates are
described as choreographies. However, to execute our choreographic specifications we need
to provide a correct compilation from choreographies to lower level, executable programs.
Such task is very challenging and, in particular, we need to make sure that our compilation
generates correct behaviours that avoid inconsistencies on updates.
Contributions. In this paper, we present a correctness-by-construction approach to solve the
problem of dynamic updates of distributed applications. We provide a general solution that
comprises:
• the definition of a Dynamic Interaction-Oriented Choreography language, called DIOC, to
program distributed applications and supporting runtime code update (§ 2);
• the definition of a Dynamic Process-Oriented Choreography language, called DPOC.
DPOCs are based on standard send and receive primitives but they are tailored to program
updatable systems. We introduce DPOCs to describe implementations corresponding to
DIOCs. (§ 3);
• the definition of a behaviour-preserving projection function to compile DIOCs into DPOCs
(§ 4);
• the proof of the correctness of the projection function (§ 7). Correctness is guaranteed
even in a scenario where the set of available updates dynamically changes, at any moment
and without notice;
• one instantiation of our theory into a development framework for adaptive distributed
applications called AIOCJ (§ 8). In AIOCJ updates are embodied into adaptation rules,
whose application is not purely non-deterministic (as in DIOCs), but depends on the
state of the system and of its environment. AIOCJ comprises an Integrated Development
Environment, a compiler from choreographies to executable programs, and a runtime
environment to support their execution and update.
This paper integrates and extends material from [18], which outlines the theoretical aspects,
and [19], which describes the AIOCJ framework. Main extensions include the full semantics
of DIOC and DPOC, detailed proofs and technical definitions, and a thorough description
of the example. Furthermore, both the presentation and the technical development have
been deeply revised and refined.
2. Dynamic Interaction-Oriented Choreographies
In this section we introduce the syntax of DIOCs, we illustrate the constructs of the DIOC
language with a comprehensive example, and we finally present the semantics of DIOCs.
2.1. DIOC Syntax. DIOCs rely on a set of Roles, ranged over by R, S, . . . , to identify the
participants in the choreography. We call them roles to highlight that they have a specific
duty in the choreography. Each role has its local state.
Roles exchange messages over public channels, also called operations, ranged over by o.
We denote with Expr the set of expressions, ranged over by e. We deliberately do not give a
formal definition of expressions and of their typing, since our results do not depend on it.
We only require that expressions include at least values, belonging to a set Val ranged over
DYNAMIC CHOREOGRAPHIES
5
by v, and variables, belonging to a set Var ranged over by x, y, . . . . We also assume a set of
boolean expressions ranged over by b.
The syntax of DIOC processes, ranged over by I, I 0 , . . ., is defined as follows:
I ::=
|
|
|
|
o : R(e) → S(x)
I; I 0
I|I 0
x@R = e
1
(interaction)
(sequence)
(parallel)
(assignment)
(inaction)
|
|
|
|
0
if b@R {I} else {I 0 }
while b@R {I}
scope @R {I}
(end)
(conditional)
(while)
(scope)
Interaction o : R(e) → S(x) means that role R sends a message on operation o to role S (we
require R 6= S). The sent value is obtained by evaluating expression e in the local state of R
and it is then stored in the local variable x of S. Processes I; I 0 and I|I 0 denote sequential
and parallel composition, respectively. Assignment x@R = e assigns the evaluation of
expression e in the local state of R to its local variable x. The empty process 1 defines
a DIOC that can only terminate. 0 represents a terminated DIOC. It is needed for the
definition of the operational semantics and it is not intended to be used by the programmer.
We call initial a DIOC process where 0 never occurs. The conditional if b@R {I} else {I 0 }
and the iteration while b@R {I} are guarded by the evaluation of the boolean expression
b in the local state of R.The construct scope @R {I} delimits a subterm I of the DIOC
process that may be updated in the future. In scope @R {I}, role R is the coordinator
of the update: it ensures that either none of the participants update, or they all apply the
same update.
A Running Example. We report in Figure 2 a running example of a DIOC process that
extends the one presented in the Introduction: the example features a Buyer that orders
a product from a Seller, and a Bank that supports the payment from the Buyer to the
Seller. The DIOC process describes the behaviour of all of them. In this sense, the DIOC
implements a protocol they agreed upon to integrate their business. The DIOC protocol
also interacts with functionalities that are only available at the premises of some of the
participants (for instance, the function getPrice is provided by the Seller IT system and
may be only locally available). For this reason, it is important that the execution of the
DIOC is distributed across the participants.
At Lines 1–2 the Buyer initialises its local variables price ok and continue. These
variables control the while loop used by the Buyer to ask the price of some product to the
Seller. The loop execution is controlled by the Buyer, but it impacts also the behaviour
of the Seller. We will see in § 4 that this is done using automatically generated auxiliary
communications. A similar approach is used to manage conditionals. At Line 4, the
Buyer takes the name of the product from the user with function getInput, which models
interaction with the user, and proceeds to send it to the Seller on operation priceReq (Line
5). The Seller computes the price of the product calling the function getPrice (Line 7)
and, via operation offer , it sends the price to the Buyer (Line 8), that stores it in a local
variable prod price. These last two operations are performed within a scope and therefore
they can be updated at runtime to implement some new business policies (e.g., discounts).
At Lines 10–12 the Buyer checks whether the user is willing to buy the product, and, if
(s)he is not interested, whether (s)he wants to ask prices for other products. If the Buyer
accepts the offer of the Seller, the Seller sends to the Bank the payment details (Line 16),
6
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
price ok@Buyer = false;
continue@Buyer = true;
while( !price ok and continue )@Buyer{
prod@Buyer = getInput();
priceReq : Buyer(prod) → Seller(order);
scope @Seller{
order price@Seller = getPrice(order);
offer : Seller(order price) → Buyer(prod price)
};
price ok@Buyer = getInput();
if ( !price ok )@Buyer{
continue@Buyer = getInput()
}
};
if ( price ok )@Buyer{
payReq : Seller( payDesc(order price) ) → Bank(desc);
scope @Bank{
pay : Buyer( payAuth(prod price) ) → Bank(auth)
};
payment ok@Bank = makePayment(desc, auth);
if ( payment ok )@Bank{
conf irm : Bank(null) → Seller( )
|
conf irm : Bank(null) → Buyer( )
} else {
abort : Bank(null) → Buyer( )
}
}
Figure 2: DIOC process for Purchase Scenario.
computed using function payDesc. Next, the Buyer authorises the payment via operation
pay, computing the payment authorisation form using function payAuth. Since the payment
may be critical for security reasons, the related communication is enclosed in a scope (Lines
17–19), thus allowing the introduction of more refined procedures later on. After the scope
successfully terminates, at Line 20 the Bank locally executes the actual payment by calling
function makePayment. The function makePayment abstracts an invocation to the Bank IT
system. We show in § 8 that, using functions, one can integrate existing service-oriented
software into a choreographic program.
Finally, the Bank acknowledges the payment to the Seller and the Buyer in parallel
(Lines 22–24). If the payment is not successful, the failure is notified explicitly only to the
Buyer. Note that at Lines 1–2 the annotation @Buyer means that the variables price ok
and continue belong to the Buyer. Similarly, at Line 3, the annotation @Buyer means
that the guard of the while loop is evaluated by the Buyer. The term @Seller at Line 6 is
part of the scope construct and indicates the Seller as coordinator of the update.
DYNAMIC CHOREOGRAPHIES
7
2.2. Annotated DIOCs and their Semantics. In the remainder of the paper, we define
our results on an annotated version of the DIOC syntax.Annotations are numerical indexes
i ∈ N assigned to DIOC constructs. We only require indexes to be distinct. Any annotation
that satisfies this requirement provides the same result. Indeed, programmers do not need
to annotate DIOCs: the annotation with indexes is mechanisable and can be performed
by the language compiler1. Indexes are used both in the proof of our results and in the
projection to avoid interferences between different constructs. From now on we consider
only well-annotated DIOCs, defined as follows.
Definition 2.1 (Well-annotated DIOC). Annotated DIOC processes are obtained by
indexing every interaction, assignment, conditional, while loop, and scope in a DIOC process
with a positive natural number i ∈ N, resulting in the following grammar:
I ::=
|
|
|
|
i: o : R(e) → S(x)
I; I 0
I|I 0
i: x@R = e
1
|
|
|
|
0
i: if b@R {I} else {I 0 }
i: while b@R {I}
i: scope @R {I}
A DIOC process is well annotated if all its indexes are distinct.
DIOC processes do not execute in isolation: they are equipped with a global state Σ
and a set of available updates I, i.e., a set of DIOCs that may replace scopes. Set I may
change at runtime. A global state Σ is a map that defines the value v of each variable x
in a given role R, namely Σ : Roles × Var → Val . The local state of role R is denoted as
ΣR : Var → Val and it verifies that ∀x ∈ Var : Σ(R, x) = ΣR (x). Expressions are always
evaluated by a given role R: we denote the evaluation of expression e in local state ΣR as
[[e]]ΣR . We assume that [[e]]ΣR is always defined (e.g., an error value is given as a result if
evaluation is not possible) and that for each boolean expression b, [[b]]ΣR is either true or
false.
Remark 2.2. The above assumption on expressions is needed for our results. To satisfy
it, when programming, one should prevent runtime errors and notify abnormal events to
partners using normal constructs to guarantee error management or smooth termination
(see e.g., Lines 21–26 in Figure 2). A more elegant way to deal with errors would be to
include in the language well-known constructs, such as try-catch, which are however left as
future work. This could be done by adapting the ideas presented in [9].
Definition 2.3 (DIOC systems). A DIOC system is a triple hΣ, I, Ii denoting a DIOC
process I equipped with a global state Σ and a set of updates I.
Definition 2.4 (DIOC systems semantics). The semantics of DIOC systems is defined
as the smallest labelled transition system (LTS) closed under the rules in Figure 3, where
symmetric rules for parallel composition have been omitted.
The rules in Figure 3 describe the behaviour of a DIOC system by induction on the
structure of its DIOC process, with a case analysis on its topmost operator. We use µ to
range over labels. The possible values for µ are described below.
µ ::=
|
|
o : R(v) → S(x) (interaction)
I
(update)
I
(change updates)
1In fact, the AIOCJ compiler implements such a feature.
|
|
|
τ
(silent)
no-up
(no update)
√
(termination)
8
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
[[e]]ΣR = v
bDIOC |Interaction e
o:R(v)→S(x)
hΣ, I, i: o : R(e) → S(x)i −−−−−−−−→ hΣ, I, i: x@S = vi
[[e]]ΣR = v
bDIOC |Assign e
τ
hΣ, I, i: x@R = ei −
→ hΣ[v/x, R], I, 1i
√
µ
hΣ, I, Ii −
→ hΣ, I0 , I 0 i µ 6=
bDIOC |Sequence e
µ
hΣ, I, I; J i −
→ hΣ, I0 , I 0 ; J i
√
µ
hΣ, I, Ii −
→ hΣ, I, I 0 i
hΣ, I, J i −
→ hΣ, I, J 0 i
µ
hΣ, I, I; J i −
→ hΣ, I, J 0 i
µ
hΣ, I, Ii −
→ hΣ, I0 , I 0 i µ 6=
µ
hΣ, I, I k J i −
→
hΣ, I0 , I 0
√
k Ji
bDIOC |Parallel e
√
√
hΣ, I, Ii −
→ hΣ, I, I 0 i
hΣ, I, J i −
→ hΣ, I, J 0 i
√
hΣ, I, I k J i −
→
hΣ, I, I 0
k
bDIOC |Seq-end e
J 0i
[[b]]ΣR = true
bDIOC |Par-end e
τ
bDIOC |If-then e
τ
bDIOC |If-else e
hΣ, I, i: if b@R {I} else {I 0 }i −
→ hΣ, I, Ii
[[b]]ΣR = false
hΣ, I, i: if b@R {I} else {I 0 }i −
→ hΣ, I, I 0 i
[[b]]ΣR = true
τ
hΣ, I, i: while b@R {I}i −
→ hΣ, I, I; i: while b@R {I}i
[[b]]ΣR = false
τ
hΣ, I, i: while b@R {I}i −
→ hΣ, I, 1i
roles(I 0 ) ⊆ roles(I) I 0 ∈ I
bDIOC |While-exit e
connected(I 0 ) freshIndexes(I 0 )
I0
hΣ, I, i: scope @R {I}i −→
hΣ, I, I 0 i
no-up
hΣ, I, i: scope @R {I}i −−−−→ hΣ, I, Ii
bDIOC |End e
√
hΣ, I, 1i −
→ hΣ, I, 0i
bDIOC |While-unfold e
I0
hΣ, I, Ii −
→
bDIOC |Up e
bDIOC |NoUp e
hΣ, I0 , Ii
bDIOC |Change-Updates e
Figure 3: Annotated DIOC system semantics.
Rule bDIOC |Interaction e executes a communication from R to S on operation o, where R
sends to S the value v of an expression e. The communication reduces to an assignment
that inherits the index i of the interaction. The assignment stores value v in variable x
of role S. Rule bDIOC |Assign e evaluates the expression e in the local state ΣR and stores
the resulting value v in the local variable x in role R ([v/x, R] represents the substitution).
DYNAMIC CHOREOGRAPHIES
9
roles(i: o : R(e) → S(x)) = {R, S}
roles(1) = roles(0) = ∅
roles(i: x@R = e) = {R}
roles(I; I 0 ) = roles(I|I 0 ) = roles(I) ∪ roles(I 0 )
roles(i: if b@R {I} else {I 0 }) = {R} ∪ roles(I) ∪ roles(I 0 )
roles(i: while b@R {I}) = roles(i: scope @R {I}) = {R} ∪ roles(I)
Figure 4: Auxiliary function roles.
Rule bDIOC |Sequence e executes a step in the first process of a sequential composition, while
Rule bDIOC |Seq-end e acknowledges the termination of the first process, starting the second
one. Rule bDIOC |Parallel e allows a process in a parallel composition to compute, while Rule
bDIOC |Par-end e synchronises the termination of two parallel processes. Rules bDIOC |If-then e
and bDIOC |If-else e evaluate the boolean guard of a conditional, selecting the “then” and
the “else” branch, respectively. Rules bDIOC |While-unfold e and bDIOC |While-exit e correspond
respectively to the unfolding of a while loop when its condition is satisfied and to its
termination otherwise. Rule bDIOC |Up e and Rule bDIOC |NoUp e deal with updates: the
former applies an update, while the latter allows the body of the scope to be executed
without updating it. More precisely, Rule bDIOC |Up e models the application of the update
I 0 to the scope scope @R {I} which, as a result, is replaced by the DIOC process I 0 .
In the conditions of the rule, we use the function roles and the predicates connected and
freshIndexes. Function roles(I), defined in Figure 4, computes the roles of a DIOC process I.
The condition of the rule requires that the roles of the update are a subset of the roles of
the body of the scope. Predicate connected(I 0 ) holds if I 0 is connected. Connectedness is a
well-formedness property of DIOCs and is detailed in § 6. Roughly, it ensures that roles
involved in a sequential composition have enough information to enforce the correct sequence
of actions. Predicate freshIndexes(I 0 ) holds if all indexes in I 0 are fresh with respect to all
indexes already present in the target DIOC2. This is needed to avoid interferences between
communications inside the update and communications in the original DIOC. This problem
is discussed in more details in § 4, Example 4.3. Rule bDIOC |NoUp e, used when no update is
applied, removes the scope boundaries and starts the execution of the body of the scope.
Rule bDIOC |End e terminates the execution of an empty process. Rule bDIOC |Change-Updates e
allows the set I of available updates to change. This rule is always enabled and models the
fact that the set of available updates is not controlled by the system, but by the external
world: the set of updates can change at any time, the system cannot forbid or delay these
changes, and the system is not notified when they happen. Label I is needed to make the
changes to the set of available updates observable (cf. Definition 7.1).
Remark 2.5. Whether to update a scope or not, and which update to apply if many are
available, is completely non-deterministic. We have adopted this view to maximise generality.
However, for practical applications it is also possible to reduce the non-determinism using
suitable policies to decide when and whether a given update applies. One of such policies is
defined in AIOCJ (see § 8).
2We do not give a formal definition of freshIndexes(I 0 ) to keep the presentation simple. However, freshness
of indexes can be formally ensured using restriction as in π-calculus [46].
10
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
We can finally provide the definition of DIOC traces and weak DIOC traces, which
we use to express our results of behavioural correspondence between DIOCs and DPOCs.
Intuitively, in DIOC traces all the performed actions are observed, whilst in weak DIOC
traces silent actions τ are not visible.
Definition 2.6 (DIOC traces). A (strong) trace of a DIOC system hΣ1 , I1 , I1 i is a sequence
(finite or infinite) of labels µ1 , µ2 , . . . such that there is a sequence of DIOC system transitions
µ1
µ2
hΣ1 , I1 , I1 i −→ hΣ2 , I2 , I2 i −→ . . . .
A weak trace of a DIOC system hΣ1 , I1 , I1 i is a sequence of labels µ1 , µ2 , . . . obtained
by removing all silent labels τ from a trace of hΣ1 , I1 , I1 i.
3. Dynamic Process-Oriented Choreographies
In this section we define the syntax and semantics of DPOCs, the target language of our
projection from DIOCs. We remind that DIOCs are not directly executable since their basic
primitives describe distributed interactions. On the contrary, mainstream languages like Java
and C, used for implementation, describe distributed computations using local behaviours
and communication/synchronisation primitives, such as message send and message receive. In
order to describe implementations corresponding to DIOCs we introduce the DPOC language,
a core language based on this kind of primitives, but tailored to program updatable systems.
Indeed, differently from DIOC constructs, DPOC constructs are locally implementable in
any mainstream language. In AIOCJ (see § 8) we implement the DPOC constructs in the
Jolie [36] language.
3.1. DPOC syntax. DPOCs include processes, ranged over by P, P 0 , . . ., describing the
behaviour of participants. (P, Γ)R denotes a DPOC role named R, executing process P in a
local state Γ. Networks, ranged over by N , N 0 , . . ., are parallel compositions of DPOC roles
with different names. DPOC systems, ranged over by S, are DPOC networks equipped with
a set of updates I, namely pairs hI, N i.
DPOCs, like DIOCs, feature operations o. Here we call them programmer-specified
operations to mark the difference with respect to auxiliary operations, ranged over by o∗ .
We use o? to range over both programmer-specified and auxiliary operations. Differently
from communications on programmer-specified operations, communications on auxiliary
operations have no direct correspondent at the DIOC level. Indeed, we introduce auxiliary
operations in DPOCs to implement the synchronisation mechanisms needed to realise the
global constructs of DIOCs (conditionals, while loops, and scopes) at DPOC level.
Like DIOC constructs, also DPOC constructs are annotated using indexes. However,
in DPOCs we use two kinds of indexes: normal indexes i ∈ N and auxiliary indexes of the
forms iT , iF , i? , and iC where i ∈ N. Auxiliary indexes are introduced by the projection,
described in § 4, and are described in detail there. We let ι range over DPOC indexes.
In DPOCs, normal indexes are also used to prefix the operations in send and receive
primitives3. Thus, a send and a receive can interact only if they are on the same operation
and they are prefixed by the same normal index. This is needed to avoid interferences
3In principle, one may use just indexes and drop operations. However, explicit operations are the standard
for communication-based systems, in particular in the area of Web Services, as they come handy to specify
and to debug such systems.
DYNAMIC CHOREOGRAPHIES
11
between concurrent communications, in particular when one of them comes from an update.
We describe in greater detail this issue in § 4, Example 4.3.
The syntax of DPOCs is the following.
P ::=
|
|
|
|
|
ι: i.o? : x from R
ι: i.o? : e to R
i: i.o∗ : X to R
P;P0
P | P0
ι: x = e
X ::=
no
(receive)
(send)
(send-update)
(sequence)
(parallel)
(assignment)
|
P
|
|
|
|
|
|
1
(inaction)
0
(end)
i: if b {P } else {P 0 }
(conditional)
i: while b {P }
(while)
i: scope @R {P } roles {S} (scope-coord)
i: scope @R {P }
(scope)
N ::=
(P, Γ)R
|
N k N0
DPOC processes include receive action ι : i.o? : x from R on a specific operation
(either programmer-specified or auxiliary) of a message from role R to be stored in
variable x, send action ι: i.o? : e to R of the value of an expression e to be sent to role
R, and higher-order send action i : i.o∗ : X to R of the higher-order argument X to be
sent to role R. Here X may be either a DPOC process P , which is the new code for a
scope in R, or a token no, notifying that no update is needed. P ; P 0 and P |P 0 denote
the sequential and parallel composition of P and P 0 , respectively. Processes also feature
assignment ι : x = e of the value of expression e to variable x, the process 1, that can
only successfully terminate, and the terminated process 0. DPOC processes also include
conditionals i: if b {P } else {P 0 } and loops i: while b {P }. Finally, there are two constructs
for scopes. Construct i: scope @R {P } roles {S} defines a scope with body P and set of
participants S, and may occur only inside role R, which acts as coordinator of the update.
The shorter version i: scope @R {P } is used instead inside the code of some role R1 , which
is not the coordinator R of the update. In fact, only the coordinator R needs to know the
set S of involved roles to be able to send to them their updates.
i.o?
3.2. DPOC semantics. DPOC semantics is defined in two steps: we define the semantics
of DPOC roles first, and then we define how roles interact giving rise to the semantics of
DPOC systems.
Definition 3.1 (DPOC roles semantics). The semantics of DPOC roles is defined as the
smallest LTS closed under the rules in Figure 5, where we report the rules dealing with
computation, and Figure 6, in which we define the rules related to updates. Symmetric rules
for parallel composition have been omitted.
DPOC role semantics. We use δ to range over labels. The possible values for δ are as follows:
δ ::= i.o? hvi@S : R
| i.o∗ hXi@S : R
|I
|τ
(send) | i.o? (x ← v)@S : R
(send-update) | i.o∗ (← X)@S : R
(update) | no-up
√
(silent) |
(receive)
(receive-update)
(no-update)
(termination)
12
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
[[e]]Γ = v
bDPOC |One e
√
τ
(1, Γ)R −
→ (0, Γ)R
(ι: x = e, Γ)R −
→ (1, Γ[v/x])R
bDPOC |Assign e
[[e]]Γ = v
bDPOC |Send e
? hvi@S:R
i.o
ι: i.o? : e to S, Γ R −−−−−−−→ (1, Γ)R
ι:
i.o?
i.o? (x←v)@S:R
: x from S, Γ R −−−−−−−−−−→ (ι: x = v, Γ)R
bDPOC |Recv e
bDPOC |Send-Up e
? hXi@S:R
i.o
i: i.o? : X to S, Γ R −−−−−−−−→ (1, Γ)R
√
δ
(P, Γ)R →
− (P 0 , Γ0 )R δ 6=
bDPOC |Sequence e
δ
0
0
(P ; Q, Γ)R →
− (P ; Q, Γ )R
√
δ
(P, Γ)R −
→ (P 0 , Γ)R
(Q, Γ)R →
− (Q0 , Γ0 )R
δ
(P ; Q, Γ)R →
−
(Q0 , Γ0 )
δ
(P, Γ)R →
− (P 0 , Γ0 )R
δ
(P | Q, Γ)R →
−
(P 0
(P, Γ)R −
→ (P 0 , Γ)R
√
(P | Q, Γ)R −
→
R
δ 6=
Q, Γ0 )
|
√
√
√
bDPOC |Parallel e
R
(Q, Γ)R −
→ (Q0 , Γ)R
(P 0
|
Q0 , Γ)
bDPOC |Par-end e
R
[[b]]Γ = true
τ
bDPOC |If-Then e
τ
bDPOC |If-Else e
(i: if b {P } else {P 0 }, Γ)R −
→ (P, Γ)R
[[b]]Γ = false
(i: if b {P } else {P 0 }, Γ)R −
→ (P 0 , Γ)R
[[b]]Γ = true
τ
(i: while b {P }, Γ)R −
→ (P ; i: while e {P }, Γ)R
[[b]]Γ = false
τ
bDPOC |Seq-end e
(i: while b {P }, Γ)R −
→ (1, Γ)R
bDPOC |While-unfold e
bDPOC |While-exit e
Figure 5: DPOC role semantics. Computation rules. (Update rules in Figure 6)
The semantics is in the early style. Rule bDPOC |Recv e receives a value v from role S
and assigns it to local variable x of R. Similarly to Rule bDIOC |Interact e (see Figure 3), the
reception reduces to an assignment that inherits the index i from the receive primitive.
Rules bDPOC |Send e and bDPOC |Send-Up e execute send and higher-order send actions,
respectively. Send actions evaluate expression e in the local state Γ. Rule bDPOC |One e
terminates an empty process. Rule bDPOC |Assign e executes an assignment ([v/x] represents
the substitution of value v for variable x). Rule bDPOC |Sequence e executes a step in the first
process of a sequential composition, while Rule bDPOC |Seq-end e acknowledges the termination
of the first process, starting the second one. Rule bDPOC |Parallel e allows a process in a
DYNAMIC CHOREOGRAPHIES
roles(I) ⊆ S
13
freshIndexes(I) connected(I)
bDPOC |Lead-Up e
I
(i: scope @R {P } roles {S}, Γ)R −
→
Q
i: i.sb∗i : π(I, Rj ) to Rj ; π(I, R);
Rj ∈S\{R}
!
Q
i:
Rj ∈S\{R}
i.se∗i
: from Rj , Γ
R
bDPOC |Lead-NoUp e
no-up
(i: scope @R {P } roles {S}, Γ)R −−−−→
Q
Q
i: i.sb∗i : no to Rj ; P ;
Rj ∈S\{R}
Rj ∈S\{R}
0
i.sb∗
i (←P )@S:R
!
i:
i.se∗i
R
bDPOC |Up e
0
(i: scope @S {P }, Γ)R −−−−−−−−−−−→ (P ; i:
i.sb∗
i (←no)@S:R
: from Rj , Γ
(i: scope @S {P }, Γ)R −−−−−−−−−−−→ (P ; i:
i.se∗i
i.se∗i
: ok to S, Γ)R
bDPOC |NoUp e
: ok to S, Γ)R
Figure 6: DPOC role semantics. Update rules. (Computation rules in Figure 5)
parallel composition to compute, while Rule bDPOC |Par-end e synchronises the termination of
two parallel processes. Rules bDPOC |If-then e and bDPOC |If-else e select the “then” and the
“else” branch in a conditional, respectively. Rules bDPOC |While-unfold e and bDPOC |While-exit e
model respectively the unfolding and the termination of a while loop.
The rules reported in Figure 6 deal with code updates. Rules bDPOC |Lead-Up e and
DPOC
b
|Lead-NoUp e specify the behaviour of the coordinator R of the update, respectively
when an update is performed and when no update is performed. In particular, R nondeterministically selects whether to update or not and, in the first case, which update to
apply. The coordinator communicates the selection to the other roles in the scope using
operations sb∗i . The content of the message is either the new code that the other roles need to
execute, if the update is performed, or a token no, if no update is applied. Communications
on operations sb∗i also ensure that no role starts executing the scope before the coordinator
has selected whether to update or not. The communication is received by the other roles
using Rule bDPOC |Up e if an update is selected and Rule bDPOC |NoUp e if no update is selected.
Similarly, communications on se∗i ensure that the updated code has been completely executed
before any role can proceed to execute the code that follows the scope in its original DP OC
process. Communications on se∗i carry no relevant data: they are used for synchronisation
purposes only.
As already discussed, Rule bDPOC |Lead-Up e models the fact that the coordinator R of
the update non-deterministically selects an update I. The premises of Rule bDPOC |Lead-Up e
are similar to those of Rule bDIOC |Up e (see Figure 3). Function roles is used to check that
the roles in I are included in the roles of the scope. Freshness of indexes is checked by
predicate freshIndexes, and well formedness of I by predicate connected (formally defined
later on, in Definition 6.2 in § 6). In particular, the coordinator R generates the processes
to be executed by the roles in S using the process-projection function π (detailed in § 4).
More precisely, π(I, Ri ) generates the code for role Ri . The processes π(I, Ri ) are sent via
auxiliary higher-order communications on sb∗i to the roles that have to execute them. These
14
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
communications also notify the other roles that
Q they can start executing the new code. Here,
and in the remainder of the paper, we define
Q Ri ∈S Pi as the parallel composition of DPOC
processes Pi for Ri ∈ S. We assume that
binds more tightly than sequential composition.
After the communication of the updated code to the other participants, R starts its own
updated code π(I, R). Finally, auxiliary communications se∗i are used to synchronise the end
of the execution of the update (here denotes a fresh variable to store the synchronisation
message ok). Rule bDPOC |Lead-NoUp e defines the behaviour of the coordinator R when no
update is applied. In this case, R sends a token no to the other involved roles, notifying
them that no update is applied and that they can start executing their original code. End
of scope synchronisation is the same as that of Rule bDPOC |Lead-Up e. Rules bDPOC |Up e and
bDPOC |NoUp e define the behaviour of the other roles involved in the scope. The scope waits
for a message from the coordinator. If the content of the message is no, the body of the
scope is executed. Otherwise, the content of the message is a process P 0 which is executed
instead of the body of the scope.
We highlight the importance of the coordinator R. Since the set of updates may change
at any moment, we need to be careful to avoid that the participants in the scope get code
projected from different updates. Given that only role R obtains and delivers the new code,
one is guaranteed that all the participants receive their projection of the same update.
DPOC system semantics.
Definition 3.2 (DPOC systems semantics). The semantics of DPOC systems is defined as
the smallest LTS closed under the rules in Figure 7. Symmetric rules for parallel composition
have been omitted.
We use η to range over DPOC systems labels. The possible values of η are as follows:
η ::= o? : R(v) → S(x)
(interaction)
| o∗ : R(X) → S()
(interaction-update)
| δ
(role label)
DPOC
DPOC
Rules b
|Lift e and b
|Lift-Up e lift role transitions to the system level. Rule
bDPOC |Lift-Up e also checks that the update I belongs to the set of currently available updates
I. Rule bDPOC |Synch e synchronises a send with the corresponding receive, producing an
interaction. Rule bDPOC |Synch-Up e is similar, but it deals with higher-order interactions. Note
that Rules bDPOC |Synch e and bDPOC |Synch-Up e remove the prefixes from DPOC operations
in transition labels. The labels of these transitions store the information on the occurred
communication: label o? : R1 (v) → R2 (x) denotes an interaction on operation o? from role
R1 to role R2 where the value v is sent by R1 and then stored by R2 in variable x. Label
o∗ : R1 (X) → R2 () denotes a similar interaction, but concerning a higher-order value X,
which can be either the code used in the update or a token no if no update is performed. No
receiver variable is specified, since the received value becomes part of the code of the receiving
process. Rule bDPOC |Ext-Par e allows a network inside a parallel composition to compute.
Rule bDPOC |Ext-Par-End e synchronises the termination of parallel networks. Finally, Rule
bDPOC |Change-Updates e allows the set of updates to change arbitrarily.
We now define DPOC traces and weak DPOC traces, which we later use, along with
DIOC traces and weak DIOC traces, to define our result of correctness.
DYNAMIC CHOREOGRAPHIES
δ
N →
− N0
δ 6= I
δ
hI, N 0 i
hI, N i →
−
I
N −
→ N0
bDPOC |Lift e
i.o? hvi@S:R
hI, N i −−−−−−−→ hI, N 0 i
I∈I
I
hI, N i −
→ hI, N 0 i
15
bDPOC |Lift-Up e
i.o? (x←v)@R:S
hI, N 00 i −−−−−−−−−−→ hI, N 000 i
o? :R(v)→S(x)
hI, N k N 00 i −−−−−−−−−→ hI, N 0 k N 000 i
i.o∗ (←X)@R:S
i.o∗ hXi@S:R
hI, N i −−−−−−−−→ hI, N 0 i
hI, N k
bDPOC |Synch e
hI, N 00 i −−−−−−−−−→ hI, N 000 i
o∗ :R(X)→S()
N 00 i −−−−−−−−→
η
hI, N i −
→ hI, N 0 i
hI, N 0
η 6=
k
√
η
hI, N k N 00 i −
→ hI, N 0 k N 00 i
bDPOC |Ext-Parallel e
√
√
hI, N i −
→ hI, N 0 i
hI, N 00 i −
→ hI, N 000 i
√
hI, N k N 00 i −
→ hI, N 0 k N 000 i
I0
hI, N i −
→
hI0 , N i
bDPOC |Synch-Up e
N 000 i
bDPOC |Ext-Par-End e
bDPOC |Change-Updates e
Figure 7: DPOC system semantics.
Definition 3.3 (DPOC traces). A (strong) trace of a DPOC system hI1 , N1 i is a sequence
(finite or infinite) of labels η1 , η2 , . . . with
√
ηi ∈ {τ, o? : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), , I, no-up, I}
η1
η2
such that there is a sequence of transitions hI1 , N1 i −→ hI2 , N2 i −→ . . . .
A weak trace of a DPOC system hI1 , N1 i is a sequence of labels η1 , η2 , . . . obtained by removing
all the labels corresponding to auxiliary communications, i.e., of the form o∗ : R1 (v) → R2 (x)
or o∗ : R1 (X) → R2 (), and the silent labels τ , from a trace of hI1 , N1 i.
DPOC traces do not allow send and receive actions. Indeed these actions represent
incomplete interactions, thus they are needed for compositionality reasons, but they do not
represent relevant behaviours of complete systems. Note also that these actions have no
correspondence at the DIOC level, where only whole interactions are allowed.
Remark 3.4. Contrarily to DIOCs, DPOCs in general can deadlock. For instance,
i: i.o : x from R0 , Γ R
is a deadlocked DPOC network: the process i: i.o : x from R0 is not terminated, and the
only enabled actions are changes of the set of updates (i.e., transitions with label I), which
are not actual system activities, but are taken by the environment. Notably, the DPOC
above cannot be obtained by projecting a DIOC. In fact, DPOCs generated from DIOCs are
guaranteed to be deadlock free.
16
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
4. Projection Function
We now introduce the projection function proj. Given a DIOC specification, proj returns a
network of DPOC programs that enact the behaviour defined by the originating DIOC.
We write the projection of a DIOC I as proj(I, Σ), where Σ is a global state. Informally,
the projection of a DIOC is a parallel composition of terms, one for each role of the DIOC.
The body of these roles is computed by the process-projection function π (defined below).
Given a DIOC and a role name R, the process-projection returns the process corresponding
to the local behaviour of role R. Since the roles executing the process-projections are
composed in parallel, the projection of a DIOC program results into the DPOC network of
the projected roles.
To give the formal definition of projection, we first define ki∈I Ni as the parallel
composition of networks Ni for i ∈ I.
Definition 4.1 (Projection). The projection of a DIOC process I with global state Σ is
the DPOC network defined by:
proj(I, Σ) =kS∈roles(I) (π(I, S), ΣS )S
The process-projection function that derives DPOC processes from DIOC processes is
defined as follows.
Definition 4.2 (Process-projection). Given an annotated DIOC process I and a role R
the projected DPOC process π(I, R) is defined as in Figure 8.
With a little abuse of notation, we write roles(I, I 0 ) for roles(I) ∪ roles(I 0 ). We assume
that variables xi are never used in the DIOC to be projected and we use them for auxiliary
synchronisations.
The projection is homomorphic for sequential and parallel composition, 1 and 0. The
projection of an assignment is the assignment on the role performing it and 1 on other roles.
The projection of an interaction is a send on the sender role, a receive on the receiver, and
1 on any other role. The projection of a scope is a scope on all its participants. On its
coordinator it also features a clause that records the roles of the involved participants. On
the roles not involved in the scope the projection is 1. Projections of conditional and while
loop are a bit more complex, since they need to coordinate a distributed computation. To
this end they exploit communications on auxiliary operations. In particular, cnd∗i coordinates
the branching of conditionals, carrying information on whether the “then” or the “else”
branch needs to be taken. Similarly, wb∗i coordinates the beginning of a while loop, carrying
information on whether to loop or to exit. Finally, we∗i coordinates the end of the body of
the while loop. This closing operation carries no relevant information and it is just used for
synchronisation purposes. In order to execute a conditional i: if b@R {I} else {I 0 }, the
coordinator R of the conditional locally evaluates the guard and tells the other roles which
branch to choose using auxiliary communications on cnd∗i . Finally, all the roles involved
in the conditional execute their code corresponding to the chosen branch. Execution of a
loop i: while b@R {I} is similar, with two differences. First, end of loop synchronisation
on operations we∗i is used to notify the coordinator that an iteration is terminated, and a
new one can start. Second, communication of whether to loop or to exit is more tricky than
communication on the branch to choose in a conditional. Indeed, there are two points in
the projected code where the coordinator R sends the decision: the first is inside the body
of the loop and it is used if the decision is to loop; the second is after the loop and it is
DYNAMIC CHOREOGRAPHIES
π(1, S) = 1
0
0
π(I; I , S) = π(I, S); π(I , S)
π(0, S) = 0
π(I|I 0 , S) = π(I, S)|π(I 0 , S)
π(i: x@R = e, R)
= i: x = e
π(i: x@R = e, S) and S 6= R
= 1
π(i: o : R1 (e) → R2 (x), R1 )
= i: i.o : e to R2
π(i: o : R1 (e) → R2 (x), R2 )
= i: i.o : x from R1
π(i: o : R1 (e) → R2 (x), S)
and S 6∈ {R1 , R2 }
= 1
π(i: if b@R {I} else {I 0 }, R)
=
i: if b
17
Q
iT: i.cnd∗i : true to R0
!
;
R0 ∈roles(I,I 0 )\{R}
π(I,
R)
!
Q
iF: i.cnd∗i : false to R0 ;
else
R0 ∈roles(I,I 0 )\{R}
π(I 0 , R)
π(i: if b@R {I} else {I 0 }, S)
and S ∈ roles(I, I 0 ) \ {R}
= i?: i.cnd∗i : xi from R; i: if xi {π(I, S)} else {π(I 0 , S)}
π(i: if b@R {I} else {I 0 }, S)
and S 6∈ roles(I, I 0 ) ∪ {R}
= 1
π(i: while b@R {I}, R)
π(i: while b@R {I}, S)
and S ∈ roles(I) \ {R}
n
i: while b
!
Q
0
∗
iT: i.wbi : true to R ;
R0 ∈roles(I)\{R}
Q
=
iC: i.we∗i : from R0
R0 ∈roles(I)\{R}
o
Q
iF: i.wb∗i : false to R0
;
0 ∈roles(I)\{R}
R
∗
i?: i.wbi : xi from
R;
π(I, S);
=
∗
i: while xi iC: i.wei : ok to R;
i?: i.wb∗i : xi from R
π(i: while b@R {I}, S)
and S 6∈ roles(I) ∪ {R}
= 1
π(i: scope @R {I}, R)
= i: scope @R {π(I, R)} roles {roles(I)}
π(i: scope @R {I}, S)
and S ∈ roles(I) \ {R}
= i: scope @R {π(I, S)}
π(i: scope @R {I}, S)
and S 6∈ roles(I) ∪ {R}
= 1
Figure 8: process-projection function π.
π(I, R);
18
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
used if the decision is to exit. Also, there are two points where these communications are
received by the other roles: before their loop at the first iteration, at the end of the previous
iteration of the body of the loop in the others.
One has to keep attention since, by splitting an interaction into a send and a receive
primitive, primitives corresponding to different interactions, but on the same operation, may
interfere.
Example 4.3. We illustrate the issue of interferences using the two DPOC processes below,
identified by their respective roles, R1 (right) and R2 (left), assuming that operations are
not prefixed by indexes. We describe only R1 as R2 is its dual. At Line 1, R1 sends a
message to R2 on operation o. In parallel with the send, R1 had a scope (Lines 3–5) that
performed an update. The new code (Line 4) contains a send on operation o to role R2 .
Since the two sends and the two receives share the same operation o and run in parallel,
they can interfere with each other.
process R1
1. 1: o : e1 to R2
2. |
3. // update auxiliary code
4. 2: o : e2 to R2
5. // update auxiliary code
process R2
1. 1: o : x1 from R2
2. |
3. // update auxiliary code
4. 2: o : x2 from R2
5. // update auxiliary code
Note that, since updates come from outside and one cannot know in advance which
operations they use, this interference cannot be statically avoided.
For this reason, in § 3 we introduced indexes to prefix DPOC operations.
A similar problem may occur also for auxiliary communications. In particular, imagine
to have two parallel conditionals executed by the same role. We need to avoid that, e.g.,
the decision to take the “else” branch on the first conditional is wrongly taken by some
role as a decision concerning the second conditional. To avoid this problem, we prefix
auxiliary operations using the index i of the conditional. In this way, communications
involving distinct conditionals cannot interact. Note that communications concerning the
same conditional (or while loop) may share the same operation name and prefix. However,
since all auxiliary communications are from the coordinator of the construct to the other
roles involved in it, or vice versa, interferences are avoided.
We now describe how to generate indexes for statements in the projection. As a general
rule, all the DPOC constructs obtained by projecting a DIOC construct with index i have
index i. The only exceptions are the indexes of the auxiliary communications of the projection
of conditionals and while loops.
Provided i is the index of the conditional: i ) in the projection of the coordinator we
index the auxiliary communications for selecting the “then” branch with index iT , the
ones for selecting the “else” branch with index iF ; ii ) in the projection of the other roles
involved in the conditional we assign the index i? to the auxiliary receive communications.
To communicate the evaluation of the guard of a while loop we use the same indexing
scheme (iT , iF , and i? ) used in the projection of conditional. Moreover, all the auxiliary
communications for end of loop synchronisation are indexed with iC .
DYNAMIC CHOREOGRAPHIES
19
5. Running Example: Projection and Execution
In this section we use our running example (see Figure 2) to illustrate the projection and
execution of DIOC programs.
5.1. Projection. Given the code in Figure 2, we need to annotate it to be able to project it
(we remind that in § 4 we defined our projection function on well-annotated DIOCs). Since
we wrote one instruction per line in Figure 2, we annotate every instruction using its line
number as index. This results in a well-annotated DIOC.
From the annotated DIOC, the projection generates three DPOC processes for the
Seller, the Buyer, and the Bank, respectively reported in Figures 9, 11, and 10. To improve
readability, we omit some 1 processes. In the projection of the program, we also omit to
write the index that prefixes the operations since it is always equal to the numeric part of
the index of their correspondent construct. Finally, we write auxiliary communications in
grey.
5.2. Runtime Execution. We now focus on an excerpt of the code to exemplify how
updates are performed at runtime. We consider the code of the scope at Lines 6–9 of
Figure 2. In this execution scenario we assume to introduce in the set of available updates
the update presented in Figure 1, which enables the use of a fidelity card to provide a price
discount. Below we consider both the DIOC and the DPOC level, dropping some 1s to
improve readability.
Since we describe a runtime execution, we assume that the Buyer has just sent the
name of the product (s)he is interested in to the Seller (Line 5 of Figure 2). The annotated
DIOCs we execute is the following.
6: scope @Seller{
7: order price@Seller = getPrice(order);
3
8: offer : Seller(order price) → Buyer(prod price)
4 }
1
2
At runtime we apply Rule bDIOC |Up e that substitutes the scope with the new code. The
replacement is atomic. Below we assume that the instructions of the update are annotated
with indexes corresponding to their line number plus 30.
1
2
3
4
5
6
7
8
9
31: cardReq : Seller(null) → Buyer( );
32: card id@Buyer = getInput();
33: card : Buyer(card id) → Seller(buyer id);
34: if isValid(buyer id)@Seller{
35: order price@Seller = getPrice(order) ∗ 0.9
} else {
37: order price@Seller = getPrice(order)
};
39: offer : Seller(order price) → Buyer(prod price)
Let us now focus on the execution at DPOC level, where the application of updates
is not atomic. The scope is distributed between two participants. The first step of the
20
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
3? : wb∗3 : x3 from Buyer;
3: while(x3 ){
5: priceReq : order from Buyer;
6: scope @Seller{
7: order price = getPrice(order);
8: offer : order price to Buyer
} roles { Seller, Buyer };
3C : we∗3 : ok to Buyer;
3? : wb∗3 : x3 from Buyer
};
15? : cnd∗15 : x15 from Buyer;
15: if (x15 ){
16: payReq : payDesc(order price) to Bank;
21? : cnd∗21 : x21 from Bank;
21: if (x21 ){
22: conf irm : from Bank }
}
Figure 9: Seller DPOC Process.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
15? : cnd∗15
: x15 from Buyer;
15: if (x15 ){
16: payReq : desc from Seller;
17: scope @Bank{
18: pay : auth from Buyer
} roles {Buyer, Bank};
20: payment ok = makePayment(desc, auth);
21: if (payment ok){
{
21T : cnd∗21 : true to Seller
| 21T : cnd∗21 : true to Buyer
} ; {
22: conf irm : null to Seller
| 24: conf irm : null to Buyer
}
} else {
{
21F : cnd∗21 : false to Seller
| 21F : cnd∗21 : false to Buyer
};
26: abort : null to Buyer
}
}
Figure 10: Bank DPOC Process.
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
1: price ok = false;
2: continue = true;
3: while(!price ok and continue){
3T : wb∗3 : true to Seller;
4: prod = getInput();
5: priceReq : prod to Seller;
6: scope @Seller{
7: offer : prod price from Seller
};
10: price ok = getInput();
11: if (!price ok){
12: continue = getInput()
};
3C : we∗3 : from Seller
};
3F : wb∗3 : false to Seller;
15: if (price ok){
{
15T : cnd∗15 : true to Seller
|15T : cnd∗15 : true to Bank
};
17: scope @Bank{
19: pay : payAuth(prod price) to Bank
};
21? : cnd∗21 : x21 from Bank;
21: if (x21 ){
24: conf irm : from Bank
} else {
26: abort : from Bank}
}
}
Figure 11: Buyer DPOC Process.
DYNAMIC CHOREOGRAPHIES
21
update protocol is performed by the Seller, since (s)he is the coordinator of the update.
The DPOC description of the Seller before the update is:
6: scope @Seller{
7: order price = getPrice(order);
8: offer : order price to Buyer
} roles {Seller, Buyer}
When the scope construct is enabled, the Seller non-deterministically selects whether
to update or not and, in the first case, which update to apply. Here, we assume that the
update using the code in Figure 1 is selected. Below we report on the left the reductum of
the projected code of the Seller after the application of Rule bDPOC |Lead-Up e. The Seller
sends to the Buyer the code — denoted as PB and reported below on the right — obtained
projecting the update on role Buyer.
1
2
3
4
5
6
7
8
9
10
6: sb∗6 : PB to Buyer;
31: cardReq : null to Buyer;
33: card : buyer id from Buyer;
34: if isValid(buyer id){
PB := 31: cardReq : null from Seller;
35: order price = getPrice(order) ∗ 0.9
32: card id = getInput();
} else {
33: card : card id to Seller;
37: order price = getPrice(order)
39: offer : prod price from Seller
};
39: offer : order price to Buyer;
6: se∗6 : from Buyer;
Above, at Line 1 the Seller requires the Buyer to update, sending to him the new
DPOC fragment to execute. Then, the Seller starts to execute its own updated DPOC.
At the end of the execution of the new DPOC code (Line 10) the Seller waits for the
notification of termination of the DPOC fragment executed by the Buyer.
Let us now consider the process-projection of the Buyer, reported below.
6: scope @Seller{
8: offer : order price from Seller
}
At runtime, the scope waits for the arrival of a message from the coordinator of the update.
In our case, since we assumed that the update is applied, the Buyer receives using Rule
bDPOC |Up e the DPOC fragment PB sent by the coordinator. In the reductum, PB replaces
the scope, followed by the notification of termination to the Seller.
31: cardReq : null from Seller;
32: card id = getInput();
33: card : card id to Seller;
39: offer : prod price from Seller
6: se∗6 : ok to Seller
22
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Consider now what happens if no update is applied. At DIOC level the Seller applies
Rule bDIOC |NoUp e, which removes the scope and runs its body. At DPOC level, the update
is not atomic. The code of the Seller is the following one.
6: sb∗6 : no to Buyer;
2 7: order price = getPrice(order);
3 8: offer : order price to Buyer;
4 6: se∗6 : from Buyer;
1
Before executing the code inside the scope, the Seller notifies the Buyer that (s)he can
proceed with her execution (Line 1). Like in the case of update, the Seller also waits for
the notification of the end of execution from the Buyer (Line 4).
Finally, we report the DPOC code of the Buyer after the reception of the message that
no update is needed. Rule bDPOC |NoUp e removes the scope and adds the notification of
termination (Line 2 below) to the coordinator at the end.
1
2
7: offer : prod price from Seller;
6: se∗6 : ok to Seller;
6. Connected DIOCs
We now give a precise definition of the notion of connectedness that we mentioned in § 2.2
and § 3.2. In both DIOC and DPOC semantics we checked such a property of updates with
predicate connected, respectively in Rule bDIOC |Up e (Figure 3) and Rule bDPOC |Up-Lead e
(Figure 6).
To give the intuition of why we need to restrict to connected updates, consider the
scenario below of a DIOC (left side) and its projection (right side).
op1 : A(e1 ) → B(x);
op2 : C(e2 ) → D(y)
projection
======⇒
process A
op1 : e1 to B
process B
op1 : x from A
process C
op2 : e2 to D
process D
op2 : y from C
DIOCs can express interactions that, if projected as described in § 4, can behave differently
with respect to the originating DIOC. Indeed, in our example we have a DIOC that composes
in sequence two interactions: an interaction between A and B on operation op1 followed by
an interaction between C and D on operation op2 . The projection of the DIOC produces
four processes (identified by their role): A and C send a message to B and D, respectively.
Dually, B and D receive a message form A and C, respectively. In the example, at the level
of processes we lose the global order among the interactions: each projected process runs its
code locally and it is not aware of the global sequence of interactions. Indeed, both sends
and both receives are enabled at the same time. Hence, the semantics of DPOC lets the two
interactions interleave in any order. It can happen that the interaction between C and D
occurs before the one between A and B, violating the order of interactions prescribed by
the originating DIOC.
Restricting to connected DIOCs avoids this kind of behaviours. We formalise connectedness as an efficient (see Theorem 6.3) syntactic check. We highlight that our definition of
connectedness does not hamper programmability and it naturally holds in most real-world
DYNAMIC CHOREOGRAPHIES
23
scenarios (the interested reader can find in the website of the AIOCJ project [1] several such
scenarios).
Remark 6.1. There exists a trade-off between efficiency and ease of programming with
respect to the guarantee that all the roles are aware of the evolution of the global computation.
This is a common element of choreographic approaches, which has been handled in different
ways, e.g., i ) by restricting the set of well-formed choreographies to only those on which
the projection preserves the order of actions [10]; ii ) by mimicking the non-deterministic
behaviour of process-level networks at choreography level [13]; or iii ) by enforcing the order
of actions with additional auxiliary messages between roles [31].
Our choice of preserving the order of interactions defined at DIOC level follows the same
philosophy of [10], whilst for scopes, conditionals, and while loops we enforce connectedness
with auxiliary messages as done in [31]. We remind that we introduced auxiliary messages
for coordination both in the semantics of scopes at DPOC level (§ 3.2) and in the projection
(§ 4). We choose to add such auxiliary messages to avoid to impose strong constraints on the
form of scopes, conditionals, and while loops, which in the end would pose strong limitations
to the programmers of DIOCs. On the other hand, for sequential composition we choose to
restrict the set of allowed DIOCs by requiring connectedness, which ensures that the order
of interactions defined at DIOC level is preserved by projection.
As discussed above, the execution of conditionals and while loops rely on auxiliary
communications to coordinate the different roles. Some of these communications may be
redundant. For instance, in Figure 9, Line 8, the Seller notifies to the Buyer that (s)he
has completed her part of the while loop. However, since her last contribution to the while
loop is the auxiliary communication for end of scope synchronisation, the Buyer already
has this information. Hence, Line 8 in Figure 9, where the notification is sent, and Line 14
in Figure 11, where the notification is received, can be safely dropped. Removing redundant
auxiliary communications can be automatised using a suitable static analysis. We leave this
topic for future work.
To formalise connectedness we introduce, in Figure 12, the auxiliary functions transI and
transF that, given a DIOC process, compute sets of pairs representing senders and receivers
of possible initial and final interactions in its execution. We represent one such pair as
R → S. Actions located at R are represented as R → R. For instance, given an interaction
i : o : R(e) → S(x) both its transI and transF are {R → S}. For conditional, transI(i :
if b@R {I} else {I 0 }) = {R → R} since the first action executed is the evaluation of the
guard by role R. The set transF(i: if b@R {I} else {I 0 }) is normally transF(I) ∪ transF(I 0 ),
since the execution terminates with an action from one of the branches. If instead the
branches are both empty then transF is {R → R}, representing guard evaluation.
Finally, we give the formal definition of connectedness.
Definition 6.2 (Connectedness). A DIOC process I is connected if each subterm I 0 ; I 00 of
I satisfies
∀ R1 → R2 ∈ transF(I 0 ), ∀ S1 → S2 ∈ transI(I 00 ) . {R1 , R2 } ∩ {S1 , S2 } =
6 ∅
Connectedness can be checked efficiently.
Theorem 6.3 (Connectedness-check complexity).
The connectedness of a DIOC process I can be checked in time O(n2 log(n)), where n is the
number of nodes in the abstract syntax tree of I.
24
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
transI(i: o : R(e) → S(x))
= transF(i: o : R(e) → S(x)) = {R → S}
transI(i: x@R = e)
= transF(i: x@R = e) = {R → R}
transI(1)
= transI(0) = transF(1) = transF(0) = ∅
transI(I|I 0 )
= transI(I) ∪ transI(I 0 )
transF(I|I 0 )
= transF(I) ∪ transF(I 0 )
(
transI(I 0 ) if transI(I) = ∅
=
transI(I) otherwise
(
transF(I) if transF(I 0 ) = ∅
=
transF(I 0 ) otherwise
transI(I; I 0 )
transF(I; I 0 )
transI(i: if b@R {I} else {I 0 })
= transI(i: while b@R {I}) = {R → R}
(
{R → R}
if transF(I) ∪ transF(I 0 ) = ∅
transF(i: if b@R {I} else {I 0 }) =
transF(I) ∪ transF(I 0 ) otherwise
if transF(I) = ∅
{R → R}
S
0
transF(i: while b@R {I})
=
{R → R } otherwise
R0 ∈roles(I)r{R}
transI(i: scope @R {I})
= {R → R}
if roles(I) ⊆ {R}
{R → R}
S
0
transF(i: scope @R {I})
=
{R → R} otherwise
R0 ∈roles(I)r{R}
Figure 12: Auxiliary functions transI and transF.
The proof of the theorem is reported in Appendix A.
We remind that we allow only connected updates. Indeed, replacing a scope with a
connected update always results in a deadlock- and race-free DIOC. Thus, one just needs to
statically check connectedness of the starting program and of the updates, and there is no
need to perform expensive runtime checks on the whole application after updates have been
performed.
7. Correctness
In the previous sections we have presented DIOCs, DPOCs, and described how to derive
a DPOC from a given DIOC. This section presents the main technical result of the paper,
namely the correctness of the projection. Moreover, as a consequence of the correctness, in
Section 7.2 we prove that properties like deadlock freedom, termination, and race freedom
are preserved by the projection.
Correctness here means that a connected DIOC and its projected DPOC are weak
system bisimilar. Weak system bisimilarity is formally defined as follows.
Definition 7.1 (Weak System Bisimilarity). A weak system bisimulation is a relation R
between DIOC systems and DPOC systems such that if (hΣ, I, Ii , hI0 , N i) ∈ R then:
DYNAMIC CHOREOGRAPHIES
µ
η1
η
25
µ
k
−
→ hI000 , N 000 i with
• if hΣ, I, Ii −
→ hΣ00 , I00 , I 00 i then hI0 , N i −→, . . . , −→
∗
∗
∀ i ∈ [1 . . . k], ηi ∈ {o : R1 (v) → R2 (x), o : R1 (X) → R2 (), τ } and
(hΣ00 , I00 , I 00 i , hI000 , N 000 i) ∈ R;
√
η
• if hI0 , N i −
→ hI000 , N 000 i with η ∈ {o? : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), , I,
no-up, I000 , τ } then one of the following two conditions holds:
η
– hΣ, I, Ii −
→ hΣ00 , I0 , I 00 i and (hΣ00 , I00 , I 00 i , hI000 , N 000 i) ∈ R or
– η ∈ {o∗ : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), τ } and (hΣ, I, Ii , hI000 , N 000 )i ∈ R
A DIOC system hΣ, I, Ii and a DPOC system hI0 , N i are weak system bisimilar iff there
exists a weak system bisimulation R such that (hΣ, I, Ii , hI0 , N i) ∈ R.
In the proof, we provide a relation R which relates each well-annotated connected DIOC
system with its projection and show that it is a weak system bisimulation. Such a relation
is not trivial since events that are atomic in the DIOC, e.g., the evaluation of the guard of a
conditional, including the removal of the discarded branch, are not atomic at DPOC level.
In the case of conditional, the DIOC transition is mimicked by a conditional performed
by the role evaluating the guard, a set of auxiliary communications sending the value of
the guard to the other roles, and local conditionals based on the received value. These
mismatches are taken care by function upd (Definition 7.18). This function needs also to
remove the auxiliary communications used to synchronise the termination of scopes, which
have no counterpart after the DIOC scope has been consumed. However, we have to record
the impact of the mentioned auxiliary communications on the possible executions. Thus we
define an event structure for DIOC (Definition 7.10) and one for DPOC (Definition 7.11)
and we show that the two are related (Lemma 7.15).
Thanks to the existence of a bisimulation relating each well-annotated connected DIOC
system with its projection we can prove that the projection is correct. Formally:
Theorem 7.2 (Correctness). For each initial, connected DIOC process I, each state Σ,
each set of updates I, the DIOC system hΣ, I, Ii and the DPOC system hI, proj(I, Σ)i are
weak system bisimilar.
As a corollary of the result above, a DIOC system and its projection are also trace
equivalent. Trace equivalence is defined as follows.
Definition 7.3 (Trace equivalence). A DIOC system hΣ, I, Ii and a DPOC system hI, N i
are (weak) trace equivalent iff their sets of (weak) traces coincide.
The following lemma shows that, indeed, weak system bisimilarity implies weak trace
equivalence.
Lemma 7.4. Let hΣ, I, Ii be a DIOC system and hI0 , N i a DPOC system.
If hΣ, I, Ii ∼ hI0 , N i then the DIOC system hΣ, I, Ii and the DPOC system hI0 , N i are weak
trace equivalent.
Proof. The proof is by coinduction. Take a DIOC trace µ1 , µ2 , . . . of the DIOC system.
From bisimilarity, the DPOC system has a sequence of transitions with labels η1 , . . . , ηk , µ1
where η1 , . . . , ηk are weak transitions. Hence, the first label in the weak trace is µ1 . After
the transition with label µ1 , the DIOC system and the DPOC system are again bisimilar.
By coinductive hypothesis, the DPOC system has a weak trace µ2 , . . .. By composition the
DPOC system has a trace µ1 , µ2 , . . . as desired. The opposite direction is similar.
26
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Hence the following corollary holds.
Corollary 7.5 (Trace Equivalence). For each initial, connected DIOC process I, each state
Σ, each set of updates I, the DIOC system hΣ, I, Ii and the DPOC system hI, proj(I, Σ)i are
weak trace equivalent.
Proof. It follows from Theorem 7.2 and Lemma 7.4.
The following section presents the details of the proof of Theorem 7.2. The reader not
interested in such details can safely skip § 7.1 and go to § 7.2.
7.1. Detailed Proof of Correctness. In our proof strategy we rely on the distinctness of
indexes of DIOC constructs that, unfortunately, is not preserved by transitions due to while
unfolding.
Example 7.6. Consider the DIOC i : while b@R {j : x@R = e}. If the condition b
evaluates to true, in one step the application of Rule bDPOC |While-unfold e produces the
DIOC I below
j: x@R = e; i: while b@R {j: x@R = e}
where the index j occurs twice.
To solve this problem, instead of using indexes, we rely on global indexes built on top of
indexes. Global indexes can be used both at the DIOC level and at the DPOC level and
their distinctness is preserved by transitions.
Definition 7.7 (Global index). Given an annotated DIOC process I, or an annotated
DPOC network N , for each annotated construct with index ι we define its global index ξ as
follows:
• if the construct is not in the body of a while loop then ξ = ι;
• if the innermost while construct that contains the considered construct has global index
ξ 0 then the considered construct has global index ξ = ξ 0 : ι.
Example 7.8. Consider the DIOC I in Example 7.6. The first assignment with index j
also has global index j, while the second assignment with index j has global index i: j, since
this last assignment is inside a while loop with global index i.
Lemma 7.9 (Distinctness of Global Indexes). Given a well-annotated DIOC I, a global
η1
ηn
state Σ, and a set of updates I, if hΣ, I, Ii −→ . . . −→ hΣ0 , I0 , I 0 i then all global indexes in I 0
are distinct.
Proof. The proof is by induction on the number n of transitions. Details are in Appendix B.
Using global indexes we can now define event structures corresponding to the execution
of DIOCs and DPOCs. We start by defining DIOC events. Some events correspond to
transitions of the DIOC, and we say that they are enabled when the corresponding transition
is enabled, executed when the corresponding transition is executed.
Definition 7.10 (DIOC events). We use ε to range over events, and we write [ε]R to
highlight that event ε is performed by role R. An annotated DIOC I contains the following
events:
DYNAMIC CHOREOGRAPHIES
27
Communication events: a sending event ξ : o@R2 in role R1 and a receiving event
ξ : o@R1 in role R2 for each interaction i: o : R1 (e) → R2 (x) with global index ξ; we also
denote the sending event as fξ or [fξ ]R1 and the receiving event as tξ or [tξ ]R2 . Sending and
receiving events correspond to the transition executing the interaction.
Assignment events: an assignment event εξ in role R for each assignment i: x@R = e
with global index ξ; the event corresponds to the transition executing the assignment.
Scope events: a scope initialisation event ↑ξ and a scope termination event ↓ξ for
each scope i: scope @R {I} with global index ξ. Both these events belong to all the roles
in roles(I). The scope initialisation event corresponds to the transition performing or not
performing an update on the given scope. The scope termination event is just an auxiliary
event (related to the auxiliary interactions implementing the scope termination).
If events: a guard if-event εξ in role R for each construct i: if b@R {I} else {I 0 } with
global index ξ; the guard-if event corresponds to the transition evaluating the guard of the
condition.
While events: a guard while-event εξ in role R for each construct i: while b@R {I}
with global index ξ; the guard-while event corresponds to the transition evaluating the guard
of the while loop.
Function events(I) denotes the set of events of the annotated DIOC I. A sending and a
receiving event with the same global index ξ are called matching events. We denote with ε
an event matching event ε. A communication event is either a sending event or a receiving
event. A communication event is unmatched if there is no event matching it.
As a corollary of Lemma 7.9 events have distinct names. Note also that, for each while
loop, there are events corresponding to the execution of just one iteration of the loop. If
unfolding is performed, new events are created.
Similarly to what we have done for DIOC, we can define events for DPOC as follows.
Definition 7.11 (DPOC events). An annotated DPOC network N contains the following
events:
Communication events: a sending event ξ : o? @R2 in role R1 for each send ι : i.o? :
e to R2 with global index ξ in role R1 ; and a receiving event ξ : o? @R1 in role R2 for
each receive ι: i.o? : x from R1 with global index ξ in role R2 ; we also denote the sending
event as fξ or [fξ ]R1 ; and the receiving event as tξ or [tξ ]R2 . Sending and receiving events
correspond to the transitions executing the corresponding communication.
Assignment events: an assignment event εξ in role R for each assignment i: x = e with
global index ξ; the event corresponds to the transition executing the assignment.
Scope events: a scope initialisation event ↑ξ and a scope termination event ↓ξ for each
i: scope @R {P } roles {S} or i: scope @R {P } with global index ξ. Scope events with
the same global index coincide, and thus the same event may belong to different roles;
the scope initialisation event corresponds to the transition performing or not performing
an update on the given scope for the role leading the update. The scope termination
event is just an auxiliary event (related to the auxiliary interactions implementing the
scope termination).
If events: a guard if-event εξ in role R for each construct i: if b {P } else {P 0 } with global
index ξ; the guard-if event corresponds to the transition evaluating the guard of the
condition.
28
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
While events: a guard while-event εξ in role R for each construct i: while b {P } with
global index ξ; the guard-while event corresponds to the transition evaluating the guard
of the while loop.
Let events(N ) denote the set of events of the network N . A sending and a receiving event
with either the same global index ξ or with global indexes differing only for replacing index
i? with iT or iF are called matching events. We denote with ε an event matching event ε.
With a slight abuse of notation, we write events(P ) to denote events originated by
constructs in process P , assuming the network N to be understood. We use the same syntax
for events of DIOCs and of DPOCs. Indeed, the two kinds of events are strongly related (cf.
Lemma 7.15).
We define below a causality relation ≤DIOC among DIOC events based on the constraints
given by the semantics on the execution of the corresponding transitions.
Definition 7.12 (DIOC causality relation). Let us consider an annotated DIOC I. A
causality relation ≤DIOC ⊆ events(I) × events(I) is the minimum reflexive and transitive
relation satisfying:
Sequentiality: let I 0 ; I 00 be a subterm of DIOC I. If ε0 is an event in I 0 and ε00 is an
event in I 00 , then ε0 ≤DIOC ε00 .
Scope: let i: scope @R {I 0 } be a subterm of DIOC I. If ε0 is an event in I 0 then
↑ξ ≤DIOC ε0 ≤DIOC ↓ξ .
Synchronisation: for each interaction the sending event precedes the receiving event.
If: let i: if b@R {I 00 } else {I 00 } be a subterm of DIOC I, let εξ be the guard if-event
in role R, then for every event ε0 in I 0 and for every event ε00 in I 00 we have εξ ≤DIOC ε0
and εξ ≤DIOC ε00 .
While: let i: while b@R {I 0 } be a subterm of DIOC I, let εξ be the guard while-event
in role R, then for every event ε0 in I 0 we have εξ ≤DIOC ε0 .
As expected, the relation ≤DIOC is a partial order.
Lemma 7.13. Let us consider an annotated DIOC I. The relation ≤DIOC among events
of I is a partial order.
Proof. A partial order is a relation which is reflexive, transitive, and antisymmetric. Reflexivity and transitivity follow by definition. We show antisymmetry by showing that ≤DIOC
does not contain any cycle. The proof is by structural induction on the DIOC I. The base
cases are interaction, assignment, 1, and 0, which are all trivial. In the case of sequence,
I 0 ; I 00 , by inductive hypothesis there are no cycles among the events of I 0 nor among events
of I 00 . Since all events of I 0 precede all events of I 00 , there are no cycles among the events
of I 0 ; I 00 . In the case of parallel composition, I 0 |I 00 , there is no relation between events in
I 0 and in I 00 , hence the thesis follows. In the case of conditional i: if b@R {I 0 } else {I 00 }
there are no relations between events in I 0 and in I 00 and all the events follow the guard-if
event. Hence the thesis follows. The cases of while and scope are similar.
We can now define a causality relation ≤DP OC among DP OC events.
Definition 7.14 (DPOC causality relation). Let us consider an annotated DPOC network
N . A causality relation ≤DP OC ⊆ events(N ) × events(N ) is the minimum reflexive and
transitive relation satisfying:
Sequentiality: Let P 0 ; P 00 be a subterm of DPOC network N . If ε0 is an event in P 0
and ε00 is an event in P 00 then ε0 ≤DP OC ε00 .
DYNAMIC CHOREOGRAPHIES
29
Scope: Let i: scope @R {P } roles {S} or i: scope @R {P } be a subterm of DPOC
N with global index ξ. If ε0 is an event in P then ↑ξ ≤DP OC ε0 ≤DP OC ↓ξ .
Synchronisation: For each pair of events ε and ε0 , ε ≤ ε0 implies ε ≤DP OC ε0 .
If: Let i: if b {P } else {P 0 } be a subterm of DPOC network N with global index ξ, let
εξ be the guard if-event, then for every event ε in P and for every event ε0 in P 0 we have
εξ ≤DP OC ε and εξ ≤DP OC ε0 .
While: Let i: while b {P } be a subterm of DPOC network N with global index ξ, let
εξ be the guard while-event, then for every event ε in P we have εξ ≤DP OC ε.
On DP OC networks obtained as projections of well-annotated DIOCs the relation
≤DP OC is a partial order, as expected. However, since this result is not needed in the
remainder of the paper, we do not present its proof.
There is a relation between DIOC events and causality relation and their counterparts
at the DP OC level. Indeed, the events and causality relation are preserved by projection.
Lemma 7.15. Given a well-annotated connected DIOC process I and for each state Σ the
DPOC network proj(I, Σ) is such that:
(1) events(I) ⊆ events(proj(I, Σ));
(2) ∀ ε1 , ε2 ∈ events(I).ε1 ≤DIOC ε2 ⇒ ε1 ≤DP OC ε2 ∨ ε1 ≤DP OC ε2
Proof. The events of a DPOC obtained by projecting a DIOC I are included in the events
of the DIOC I by definition of projection. The preservation of the causality relation can be
proven by a case analysis on the condition used to derive the dependency (i.e., sequentiality,
scope, synchronisation, if and while). Details are in Appendix B.
To complete the definition of our event structure we now define a notion of conflict between
(DIOC and DPOC) events, relating events which are in different branches of the same
conditional.
Definition 7.16 (Conflicting events). Given a DIOC process I, two events ε, ε0 ∈ events(I)
are conflicting if they belong to different branches of the same conditional, i.e., there exists
a subprocess i: if b@R {I 0 } else {I 00 } of I such that ε ∈ events(I 0 ) ∧ ε0 ∈ events(I 00 ) or
ε0 ∈ events(I 0 ) ∧ ε ∈ events(I 00 ).
Similarly, given a DPOC network N , we say that two events ε, ε0 ∈ events(N ) are
conflicting if they belong to different branches of the same conditional, i.e., there exists
a subprocess i : if b {P } else {P 0 } of N such that ε ∈ events(P ) ∧ ε0 ∈ events(P 0 ) or
ε0 ∈ events(P ) ∧ ε ∈ events(P 0 ).
Similarly to what we did for DIOCs, we define below well-annotated DPOCs. Wellannotated DPOCs include all DPOCs obtained by projecting well-annotated DIOCs. As
stated in the definition below, and proved in Lemma 7.20, well-annotated DPOCs enjoy
various properties useful for our proofs.
Definition 7.17 (Well-annotated DPOC). An annotated DPOC network N is well annotated for its causality relation ≤DP OC if the following conditions hold:
C1: for each global index ξ there are at most two communication events on programmerspecified operations with global index ξ and, in this case, they are matching events;
C2: only events which are minimal according to ≤DP OC may correspond to enabled transitions;
30
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
C3: for each pair of non-conflicting sending events [fξ ]R1 and [fξ0 ]R1 on the same operation i.o? with the same target R2 such that ξ 6= ξ 0 we have [fξ ]R1 ≤DP OC [fξ0 ]R1 or
[fξ0 ]R1 ≤DP OC [fξ ]R1 ;
C4: for each pair of non-conflicting receiving events [tξ ]R2 and [tξ0 ]R2 on the same operation
i.o? with the same sender R1 such that ξ 6= ξ 0 we have [tξ ]R2 ≤ [tξ0 ]R2 or [tξ0 ]R2 ≤ [tξ ]R2 ;
C5: if ε is an event inside a scope with global index ξ then its matching events ε (if they
exist) are inside a scope with the same global index.
C6: if two events have the same index but different global indexes then one of them, let us
call it ε1 , is inside the body of a while loop with global index ξ1 and the other, ε2 , is not.
Furthermore, ε2 ≤DP OC εξ1 where εξ1 is the guarding while-event of the while loop with
global index ξ1 .
Since scope update, conditional, and iteration at the DIOC level happen in one step, while
they correspond to many steps of the projected DPOC, we introduce a function, denoted
upd, that bridges this gap. More precisely, function upd is obtained as the composition of
two functions, a function compl that completes the execution of DIOC actions which have
already started, and a function clean that eliminates all the auxiliary closing communications
of scopes (scope execution introduces in the DPOC auxiliary communications which have
no correspondence in the DIOC).
Definition 7.18 (upd function). Let N be an annotated DPOC (we drop indexes if
not relevant). The upd function is defined as the composition of a function compl and a
function clean. Thus, upd(N ) = clean(compl(N )). Network compl(N ) is obtained from N
by repeating the following operations while possible.
(1) Performing the reception of the positive evaluation of the guard of a while loop, by
replacing for every i.wb∗i : true to R0 enabled, all the terms
i.wb∗i : xi from R; while xi {P ; i.we∗i : ok to R; i.wb∗i : xi from R}
not inside another while construct, with
P ; i.we∗i : ok to R; i.wb∗i : xi from R; while xi {P ; i.we∗i : ok to R; i.wb∗i : xi from R}
and replace i.wb∗i : true to R0 with 1.
(2) Performing the reception of the negative evaluation of the guard of a while loop by
replacing, for every i.wb∗i : false to R0 enabled, all the terms
i.wb∗i : xi from R; while xi {P ; i.we∗i : ok to R; i.wb∗i : xi from R}
not inside another while construct, with 1, and replace i.wb∗i : false to R0 with 1.
(3) Performing the unfolding of a while loop by replacing every
while xi {P ; i.we∗i : ok to R; i.wb∗i : xi from R}
enabled not inside another while construct, such that xi evaluates to true in the local
state, with
P ; i.we∗i : ok to R; i.wb∗i : xi from R; while xi {P ; i.we∗i : ok to R; i.we∗i : xi from R}
(4) Performing the termination of while loop by replacing every
while xi {P ; i.we∗i : ok to R; i.wb∗i : xi from R}
enabled not inside another while construct, such that xi evaluates to false in the local
state, with 1.
DYNAMIC CHOREOGRAPHIES
31
(5) Performing the reception of the positive evaluation of the guard of a conditional by
replacing, for every i.cnd∗i : true to R0 enabled, all the terms
i.cnd∗i : xi from R; if xi {P 0 } else {P 00 }
not inside a while construct with P 0 , and replace i.cnd∗i : true to R0 with 1.
(6) Performing the reception of the negative evaluation of the guard of a conditional by
replacing, for every i.cnd∗i : false to R0 enabled, all the terms
i.cnd∗i : xi from R; if xi {P 0 } else {P 00 }
not inside a while construct, with P 00 , and replace i.cnd∗i : false to R0 with 1.
(7) Performing the selection of the “then” branch by replacing every
if xi {P 0 } else {P 00 }
enabled, such that xi evaluates to true in the local state, with P 0 .
(8) Performing the selection of the “else” branch by replacing every
if xi {P 0 } else {P 00 }
enabled, such that xi evaluates to false in the local state, with P 00 .
(9) Performing the communication of the updated code by replacing, for every
i.sb∗i : P to S enabled, all the terms
i: scope @R {P 0 }
in role S not inside a while construct with P , and replace i.sb∗i : P to S with 1.
(10) Performing the communication that no update is needed by replacing, for each
i.sb∗i : no to S enabled, all the terms
i: scope @R {P 0 }
in role S not inside a while construct with P 0 , and replace i.sb∗i : P to S with 1.
Network clean(N ) is obtained from N by repeating the following operations while possible:
• Removing the auxiliary communications for end of scope and end of while loop synchronisation by replacing each
i.se∗i : ok to R
i.se∗i : from R
iC : i.we∗i : ok to R
iC : i.we∗i : from R
not inside a while construct with 1.
Furthermore clean may apply 0 or more times the following operation:
• replace a subterm 1; P by P or a subterm 1 | P by P .
Note that function compl does not reduce terms inside a while construct. Assume, for
instance, to have an auxiliary send targeting a receive inside the body of a while loop.
These two communications should not interact since they have different global indexes. This
explains why we exclude terms inside the body of while loops.
We proceed now to prove some of the proprieties of DIOC and DPOC. The first result
states that in a well-annotated DPOC only transitions corresponding to events minimal with
respect to the causality relation ≤DP OC may be enabled.
Lemma 7.19. If N is a DPOC, ≤DP OC its causality relation and ε is an event corresponding
to a transition enabled in N then ε is minimal with respect to ≤DP OC .
32
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Proof. The proof is by contradiction. Suppose ε is enabled but not minimal, i.e., there exists
ε0 such that ε0 ≤DP OC ε. If there is more than one such ε0 consider the one such that the
length of the derivation of ε0 ≤DP OC ε is minimal. This derivation should have length one,
and following Definition 7.14 it may result from one of the following cases:
• Sequentiality: ε0 ≤DP OC ε means that ε0 ∈ events(P 0 ), ε ∈ events(P 00 ), and P 0 ; P 00 is a
subterm of N . Because of the semantics of sequential composition ε cannot be enabled.
• Scope: let i: scope @R {P } roles {S} or i: scope @R {P } be a subprocess of N with
global index ξ. We have the following cases:
– ε0 =↑ξ and ε ∈ events(P ), and this implies that ε cannot be enabled since if ε0 is enabled
then the Rule bDPOC |Up e or Rule bDPOC |NoUp e for starting the execution of the scope
have not been applied yet;
– ε0 =↑ξ and ε =↓ξ , or ε0 ∈ events(P ) and ε =↓ξ : this is trivial, since ↓ξ is an auxiliary
event and no transition corresponds to it;
• If: ε0 ≤DP OC ε means that ε0 is the evaluation of the guard of a subterm
i : if xi {P 0 } else {P 00 } and ε ∈ events(P 0 ) ∪ events(P 00 ). Event ε cannot be enabled
because of the semantics of conditionals.
• While: ε0 ≤DP OC ε means that ε0 is the evaluation of the guard of a subterm
i: while xi {P } and ε ∈ events(P ). Event ε cannot be enabled because of the semantics
of the while loop.
We now prove that all the DPOCs obtained as projection of well-annotated connected DIOCs
are well-annotated.
Lemma 7.20. Let I be a well-annotated connected DIOC process and Σ a state. Then the
projection N = proj(I, Σ) is a well-annotated DPOC network with respect to ≤DP OC .
Proof. We have to prove that proj(I, Σ) satisfies the conditions of Definition 7.17. We have
a case for each condition. Details in Appendix B.
The next lemma shows that for every starting set of updates I the DPOC N and upd(N )
have the same set of weak traces.
Lemma 7.21. Let N be a DPOC. The following properties hold:
√
η
(1) if hI, upd(N )i −
→ hI0 , N 0 i with η ∈ {o? : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), I0 , , I,
η1
ηk η
no-up, τ } then there exists N 00 such that hI, N i −→ . . . −→
−
→ hI0 , N 00 i where ηi ∈ {o∗ :
∗
R1 (v) → R2 (x), o : R1 (X) → R2 (), τ } for each i ∈ {1, . . . , k} and upd(N 00 ) = upd(N 0 ).
√
η
(2) if hI, N i −
→ hI0 , N 0 i for η ∈ {o? : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), I0 , , I, no-up, τ },
then one of the following holds:
(a) upd(N ) = upd(N 0 ) and η ∈ {o∗ : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), τ }, or
η
(b) hI, upd(N )i −
→ hI0 , N 00 i such that upd(N 0 ) = upd(N 00 ).
DYNAMIC CHOREOGRAPHIES
33
Proof.
(1) Applying the upd function corresponds to perform weak transitions, namely transitions
with labels in {o∗ : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), τ }. Some of such transitions
may not be enabled yet. Hence, N may perform the subset of the weak transitions
above which are or become enabled, reducing to some N 000 . Then, η is enabled also in
η
N 000 and we have hI, N 000 i −
→ hI0 , N 00 i. At this point we have that N 00 and N 0 may differ
only for the weak transitions that were never enabled, which can be executed by upd.
(2) There are two cases. In the first case the transition with label η is one of the transitions
executed by function upd. In this case the condition 2a holds. In the second case, the
transition with label η is not one of the transitions executed by function upd. In this
case the transition with label η is still enabled in upd(N ) and can be executed. This
leads to a network that differs from N 0 only because of transitions executed by the upd
function and case 2b holds.
√
We now prove a property of transitions with label .
√
Lemma 7.22. For each DIOC system hΣ, I, Ii that reduces with a transition labelled
then, for√each role R ∈ roles(I), the DPOC role (π(I, R), ΣR )R can reduce with a transition
labelled
and vice versa.
√
Proof. Note that a DIOC can perform a transition with label only if it is a term obtained
using sequential and/or parallel composition starting from 1 constructs. The projection has
the same shape, hence it can perform the desired transition. The other direction is similar.
We can now prove our main theorem (Theorem 7.2, restated below) for which, given a
connected well-annotated DIOC process I and a state Σ, the DPOC network obtained as
its projection has the same behaviours of I.
Theorem 7.2 (Correctness). For each initial, connected DIOC process I, each state Σ,
each set of updates I, the DIOC system hΣ, I, Ii and the DPOC system hI, proj(I, Σ)i are
weak system bisimilar.
Proof. We prove that the relation R below is a weak system bisimulation.
upd(N ) = proj(I, Σ),
events(I) ⊆ events(compl(N )),
R=
(hΣ, I, Ii , hI, N i)
∀
ε1 , ε2 ∈ events(I) .
ε1 ≤DIOC ε2 ⇒ ε1 ≤DP OC ε2 ∨ ε1 ≤DP OC ε2
where I is obtained from a well-annotated connected DIOC via 0 or more transitions
and upd(N ) is a well-annotated DPOC.
To ensure that proving that the relation above is a weak system bisimulation implies
our thesis, let us show that the pair (hΣ, I, Ii , hI, proj(I, Σ)i) from the theorem statement
belongs to R. Note that here I is well-annotated and connected, and for each such
I we have upd(proj(I, Σ)) = proj(I, Σ). From Lemma 7.20 proj(I, Σ) is well-annotated,
thus upd(proj(I, Σ)) is well-annotated. Observe that compl is the identity on proj(I, Σ),
thus from Lemma 7.15 we have that the conditions events(I) ⊆ events(compl(N )) and
∀ε1 , ε2 ∈ events(I) . ε1 ≤DIOC ε2 ⇒ ε1 ≤DP OC ε2 ∨ ε1 ≤DP OC ε2 are satisfied.
34
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
We now prove that R is a weak system bisimulation. To prove it, we show below that
it is enough to consider only the case in which N (and not upd(N )) is equal to proj(I, Σ).
Furthermore, in this case the transition of hΣ, I, Ii is matched by the first transition of
hI, proj(I, Σ)i.
Formally, for each (hΣ, I, Ii , hI, N i) where N = proj(I, Σ) we have to prove the following
simplified bisimulation clauses.
µ
µ
• if hΣ, I, Ii −
→ hΣ00 , I00 , I 00 i then hI, N i −
→ hI00 , N 000 i with (hΣ00 , I00 , I 00 i , hI00 , N 000 i) ∈ R;
√
η
• if hI, N i −
→ hI00 , N 000 i with η ∈ {o : R1 (v) → R2 (x); ; I; no-up; I00 ; τ } then
η
hΣ, I, Ii −
→ hΣ00 , I00 , I 00 i and (hΣ00 , I00 , I 00 i , hI00 , N 000 i) ∈ R.
µ
In fact, consider a general network Ng with upd(Ng ) = proj(I, Σ). If hΣ, I, Ii −
→ hΣ00 , I00 , I 00 i,
µ
then by hypothesis hI, upd(Ng )i −
→ hI00 , N 000 i. From Lemma 7.21 case 1 there exists N 00 such
η1
ηk µ
that hI, Ng i −→ . . . −→−
→ hI00 , N 00 i where ηi ∈ {o∗ : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), τ }
for each i ∈ {1, . . . , k} and upd(N 00 ) = upd(N 000 ). By hypothesis (hΣ00 , I00 , I 00 i , hI00 , N 000 i) ∈ R,
hence, by definition of R, upd(N 000 ) = proj(I 00 , Σ00 ), and therefore also upd(N 00 ) = proj(I 00 , Σ00 ).
The conditions on events hold by hypothesis since function upd has no effect on DPOC
events corresponding to DIOC events. Furthermore, only enabled interactions have been
executed, hence dependencies between DPOC events corresponding to DIOC events are
untouched.
√
η
If instead hI, Ng i −
→ hI00 , N 000 i with η ∈ {o? : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), , I,
no-up, I00 , τ } then thanks to Lemma 7.21 we have one of the following: (2a) upd(Ng ) =
η
upd(N 000 ) and η ∈ {o∗ : R1 (v) → R2 (x), o∗ : R1 (X) → R2 (), τ }, or (2b) hI, upd(Ng )i −
→
η
hI00 , N 00 i such that upd(N 000 ) = upd(N 00 ). In case (2b) we have hI, upd(Ng )i −
→ hI00 , N 00 i.
η
Then, by hypothesis, we have hΣ, I, Ii −
→ hΣ00 , I00 , I 00 i and (hΣ00 , I00 , I 00 i , hI00 , N 00 i) ∈ R. To
deduce that (hΣ00 , I00 , I 00 i , hI00 , N 000 i) ∈ R, one can proceed using the same strategy as the
case of the challenge from the DIOC above. In case (2a) the step is matched by the DIOC
by staying idle, following the second option in the definition of weak system bisimulation.
The proof is similar to the one above.
Thus, we have to prove the two simplified bisimulation clauses above. The proof is
by structural induction on the DIOC I. All the subterms of a well-annotated connected
DIOC are well-annotated and connected, thus the induction can be performed. We√consider
both challenges from the DIOC (→) and from the DPOC (←). The case for label follows
from Lemma 7.22. The case for labels I is trivial. Let us consider the other labels, namely
o : R1 (v) → R2 (x), I, no-up, and τ .
Note that no transition (at the DIOC or at the DPOC level) with one of these labels
can change the set of updates I. Thus, in the following, we will not write it. Essentially, we
will use DIOC processes and DPOC networks instead of DIOC systems and DPOC systems,
respectively. Note that DPOC networks also include the state, while this is not the case
for DIOC processes. For DIOC processes, we assume to associate to them the state Σ, and
comment on its changes whenever needed.
Case 1, 0: trivial.
Case i: x@R = e: the assignment changes the global state in the DIOC, and its projection
on the role R changes the local state of the role in the DPOC in a corresponding way.
Case i: o : R1 (e) → R2 (x): trivial. Just note that at the DPOC level the interaction gives
rise to one send and one receive with the same operation and prefixed by the same index.
DYNAMIC CHOREOGRAPHIES
35
Synchronisation between send and receive is performed by Rule bDPOC |Synch e that also
removes the index from the label.
Case I; I 0 : from the definition of the projection function we have that N =kR∈roles(I;I 0 )
(π(I, R); π(I 0 , R), ΣR )R .
µ
→: Assume that I; I 0 −
→ I 00 with µ ∈ {o : R1 (v) → R2 (x); I; no-up, τ }. There are two
µ
possibilities: either (i ) I −
→ I 000 and I 00 = I 000 ; I 0 or (ii ) I has a transition with label
√
µ
and I 0 −
→ I 00 .
In case (i ) by inductive hypothesis
µ
kR∈roles(I) (π(I, R), ΣR )R −
→ N 000 and upd(N 000 ) =kR∈roles(I) π(I 000 , R), Σ0R R
Thus
µ
kR∈roles(I) π(I, R); π(I 0 , R), ΣR R −
→ N and
000
0
upd(N ) =kR∈roles(I) π(I , R); π(I , R), Σ0R R
If roles(I 0 ) ⊆ roles(I) then the thesis follows. If roles(I 0 ) 6⊆ roles(I) then at the DPOC
level the processes in the roles in roles(I 0 ) \ roles(I) are not affected by the transition.
Note however that the projection of I on these roles is a term composed only by 1s,
and the ones corresponding to parts of I that have been consumed can be removed
by the clean part of function upd.
√
µ
In case (ii ), I has a transition with label and I 0 −
→ I 00 . By inductive hypothesis
µ
proj(I 0 , Σ) −
→ N 00 and upd(N 00 ) = proj(I 00 , Σ0 ). The thesis follows since, thanks to
µ
Lemma 7.22, proj(I; I 0 , Σ) −
→ N and upd(N ) = proj(I 00 , Σ0 ), possibly using the clean
part of function upd to remove the 1s which are no more needed.
Note that, in both the cases, conditions on events follow by inductive hypothesis.
←: Assume that
η
N =kR∈roles(I;I 0 ) π(I, R); π(I 0 , R), ΣR R −
→kR∈roles(I;I 0 ) PR , Σ0R R
with η ∈ {o : R1 (v) → R2 (x), I, no-up, τ }. We have a case analysis on η.
If η = o : R1 (v) → R2 (x) then
i.ohvi@R2 :R1
π(I; I 0 , R1 ), ΣR1 R1 −−−−−−−−→ (PR1 , ΣR1 )R1 and
π(I; I 0 , R2 ), ΣR2
i.o(x←v)@R1 :R2
R2
−−−−−−−−−−−→ (PR2 , ΣR2 )R2
The two events have the same global index since they have the same index i (otherwise
they could not synchronise) and they are both outside of any while loop (since they
are enabled), hence the global index coincides with the index. Thus, they are either
both from I or both from I 0 .
In the first case we have also
o:R1 (v)→R2 (x)
00
kR∈roles(I;I 0 ) (π(I, R), ΣR )R −−−−−−−−−−→kR∈roles(I;I 0 ) PR
, ΣR R
o:R1 (v)→R2 (x)
00 ; π(I 0 , R). Thus, by inductive hypothesis, I −
with PR = PR
−−−−−−−−−→ I 00 and
00
00
upd(kR∈roles I;I 0 (PR , ΣR )R ) = proj(I , Σ). Hence, we have that
o:R1 (v)→R2 (x)
I; I 0 −−−−−−−−−−→ I 00 ; I 0
36
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
and
00
; π(I 0 , R), ΣR R ) = proj(I 00 ; I 0 , Σ)
upd(kR∈roles I;I 0 PR
The thesis follows.
In the second case, we need to show that the interaction o : R1 (v) → R2 (x) is
enabled. Assume that this is not the case. This means that there is a DIOC event
ε corresponding to some construct in I. Because of the definition of R, ε is also a
DPOC event and ε ≤DP OC ξ : o@R2 ∨ ε ≤DP OC ξ : o@R1 . Hence, at least one of the
two events is not minimal and the corresponding transition cannot be enabled, against
our hypothesis. Therefore the interaction o : R1 (v) → R2 (x) is enabled. Thus, I has
√
o:R1 (v)→R2 (x)
a transition with label and I 0 −−−−−−−−−−→ I 00 . Thanks to Lemma 7.22
√ then both
(π(I, R1 ), ΣR1 )R1 and (π(I, R2 ), ΣR2 )R2 have a transition with label . Thus, we
have
i.ohvi@R2 :R1
π(I 0 , R1 ), ΣR1 R1 −−−−−−−−→ (PR1 , ΣR1 )R1
i.o(x←v)@R1 :R2
π(I 0 , R2 ), ΣR2 R2 −−−−−−−−−−−→ (PR2 , ΣR2 )R2 and thus
o:R1 (v)→R2 (x)
proj(I 0 , Σ) −−−−−−−−−−→kR∈roles(I 0 ) (PR , ΣR )R
The thesis follows by inductive hypothesis.
For the other cases of η, all the roles but one are unchanged. The proof of these cases
is similar to the one for interaction, but simpler.
Note that in all the above cases, conditions on events follow by inductive hypothesis.
Case I|I 0 : from the definition of the projection function we have
N =kR∈roles(I;I 0 ) (π(I, R) | π(I 0 , R), ΣR )R
→: We have a case analysis on the rule used to derive the transition. If the transition
is derived using Rule bDIOC |Parallel e and I|I 0 can perform a transition with label
µ then one of its two components can perform a transition with the same label µ
and the thesis follows by inductive hypothesis. Additional roles not occurring in the
term performing the transition are dealt with by the clean part of function upd. If
instead the transition is derived using Rule bDIOC |Par-end e then the thesis follows
from Lemma 7.22.
←: We have a case analysis on the label η of the transition. If η = o? : R1 (v) → R2 (x)
then a send and a receive on the same operation are enabled. The two events have
the same global index since they have the same index i (otherwise they could not
synchronise) and they are both outside of any while loop (since they are enabled),
hence the global index coincides with the index. Thus, they are either both from I or
both from I 0 . The thesis follows by inductive hypothesis. For the other cases of η,
only the process of one role changes. The thesis follows by inductive hypothesis. In
all the cases, roles not occurring in the term performing the transition are dealt with
by function upd.
DYNAMIC CHOREOGRAPHIES
37
Case i: if b@R {I} else {I 0 }: from the definition of projection
N = kS∈roles(I,I 0 )r{R} i? : i.cnd∗i : xi from R; i: if xi {π(I, S)} else {π(I 0 , S)}, ΣS
k
S
Y
iT : i.cnd∗i : true to R0 ; π(I, R)
i: if b
R0 ∈roles(I,I 0 )r{R}
!
Y
iF : i.cnd∗i : false to R0 ; π(I 0 , R) , ΣR
else
R
R0 ∈roles(I,I 0 )r{R}
Let us consider the case when the guard is true (the other one is analogous).
τ
→: The only possible transition from the DIOC is i: if b@R {I} else {I 0 } −
→ I. The
DPOC can match this transition by reducing to
i? : i.cnd∗i : xi from R;
0
N = kS∈roles(I,I 0 )r{R}
, ΣS
k
i: if xi {π(I, S)} else {π(I 0 , S)}
S
Y
iT : i.cnd∗i : true to R0 ; π(I, R), ΣR
R0 ∈roles(I,I 0 )r{R}
R
By applying function upd we get
upd(N 0 ) = kS∈roles(I,I 0 )r{R} (π(I, S), ΣS )S k (π(I, R), ΣR )R
Concerning events, at the DIOC level events corresponding to the guard and to the
discarded branch are removed. The same holds at the DPOC level, thus conditions
on the remaining events are inherited. This concludes the proof.
←: The only possible transition from the DPOC is the evaluation of the guard from
the coordinator. This reduces N to N 0 above and the thesis follows from the same
reasoning.
Case i: while b@R {I}: from the definition of projection
i? : i.wb∗i : xi from R; i: while xi {π(I, S);
N = kS∈roles(I)r{R}
, ΣS
k
iC : i.we∗i : ok to R; i? : i.wb∗i : xi from R}
S
!
Q
iT : i.wb∗i : true to R0 ; π(I, R);
R0 ∈roles(I)r{R}
;
i: while b
Q
iC : i.we∗i : from R0
R0 ∈roles(I)r{R}
Q
iF : i.wb∗i : false to R0 , ΣR
R0 ∈roles(I)r{R}
R
→: Let us consider the case when the guard is true. The only possible transition from
τ
the DIOC is i: while b@R {I} −
→ I; i: while b@R {I}. The DPOC can match this
38
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
transition by reducing to
0
N =kS∈roles(I)r{R}
i? : i.wb∗i : xi from R; i: while xi {π(I, S);
, ΣS
k
iC : i.we∗i : ok to R; i? : i.wb∗i : xi from R}
S
!
Q
iT : i.wb∗i : true to R0 ; π(I, R);
R0 ∈roles(I)r{R}
!
Q
iC : i.we∗i : from R0 ;
R0 ∈roles(I)r{R}
!
Q
∗
0
iT : i.wbi : true to R ; π(I, R);
R0 ∈roles
(I)r{R}
i: while b
Q
iC : i.we∗i : from R0
0
R ∈roles(I)r{R}
Q
iF : i.wb∗i : false to R0 , ΣR
0
R ∈roles(I)r{R}
By applying function upd we get
R0 ∈
;
R
k
π(I, R);
i: while b
Q
π(I, S); i? : i.wb∗i : xi from R;
i: while xi {π(I, S);
, ΣS
iC : i.we∗i : ok to R;
∗
i? : i.wbi : xi from R}
S
upd(N 0 ) =
kS∈roles(I)r{R}
Q
R0 ∈roles
Q (I)r{R}
R0 ∈roles(I)r{R}
roles(I)r{R}
!
iT : i.wb∗i : true to R0 ; π(I, R);
iC : i.we∗i : from R0
iF : i.wb∗i : false to R0
;
, ΣR
R
exactly the projection of I; i: while b@R {I}.
As far as events are concerned, in compl(N 0 ) we have all the needed events since, in
particular, we have already done the unfolding of the while in all the roles. Concerning
the ordering, at the DIOC level, we have two kinds of causal dependencies: (1) events
in the unfolded process precede the guard event; (2) the guard event precedes the
events in the body. The first kind of causal dependency is matched at the DPOC
level thanks to the auxiliary synchronisations that close the unfolded body (which
are not removed by compl) using synchronisation and sequentiality. The second kind
of causal dependency is matched thanks to the auxiliary synchronisations that start
the following iteration using synchronisation, sequentiality and while.
The case when the guard evaluates to false is simpler.
←: The only possible transition from the DPOC is the evaluation of the guard from
the coordinator. This reduces N to N 0 above and the thesis follows from the same
reasoning.
DYNAMIC CHOREOGRAPHIES
39
Case i: scope @R {I}: from the definition of the projection
N = kR0 ∈roles(I)r{R} i: scope @R {π(I, R0 )}, ΣR0 R0 k
(i: scope @R {π(I, R)} roles {roles(I)}, ΣR )R
→: Let us consider the case when the scope is updated. At the DIOC level all the
possible transitions have label of the form I 0 and are obtained by applying Rule
bDIOC |Up e. Correspondingly, at the DPOC level one applies Rule bDPOC |Lead-Up e to
the coordinator of the update, obtaining
N 0 = kR0 ∈roles(I)r{R} i: scope @R {π(I, R0 )}, ΣR0 R0 k
!
Q
∗
0
0
0
i.sbi : π(I , R ) to R ;
R0 ∈roles(I)r{R}
, ΣR
π(I 0 , R);
Q
i.se∗i : from R0
R0 ∈roles(I)r{R}
R
By applying the upd function we get:
upd(N 0 ) = kR0 ∈roles(I)r{R} π(I 0 , R0 ), ΣR0 R0 k π(I 0 , R), ΣR R
This is exactly the projection of the DIOC obtained after applying the rule bDIOC |Up e.
The conditions on events are inherited. Observe that the closing event of the scope is
replaced by events corresponding to the auxiliary interactions closing the scope. This
allows us to preserve the causality dependencies also when the scope is inserted in a
context.
The case of Rule bDIOC |NoUp e is simpler.
←: The only possible transitions from the DPOC are the ones of the coordinator of the
update checking whether to apply an update or not. This reduces N to N 0 above
and the thesis follows from the same reasoning.
7.2. Deadlock freedom, termination, and race freedom. Due to the fact that the
projection preserves weak traces, we have that trace-based properties of the DIOC are
inherited by the DPOC. A first example of such properties is deadlock freedom.
Definition 7.23 (Deadlock freedom). An internal DIOC (resp. DPOC) trace is obtained
by removing transitions labelled I from a DIOC (resp. DPOC) trace. √
A DIOC (resp. DPOC)
system is deadlock free if all its maximal finite internal traces have as last label.
Intuitively, internal traces are needed since labels I do not correspond to activities of
the
application
and may be executed also after application termination. The fact that after
√
a only changes in the set of available updates are possible is captured by the following
lemma.
Lemma √7.24. For each initial, connected DIOC I, state Σ, and set of updates I, if
hΣ, I, Ii −
→ hΣ0 , I0 , I 0 i then each transition of hΣ0 , I0 , I 0 i has label I00 for some I00 .
Proof. The proof is by case analysis on the rules which can derive a transition with label
All the cases are easy.
√
.
40
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Since by construction initial DIOCs are deadlock free we have that also the DPOC
obtained by projection is deadlock free.
Corollary 7.25 (Deadlock freedom). For each initial, connected DIOC I, state Σ, and set
of updates I the DPOC system hI, proj(I, Σ)i is deadlock free.
Proof. Let us first prove that for each initial, connected DIOC I, state Σ, and set of updates
I, the DIOC system hΣ, I, Ii
√ is deadlock free. This amounts to prove that all its maximal
finite internal traces have as last label. For each trace, the proof is by induction on its
length, and for each length by structural induction on I. The proof is based on the fact
that I is initial. The induction considers a reinforced hypothesis, saying also that:
√
•
may occur only as the last label of the internal trace;
• all the DIOC systems in the sequence of transitions generating the trace, but the last one,
are initial.
We have a case analysis on the top-level operator in I. Note that in all the cases, but 0, at
least a transition is derivable.
Case 0: not allowed since we assumed an initial DIOC.
√
Case 1: trivial because by Rule bDIOC |End e and Lemma 7.24 its only internal trace is .
Case x@R = e: the only applicable rule is bDIOC |Assign e that in one step leads to a 1
process. The thesis follows by inductive hypothesis on the length of the trace.
Case o? : R1 (e) → R2 (x): the only applicable rule is bDIOC |Interaction e, which leads to an
assignment. Then the thesis follows by inductive hypothesis on the length of the trace.
Case I; I 0 : the first transition can be derived either by Rule bDIOC |Sequence e or Rule
bDIOC |Seq-end e. In the first case the thesis follows by induction on the length of the
trace. In the second case the trace coincides with a trace of I 0 , and the thesis follows by
structural induction.
Case I|I 0 : the first transition can be derived either by Rule bDIOC |Parallel e or by Rule
[Par-End]. In the first case the thesis follows by induction on the length
√ of the trace.
In the second case the thesis follows by Lemma 7.24, since the label is .
Case if b@R {I} else {I 0 }: the first transition can be derived using either Rule
bDIOC |If-then e or Rule bDIOC |If-else e. In both the cases the thesis follows by induction
on the length of the trace.
Case while b@R {I}: the first transition can be derived using either Rule
bDIOC |While-unfold e or Rule bDIOC |While-exit e. In both the cases the thesis follows by
induction on the length of the trace.
Case scope @R {I}: the first transition can be derived using either Rule bDIOC |Up e or
Rule bDIOC |NoUp e. In both the cases the thesis follows by induction on the length of the
trace.
The weak internal traces of the DIOC coincide with the weak internal traces of the√DPOC
by Theorem 7.2, thus also the finite weak internal traces of √
the DPOC end with . The
same holds for the finite strong internal traces, since label
is preserved when
√ moving
between strong and weak traces, and no transition can be added after the
thanks to
Lemma 7.24.
DPOCs also inherit termination from terminating DIOCs.
Definition 7.26 (Termination). A DIOC (resp. DPOC) system terminates if all its internal
traces are finite.
DYNAMIC CHOREOGRAPHIES
41
Note that if arbitrary sets of updates are allowed, then termination of DIOCs that
contain at least a scope is never granted. Indeed, the scope can always be replaced by a
non-terminating update or it can trigger an infinite chain of updates. Thus, to exploit this
result, one should add constraints on the set of updates ensuring DIOC termination.
Corollary 7.27 (Termination). If the DIOC system hΣ, I, Ii terminates and I is connected
then the DPOC system hI, proj(I, Σ)i terminates.
Proof. It follows from the fact that only a finite number of auxiliary actions are added when
moving from DIOCs to DPOCs.
Other interesting properties derived from weak trace equivalence are freedom from races
and orphan messages. A race occurs when the same receive (resp. send) may interact with
different sends (resp. receives). In our setting, an orphan message is an enabled send that is
never consumed by a receive. Orphan messages are more relevant in asynchronous systems,
where a message may be sent, and stay forever in the network, since the corresponding
receive operation may never become enabled. However, even in synchronous systems
orphan messages should be avoided: the message is not communicated since the receive
is not available, hence a desired behaviour of the application never takes place due to
synchronisation problems. Trivially, DIOCs avoid races and orphan messages since send and
receive are bound together in the same construct. Differently, at the DPOC level, since all
receive of the form ι: i.o? : x from R1 in role R2 may interact with the sends of the form
ι: i.o? : e to R2 in role R1 , races may happen. However, thanks to the correctness of the
projection, race freedom holds also for the projected DPOCs.
Corollary 7.28 (Race freedom). For each initial, connected DIOC I, state Σ, and set
η1
ηn
0
?
∗
of updates I, if hI,
√ proj(I, Σ)i −→ · · · −→ hI , N i, where ηi ∈ {τ, o : R1 (v) → R2 (x), o :
R1 (X) → R2 (), , I, no-up, I} for each i ∈ {1, . . . , n}, then in N there are no two sends
(resp. receives) which can interact with the same receive (resp. send).
Proof. We have two cases, corresponding respectively to programmer-specified and auxiliary
operations.
For programmer-specified operations, thanks to Lemma 7.20, case C1, for each global
index ξ there are at most two communication events with global index ξ. The corresponding
DPOC terms can be enabled only if they are outside of the body of a while loop. Hence,
their index coincides with their global index. Since the index prefixes the operation, then no
interferences with other sends or receives are possible.
For auxiliary operations, the reasoning is similar. Note, in fact, that sends or receives with
the same global index can be created only by a unique DIOC construct, but communications
between the same pair of roles are never enabled together. This can be seen by looking at
the definition of the projection.
As far as orphan messages are concerned, they may appear in infinite DPOC computations since a receive may not become enabled due to an infinite loop. However, as a corollary
of trace equivalence, we have that terminating DPOCs are orphan-message free.
Corollary 7.29 (Orphan-message freedom). For each
initial, connected DIOC I, state Σ,
√
η1
ηn
and set of updates I, if hI,√
proj(I, Σ)i −→ · · · −→−
→ hI0 , N i, where ηi ∈ {τ, o? : R1 (v) →
∗
R2 (x), o : R1 (X) → R2 (), , I, no-up, I}, then N contains no sends.
√
Proof. The proof is by case analysis on the rules which can derive a transition with label .
All the cases are easy.
42
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
8. Adaptable Interaction-Oriented Choreographies in Jolie
In this section we present AIOCJ (Adaptable Interaction-Oriented Choreographies in Jolie),
a development framework for adaptable distributed applications [28]. AIOCJ is one of the
possible instantiations of the theoretical framework presented in the previous sections and
it gives a tangible proof of the expressiveness and feasibility of our approach. We say that
AIOCJ is an instance of our theory because it follows the theory, but it provides mechanisms
to resolve the non-determinism related to the choice of whether to update or not and on
which update to select. Indeed, in AIOCJ updates are chosen and applied according to the
state of the application and of its running environment. To this end, updates are embodied
into adaptation rules, which specify when and whether a given update can be applied, and to
which scopes. AIOCJ also inherits all the correctness guarantees provided by our theory, in
particular: i ) applications are free from deadlocks and races by construction, ii ) applications
remain correct after any step of adaptation. As in the theory, adaptation rules can be added
and removed while applications are running.
Remark 8.1. In order for the correctness guarantees to be provided, one needs to satisfy
the required assumptions, in particular the fact that functions never block and always return
a value (possibly an error notification). If this condition is not satisfied, at the theoretical
level, the behaviour of the source DIOC and the behaviour of its projection may differ,
as described in Remark 2.2. The behaviour of the corresponding AIOCJ application can
be different from both of them. Usually, the violation of the assumption above leads the
application to crash. This allows the programmer to realise that the assumption has been
violated.
Below we give a brief overview of the AIOCJ framework by introducing its components:
the Integrated Development Environment (IDE), the AIOCJ compiler, and the Runtime
Environment.
Integrated Development Environment. AIOCJ supports the writing of programs and adaptation rules in the Adaptable Interaction-Oriented Choreography (AIOC) language, an
extension of the DIOC language. We discuss the main novelties of the AIOC language in
§ 8.1. AIOCJ offers an integrated environment for developing programs and adaptation
rules that supports syntax highlighting and on-the-fly syntax checking. Since checking
for connectedness (see § 6) of programs and adaptation rules is polynomial (as proven by
Theorem 6.3), the IDE also performs on-the-fly checks on connectedness of programs and
rules.
Compiler . The AIOCJ IDE also embeds the AIOCJ compiler, which implements the procedure for projecting AIOCs and adaptation rules into distributed executable code. The
implementation of the compiler is based on the rules for projecting DIOCs, described in § 4.
The target language of the AIOCJ compiler is Jolie [26, 35], a Service-Oriented language
with primitives similar to those of our DPOC language. Jolie programs are also called
services. Given an AIOC program, the AIOCJ compiler produces one Jolie service for each
role in the source AIOC. The compilation of an AIOC rule produces one Jolie service for
each role and an additional service that describes the applicability condition of the rule. All
these services are enclosed into an Adaptation Server, described below.
DYNAMIC CHOREOGRAPHIES
43
Runtime Environment. The AIOCJ runtime environment comprises a few Jolie services
that support the execution and adaptation of compiled programs. The main services of
the AIOCJ runtime environment are the Adaptation Manager, Adaptation Servers, and the
Environment. The compiled services interact both among themselves and with an Adaptation
Manager, which is in charge of managing the adaptation protocol. Adaptation Servers
contain adaptation rules, and they can be added or removed dynamically, thus enabling
dynamic changes in the set of rules, as specified by Rule bDIOC |Change-Updates e. When
started, an Adaptation Server registers itself at the Adaptation Manager. The Adaptation
Manager invokes the registered Adaptation Servers to check whether their adaptation rules
are applicable. In order to check whether an adaptation rule is applicable, the corresponding
Adaptation Server evaluates its applicability condition. Applicability conditions may refer
to the state of the role which coordinates the update, to properties of the scope, and to
properties of the environment (e.g., time, temperature, etc.), stored in the Environment
service.
In the remainder of this Section we detail the grammar of the AIOC language used by
AIOCJ in § 8.1, we illustrate the use of AIOCJ on a simple example in § 8.2, we discuss
some relevant implementation aspects of AIOCJ in § 8.3, and we present some guidelines on
how to use scopes in AIOCJ in § 8.4.
8.1. DIOC Language Extensions in AIOCJ. AIOCs extend DIOCs with:
• the definition of adaptation rules, instead of updates, that include the information needed
to evaluate their applicability condition;
• the definition of constructs to express the deployment information needed to implement
real-world distributed applications.
Below we describe in detail, using the Extended Backus-Naur Form [3], the new or refined
constructs introduced by the AIOC language.
Function inclusions. The AIOC language can exploit functionalities provided by external
services via the include construct. The syntax is as follows.
Include ::= include FName [ , FName ]* from " URL " [ with PROTOCOL ]
This allows one to reuse existing legacy code and to interact with third-party external
applications. As an example, the Seller of our running example can exploit an external
database to implement the functionality for price retrieval getPrice, provided that such a
functionality is exposed as a service. If the service is located at "socket://myService:8000"
and accessible via the "SOAP" protocol we enable its use with the following inclusion:
include getPrice from " socket :// myService :8000 " with " SOAP "
Similarly, the Bank IT system is integrated in the example by invoking the function
makePayment, also exposed as a service.
This feature enables a high degree of integration since AIOCJ supports all protocols
provided by the underlying Jolie language, which include TCP/IP, RMI, SOAP, XML/RPC,
and their encrypted correspondents over SSL.
External services perfectly fit the theory described in previous sections, since they are
seen as functions, and thus introduced in expressions. Notice, in particular, that Remark 8.1
applies.
Adaptation rules. Adaptation rules extend DIOC updates, and are a key ingredient of
the AIOCJ framework. The syntax of adaptation rules is as follows:
44
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Rule ::= rule {
[ Include ]*
on { Condition }
do { Choreography }
}
The applicability condition of the adaptation rule is specified using the keyword on,
while the code to install in case adaptation is performed (which corresponds to the DIOC
update) is specified using the keyword do. Optionally, adaptation rules can include functions
they rely on.
The Condition of an adaptation rule is a propositional formula which specifies when the
rule is applicable. To this end, it can exploit three sources of information: local variables of
the coordinator of the update, environmental variables, and properties of the scope to which
the adaptation rule is applied. Environmental variables are meant to capture contextual
information that is not under the control of the application (e.g., temperature, time, available
resources, . . . ). To avoid ambiguities, local variables of the coordinator are not prefixed,
environment variables are prefixed by E, and properties of the scope are prefixed by N.
For example, if we want to apply an adaptation rule only to those scopes whose property
name is equal to the string "myScope" we can use as applicability condition the formula
N.name == "myScope".
Scopes. As described above, scopes in AIOC also feature a set of properties (possibly
empty). Scope properties describe the current implementation of the scope, including both
functional and non-functional properties. Such properties are declared by the programmer,
and the system only uses them to evaluate the applicability condition of adaptation rules,
to decide whether a given adaptation rule can be applied to a given scope. Thus the syntax
for scopes in AIOC is:
scope @ Role { Choreography }
[ prop { Properties } ]?
where clause prop introduces a list of comma-separated assignments of the form
N.ID = Expression.
For instance, the code
scope @ R {
// AIOC code
} prop { N . name = " myScope " }
specifies that the scope has a property name set to the string "myScope", thus satisfying the
applicability condition N.name == "myScope" discussed above.
Programs. AIOC programs have the following structure.
Program ::=
[ Include ]*
preamble {
starter : Role
[ Location ]*
}
aioc { Choreography }
DYNAMIC CHOREOGRAPHIES
45
where Include allows one to include external functionalities, as discussed above, and
keyword aioc introduces the behaviour of the program, which is a DIOC apart for the fact
that scopes may define properties, as specified above.
The keyword preamble introduces deployment information, i.e., the definition of the
starter of the AIOC and the Location of participants.
The definition of a starter is mandatory and designs which role is in charge of waiting
for all other roles to be up and running before starting the actual computation. Any role
can be chosen as starter, but the chosen one needs to be launched first when running the
distributed application.
Locations define where the participants of the AIOC will be deployed. They are specified
using the keyword location:
Location ::= location @ Role : " URL "
where Role is the name of a role (e.g., Role1) and URL specifies where the service
can be found (e.g., "socket://Role1:8001"). When not explicitly defined, the projection
automatically assigns a distinct local TCP/IP location to each role.
8.2. AIOCJ Workflow. Here we present a brief description of how a developer can write
an adaptable distributed system in AIOCJ, execute it, and change its behaviour at runtime
by means of adaptation rules. For simplicity, we reuse here the minimal example presented
in the Introduction, featuring a scope that encloses a price offer from the Seller to the
Buyer and an update — here an adaptation rule — that provides a discount for the Buyer.
We report the AIOC program in the upper part and the adaptation rule in the lower part of
Figure 13.
At Line 1 of the AIOC program we have the inclusion of function getPrice, which is
provided by a service within the internal network of the Seller, reachable as a TCP/IP node
at URL "storage.seller.com" on port "80". At Line 2 of the AIOC program we have the
preamble. The preamble specifies deployment information, and, in particular, defines the
starter, i.e., the service that ensures that all the participants are up and running before
starting the actual computation. No locations are specified, thus default ones are used.
The actual code is at Lines 4–7, where we declare a scope. At Line 7 we define a property
scope_name of this scope with value "price_inquiry".
1 and execution ○
2 of the AIOC. From
Figure 14 depicts the process of compilation ○
left to right, we write the AIOC and we compile it into a set of executable Jolie services
(Buyer service and Seller service). To execute the projected system, we first launch the
Adaptation Manager and then the two compiled services, starting from the Seller, which is
the starter. Since there is no compiled adaptation rule, the result of the execution is the
offering of the standard price to the Buyer.
Now, let us suppose that we want to adapt our system to offer discounts in the Fall
season. To do that, we can write the adaptation rule shown in the lower part of Figure 13.
Since we want to replace the scope for the price_inquiry, we define, at Line 4, that this
rule applies only to scopes with property scope_name set to "price_inquiry". Furthermore,
since we want the update to apply only in the Fall season, we also specify that the environment
variable E.season should match the value "Fall". The value of E.season is retrieved from the
Environment Service.
The rule uses two functions. At Line 2, it includes function getPrice, which is the one
used also in the body of the scope. The other function, isValid, is included at Line 3 and
46
1
2
3
4
5
6
7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
include getPrice from " socket :// storage . seller . com :80 "
preamble { starter : Seller }
aioc {
scope @ Seller {
order_price @ Seller = getPrice ( order ) ;
offer : Seller ( order_price ) -> Buyer ( prod_price )
} prop { N . scope_name = " price_inquiry " } }
rule {
include getPrice from " socket :// storage . seller . com :80 "
include isValid from " socket :// discounts . seller . com :80 "
on { N . scope_name == " price_inquiry " and E . season == " Fall " }
do {
cardReq : Seller ( _ ) -> Buyer ( _ ) ;
card_id @ Buyer = getInput ( " Insert your customer card ID " ) ;
cardRes : Buyer ( card_id ) -> Seller ( buyer_id ) ;
if isValid ( buyer_id ) @ Seller {
order_price @ Seller = getprice ( order )*0.9 ;
} else {
order_price @ Seller = getPrice ( order )
};
offer : Seller ( order_price ) -> Buyer ( prod_price )
}
}
Figure 13: An AIOC program (upper part) and an applicable adaptation rule (lower part).
it is provided by a different TCP/IP node, located at URL "discounts.seller.com" on port
"80" within the internal network of the Seller.
The body of the adaptation rule (Lines 6–14) specifies the same behaviour described
in Figure 1 in the Introduction: the Seller asks the Buyer to provide its card_id, which
the Seller uses to provide a discount on the price of the ordered product. We depict
the inclusion of the new adaptation rule (outlined with dashes) and the execution of the
3 of Figure 14. From right to left, we write the rule and we compile
adaptation at point ○
it. The compilation of a (set of) adaptation rule(s) in AIOCJ produces a service, called
Adaptation Server (also outlined with dashes), that the Adaptation Manager can query to
fetch adaptation rules at runtime. The compilation of the adaptation rule can be done while
the application is running. After the compilation, the generated Adaptation Server is started
and registers on the Adaptation Manager. Since the rule relies on the environment to check
its applicability condition, we also need the Environment service to be running. In order for
the adaptation rule to apply we need the environment variable season to have value "Fall".
8.3. Implementation. AIOCJ is composed of two elements: the AIOCJ Integrated Development Environment (IDE), named AIOCJ − ecl, and the adaptation middleware that
enables AIOC programs to adapt, called AIOCJ − mid.
DYNAMIC CHOREOGRAPHIES
47
Runtime Environment
2
Adaptation
Manager
Compilation on
role Buyer
AIOC
Buyer service
3
Adaptation
Server
1
Compilation
Adaptation
Rules
Compilation on
role Seller
AIOC Language
Seller service
Environment
service
Jolie Language
AIOC Language
Figure 14: Representation of the AIOCJ framework — Projection and execution of the
example in Figure 13.
Figure 15: Check for connectedness.
AIOCJ − ecl is a plug-in for Eclipse [21] based on Xtext [49]. Starting from a grammar,
Xtext generates the parser for programs written in the AIOC language. Result of the parsing
is an abstract syntax tree (AST) we use to implement i ) the checker of connectedness for
AIOC programs and adaptation rules and ii ) the generation of Jolie code for each role. Since
the check for connectedness has polynomial computational complexity (cf. Theorem 6.3)
it is efficient enough to be performed while editing the code. Figure 15 shows AIOCJ − ecl
notifying the error on the first non-connected instruction (Line 13).
48
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
As already mentioned, we chose Jolie as target language of the compilation of AIOCJ
because its semantics and language constructs naturally lend themselves to translate our
theoretical results into practice. Indeed, Jolie supports architectural primitives like dynamic
embedding, aggregation, and redirection, which ease the compilation of AIOCs.
Each scope at the AIOC level is projected into a specific sub-service for each role. The
roles run the projected sub-services by embedding them and access them via redirection. In
this way, we implement adaptation by disabling the default sub-service and by redirecting
the execution to a new one, obtained from the Adaptation server.
When at runtime the coordinator of the update reaches the beginning of that scope, it
queries the Adaptation Manager for adaptation rules to apply. The Adaptation Manager
queries each Adaptation Server sequentially, based on their order of registration. On its
turn, each Adaptation Server checks the applicability condition of each of its rules. The first
rule whose applicability condition holds is applied. The adaptation manager sends to the
coordinator the updates that are then distributed to the other involved roles. In each role,
the new code replaces the old one.
To improve the performance, differently from the theory, in AIOCJ adaptation rules are
compiled statically, and not by the coordinator of the update when the update is applied.
8.4. AIOCJ Practice. AIOCJ provides scopes as a way to specify which parts of the
application can be updated. As a consequence, deciding which parts of the code to enclose
in scopes is a relevant and non trivial decision. Roughly, one should enclose into scopes parts
of the code that may need to be updated in the future. This includes for instance parts of
the code which model business rules, hence may change according to business needs, parts
of the code which are location dependent, hence may be updated to configure the software
for running in different locations, parts of the code which are relevant for performance or
security reasons, hence may be updated to improve the performance or the security level.
One may think that a simple shortcut would be to have either a big scope enclosing the
whole code of the application or many small scopes covering all the code, but both these
solutions have relevant drawbacks. Indeed, if one replaces a big scope in order to change a
small piece of it, (s)he has to provide again all the code, while one would like to provide only
the part that needs to change. Furthermore, update for the big scope is no more available
as soon as its execution starts, which may happen quite earlier than when the activity to
be updated starts. Using small scopes also is not a solution, since an update may need to
cover more than one such scope, but there is no way to ensure that multiple scopes are
updated in a coordinated way. Also, both the approaches have performance issues: in the
first case large updates have to be managed, in the second case many checks for the updates
are needed, causing a large overhead.
In order to write updates for a given scope, one needs to have some knowledge about
the scope that the update is going to replace, since the update will run in the context where
the scope is. This includes, for instance, knowing which variables are available, their types
and their meaning, and in which variables results are expected. This information can be
either tracked in the documentation of the application, or added to the scope properties.
In this way, the adaptation rule embodying the update can check this information before
applying the update (see for instance the approach of [29]). The fact that this information
is needed for all scopes is another point against the approach of using many small scopes.
DYNAMIC CHOREOGRAPHIES
49
9. Related work and discussion
Recently, languages such as Rust [44] or SCOOP [41] have been defined to provide highlevel primitives to program concurrent applications avoiding by construction some of the
risks of concurrent programming. For instance, Rust ensures that there is exactly one
binding to any given resource using ownership and borrowing mechanisms to achieve memory
safety. SCOOP instead provides the separate keyword to enable programming patterns
that enforce data-race freedom. In industry, Rust has started to be used for the development
of complex concurrent/distributed systems (e.g., the file storage system of Dropbox [33] or
the parallel rendering engine of Mozilla [51]). However, industry has not yet widely adopted
any language-based solution trading expressive power of the language for safety guarantees.
Inspired by the languages above, we have presented high-level primitives for the dynamic
update of distributed applications. We guarantee the absence of communication deadlocks
and races by construction for the running distributed application, even in presence of updates
that were unknown when the application was started. More generally, the compilation of a
DIOC specification produces a set of low-level DPOCs whose behaviour is compliant with
the behaviour of the originating DIOC.
As already remarked, our theoretical model is very general. Indeed, whether to update a
scope or not and which update to apply if many are available is completely non deterministic.
Choosing a particular policy for decreasing the non-determinism of updates is orthogonal
with respect to our results. Our properties are preserved by any such policy, provided that
the same policy is applied both at the DIOC and at the DPOC level. AIOCJ presents
a possible instantiation of our general approach with a concrete mechanism for choosing
updates according to the state of the application and of its running environment.
Our work is on the research thread of choreographies and multiparty session types [12,
24, 13, 14, 5, 45]. One can see [25] for a description of the field. Like choreographies and
multiparty session types, DIOCs target communication-centred concurrent and distributed
applications, avoiding deadlocks and communication races. Thanks to the particular structure
of the language, DIOCs provide the same guarantees without the need of types. Below we
just discuss the approaches in this research thread closest to ours. The two most related
approaches we are aware of are based on multiparty session types, and deal with dynamic
software updates [2] and with monitoring of self-adaptive systems [16]. The main difference
between [2] and our approach is that [2] targets concurrent applications which are not
distributed. Indeed, it relies on a check on the global state of the application to ensure
that the update is safe. Such a check cannot be done by a single role, thus it is impractical
in a distributed setting. Furthermore, the language in [2] is much more constrained than
ours, e.g., requiring each pair of participants to interact on a dedicated pair of channels, and
assuming that all the roles that are not the sender or the receiver within a choice behave
the same in the two branches. The approach in [16] is very different from ours, too. In
particular, in [16], all the possible behaviours are available since the very beginning, both at
the level of types and of processes, and a fixed adaptation function is used to switch between
them. This difference derives from the distinction between self-adaptive applications, as
they discuss, and applications updated from the outside, as in our case.
We also recall [20], which uses types to ensure safe adaptation. However, [20] allows
updates only when no session is active, while we change the behaviour of running DIOCs.
We highlight that, contrarily to our approach, none of the approaches above has been
implemented.
50
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Various tools [17, 4, 8] support adaptation exploiting automatic planning techniques in
order to elaborate, at runtime, the best sequence of activities to achieve a given goal. These
techniques are more declarative than ours, but, to the best of our knowledge, they are not
guaranteed to always find a plan to adapt the application.
[38] presents an approach to safe dynamic software updates for concurrent systems.
According to [38], safe means that all the traces are version consistent, namely that for each
trace there is a version of the application able to produce it. In our case, we want the effects
of the update to be visible during the computation, hence we do not want this property to
hold. Instead, for us safe means that the DIOC and the projected DPOC are weak system
bisimilar, and that they are both free from deadlocks and races.
Among the non-adaptive languages, Chor [13] is the closest to ours. Indeed, like ours,
Chor is a choreographic language that compiles to Jolie. Actually, AIOCJ shares part of
the Chor code base. Our work shares with [37] the interest in choreographies composition.
However, [37] uses multiparty session types and only allows static parallel composition, while
we replace a term inside an arbitrary context at runtime.
Extensions of multiparty session types with error handling [11, 9] share with us the
difficulties in coordinating the transition from the expected pattern to an alternative one, but
in their case the error recovery pattern is known since the very beginning, thus considerably
simplifying the analysis.
We briefly compare to some works that exploit choreographic descriptions for adaptation,
but with very different aims. For instance, [27] defines rules for adapting the specification
of the initial requirements for a choreography, thus keeping the requirements up-to-date in
presence of run-time changes. Our approach is in the opposite direction: we are not interested
in updating the system specification tracking system updates, but in programming and
ensuring correctness of adaptation itself. Other formal approaches to adaptation represent
choreographies as annotated finite state automata. In [43] choreographies are used to
propagate protocol changes to the other peers, while [48] presents a test to check whether
a set of peers obtained from a choreography can be reconfigured to match a second one.
Differently from ours, these works only provide change recommendations for adding and
removing message sequences.
An approach close to ours is multi-tier programming [39, 6, 15], where the programmer
writes a single program that is later automatically distributed to different tiers. This is
similar to what we do by projecting a DIOC on the different roles. In particular, [39]
considers functional programs with role annotations and splits them into different sequential
programs to be run in parallel. Multi-tier approaches such as HipHop [6] or Link [15]
have been implemented and used to develop three tier web applications. Differently from
ours, these works abstract away communication information as much as possible, while
we emphasise this kind of information. Furthermore, they do not take into account code
update. It would be interesting to look for cross-fertilisation results between our approach
and multi-tier programming. For instance, one could export our techniques to deal with
runtime updates into multi-tier programming. In the other direction, we could abstract
away communications and let the compiler manage them according to the programmed flow
of data, as done in multi-tier programming. For instance, in Figure 2 the communications
at Lines 22, 24, and 26 could be replaced by local assignments, since the information on the
chosen branch is carried by the auxiliary communications. In general, communications seem
not strictly needed from a purely technical point of view, however they are relevant for the
systems we currently consider. Furthermore, the automatic introduction of communications
DYNAMIC CHOREOGRAPHIES
51
in an optimised way is not always trivial since the problem is NP-hard even for languages
that do non support threads [40].
On a broader perspective, our theory can be used to inject guarantees of freedom
from deadlocks and races into many existing approaches to adaptation, e.g., the ones in
the surveys [32, 22]. However, this task is cumbersome, due to the huge number and
heterogeneity of those approaches. For each of them the integration with our techniques is
far from trivial. Nevertheless, we already started it. Indeed, AIOCJ follows the approach to
adaptation described in [29]. However, applications in [29] are not distributed and there
are no guarantees on the correctness of the application after adaptation. Furthermore, in
the website of the AIOCJ project [1], we give examples of how to integrate our approach
with distributed [42] and dynamic [50] Aspect-Oriented Programming (AOP) and with
Context-Oriented Programming (COP) [23]. In general, we can deal with cross-cutting
concerns like logging and authentication, typical of AOP, viewing pointcuts as empty scopes
and advices as updates. Layers, typical of COP, can instead be defined by updates which can
fire according to contextual conditions. As future work, besides applying the approach to
other adaptation mechanisms, we also plan to extend our techniques to deal with multiparty
session types [12, 24, 13, 14]. The main challenge here is to deal with multiple interleaved
sessions whilst each of our choreographies corresponds to a single session. An initial analysis
of the problem is presented in [7].
Also the study of more refined policies for rule selection, e.g., based on priorities, is a
topic for future work.
References
[1] AIOCJ website. http://www.cs.unibo.it/projects/jolie/aiocj.html.
[2] G. Anderson and J. Rathke. Dynamic software update for message passing programs. In APLAS, volume
7705 of LNCS, pages 207–222. Springer, 2012.
[3] J. W. Backus. The syntax and semantics of the proposed international algebraic language of the zurich
ACM-GAMM conference. In IFIP Congress, pages 125–131, 1959.
[4] L. Baresi, A. Marconi, M. Pistore, and A. Sirbu. Corrective Evolution of Adaptable Process Models. In
BMMDS/EMMSAD, volume 147 of LNBIP, pages 214–229. Springer, 2013.
[5] S. Basu, T. Bultan, and M. Ouederni. Deciding choreography realizability. In POPL, pages 191–202.
ACM, 2012.
[6] G. Berry and M. Serrano. Hop and HipHop: Multitier Web Orchestration. In ICDCIT, volume 8337 of
LNCS, pages 1–13. Springer, 2014.
[7] M. Bravetti et al. Towards global and local types for adaptation. In SEFM Workshops, volume 8368 of
LNCS, pages 3–14. Springer, 2013.
[8] A. Bucchiarone, A. Marconi, C. A. Mezzina, M. Pistore, and H. Raik. On-the-fly adaptation of dynamic
service-based systems: Incrementality, reduction and reuse. In ICSOC, volume 8274 of LNCS, pages
146–161. Springer, 2013.
[9] S. Capecchi, E. Giachino, and N. Yoshida. Global Escape in Multiparty Sessions. In Proc. of FSTTCS
2010, volume 8 of LIPIcs, pages 338–351. Schloss Dagstuhl, 2010.
[10] M. Carbone, K. Honda, and N. Yoshida. Structured Communication-Centred Programming for Web
Services. In ESOP, volume 4421 of LNCS, pages 2–17. Springer, 2007.
[11] M. Carbone, K. Honda, and N. Yoshida. Structured Interactional Exceptions in Session Types. In Proc.
of CONCUR’08, volume 5201 of LNCS, pages 402–417. Springer, 2008.
[12] M. Carbone, K. Honda, and N. Yoshida. Structured communication-centered programming for web
services. ACM Trans. Program. Lang. Syst., 34(2):8, 2012.
[13] M. Carbone and F. Montesi. Deadlock-Freedom-by-Design: Multiparty Asynchronous Global Programming. In POPL, pages 263–274. ACM, 2013.
52
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
[14] G. Castagna, M. Dezani-Ciancaglini, and L. Padovani. On global types and multi-party session. Logical
Methods in Computer Science, 8(1), 2012.
[15] E. Cooper, S. Lindley, P. Wadler, and J. Yallop. Links: Web Programming Without Tiers. In FMCO,
volume 4709 of LNCS, pages 266–296. Springer, 2006.
[16] M. Coppo, M. Dezani-Ciancaglini, and B. Venneri. Self-adaptive multiparty sessions. Service Oriented
Computing and Applications, 9(3-4):249–268, 2015.
[17] G. Cugola, C. Ghezzi, and L. S. Pinto. DSOL: a declarative approach to self-adaptive service orchestrations. Computing, 94(7):579–617, 2012.
[18] M. Dalla Preda, M. Gabbrielli, S. Giallorenzo, I. Lanese, and J. Mauro. Dynamic choreographies - safe
runtime updates of distributed applications. In COORDINATION, volume 9037 of LNCS, pages 67–82.
Springer, 2015.
[19] M. Dalla Preda, S. Giallorenzo, I. Lanese, J. Mauro, and M. Gabbrielli. AIOCJ: A choreographic
framework for safe adaptive distributed applications. In SLE, volume 8706 of LNCS, pages 161–170.
Springer, 2014.
[20] C. Di Giusto and J. A. Pérez. Disciplined structured communications with consistent runtime adaptation.
In SAC, pages 1913–1918. ACM, 2013.
[21] Eclipse website. http://www.eclipse.org/.
[22] C. Ghezzi, M. Pradella, and G. Salvaneschi. An evaluation of the adaptation capabilities in programming
languages. In SEAMS, pages 50–59. ACM, 2011.
[23] R. Hirschfeld, P. Costanza, and O. Nierstrasz. Context-oriented Programming. Journal of Object
Technology, 7(3):125–151, 2008.
[24] K. Honda, N. Yoshida, and M. Carbone. Multiparty Asynchronous Session Types. In POPL, pages
273–284. ACM, 2008.
[25] H. Hüttel et al. Foundations of session types and behavioural contracts. ACM Computing Surveys, 2016.
[26] Jolie website. http://www.jolie-lang.org/.
[27] I. Jureta, S. Faulkner, and P. Thiran. Dynamic requirements specification for adaptable and open
service-oriented systems. In ICSOC, volume 4749 of LNCS, pages 270–282. Springer, 2007.
[28] J. O. Kephart and D. M. Chess. The vision of autonomic computing. Computer, 36(1):41–50, 2003.
[29] I. Lanese, A. Bucchiarone, and F. Montesi. A Framework for Rule-Based Dynamic Adaptation. In TGC,
volume 6084 of LNCS, pages 284–300. Springer, 2010.
[30] I. Lanese, C. Guidi, F. Montesi, and G. Zavattaro. Bridging the Gap between Interaction- and ProcessOriented Choreographies. In SEFM, pages 323–332. IEEE, 2008.
[31] I. Lanese, F. Montesi, and G. Zavattaro. Amending choreographies. In WWV, volume 123, pages 34–48.
EPTCS, 2013.
[32] L. A. F. Leite et al. A systematic literature review of service choreography adaptation. Service Oriented
Computing and Applications, 7(3):199–216, 2013.
[33] C. Metz. The epic story of dropbox’s exodus from the amazon cloud empire. Wired, http://www.wired.
com/2016/03/epic-story-dropboxs-exodus-amazon-cloud-empire/.
[34] F. Montesi. Kickstarting choreographic programming. In WS-FM, pages 3–10. Springer, 2015.
[35] F. Montesi, C. Guidi, and G. Zavattaro. Composing services with JOLIE. In ECOWS, pages 13–22.
IEEE, 2007.
[36] F. Montesi, C. Guidi, and G. Zavattaro. Service-oriented programming with Jolie. In Web Services
Foundations, pages 81–107. Springer, 2014.
[37] F. Montesi and N. Yoshida. Compositional choreographies. In CONCUR, volume 8052 of LNCS, pages
425–439. Springer, 2013.
[38] I. Neamtiu and M. W. Hicks. Safe and timely updates to multi-threaded programs. In PLDI, pages
13–24. ACM, 2009.
[39] M. Neubauer and P. Thiemann. From sequential programs to multi-tier applications by program
transformation. In POPL, pages 221–232. ACM, 2005.
[40] M. Neubauer and P. Thiemann. Placement inference for a client-server calculus. In ICALP, volume 5126
of LNCS, pages 75–86. Springer, 2008.
[41] P. Nienaltowski. Practical framework for contract-based concurrent object-oriented programming. PhD
thesis, ETH Zurich, 2007.
[42] R. Pawlak et al. JAC: an aspect-based distributed dynamic framework. Software: Practice and Experience,
34(12):1119–1148, 2004.
DYNAMIC CHOREOGRAPHIES
53
[43] S. Rinderle, A. Wombacher, and M. Reichert. Evolution of Process Choreographies in DYCHOR. In
OTM Conferences (1), volume 4275 of LNCS, pages 273–290. Springer, 2006.
[44] Rust website. http://www.rust-lang.org/.
[45] G. Salaün, T. Bultan, and N. Roohi. Realizability of choreographies using process algebra encodings.
IEEE T. Services Computing, 5(3):290–304, 2012.
[46] D. Sangiorgi and D. Walker. The pi-calculus: a Theory of Mobile Processes. Cambridge university press,
2003.
[47] Scribble website. http://www.jboss.org/scribble.
[48] A. Wombacher. Alignment of choreography changes in BPEL processes. In IEEE SCC, pages 1–8. IEEE,
2009.
[49] Xtext website. http://www.eclipse.org/Xtext/.
[50] Z. Yang, B. H. C. Cheng, R. E. K. Stirewalt, J. Sowell, S. M. Sadjadi, and P. K. McKinley. An
aspect-oriented approach to dynamic adaptation. In WOSS, pages 85–92. ACM, 2002.
[51] S. Yegulalp. Mozilla’s Rust-based Servo browser engine inches forward. InfoWorld, http://www.infoworld.
com/article/2905688/applications/mozillas-rust-based-servo-browser-engine-inches-forward.
html.
54
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
Appendix A. Proof of Theorem 6.3
In order to prove the bound on the complexity of the connectedness check we use the lemma
below, showing that the checks to verify the connectedness for a single sequence operator
can be performed in linear time on the size of the sets generated by transI and transF.
Lemma A.1. Given S, S 0 sets of multisets of two elements, checking if ∀ s ∈ S . ∀s0 ∈
S 0 . s ∩ s0 6= ∅ can be done in O(n) steps, where n is the maximum of |S| and |S 0 |.
Proof. Without loss of generality, we can assume that |S| ≤ |S 0 |. If |S| ≤ 9 then the check
can be performed in O(n) by comparing all the elements in S with all the elements in S 0 . If
|S| > 9 then at least 4 distinct elements appear in the multisets in S since the maximum
number of multisets with cardinality 2 obtained by 3 distinct elements is 9. In this case the
following cases cover all the possibilities:
• there exist distinct elements a, b, c, d s.t. {a, b}, {a, c}, and {a, d} belong to S. In this case
for the check to succeed all the multisets in S 0 must contain a, otherwise the intersection
of the multiset not containing a with one among the multisets {a, b}, {a, c}, and {a, d}
is empty. Similarly, since |S 0 | > 9, for the check to succeed all the multisets in S must
contain a. Hence, if {a, b}, {a, c}, and {a, d} belong to S then the check succeeds iff a
belongs to all the multisets in S and in S 0 .
• there exist distinct elements a, b, c, d s.t. {a, b} and {c, d} belong to S. In this case the
check succeeds only if S 0 is a subset of {{a, c}, {a, d}, {b, c}, {b, d}}. Since |S 0 | > 9 the
check can never succeed.
• there exist distinct elements a, b, c s.t. {a, a} and {b, c} belong to S. In this case the
check succeeds only if S 0 is a subset of {{a, b}, {a, c}}. Since |S 0 | > 9 the check can never
succeed.
• there exist distinct elements a, b s.t. {a, a} and {b, b} belong to S. In this case the check
succeeds only if S 0 is a subset of {{a, b}}. Since |S 0 | > 9 the check can never succeed.
Summarising, if |S| > 9 the check can succeed iff all the multisets in S and in S 0 share a
common element. The existence of such an element can be verified in time O(n).
Theorem 6.3 (Connectedness-check complexity).
The connectedness of a DIOC process I can be checked in time O(n2 log(n)), where n is the
number of nodes in the abstract syntax tree of I.
Proof. To check the connectedness of I we first compute the values of the functions transI
and transF for each node of the abstract syntax tree (AST). We then check for each sequence
operator whether connectedness holds.
The functions transI and transF associate to each node a set of pairs of roles. Assuming an
implementation of the data set structure based on balanced trees (with pointers), transI and
transF can be computed in constant time for interactions, assignments, 1, 0, and sequence
constructs. For while and scope constructs computing transF(I 0 ) requires the creation of
balanced trees having an element for every role of I 0 . Since the roles are O(n), transF(I 0 )
can be computed in O(n log(n)). For parallel and if constructs a union of sets is needed. The
union costs O(n log(n)) since each set generated by transI and transF contains at maximum
n elements.
Since the AST contains n nodes, the computation of the sets generated by transI and
transF can be performed in O(n2 log(n)).
To check connectedness we have to verify that for each node I 0 ; I 00 of the AST ∀R1 →
R2 ∈ transF(I 0 ), ∀S1 → S2 ∈ transI(I 00 ) . {R1 , R2 } ∩ {S1 , S2 } 6= ∅. Since transF(I 0 ) and
DYNAMIC CHOREOGRAPHIES
55
transI(I 00 ) have O(n) elements, thanks to Lemma A.1, checking if I 0 ; I 00 is connected costs
O(n). Since in the AST there are less than n sequence operators, checking the connectedness
on the whole AST costs O(n2 ).
The complexity of checking the connectedness of the entire AST is therefore limited by
the cost of computing functions transI and transF and of checking the connectedness. All
these activities have a complexity of O(n2 log(n)).
Appendix B. Proofs of Section 7.1
Lemma 7.9 (Distinctness of Global Indexes). Given a well-annotated DIOC I, a global
η1
ηn
state Σ, and a set of updates I, if hΣ, I, Ii −→ . . . −→ hΣ0 , I0 , I 0 i then all global indexes in I 0
are distinct.
Proof. The proof is by induction on the number n of transitions, using a stronger inductive
hypothesis: indexes are distinct but, possibly, inside DIOC subterms of the form I; i :
while b@R {I 0 }. In this last case, the same index can occur both in I and in I 0 , attached
to constructs with different global indexes. The statement of the Lemma follows directly:
first, distinct indexes imply distinct global indexes; second, global indexes of I and of I 0 are
distinct, since I 0 is inside the while loop whilst I is not.
In the base case (n = 0), thanks to well annotatedness, indexes are
√ always distinct. The
inductive case follows directly by induction for transitions with label . Otherwise, we √
have
a case analysis on the only axiom which derives a transition with label different from .
The only difficult cases are bDIOC |While-unfold e and bDIOC |Up e.
In the case of Rule bDIOC |While-unfold e, note that the while is enabled, hence it cannot
be part of a term of the form I; i: while b@R {I 0 }. Hence, indexes of the body of the while
loop do not occur elsewhere. As a consequence, after the transition no clashes are possible
with indexes in the context. Note also that indexes of the body of the loop are duplicated,
but the resulting term has the form I; i: while b@R {I 0 }, thus global indexes are distinct
by construction.
The case bDIOC |Up e follows thanks to the condition freshIndexes(I 0 ).
Lemma 7.15. Given a well-annotated connected DIOC process I and for each state Σ the
DPOC network proj(I, Σ) is such that:
(1) events(I) ⊆ events(proj(I, Σ));
(2) ∀ ε1 , ε2 ∈ events(I).ε1 ≤DIOC ε2 ⇒ ε1 ≤DP OC ε2 ∨ ε1 ≤DP OC ε2
Proof.
(1) By definition of projection.
(2) Let ε1 ≤DIOC ε2 . We have a case analysis on the condition used to derive the dependency.
Sequentiality: Consider I = I 0 ; I 00 . If events are in the same role the implication
follows from the sequentiality of the ≤DP OC .
Let us show that there exists an event ε00 in an initial interaction of I 00 such that either
ε00 ≤DP OC ε2 or ε00 ≤DP OC ε2 . The proof is by induction on the structure of I 00 . The
only difficult case is sequential composition. Assume I 00 = I1 ; I2 . If ε2 ∈ events(I1 )
the thesis follows from inductive hypothesis. If ε2 ∈ events(I2 ) then by induction
there exists an event ε3 in an initial interaction of I2 such that ε3 ≤DP OC ε2 or
ε3 ≤DP OC ε2 . By synchronisation (Definition 7.14) we have that ε3 ≤DP OC ε2 or
ε3 ≤DP OC ε2 . By connectedness we have that ε3 or ε3 are in the same role of an
56
M. DALLA PREDA, M. GABBRIELLI, S. GIALLORENZO, I. LANESE, AND J. MAURO
event ε4 in I 0 . By sequentiality (Definition 7.14) we have that ε4 ≤DP OC ε3 or
ε4 ≤DP OC ε3 . By synchronisation we have that ε4 ≤DP OC ε3 or ε4 ≤DP OC ε3 . The
thesis follows from the inductive hypothesis on ε4 and by transitivity of ≤DP OC .
Let us also show that there exists a final event ε000 ∈ events(I 0 ) such that ε1 ≤DP OC ε000
or ε1 ≤DP OC ε000 . The proof is by induction on the structure of I 0 . The only difficult
case is sequential composition. Assume I 0 = I1 ; I2 . If ε1 ∈ events(I2 ) the thesis
follows from inductive hypothesis. If ε1 ∈ events(I1 ) then the proof is similar to the
one above, finding a final event in I1 and applying sequentiality, synchronisation,
and transitivity.
The thesis follows from the two results above again by sequentiality, synchronisation,
and transitivity.
Scope: it means that either (a) ε1 =↑ξ and ε2 is an event in the scope or (b) ε1 =↑ξ and
ε2 =↓ξ , or (c) ε1 is an event in the scope and ε2 =↓ξ . We consider case (a) since case
(c) is analogous and case (b) follows by transitivity. If ε2 is in the coordinator then
the thesis follows easily. Otherwise it follows thanks to the auxiliary synchronisations
with a reasoning similar to the one for sequentiality.
Synchronisation: it means that ε1 is a sending event and ε2 is the corresponding
receiving event, namely ε1 = ε2 . Thus, since ε2 ≤DP OC ε2 then ε2 ≤DP OC ε2 .
If: it means that ε1 is the evaluation of the guard and ε2 is an event in one of the two
branches. Thus, if ε2 is in the coordinator then the thesis follows easily. Otherwise it
follows thanks to the auxiliary synchronisations with a reasoning similar to the one
for sequentiality.
While: it means that ε1 is the evaluation of the guard and ε2 is in the body of the
while loop. Thus, if ε2 is in the coordinator then the thesis follows easily. Otherwise
it follows thanks to the auxiliary synchronisations with a reasoning similar to the
one for sequentiality.
Lemma 7.20. Let I be a well-annotated connected DIOC process and Σ a state. Then the
projection N = proj(I, Σ) is a well-annotated DPOC network with respect to ≤DP OC .
Proof. We have to prove that proj(I, Σ) satisfies the conditions of Definition 7.17 of wellannotated DPOC:
C1: For each global index ξ there are at most two communication events on programmerspecified operations with global index ξ and, in this case, they are matching events.
The condition follows by the definition of the projection function, observing that in
well-annotated DIOCs, each interaction has its own index, and different indexes are
mapped to different global indexes.
C2: Only events which are minimal according to ≤DP OC may correspond to enabled
transitions. This condition follows from Lemma 7.19.
C3: For each pair of non-conflicting sending events [fξ ]R and [fξ0 ]R on the same operation
o? and with the same target R0 such that ξ =
6 ξ 0 we have [fξ ]R ≤DP OC [fξ0 ]R or
[fξ0 ]R ≤DP OC [fξ ]R . Note that the two events are in the same role R, thus without loss of
generality we can assume that there exist two processes P, P 0 such that [fξ ]R ∈ events(P )
and [fξ0 ]R ∈ events(P 0 ) and there is a subprocess of N of one of the following forms:
• P ; P 0 : the thesis follows by sequentiality (Definition 7.14);
• P |P 0 : this case can never happen for the reasons below. For events on programmerspecified operations this follow by the definition of projection, since the prefixes of
the names of operations are different. For events on auxiliary operations originated
DYNAMIC CHOREOGRAPHIES
57
by the same construct this follows since all the targets are different. For events on
auxiliary operations originated by different constructs this follows since the prefixes of
the names of the operations are different.
• if b {P } else {P 0 }: this case can never happen since the events are non-conflicting
(Definition 7.16).
C4: Similar to the previous case, with receiving events instead of sending events.
C5: If ε is an event inside a scope with global index ξ then its matching events ε (if they
exist) are inside a scope with the same global index. This case holds by definition of the
projection function.
C6: If two events have the same index but different global indexes then one of them, let us
call it ε1 , is inside the body of a while loop with global index ξ1 and the other, ε2 , is not.
Furthermore, ε2 ≤DP OC εξ1 where εξ1 is the guarding while-event of the while loop with
global index ξ1 . By definition of well-annotated DIOC and of projection the only case
where there are two events with the same index but different global indexes is for the
auxiliary communications in the projection of the while construct, where the conditions
hold by construction.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view a
copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a letter
to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or Eisenacher Strasse 2, 10777 Berlin, Germany
| 2cs.AI
|
Quasi-isometries Between Groups with Two-Ended
Splittings
Christopher H. Cashen and Alexandre Martin
arXiv:1601.07147v1 [math.GR] 26 Jan 2016
April 10, 2018
Abstract
We construct ‘structure invariants’ of a one-ended, finitely presented group that
describe the way in which the factors of its JSJ decomposition over two-ended subgroups fit together.
For groups satisfying two technical conditions, these invariants reduce the problem of quasi-isometry classification of such groups to the problem of relative quasiisometry classification of the factors of their JSJ decompositions. The first condition
is that their JSJ decompositions have two-ended cylinder stabilizers. The second
is that every factor in their JSJ decompositions is either ‘relatively rigid’ or ‘hanging’. Hyperbolic groups always satisfy the first condition, and it is an open question
whether they always satisfy the second.
The same methods also produce invariants that reduce the problem of classification of one-ended hyperbolic groups up to homeomorphism of their Gromov boundaries to the problem of classification of the factors of their JSJ decompositions up
to relative boundary homeomorphism type.
1
Introduction
Gromov proposed a program of classifying finitely generated groups up to the geometric
equivalence relation of quasi-isometry [20]. A natural approach to this problem is to
first try to decompose the group into “smaller” pieces by means of a graph of groups
decomposition, and then reduce the quasi-isometry classification problem to the problem
of understanding the quasi-isometry types of the various vertex groups and the way these
subgroups fit together.
The simplest such decomposition is the decomposition of a group as a graph of groups
over finite subgroups. Stallings’s Theorem [42, 43] asserts that a finitely generated group
splits as an amalgamated product or an HNN extension over a finite group if and only if
it has more than one end. In particular, the existence of a splitting as a graph of groups
with finite edge groups is a quasi-isometry invariant. Finitely presented groups admit a
maximal splitting over finite subgroups [16], and a theorem of Papasoglu and Whyte [37]
says that the collection of quasi-isometry types of one-ended vertex groups of a maximal
decomposition of an infinite-ended, finitely presented group is a complete quasi-isometry
invariant. This reduces the quasi-isometry classification problem to one-ended groups.
1
We push this program to its next logical step, which is to decompose one-ended groups
over two-ended subgroups. Papasoglu [36] shows that the existence of a splitting of a
finitely presented one-ended group over two-ended subgroups is quasi-isometry invariant,
provided that the group is not commensurable to a surface group. Moreover, such a
group admits a maximal decomposition as a graph of groups over two-ended subgroups,
known as a JSJ decomposition, and Papasoglu’s results imply that quasi-isometries respect (in a certain sense that will be made precise in Section 2.3.2) the JSJ decomposition. In particular, the quasi-isometry types of the non-elementary vertex groups of the
JSJ decomposition are invariant under quasi-isometries. Even more is true: In a nonelementary vertex group of the JSJ we see the collection of conjugates of the two-ended
subgroups corresponding to incident edges of the JSJ decomposition. Quasi-isometries
of the group must preserve the vertex groups together with their patterns of conjugates
of incident edge subgroups. We say a quasi-isometry must preserve the relative quasiisometry type of the vertex. Such patterns have been exploited before [5, 13, 32, 33] to
produce quasi-isometry invariants from various ‘pattern rigidity’ phenomena. However,
the quasi-isometry types, or relative quasi-isometry types, of the vertex groups alone
do not give complete quasi-isometry invariants, because not all vertex groups have the
same coarse intersections: Vertices that are adjacent in the Bass-Serre tree of the decomposition have vertex groups that intersect in two-ended subgroups, while the coarse
intersection of non-adjacent vertices may or may not be bounded.
In this paper, we produce further quasi-isometry invariants of a finitely presented
one-ended group from an appropriate JSJ decomposition over two-ended subgroups. Under a mild technical restriction, which is known to hold for large classes of groups (see
Section 5.2 for a discussion), we show that our invariants give complete quasi-isometry
invariants, and thus reduce the quasi-isometry classification problem for such groups to
relative versions of these problems in the vertex groups of their JSJ decomposition.
In the case of one-ended hyperbolic groups, a weaker classification is possible, namely
the classification up to homeomorphisms of Gromov boundaries. Indeed, recall that
quasi-isometric hyperbolic groups have homeomorphic Gromov boundaries at infinity,
but there are examples of hyperbolic groups with homeomorphic boundary that are
not quasi-isometric. For a hyperbolic group, the existence of a splitting over a finite
subgroup amounts to having a disconnected Gromov boundary, and is thus detected
by the homeomorphism type of the boundary. Paralleling the results of Papasoglu and
Whyte for quasi-isometries, Martin and Świątkowski [31] show that hyperbolic groups
with infinitely many ends have homeomorphic boundaries if and only if they have the
same sets of homeomorphism types of boundaries of one-ended factors, reducing the
classification problem to the case of one-ended hyperbolic groups.
For a one-ended hyperbolic group G whose boundary is not a circle, Bowditch shows
that the there exists a splitting over a two-ended subgroup if and only if there exists
a cut pair in the boundary [9]. From the structure of cut pairs in the boundary, he
deduces the existence of a simplicial tree on which Homeo(∂G) acts by isomorphisms.
Paralleling the quasi-isometry case, the action of Homeo(∂G) on this tree preserves the
2
relative boundary homeomorphism types of vertex stabilizers in G.
Using the same approach as for our quasi-isometry classification, we construct a
complete system of invariants of the homeomorphism type of the Gromov boundary of
a hyperbolic group, which completely reduces the classification problem for hyperbolic
groups to relative versions of this problem in the vertex groups of their JSJ decomposition.
Our results rely heavily on the existence of a canonical choice of tree that is preserved
by quasi-isometries, respectively, homeomorphisms of the boundary, with the additional
property that any such map induces maps of the same nature at the level of the vertex
groups.
In the case of hyperbolic groups, the tree is Bowditch’s canonical JSJ tree constructed
from cut pairs in the boundary. For a more general finitely presented one-ended group
G, let Γ be a JSJ decomposition over two-ended subgroups. Let T := T (Γ) be the
Bass-Serre tree of Γ. Commensurability of stabilizers defines an equivalence relation on
edges of T whose equivalence classes are called cylinders. Guirardel and Levitt [23] show
that the dual tree to the covering of T by cylinders defines a new tree Cyl(T ), the tree
of cylinders of T , with a cocompact G–action, and, in fact, this tree is independent of
Γ, so it makes sense to call it the tree of cylinders of G, and denote it Cyl(G). It follows
from Papasoglu’s results, see Theorem 2.9, that a quasi-isometry between finitely presented one-ended groups induces an isomorphism between their trees of cylinders, and
restricts to give a quasi-isometry of each vertex group Gv of the tree of cylinders that
coarsely preserves the pattern Pv of edge groups Ge , for edges e incident to v. When the
edge stabilizer of Cyl(G) in G are two-ended then the quotient graph of groups gives a
canonical decomposition of G over two-ended subgroups. Our results are strongest when
additionally the cylinder stabilizers are two-ended. This is the case, in particular, when
G is hyperbolic, in which case Cyl(G) is equivariantly isomorphic to Bowditch’s tree.
Before giving more details about the ideas behind our classifications, we restrict to
the case that G is a one-ended hyperbolic group, and we adopt some notation that will
allow us to to discuss simultaneously the quasi-isometry and boundary homeomorphism
cases. Let Map((X, P X ), (Y, P Y )) denote alternately:
• The set of quasi-isometries from X to Y taking the collection of coarse equivalence
classes of subsets P X to the collection of coarse equivalence classes of subsets P Y .
• The set of boundary homeomorphisms from ∂X to ∂Y taking the collection of
subsets ∂P X to the collection of subsets ∂P Y .
Similarly, Map(G) is either QI(G) or Homeo(∂G). An element of Map(· · · ) will be
referred to as a Map–equivalence.
The idea behind our classification result is the following. First remark that a finitely
presented one-ended group is quasi-isometric to a complex of spaces over the JSJ tree
of cylinders (this will be recalled in Section 2.4). Moreover, such a decomposition as
a complex of spaces is compatible with Map-equivalences, in that a Map-equivalence
3
between two finitely presented one-ended groups coarsely preserves the structure of complex of spaces, as follows from the aforementioned results of Papasoglu and Bowditch.
To determine whether two groups are Map-equivalent, we thus want to decide whether
their JSJ trees of cylinders are isomorphic and then try to promote such an isomorphism
to a Map-equivalence between the complexes of spaces, hence, between the groups. In
full generality, a Map-equivalence not only induces a simplicial isomorphism between the
trees of cylinders, but also preserves extra information about the vertex groups that we
want to take into account: Map–class of vertex groups, relative Map–class with respect to
incident edge groups, etc. In other words, we decorate the vertices of the trees of cylinders
with these additional pieces of information. What we then want is to understand when
there exists a decoration-preserving isomorphism between the tree of cylinders of two
groups. We want to add enough additional information so that a decoration-preserving
isomorphism between the trees of cylinders can be promoted to a Map–equivalence between the associated groups. Moreover, we would like to have an algorithm telling us
when two such decorated trees of cylinders are isomorphic.
This latter point will be dealt with by generalizing to cocompact decorated trees a
theorem from graph theory giving a necessary and sufficient condition for the universal
covers of two graphs to be isomorphic [27], see Section 4.2. To such a decorated tree we
will associate a structure invariant that completely determines the tree up to decorationpreserving isomorphism.
To now get an intuition of the decorations we consider in this article, let us start
from the decoration of the tree of cylinders that associates to each vertex v the relative
Map-equivalence Map(Gv , Pv ) mentioned earlier, where Pv is the peripheral structure
coming from the incident edge groups, and let us try to promote a decoration-preserving
isomorphism χ between trees of cylinders to a Map-equivalence between groups. The
goal is to choose, for each vertex v, a Map–equivalence between v and χ(v), and piece
them together to get Map-equivalence between groups.
The first potential problem is a realization problem. We know that v and χ(v) are
Map–equivalent, but we also need to know that there is such a Map–equivalence that
matches up peripheral subsets in the same way that χ matches up edges incident to v
and χ(v).
The second potential problem is that the vertex Map–equivalences must agree when
their domains overlap. For boundary homeomorphisms the overlap is just the boundary
of an edge space, which is a pair of points, so this a matter of choosing consistent
orientations on the edge spaces.
Roughly, this is the phenomenon underlying the fact that two HNN extensions of the
form
G+ := hG, t|tut̄ = vi and G− := hG, t|tut̄ = v̄i
might not be Map-equivalent in general, for a group G and elements u, v of G.
It is worth noting that the orientation obstruction automatically disappears if all the
edge groups contain an infinite dihedral group.
For the classification up to boundary homeomorphisms, these three obstructions,
4
relative type, realizability, and orientation, are essentially the only obstructions to constructing a Map–equivalence between the groups. We use these considerations to produce
a finer decoration that yields a structure invariant that is a complete invariant for boundary homeomorphism type, see Theorem 7.1.
In the case of quasi-isometries, orientation of the edge spaces is not enough; we must
choose Map–equivalence of the vertex spaces that agree all along the length of shared
edge spaces. There are two cases in which we can decide if this is possible. The first
is that vertex spaces are extremely flexible so that we have a lot of freedom to choose
Map–equivalences and make them agree on edge spaces. This is the case for the socalled ‘hanging’ vertices. The other case is the opposite one, in which the vertex space
is extremely ‘rigid’ and we have very little choice about how to choose the maps. In
this case we define an invariant called a stretch factor that we then incorporate into
the decoration. If all vertices are either rigid or hanging then the decoration that takes
into account vertex relative quasi-isometry type, realizability, orientation, and stretch
factors yields a complete quasi-isometry invariant, see Theorem 8.5. This notion of relative quasi-isometric rigidity is known to hold for many classes of groups, see Section 5.2.
The classifications we provide are inherently more technical than the classifications for
splittings over finite subgroups by Papasoglu and Whyte and Martin and Świątkowski.
Underlying those theorems is the fact that in splittings over finite groups the vertex
groups of the canonical decomposition are essentially independent of one another, so
only the pieces of the decomposition matter. Our classification must handle not just the
pieces, but also their complex interactions.
1.1
Applications
Our main results give invariants that reduce classification problems for finitely presented,
one-ended groups to a relative version of the classification problem on the vertex groups
of a JSJ decomposition.
Note that vertex groups of a two-ended JSJ decomposition of a one-ended group
are not necessarily one-ended, but we really have achieved a significant reduction of
the classification problem because the problem of classification of relative quasi-isometry
types is far more constrained than the general problem. The most striking example of
this phenomenon is a theorem of Cashen and Macura [14] that says that if a free group
appears as a vertex group of a JSJ decomposition then it is either hanging or it is quasiisometrically rigid relative to the peripheral structure induced by incident edge groups.
Moreover, in the rigid case the relative quasi-isometry question reduces to a relative
isometry question in certain associated cube complexes, which seems far more tractable.
A sample application of our main results is a complete description, in terms of [14], of
the quasi-isometry and boundary homeomorphism types of one-ended hyperbolic groups
that split as graphs of groups1 with free vertex groups and cyclic edge groups.
Consider the following example:
1
Such a decomposition can be algorithmically improved to a JSJ decomposition of the same type [12].
5
Example 1.1. Let Gi = a, b, t | tui t = vi , where ui and vi are words in ha, bi given
below. In each case Gi should be thought of as an HNN extension of ha, bi over Z with
stable letter t.
Let u0 := a, v0 := abab2 , u1 := ab, v1 := a2 b2 , u2 := ab2 , and v2 := a2 b.
Then G0 , G1 , and G2 are pairwise non-quasi-isometric, but have homeomorphic Gromov boundaries.
The quasi-isometry claim follows because the vertex groups in each case are quasiisometrically rigid relative to the incident edge groups, and the stretch factor for Gi is
the ratio of the word lengths of ui and vi , which are different for each i; see see Section 3
and Section 5.7. The groups have homeomorphic boundaries because the stretch factors
are the only difference, and boundary homeomorphism is not sensitive to stretch factors;
see Section 3 and Example 7.2.
Our methods produce interesting invariants even in cases that the groups are not
hyperbolic or that the relative problems for the vertex groups is not be completely understood:
Example 1.2. Let M be the mapping class group of a non-sporadic hyperbolic surface,
with a fixed finite generating set. Let g0 , g1 , g00 , g10 ∈ M be pseudo-Anosov elements
that are not proper powers. Let G := M ∗Z M and G0 := M ∗Z M be the amalgamated
product groups obtained by identifying g0 with g1 and g00 with g10 , respectively. Then G
and G0 are not quasi-isometric if the ratio of translation lengths of g0 and g1 is different,
up to inversion, from the ratio of translation lengths of g00 and g10 .
This follows because mapping class groups are quasi-isometrically rigid and the ratios
of the translation lengths in this example are the stretch factors; see Section 5.2 and
Section 5.
Acknowledgements
The first author thanks Mladen Bestvina, Jason Behrstock, and Gilbert Levitt for interesting conversations related to this work. In particular, the idea that a version of
Leighton’s Theorem would provide a concise description of our structure invariants was
suggested by Bestvina, and Example 1.2 was suggested by Behrstock.
Both authors were supported by the European Research Council (ERC) grant of
Goulnara Arzhantseva, grant agreement #259527 and the Erwin Schrödinger Institute
workshop “Geometry of Computation in Groups”. The first author is partially supported
by the Austrian Science Fund (FWF):M1717-N25. The second author is partially supported by the Austrian Science Fund (FWF):M1810-N25.
2
Preliminaries
We assume familiarity with standard concepts such as Cayley graphs, ends of spaces,
and (Gromov) hyperbolic geometry. See [10] for background.
6
A group is virtually cyclic if it has an infinite cyclic subgroup of finite index. A group
is non-elementary is it is neither finite nor virtually cyclic. A standard exercise is to
show that a finitely generated, two-ended group is virtually cyclic.
2.1
Coarse geometry
Throughout the paper the qualifier ‘coarse’ is used to indicate ‘up to additive error’.
Let (X, dX ) and (Y, dY ) be metric spaces. Subsets of X are coarsely equivalent if they
are bounded Hausdorff distance from one another. A subset A is coarsely contained in B
if A is coarsely equivalent to a subset of B. Two maps φ and φ0 from X to Y are said to be
coarsely equivalent or bounded distance from each other if supx∈X dY (φ(x), φ0 (x)) < ∞.
A map φ : X → Y is coarsely surjective if its image is coarsely equivalent to Y .
A map φ : X → Y is a controlled embedding 2 if there exist unbounded, non-decreasing
real functions ρ0 and ρ1 such that for all x and x0 in X we have:
ρ0 (dX (x, x0 )) ≤ dY (φ(x), φ(x0 )) ≤ ρ1 (dX (x, x0 ))
There are several classes of controlled embeddings that have special names. If ρi (r) :=
Mi r for i ∈ {0, 1} then φ is
• an isometric embedding if M0 = 1 = M1 ,
• an M –similitude if M0 = M = M1 ,
• an M –biLipschitz embedding if M0−1 = M = M1 .
If such a map is surjective then it is called, respectively, an isometry, similarity, or
biLipschitz equivalence.
For each of these, we can add the qualifier ‘coarse’ and allow an additive error, so that
ρ0 (r) := M0 r − A and ρ1 (r) := M1 r + A. For instance, φ is an (M, A)–coarse biLipschitz
embedding if ρi is as above with M0 = M −1 and M1 = M . In the ‘coarse’ cases we drop
the term ‘embedding’ if the map is coarsely surjective.
An (M, A)–coarse biLipschitz embedding is more commonly called an (M, A)–quasiisometric embedding.
Let QIsom(X, Y ) denote the set of quasi-isometries from X to Y , and let CIsom(X, Y )
denote the set of coarse isometries from X to Y . For a coarsely surjective map φ : X → Y ,
a coarse inverse is a map φ : Y → X such that φ ◦ φ is coarsely equivalent to IdX and
φ ◦ φ is coarsely equivalent to IdY . If φ ∈ QIsom(X, Y ) then all coarse inverses of φ are
coarsely equivalent and belong to QIsom(Y, X). If φ ∈ CIsom(X, Y ) then every coarse
inverse of φ belongs to CIsom(Y, X).
Let I(X, Y ), CI(X, Y ), and QI(X, Y ) denote, respectively, the sets Isom(X, Y ),
CIsom(X, Y ), and QIsom(X, Y ), modulo coarse equivalence. When Y = X we shorten
the notation to I(X), CI(X), QI(X), and each of these form a group under composition.
2
This is more commonly called a ‘coarse embedding’.
7
A subset of CI(X, Y ) or QI(X, Y ) is said to be uniform if there exists a C such
that every element of the subset is an equivalence class of maps containing a C–coarse
isometry or a (C, C)–quasi-isometry, respectively.
Quasi-isometries respect coarse equivalence of subsets. If P is a set of coarse equivalence classes of subsets of X, and P 0 is a set of coarse equivalence classes of subsets of
Y , let QI((X, P), (Y, P 0 )) be the subset of QI(X, Y ) consisting of quasi-isometries that
induce bijections between P and P 0 . Similarly, QI((X, P)) := QI((X, P), (X, P)) is a
subgroup of QI(X).
If φ : X → Y is a quasi-isometry, define φ∗ : QI(X) → QI(Y ) by φ∗ (ψ) := φ ◦ ψ ◦ φ.
A subset Y ⊂ X is coarsely connected if there exists some A such that for every y
and y 0 ∈ Y there is, for some k, a chain y0 := y, y1 , . . . , yk := y 0 of (k + 1) points yi ∈ Y
such that d(yi , yi+1 ) ≤ A.
A geodesic is an isometric embedding of a connected subset of R. A coarse geodesic
is a coarse geodesic embedding of a coarsely connected subset of R. A quasi-geodesic is
a quasi-isometric embedding of a coarsely connected subset of R.
The space X is said to be geodesic, A–coarse geodesic, or (M, A)–quasi-geodesic if
for every pair of points in X there exists, respectively, a geodesic, A–coarse geodesic, or
(M, A)–quasi-geodesic connecting them.
Let JXK denote the set of proper geodesic metric spaces quasi-isometric to X. If P
is a set of coarse equivalence classes of subsets of X, let J(X, P)K denote the set of pairs
(Y, P 0 ) where Y is a geodesic metric space and P 0 is a collection of coarse equivalence
classes of subsets of Y such that there exists a quasi-isometry from X to Y that induces
a bijection from P to P 0 . We call JXK the quasi-isometry type of X and J(X, P)K the
relative quasi-isometry type of (X, P).
If L is a path connected subset of a geodesic metric space (X, dX ), let dL denote
the induced length metric on L. A quasi-line in X is a path connected subset L such
that (L, dL ) is quasi-isometric to R and the inclusion map of L into X is a controlled
embedding.
We define a peripheral structure P on geodesic metric space X to be a collection of
coarse equivalence classes of quasi-lines. In particular, if G is a finitely generated group
and H is a finite collection of two-ended subgroups of G, then H induces a peripheral
structure consisting of distinct coarse equivalence classes of conjugates of elements of H.
2.2
Graphs of groups
Let Γ be a finite oriented graph. Let VΓ be the set of vertices of Γ, and let E + Γ be the set
of oriented edges. For e ∈ E + Γ, let ι(e) be its initial vertex, and let τ (e) be its terminal
vertex. For each e ∈ E + Γ formally define e to be an inverse edge to e with ι(v) := τ (e),
τ (e) := ι(e), and ē¯ := e. The inverse edge e should be thought of as e traversed against
its given orientation. Let E − Γ denote the set of inverse edges, and EΓ := E + Γ ∪ E − Γ.
A graph of groups Γ := (Γ, {Gγ }γ∈VΓ∪E + Γ , {εe }e∈E + Γ ) consists of a finite directed
graph Γ, groups Gγ for γ ∈ VΓ ∪ E + Γ such that Ge < Gι(e) for e ∈ E + Γ, and injections
εe : Ge ,→ Gτ (e) for e ∈ E + Γ.
8
For symmetry in the notation it is convenient to define Gē := Ge for each e ∈ E − Γ
and let εē denote the inclusion of Gē := Ge into Gτ (ē) := Gι(e) .
A graph of groups Γ has an associated fundamental group G = G(Γ) obtained by
amalgamating the vertex groups over the edge groups [41]. We say that Γ is a graph of
groups decomposition of G.
The Bass-Serre tree T := T (Γ) of Γ is the tree on which G acts without edge inversions, such that G\T = Γ and such that the stabilizer Gt of t ∈ VT ∪ ET is a conjugate
in G of the group Gt , where t is the image of t under the quotient map T → Γ.
Throughout we use the notation t to denote the image of t in Γ. Conversely, for each
γ ∈ VΓ ∪ EΓ we choose some lift γ
e of γ to T . Given a maximal subtree in Γ we can, and
do, choose lifts of vertices and edge in the subtree to get a subtree in T .
Definition 2.1. Given a vertex group Gv of Γ, the peripheral structure coming from
incident edge groups, Pv , is the set of distinct coarse equivalence classes in Gv of Gv –
conjugates of the images of the edge injections εe : Ge ,→ Gv for edges e ∈ EΓ with
τ (e) = v.
Definition 2.2. Given a vertex v of T (Γ), the peripheral structure coming from incident
edge groups, Pv , is the set of distinct coarse equivalence classes in the stabilizer subgroup
Gv of v of stabilizers of incident edge groups.
It is immediate from the definitions that the quotient by the G–action identifies
(Gv , Pv ) and (Gv , Pv ).
We are interested in graphs of groups in which the edge groups are two-ended, hence
virtually cyclic. Commensurability of edge stabilizers defines an equivalence relation on
the edges of the Bass-Serre tree T of such a splitting. The equivalence classes of edges
are called cylinders. Every cylinder is a subtree of T [23, Lemma 4.2]. It follows that we
get another tree Cyl(T ), called the tree of cylinders of T , by taking the dual tree to the
covering of T by cylinders.
Definition 2.3. A graph of groups with two-ended edge groups is k–acylindrical if the
cylinders of its Bass-Serre tree have diameter at most k. It is acylindrical if there exists
k such that it is k–acylindrical.
Let C be a cylinder in T and let Stab(C) be the stabilizer of C in G. Choose an
infinite order element z in Ge for some edge e ∈ C. For any element g ∈ Stab(C), Ge and
Gge are commensurable, virtually cyclic groups. Since hzi is a finite index subgroup of
Ge , there exist non-zero a and b such that gz a g −1 = z b . Define ∆(g) = ab . This defines
a homomorphism ∆ : Stab(C) → Q∗ , called the modular homomorphism of C, that is
independent of the choice of z.
Definition 2.4. A cylinder is called unimodular if the image of its modular homomorphism is in {−1, 1}.
A graph of groups with two-ended edge groups is unimodular if all of its cylinders
are unimodular.
9
ˆ on pairs of edges of a cylinder C:
We also define a modulus ∆
Definition 2.5. Let e0 and e1 be edges in C. Let hz0 i < Ge0 and hz1 i < Ge1 be infinite
ˆ 0 , e1 ) = [hz1 i:hz0 i∩hz1 i] .
cyclic subgroups of minimal index. Define ∆(e
[hz0 i:hz0 i∩hz1 i]
ˆ does not depend on the choice of minimal index infinite
It is easy to check that ∆
ˆ ge) = |∆(g)|.
cyclic subgroups, and that ∆(e,
2.3
2.3.1
The JSJ tree of cylinders
Definitions
A JSJ decomposition of a finitely presented, one-ended group G over two-ended subgroups
is a graph of groups with two-ended edge groups that encodes all splittings of G over twoended subgroups. Equivalent descriptions of such decompositions appear in Dunwoody
and Sageev [17], Fujiwara and Papasoglu [19], and Guirardel and Levitt [22]. See also
Rips and Sela [39].
Following Papasoglu [36], we will give a geometric description of JSJ decompositions.
First, we need some terminology.
A quasi-line L is separating if its complement has at least two essential components,
that is, components that are not contained in any finite neighborhood of L. In particular,
if G splits over a two-ended subgroup then that two-ended subgroup is bounded distance
from a separating quasi-line. Separating quasi-lines cross if each travels arbitrarily deeply
into two different essential complementary components of the other.
Let Gv be a vertex group in a graph of groups decomposition. Let Pv be the peripheral
structure on Gv coming from incident edge groups. Let Σ be a hyperbolic pair of pants.
Let P∂Σ be the peripheral structure on the universal cover Σ̃ of Σ consisting of the coarse
equivalence classes of the components of the preimages of the boundary curves.
Definition 2.6. We say v is hanging 3 if (Gv , Pv ) is quasi-isometric to (Σ̃, P∂Σ ). We
say v is rigid if it is not two-ended, not hanging, and does not split over a two-ended
subgroup relative to its incident edge groups.
The rigid and hanging terminology extends to vertices in T (Γ) in the obvious way.
Definition 2.7. Let G be a finitely presented one-ended group that is not commensurable
to a surface group. A JSJ decomposition of G is a (possibly trivial) graph of groups
decomposition Γ with two-ended edge groups satisfying the following conditions:
(a) Every vertex group is either two-ended, hanging, or rigid.
(b) If v is a valence one vertex with two-ended vertex group then the incident edge group
does not surject onto Gv .
(c) Every cylinder in the Bass-Serre tree of Γ that contains exactly two hanging vertices
also contains a rigid vertex.
3
After the ‘quadratically hanging’ vertex groups of Rips and Sela [38].
10
This definition is equivalent to those cited above. The essential facts are that:
1. Hanging vertices contain crossing pairs of separating quasi-lines.
2. Every pair of crossing separating quasi-lines is coarsely contained in a conjugate of
a hanging vertex group.
3. A separating quasi-line that is not crossed by any other separating quasi-line is
coarsely equivalent to a conjugate of an edge group.
4. Every edge group is coarsely equivalent to a separating quasi-line that is not crossed
by any other separating quasi-line.
Remark. Condition (c) implies that the hanging vertex groups are maximal hanging,
which is necessary for item 4.
Remark. The case that a vertex group is the fundamental group of a pair of pants and
the incident edge groups glue on to the boundary curves is called ‘rigid’ in the usual
JSJ terminology because there are no splittings of the pair of pants group relative to
the boundary subgroups. Algebraically, such a vertex behaves like our rigid vertices, but
geometrically this is a hanging vertex.
In general a group does not have a unique JSJ decomposition, but rather a deformation space of JSJ decompositions [18, 22]. Furthermore, all JSJ decompositions are in
the same deformation space, which means any one can be transformed into any other by
meas of a finite sequence of moves of a prescribed type. The tree of cylinders of a decomposition depends only on the deformation space [23, Theorem 1], up to G–equivariant
isomorphism, so there is a unique JSJ tree of cylinders.
Definition 2.8. Let G be a finitely presented one-ended group. The JSJ tree of cylinders
Cyl(G) of G is the tree of cylinders of the Bass-Serre tree of any JSJ decomposition of
G over two-ended subgroups.
The quotient graph of groups G\Cyl(G) gives a canonical decomposition of G. It is
canonical in the sense that its Bass-Serre tree is G–equivariantly isomorphic to the tree
of cylinders of any JSJ decomposition of G over two-ended subgroups. However, such
a graph of cylinders is not necessarily a JSJ decomposition, and it does not even have
two-ended edge groups, in general. We return to this issue in Section 2.3.4.
2.3.2
Quasi-isometry invariance of the JSJ tree of cylinders
Since quasi-isometries coarsely preserve quasi-lines, and preserve the crossing and separating properties of quasi-lines, the following version of quasi-isometry invariance of JSJ
decompositions follows from Papasoglu’s work:
Theorem 2.9 (cf [36, Theorem 7.1]). Let G and G0 be finitely presented one-ended
groups. Suppose φ : G → G0 is a quasi-isometry. Then there is a constant C such that φ
induces an isomorphism φ∗ : Cyl(G) → Cyl(G0 ) that preserves vertex type — cylinder,
hanging, or rigid — and for v ∈ V Cyl(G) takes Gv to within distance C of G0φ∗ (v) .
11
Proof. Two-ended subgroups of G are bounded distance from each other if and only if
they are commensurable, so cylinders are coarse equivalence classes of edge and twoended vertex stabilizers. These are exactly the coarse equivalence classes of separating
quasi-lines that are not crossed by other separating quasi-lines, so quasi-isometries induce
a bijection between cylinders. The remaining statements are from Papasoglu [36, Theorem 7.1] for the rigid and hanging vertex stabilizers and Vavrichek [45] for the cylinder
stabilizers.
Corollary 2.10. If v is a rigid or hanging vertex in Cyl(G) then
φv := πφ∗ (v) ◦ φ|Gv ∈ QI((Gv , Pv ), (G0φ∗ (v) , Pφ∗ (v) ))
where πφ∗ (v) takes the image of φ|Gv to G0φ∗ (v) by closest point projection.
Remark. πφ∗ (v) is coarsely well defined since φ(Gv ) is within distance C of G0φ∗ (v) .
2.3.3
Boundary homeomorphism invariance of the JSJ tree of cylinders
In the case of a one-ended hyperbolic group that is not cocompact Fuchsian, Bowditch
constructed a canonical JSJ splitting of the group directly from the combinatorics of the
local cut points of the Gromov boundary of the group [9]. In this case, he proves that
the JSJ tree is unique, and such a tree is thus equivariantly isomorphic to the JSJ tree
of cylinders of the group. As a homeomorphism between the Gromov boundaries of two
hyperbolic groups preserves the topology, we get the following:
Theorem 2.11 (cf [9]). Let G and G0 be one-ended hyperbolic groups that are not cocompact Fuchsian. Suppose ρ : ∂G → ∂G0 is a homeomorphism between their Gromov
boundaries. Then ρ induces an isomorphism ρ∗ : Cyl(G) → Cyl(G0 ) that preserves vertex type — cylinder, hanging, or rigid — and for v ∈ V Cyl(G) the homeomorphism ρ
restricts to a homeomorphism ρ|∂Gv : ∂Gv → ∂G0ρ∗ (v) .
Corollary 2.12. For every vertex v ∈ Cyl(G):
ρv := ρ|∂Gv ∈ Homeo((∂Gv , ∂Pv ), (∂G0ρ∗ (v) , ∂Pρ0 ∗ (v) ))
2.3.4
Improved invariants from restrictions on the JSJ tree of cylinders
The JSJ tree of cylinders Cyl(G) of a finitely presented one-ended group G suffices for
the definition of the basic quasi-isometry invariants of Section 4. These are far from
complete invariants, however.
We can refine the invariants in restricted classes of groups. For instance, if the edges of
Cyl(G) have two-ended stabilizers in G then we can define stretch factors as in Section 5.
When the cylinder stabilizers are two-ended then the full power of Section 6 can be
brought to bear. In this case the graph of cylinders is a canonical JSJ decomposition of
G over two-ended subgroups. This is the case of chief interest for this paper, and this is
always the case if G is hyperbolic.
12
If the cylindrical vertices of Cyl(G) have two-ended stabilizers then they are all finite valence in Cyl(G). Furthermore, if Γ is a JSJ decomposition of G over two-ended
subgroups and v ∈ T (Γ) is a vertex whose stabilizer is rigid or hanging then v belongs
to more than one cylinder, so Cyl(G) has a vertex corresponding to v with the same
stabilizer subgroup in G. Thus, Cyl(G) is bipartite, with one part, VC , consisting of
finite valence cylindrical vertices, one for each cylinder of T (Γ), and the other part consisting of the sets of vertices VH and VR , which are all of infinite valence, corresponding
to hanging and rigid vertices of T (Γ), respectively.
See [23, Proposition 5.2] for a general result about when the tree of cylinders gives a
JSJ decomposition.
2.4
Trees of spaces
Let T be an oriented simplicial tree. For each vertex v ∈ T let Xv be a metric space. For
each edge e ∈ E + T let Xe be a subspace of Xι(e) , and let αe : Xe → Xτ (e) be a map such
that for Xē := αe (Xe ) there exists a map αē : Xē → Xι(e) such that αē ◦ αe is bounded
distance from IdXe and αe ◦ αē is bounded distance from IdXē .
Let X be the quotient of the set
a
a a
Xv t
{x} × e
v∈VT
e∈E + T x∈Xe
by the identifications x ∼ (x, ι(e)) and αe (x) ∼ (x, τ (e)). We call
X := X(T, {Xt }t∈VT ∪ET , {αe }e∈ET )
a tree of spaces over T . The Xv are called vertex spaces and the Xe are called edge
spaces. The sets {x} × e we call rungs, and metrize them as unit intervals. The maps αe
are called attaching maps.
We say X has locally finite edge patterns if for every vertex v, every x ∈ Xv , and
v
every R ≥ 0, there are finitely many e ∈ ET such that Xe intersects B R (x) := {y ∈ Xv |
dXv (x, y) ≤ R}.
A k–chain is a sequence of points x0 , x00 , x1 , . . . , x0k such that xi and x0i are contained
in a common vertex space or rung for all 0 ≤ i ≤ k and x0i and xi+1 are contained in
0
aPcommon vertex space or rung
for 0 ≤ i < k. The length of a k–chain is d(x0 , x0 ) +
0
0
0<i≤k d(xi−1 , xi ) + d(xi , xi ) , where the distance terms are interpreted in the vertex
space or rung containing the pair. The quotient pseudo-metric on X defines the distance
between two points to be the infinum of lengths of chains joining them. When x and x0
are points in vertex spaces then for a k–chain joining them we have that x0i and xi+1 are
the endpoints of a rung for each 0 ≤ i < k, so the length of the chain is at least k. The
following properties follow easily from this observation.
Lemma 2.13. For a tree of spaces X:
• The quotient pseudo-metric is a metric.
13
• A ball of radius at most 1 in a vertex space is isometrically embedded in X.
• The rungs are isometrically embedded.
Lemma 2.14. If X is a tree of spaces over T such that each vertex space is proper and
geodesic, each edge space is closed and discrete, and edge patterns are locally finite in
each vertex space, then X is a proper geodesic space.
Proof. For properness it suffices to show a closed ball centered at a point in a vertex
space is compact, since a ball centered at an interior point in a rung is contained in the
union of balls of radius 1 larger centered at the endpoints of the rung.
v
Take R > 0 and a point x ∈ Xv for some vertex v. Let N0 := B R (x). Since edge
v
patterns are locally finite, there are finitely many edges e such that Xe intersects B R (x).
v
v
For Xe ∩ B R (x) 6= ∅, the set Xe ∩ B R (x) is finite, since Xe is closed and discrete and
v
τ (e)
v
B R (x) is compact, so B R−1 (αe (Xe ∩ B R (x))) is compact.
Let N1,1 , . . . , N1,k1 be these finitely many compact sets. For each N1,i there are finitely
many e such that Xe intersects N1,i . Let N1,i,1 , . . . , N1,i,k1,i be the finitely many closed
R − 2 balls about the sets αe (Xe ∩ N1,i ) in their respective vertex spaces. Continue in
this way for R steps. Let N be the union of the finitely many N∗ , along with the finitely
many rungs connecting them. This is a finite union of compact sets, so it is compact,
and it contains B R (x), since it contains every chain of length less than or equal to R
starting from x. Thus, X is proper.
To see that X is geodesic, first consider two points x and x0 contained in vertex
spaces. By definition of the metric, there exists a sequence of chains joining x and x0
with length decreasing to d(x, x0 ). Since x and x0 are in vertex spaces, the length of a
k–chain joining them is at least k, so there is a subsequence of approximating chains
that are all k–chains for some fixed k ≤ d(x, x0 ). Let x0,i := x, x00,i , . . . , x0k,i := x0 be
the i–th k–chain in the sequence. Set x0 := x. Now, (x00,i ) is a bounded sequence in
Xv0 . Since edge spaces are closed and discrete and edge patterns are locally finite, (x00,1 )
contains a constant subsequence. Pass to a subsequence of chains such that (x00,i ) is
constant and define x00 to be its value. Since edge patterns are locally finite, x00 belongs
to finitely many Xe . Since attaching maps are coarsely invertible and edge spaces are
closed and discrete, there are only finitely many rungs with x00 as an endpoint. Therefore,
the sequence (x1,i ) takes only finitely many values. Choose one, define it to be x1 , and
pass to the subsequence with x1,i = x1 . Repeat this process k − 1 times, passing each
time to a subsequence of chains with one additional constant coordinate. We are left with
a constant sequence of chains with length decreasing to d(x, x0 ), so the length is equal to
d(x, x0 ). Each vertex space and rung are geodesic spaces, so we connect successive points
in the chain by geodesics in the vertex space or rung containing them to get a path with
length equal to length of the chain, which is therefore a geodesic from x to x0 .
If x0 is an interior point of a rung and x is in a vertex space then find geodesics from
x to the two endpoints of the rung and concatenate with the subsegment of the rung
leading to x0 . The shorter of these two paths is a geodesic. A similar argument works if
both x and x0 are interior points of rungs.
14
2.5
Algebraic trees of spaces
Let Γ be a graph of finitely generated groups. In this section we construct a tree of spaces
over T := T (Γ) that is quasi-isometric to G := G(Γ). The idea is to take the vertex
spaces to be Cayley graphs of the vertex stabilizers and use the edge injections of Γ to
define the attaching maps. The construction is standard, but there is some bookkeeping
involved that will be useful in Section 8.2.
For each v ∈ VΓ, choose a finite generating set for Gv and coset representatives h(v,i)
for G/Gv . For each e ∈ EΓ choose coset representatives g(e,i) of Gι(e) /Ge .
For each v ∈ VΓ, choose a lift ve ∈ VT . For each edge e ∈ EΓ, choose a lift ee ∈ ET
g Define fe := h
¯e = ē
with ι(e
e) = ι(e).
e. Given a maximal
(ι(e),j) g(e,i) for i and j such that fe e
subtree of Γ it is possible to choose lifts so that fe = 1 for all edges e such that e or e
belongs to the maximal subtree.
For t ∈ VT ∪ ET , let t ∈ Γ denote the image of t under the quotient by the G–action.
e ∈ T.
Let v be a vertex of T . There is a representative h(v,i) such that v = h(v,i) v
Define Yv to be a copy of the Cayley graph of Gv with respect to the chosen generating
set, which we identify with the coset h(v,i) Gv via left multiplication by h(v,i) .
Take the edge spaces to be cosets of the edge stabilizers of Γ, and define attaching
maps via containment. Specifically, for an edge e = h(v,i) g(e,j) ee ∈ T with ι(e) = v we
define Ye := h(v,i) g(e,j) Ge ⊂ h(v,i) Gv = Yv with attaching map:
−1
αe (x) := h(v,i) g(e,j) fe εe (g(e,j)
h−1
(v,i) x) ⊂ Yτ (e)
Let Y := Y (T, {Yt }, {αe }) be the resulting tree of spaces over T , which we call an
algebraic tree of spaces for Γ.
Lemma 2.15. An algebraic tree of spaces Y for Γ is quasi-isometric to G(Γ).
Proof. By Lemma 2.14, Y is a proper geodesic metric space. The group G(Γ) acts by left
multiplication, properly discontinuously and cocompactly by isometries on Y , so they are
quasi-isometric, by the Milnor-Švarc Lemma.
2.6
2.6.1
Trees of maps
Trees of quasi-isometries
Proposition 2.16. Suppose X and X 0 are trees of spaces over T and T 0 , respectively.
Suppose that χ : T → T 0 is an isomorphism. Suppose there exists M ≥ 1 and A ≥ 0 such
that:
0
• For each v ∈ VT there is an (M, A)–quasi-isometry φv : Xv → Xχ(v)
and a quasi0
isometry inverse φ̄v : Xχ(v) → Xv .
0
0
• For every e ∈ ET , the space φι(e) (Xe ) is A–coarsely equivalent to Xχ(e)
in Xχ(ι(e))
,
0
and the space φ̄ι(e) (Xχ(e) ) is A–coarsely equivalent to Xe in Xι(e) .
15
0
• For every e ∈ ET and every x ∈ Xe there exists a point x0 ∈ Xχ(e)
such that
0
0
0
d(φι(e) (x), x ) ≤ A and d(φτ (e) (αe (x)), αe (x )) ≤ A.
0
• For every e ∈ ET and every x0 ∈ Xχ(e)
there exists a point x ∈ Xe such that
0
0
0
d(φ̄ι(e) (x ), x) ≤ A and d(φ̄τ (e) (αχ(e) (x )), αe (x)) ≤ A.
Then there is a quasi-isometry φ : X → X 0 with φ|Xv = φv for each vertex v ∈ T .
Definition 2.17. A collection of quasi-isometries (φv ) satisfying the conditions given in
Proposition 2.16 is called a tree of quasi-isometries over χ compatible with X and X 0 .
Proof of Proposition 2.16. It suffices to consider the unions of the vertex spaces, which
form coarsely dense subsets of X and X 0 , and define φ by φ|Xv := φv . Define a coarse
inverse φ̄ to φ by φ̄|X 0 := φ̄v .
χ(v)
Suppose x and x0 are points in vertex spaces of X, and let x0 , x00 , x1 , . . . , x0k be a
k–chain joining them with length at most d(x, x0 ) + 1. Let Xvi be the vertex space
containing xi and x0i , and let ei be the edge such that x0i−1 and xi are endpoints of a
rung over ei .
Let y0 := φv0 (x0 ) and yk0 := φvk (x0k ). By hypothesis, for each x0i there exists a
0
0
yi ∈ Xχ(e
such that d(φι(e) (x0i ), yi0 ) ≤ A and d(φτ (e) (xi+1 ), yi+1 ) ≤ A, where yi+1 :=
i)
0
αχ(e)
(yi0 ). This gives a k–chain y0 , y00 , . . . , yk0 joining φ(x) and φ(x0 ) whose length is:
dX 0
χ(v0 )
(y0 , y00 ) +
X
0
d(yi−1
, yi ) + dX 0
χ(vi )
0<i≤k
(yi , yi0 )
≤ dX 0
(φv0 (x0 ), φv0 (x00 )) + A+
χ(v0 )
X
1 + 2A + dX 0 (φvi (xi ), φvi (x0i ))
χ(vi )
0<i≤k
≤ M dXv0 (x0 , x00 ) + 2A +
X
1 + 3A + M dXvi (xi , x0i )
0<i≤k
= 2A + k(1 + 3A) + M
X
dXvi (xi , x0i )
0≤i≤k
X
≤ 2A + max{1 + 3A, M } k +
dXvi (xi , x0i )
0≤i≤k
≤ 2A+
max{1 + 3A, M } dXv0 (x0 , x00 ) +
X
d(x0i−1 , xi ) + dXvi (xi , x0i )
0<i≤k
0
≤ 2A + max{1 + 3A, M }(d(x, x ) + 1)
Therefore, d(φ(x), φ(x0 )) ≤ max{1 + 3A, M }d(x, x0 ) + (2A + max{1 + 3A, M }), so φ
is coarsely Lipschitz. A similar computation shows φ̄ is coarsely Lipschitz, so φ is a
quasi-isometry.
16
Xι(e)
φι(e)
O
?
Xe
φe
?
/ X0
χ(e)
α0χ(e)
αe
Xτ (e)
/ X0
ι(χ(e))
O
φτ (e)
/ X0
τ (χ(e))
Figure 1: Commuting diagram for Corollary 2.18.
Corollary 2.18. Suppose χ : T → T 0 is an isomorphism of trees and X and X 0 are trees
of spaces over T and T 0 , respectively. Suppose there are M ≥ 1 and A ≥ 0 and (M, A)–
0
0
quasi-isometries φv : Xv → Xχ(v)
for each vertex and φe : Xe → Xχ(e)
for each edge such
that the diagram in Figure 1 commutes up to uniformly bounded error. Then (φv ) is a
tree of quasi-isometries over χ compatible with X and X 0 , so X is quasi-isometric to X 0 .
2.6.2
Gromov boundaries and trees of homeomorphisms
In this section, let X be a proper geodesic hyperbolic tree of spaces over a tree T ,
such that the vertex spaces are proper geodesic hyperbolic spaces, the edge spaces are
uniformly quasi-convex in X, and the attaching maps are uniform quasi-isometries.
For example, if Γ is a finite acylindrical graph of hyperbolic groups such that the
edge injections are quasi-isometric embeddings then the algebraic tree of spaces has this
structure [4, 24]. If, moreover, the edge groups are two-ended then the edge injections
are automatically quasi-isometric embeddings.
Quasi-convexity of edge spaces implies quasi-convexity of vertex spaces, so the Gromov boundary of each vertex space and
` edge space embeds into the boundary of X.
Consider the space ∂X := ∂T t v∈VT ∂Xv modulo identifying ∂X
` e ⊂ ∂Xι(e) with
∂Xē ⊂ ∂Xτ (e) via ∂ψe for each edge e. Let ∂Stab denote the image of v∈VT ∂Xv in ∂X.
Lemma 2.19. The inclusion of ∂X into the Gromov boundary of X is a surjection.
Proof. Pick a basepoint p ∈ X and let ζ : [0, ∞] → X̄ be a geodesic ray based at p.
Consider the projection π(p) of p to T .
If π ◦ζ crosses some edge e infinitely many times then there is an unbounded sequence
of times t1 , t2 , . . . such that ζ(ti ) ∈ Xe . Since the edge space is quasi-convex, ζ stays
bounded distance from Xe , so converges to a point in ∂Xe .
A similar argument shows that if π ◦ ζ visits a vertex v infinitely many times or is
eventually constant at v then ζ converges to a point in ∂Xv .
The remaining possibility is that π ◦ ζ limits to a point of ∂T . We claim there is a
unique asymptotic class of geodesic rays in X with π ◦ζ → η ∈ ∂T . Suppose ζ 0 is another
17
such ray. Let v0 be the vertex such that p ∈ Xv0 . Let e be an edge in T on the geodesic
from v0 to η. Let e0 be another edge on the geodesic from v0 to η that is distance at least
δ + Q from e, where δ is the hyperbolicity constant for X and Q is the quasi-convexity
constant for edge spaces. Let x be a point of ζ ∩ Xe and y a point of ζ ∩ Xe0 . Let x0 be
a point of ζ 0 ∩ Xe and y 0 a point of ζ 0 ∩ Xe0 . Consider a geodesic triangle whose sides
are ζ|[p,y] , ζ 0 |[p,y0 ] , and a geodesic ζ 00 from y to y 0 . By quasi-convexity, d(x, ζ 00 ) > δ and
d(x0 , ζ 00 ) > δ. Therefore, hyperbolicity implies d(x, ζ 0 ) ≤ δ and d(x0 , ζ) ≤ δ. This implies
ζ and ζ 0 are asymptotic.
Now we define a topology on ∂X and show it is equivalent to the Gromov topology.
Definition 2.20 (domains). The domain D(η) of a point η ∈ ∂T is the singleton {η}.
The domain D(ξ) of a point ξ of ∂Stab X is the subtree of T spanned by those vertices v
of T such that ∂Xv contains a point in the equivalence class ξ.
Remark 2.21. It can be proved that domains of points of ∂Stab X have a uniformly
bounded number of edges, see [30].
We can now define neighborhoods of points of ∂X, starting with points of ∂T .
Definition 2.22 (neighborhoods of points of ∂T ). Let η be a point of ∂T , and U be a
neighborhood of η in T ∪ ∂T . We define the neighborhood VU (η) as the set of points of
∂X whose domain is contained in U .
Before moving to neighborhoods of points of ∂Stab X, we need a definition.
Definition 2.23 (ξ-family, cone). Let ξ be a point of ∂Stab X. For every vertex v of D(ξ),
choose a neighborhood Uv of ξ in Xv ∪ ∂Xv . Let U be the collection of sets Uv , v ∈ D(ξ),
which we call a ξ-family. We define the set ConeU (ξ), called a cone, as the set of points
w ∈ (T ∪ ∂T ) \ D(ξ) such that if e is the last edge of the geodesic from w to D(ξ) in T ,
we have ∂αe (∂Xe ) ⊂ Uτ (e) .
Definition 2.24 (neighborhoods of points of ∂Stab X). Let ξ be a point of ∂Stab X, and
U be a ξ-family. We define the neighborhood VU (ξ) as the set of points η of ∂X such
that the following holds:
• D(η) \ D(ξ) is contained in ConeU (ξ),
• for every vertex v of D(ξ) ∩ D(η), we have η ∈ Uv .
Theorem 2.25 ( [30, Corollary 9.19]). With the topology described above, the inclusion
of ∂X into the Gromov boundary of X is a homeomorphism.
Definition 2.26. Let X and X 0 be proper geodesic hyperbolic trees of quasi-convex
spaces over a trees T and T 0 , respectively. A tree of boundary homeomorphisms compatible
with X and X 0 over an isomorphism χ : T → T 0 consists of homeomorphisms ρv : ∂Xv →
0
∂Xχ(v)
for every vertex v ∈ T such that for ξ ∈ ∂Xv ∩ ∂Xw we have ρv (ξ) = ρw (ξ) ∈
0
0
−1
∂Xχ(v) ∩∂Xχ(w)
, and for ξ ∈ ∂Xv0 ∩∂Xw0 we have ρ−1
v (ξ) = ρw (ξ) ∈ ∂Xχ−1 (v) ∩∂Xχ−1 (w) .
18
Proposition 2.27. Let X and X 0 be proper geodesic hyperbolic trees of quasi-convex
spaces over trees T and T 0 , respectively, with a compatible tree of boundary homeomorphisms (ρv ) over χ ∈ Isom(T, T 0 ). Then there is a homeomorphism ρ : ∂X → ∂Y defined
by ρ|∂Xv := ρv and ρ|∂T := ∂χ.
Proof. We have a well defined map ρ and its inverse ρ−1 is defined by ρ−1 |∂Xv0 := ρ−1
χ−1 (v)
and ρ−1 |∂T := ∂χ−1 . It clear that ρ and ρ−1 are continuous with respect to the topology
given in Definition 2.22 and Definition 2.24, which is equivalent to the standard topology
by Theorem 2.25.
Theorem 2.28. Let G and G0 be one-ended hyperbolic groups with non-trivial two-ended
JSJ decompositions. Let X and X 0 be algebraic trees of spaces over the respective JSJ
trees of cylinders T := Cyl(G) and T 0 := Cyl(G0 ). Every homeomorphism ρ : ∂X → ∂X 0
splits as a tree of compatible boundary homeomorphisms over the isomorphism ρ∗ : T →
T 0 . Every tree of boundary homeomorphisms (ρv ) compatible with X and X 0 over an
isomorphism χ : T → T 0 patches together to give a homeomorphism ρ : ∂X → ∂X 0 with
ρ∗ = χ and ρ|∂Xv = ρv .
Proof. Since the edge groups are two-ended they are virtually cyclic, hence quasi-convex.
Therefore, the boundary of each vertex space embeds into the boundary of its tree of
spaces. By Theorem 2.11, ρ induces an isomorphism ρ∗ : T → T 0 and ρ(∂Xv ) = ∂Xρ0 ∗ (v) .
Since these spaces are embedded, ρ|∂Xv ∈ Homeo((∂Xv , ∂Pv ), (∂Xρ0 ∗ (v) , ∂Pρ0 ∗ (v) )), so ρ
splits as a tree of boundary homeomorphisms over ρ∗ .
The converse is Proposition 2.27.
3
Distinguishing orbits: Basic examples
Given a group G, a common strategy in understanding groups quasi-isometric to G is
to first understand self-quasi-isometries of G. We explained in the previous section that
QI(G) acts on Cyl(G), preserving the relative quasi-isometry types of vertices. A first
step towards understanding QI(G) is to understand its action on Cyl(G). In particular,
we would like to determine the QI(G)–orbits of vertices in Cyl(G). We know that there
are finitely many G–orbits of vertices in Cyl(G), and that QI(G)–orbits are unions of
G–orbits, so the problem reduces to distinguishing G–orbits that are not contained in a
common QI(G)–orbit.
In this section we show on some particular examples how one may distinguish G–
orbits. These examples motivate the technical machinery of the following sections, in
which we combine and iterate the considerations presented here.
Our examples are hyperbolic, so, as usual, we consider the cases of quasi-isometry
and boundary homeomorphism simultaneously by letting Map(G) mean either QI(G) or
Homeo(∂G), as appropriate.
In general our techniques reduce questions about graphs of groups to questions about
the vertex groups, so our results are strongest, and the examples are most illuminating,
19
when the vertex groups are well understood. First we establish such a well understood
vertex group, and then we use it to construct graph of group examples.
Consider the free group F2 = ha, bi. Let X be the Cayley graph of F2 with respect to
the generating set {a, b}. Recall that if H is a finite collection of two-ended subgroups
of F2 then there is a corresponding peripheral structure P on X consisting of distinct
coarse equivalence classes of conjugates of elements of H. Since X is a tree, the coarse
equivalence class of a two-ended subgroup contains a unique bi-infinite geodesic, so we
can think of P as an F2 –invariant collection of bi-infinite geodesics.
We claim that J(X, Pi )K = J(X, Pj )K for all i and j where Pi is the peripheral structure
induced by Hi defined as follows:
• H0 = {ha2 b2 abi}
• H3 = {hai, habab2 i}
• H1 = {hab2 i, ha2 bi}
• H4 = {hai, hbi, hababi}
• H2 = {habi, ha2 b2 i}
• H5 = {hai, hbi, habi, habi}
This follows from [14, Theorem 6.3]. The reason is that the Whitehead graph for
each of these examples is the complete graph on four vertices. For each i, there are
exactly 6 bi-infinite geodesics representing elements of Pi passing through the vertex of X
corresponding to the identity element of F2 . Given i and j, any choice of bijection between
these six elements of Pi and the six corresponding elements of Pj extends uniquely to an
element of QI((X, Pi ), (X, Pj )). Conversely, every element of QI((X, Pi ), (X, Pj )) is of
this form, up to pre- and post-composition with the F2 –action.
When F2 is a vertex group of a graph of groups and the peripheral structure coming
from incident edge groups is one of the Pi above then the vertex is rigid.
For variety, we give two more rigid examples.
• F3 = ha, b, ci relative to the peripheral structure induced by ha2 b2 c2 acbi. The
Whitehead graph for this example is the complete bipartite graph with parts of
three vertices each. This peripheral structure is not quasi-isometrically equivalent
to the Pi given above.
• π1 (M ), for M a closed hyperbolic 3–manifold, relative to the peripheral structure
induced by hgi, where g is a non-trivial, indivisible element of π1 (M ).
3.1
First example: wrong type, bad neighbors, stretch factors
Consider the graph of groups Γ shown in Figure 2. It is the graph of cylinders for the
one-ended hyperbolic group G := G(Γ), so T := T (Γ) = Cyl(G). The label below a
vertex is its name, and the label above a vertex is the associated vertex group. All of the
edge groups are infinite cyclic. Suppose that we have chosen a generator for each infinite
cyclic vertex and edge group. Each end of an edge is labelled to indicate the image of
this generator in the incident vertex group. For cyclic vertices we give a non-zero integer
indicating that the image of the edge generator is that power of the vertex generator.
20
2
ab
ab2
Z
a2 b̄2
F2 ab̄2
Z
r1
c2
F2 ab̄2
Z
r2
c3
F2 ab
Z
r3
c1
ab2
ab2
g
a2 b2 āb̄
F2 ab̄2
Z
a2 b2 āb̄
F2 ab̄2
Z
c6
F2
r8
abāb̄
c5
r5
F2
r7
c4
r4
π1 (M )
r6
F2
h
a2 b2 c2 acb
F3
r9
Figure 2: First example Γ
This label is omitted when it is equal to 1, which is the case for all edges in this particular
graph of groups.
For non-cyclic vertices we indicate the element that is the image of the chosen generator of the edge group. For economy of notation we reuse a and b as generators of all
the F2 vertex groups. These are distinct subgroups of G, so more properly we would say,
for instance, that Gr1 = ha1 , b1 i and the two incident edge injections send the generators
of their respective edge groups to a1 b21 and a1 b21 .
In this section we describe the Map(G)–orbits of vertices in T .
3.1.1
Wrong type
The vertex h is the only hanging vertex, so Ge
h is a Map(G)–orbit.
The vertices labelled ci are cylindrical, while the vertices labelled ri are rigid, so there
are no i and j such that rei and cej are in the same Map(G)–orbit.
The vertex group of r6 is not quasi-isometric or boundary homeomorphic to any of
the other vertex groups, so Gre6 is a Map(G)–orbit.
The vertex group of r9 relative to the peripheral structure induced by the incident
edge is not the same relative quasi-isometry or boundary homeomorphism type of any
other vertex, so Gre9 is a Map(G)–orbit.
3.1.2
Bad neighbors
The cylindrical vertices have finite valence in T . The valence of vertices in Gce1 is 5, while
for all other cylindrical vertices it is 2. Thus, Gce1 is a Map(G)–orbit.
Vertices in Gce2 are the only ones adjacent to vertices in the Map(G)–orbit Gre6 , so
Gce2 is a Map(G)–orbit.
21
Vertices in Gce2 are adjacent only to vertices in Gre1 and Gre6 , but we already know
that Gre6 is a Map(G)–orbit, so Gre1 is a Map(G)–orbit as well.
By the same sort of argument, we quickly conclude that Gce5 , Gre4 , Gce6 , and Gre5 are
distinct Map(G)–orbits.
Such arguments based on distinguishing vertices by their neighbors are the contents
of Section 4.
3.1.3
Stretch factors
It remains to differentiate the second and third branches of Γ. This is more subtle, as the
only difference is at vertices r2 and r3 , which do have the same relative quasi-isometry
type.
Notices that the generator of Gc3 is identified with an element of word length 6 in
Gr7 , and length 3 in Gr2 , whereas the generator of Gc4 is identified with elements of
word lengths 6 and 2. It is not at all obvious that these lengths are preserved by quasiisometries. Nevertheless, we will show that in this particular example ce3 and ce4 can be
distinguished up to quasi-isometry by the fact that the ratios 63 and 62 are not equal.
These ratios correspond to an invariant called the ‘stretch factor’, which is developed in
Section 5.
By neighbor arguments as above we can then conclude that QI(G)–orbits of T are
the same as G–orbits.
The stretch factor is only an invariant for quasi-isometries, not boundary homeomorphism. In this example we cannot, in fact, distinguish the second and third branch of Γ
up to boundary homeomorphism, which is to say, Gre2 ∪Gre3 is a single Homeo(∂G)–orbit,
as are Gce3 ∪ Gce4 and Gre7 ∪ Gre8 .
3.2
Second example: unbalanced cylinders
In general, if L is an element of a peripheral structure P on X, there is no reason to believe
that there is an element of Map((X, P)) that preserves L and reverses its orientation.
We briefly give an idea how one may construct such an example. Consider the genus
7 surface Σ in Figure 3, with a hyperbolic metric.
Let f ∈ π1 (Σ) be the element represented by the curve running around the central
e = H2 . The figure shows
hole. Let L be a component of the preimage of this curve in Σ
three curves the separate Σ into subsurfaces Σ1 , Σ2 , and Σ3 , each of which is a genus
two surface with two boundary components. Suppose each Σi contains a collection of
curves that is complicated enough so that all the complementary components are discs,
and different enough so that there does not exist a crossing-preserving bijection between
the components of the preimages of the curves of Σi contained in one component of the
e and the components of the preimages of the curves of Σj contained in
preimage of Σi in Σ
e induced by
one component of the preimage of Σj . Let P be the peripheral structure on Σ
all of the above curves in Σ, which is to say, each element of P is the coarse equivalence
e of one of the curves in Σ.
class of a component of the preimage in Σ
22
Figure 3: The surface Σ
Then there is no element of Map(π1 (Σ), P) that reverses L. The reason is that L
passes through components of the preimages of Σ1 , Σ2 , and Σ3 in order. If this order
were reversible by an element of Map((π1 (Σ), P)) we would contradict the restriction
that the pattern of curves in Σ1 , Σ2 , and Σ3 were chosen to be ‘different’.
Now consider the graph of groups Γ of Figure 4, where Σ and f are as above, and
g is a non-trivial, indecomposable element of a closed hyperbolic 3–manifold M . Then
G := G(Γ) is a one-ended hyperbolic group with T := T (Γ) = Cyl(G).
r4
π1 (M )
g
c2
r1
Z
π1 (Σ)
f
Z
−1
r3
c1
π1 (Σ)
f
r2
f
π1 (Σ)
Figure 4: Second example Γ
The graph of groups Γ has a central cylindrical vertex c1 that attaches to three rigid
vertices r1 , r2 , and r3 , each of which has local group π1 (Σ), with each attachment along a
copy of f . For each ri there are some number of other incident edges, each corresponding
to one of the curves in Σ described above, so that the peripheral structure on Gri induced
e Each of these
by edge inclusions is the peripheral structure P described above for Σ.
edges we attach to another cylindrical vertex, and then to a rigid vertex carrying a copy
of π1 (M ).
23
We consider a kind of ‘f –parity’ by counting the number of vertices adjacent to ce1
for which the generator of Gce1 is identified with f minus the number of vertices adjacent
to ce1 for which the generator of Gce1 is identified with f . Since ce1 has non-zero f –parity,
we call it an unbalanced cylinder. We claim that there is not an element of Map(G) that
reverses the orientation of an unbalanced cylinder. In this example, we then conclude
that Gre3 is a Map(G)–orbit and Gre1 ∪ Gre2 is a different Map(G)–orbit.
The proof is as follows. Since Gce1 is the only G–orbit of cylindrical vertices in T of
valence 3, if there were an element of Map(G) taking re3 to, say, re1 , then it would fix
ce1 . There is clearly an element of Map((Gr3 , Pr3 ), (Gr1 , Pr1 )) taking f to f , but, since f
is not reversible, if such a map extended to an element of Map(G) it would necessarily
reverse the orientation of Gce1 . Since f is not reversible, this would mean that every
vertex in which f is identified with the generator of Gce1 must be sent to a vertex in
which f is identified with the generator of Gce1 . This means that both re1 and re2 are sent
to re3 , contradicting the fact that Map(G) acts by isomorphisms on T .
Dealing with issues of orientation and unbalanced cylinders is the focus of Section 6.2.
4
Decorated trees and structure invariants
Let G be a group and let T be a simplicial tree of countable valence upon which G
acts cocompactly and without inverting an edge. In this section, we explain how to
associate to T an invariant, the structure invariant, that completely characterizes it up
to decoration-preserving isomorphism. The construction of this invariant generalizes the
‘degree refinement’ invariant of graph theory, which is an invariant that determines when
two finite graphs have isomorphic universal covers.
Definition 4.1. A decoration is a G–invariant map δ : T → O that assigns to each
vertex of T an ornament o ∈ O.
For simplicity, in this section we decorate only vertices. In the next section we will
also decorate edges, with the condition that δ(e) = δ(ē) for every edge. Formally, this
can be accomplished by subdividing each edge of T and decorating
` the new vertices.
Corresponding to a decoration there is a partition of T as o∈O δ −1 (o). We say
that a decoration δ 0 : T → O0 is a refinement of δ if the δ 0 –partition is finer than the
δ–partition. Equivalently, a decoration δ 0 : T → O0 is a refinement of δ : T → O if there
exists a surjective map π : Im δ 0 → Im δ such that π ◦ δ 0 = δ. We say δ 0 is a strict
refinement if the δ 0 –partition is strictly finer than the δ–partition. A refinement that is
not strict is a trivial refinement.
In this section, we will define a structure invariant for a decorated tree, such that we
can reconstruct the tree, up to decoration-preserving tree automorphism, from the data
of the structure invariant. In order to do so, we want to identify orbits in T under the
action of the group Aut(T, δ) := {χ ∈ Aut(T ) | δ ◦ χ = δ}, and then to say how they fit
together. In a nutshell, the idea is as follows.
Suppose v and w are two vertices. If there is a χ ∈ Aut(T, δ) with χ(v) = w, then,
of course, δ(v) = δ(w). Additionally, χ gives a decoration-preserving bijection from the
24
neighbors of v to the neighbors of w. Thus, for each ornament o, the number of neighbors
of v bearing o must be equal to the number of neighbors of w bearing o.
Conversely, if δ(v) = δ(w), but for some ornament o there are differing numbers
of neighbors of v bearing o and neighbors of w bearing o, then there is no decorationpreserving automorphism taking v to w, so we ought refine the decoration to distinguish
v from w. We then repeat this refinement process until vertices with the same ornament
can no longer be distinguished by the ornaments of their neighbors. This happens after
finitely many steps because G\T is compact, see Proposition 4.3. Section 4.1 formalizes
this process, which we call neighbor refinement.
4.1
Neighbor refinement
Let N := N ∪ {0, ∞}. Call O0 := O and δ0 := δ the ‘initial set of ornaments’ and the
‘initial decoration’, respectively. Beginning with i = 0, for each v ∈ VT define:
fv,i : Oi → N : o 7→ #{w ∈ δi−1 (o) | w is adjacent to v}
O
Define Oi+1 := O0 × N i , and define δi+1 (v) := (δ0 (v), fv,i ).
Lemma 4.2. For all i, the map δi+1 : T → Oi+1 is a decoration that refines δi : T → Oi .
Proof. Let v be a vertex, and let g ∈ G. Suppose δi is G–invariant. Then, δi+1 (gv) =
(δi (gv), fgv,i ) = (δi (v), fv,i ) = δi+1 (v). Since δ0 is G–invariant, all the δi are decorations
by induction.
For each i, let N (v, i, o) denote the number of neighbors of v in δi−1 (o).
Clearly, δ1 refines δ0 , since δ0 is the composition of δ1 with projection to the first
coordinate of theP
image. Suppose that δi refines δi−1 . Then for every o0 ∈ Oi−1 , we have
0
N (v, i − 1, o ) = o∈δi ◦δ−1 (o0 ) N (v, i, o).
i−1
If δi+1 (v) = δi+1 (w) then δ0 (v) = δ0 (w) and N (v, i, o) = N (w, i, o) for each o ∈ Oi .
Thus, for all o0 ∈ Oi−1 ,
X
X
N (v, i − 1, o0 ) =
N (v, i, o) =
N (w, i, o) = N (w, i − 1, o0 ),
−1
o∈δi ◦δi−1
(o0 )
−1
o∈δi ◦δi−1
(o0 )
so δi (v) = δi (w). Hence δi+1 refines δi . The lemma follows by induction.
Proposition 4.3. There exists an s ≥ 0 such that δi+1 is a strict refinement of δi for
all i + 1 ≤ s and δi+1 is a trivial refinement of δi for all i ≥ s.
−1
: Im δi+1 ⊂ Oi+1 → Oi . The
Proof. Since δi+1 refines δi there is a function δi ◦ δi+1
refinement is trivial precisely when this function is injective.
Suppose there exists an s such that δs+1 is a trivial refinement of δs . For every
−1
o ∈ Im δs+1 ⊂ Os+1 we have fv,s+1 (o) = fv,s ◦ δs ◦ δs+1
(o), and fv,s+1 (o) = 0 otherwise.
Therefore, δs+1 (v) = δs+1 (w) implies fv,s = fw,s , which implies fv,s+1 = fw,s+1 , which
implies δs+2 (v) = δs+2 (w). Thus, once one refinement in the sequence is trivial so are all
further refinements.
25
To see that eventually some refinement`is trivial, note that G–invariance implies that
for all i we have a partition of T by T = o∈Oi δi−1 (o) in which each part is a union of
G–orbits. A refinement is strict if and only if the new partition has strictly more parts
than the previous one. However, the number of parts is bounded above by the number
of G–orbits, which is finite.
Definition 4.4. The neighbor refinement of δ is the decoration δs : T → Os at which
the neighbor refinement process stabilizes.
Proposition 4.5. Let δ 0 : T → O0 be the neighbor refinement of δ : T → O. The δ 0 –
partition of T is equal to the partition into Aut(T, δ)–orbits and to the partition into
Aut(T, δ 0 )–orbits.
Proof. The partition into Aut(T, δ)–orbits is finer than the δ 0 –partition because each
refinement step is Aut(T, δ)–equivariant. The Aut(T, δ 0 )–orbit partition is finer than
the Aut(T, δ)–orbit partition since δ 0 is a refinement of δ. We show the δ 0 –partition
is finer than the Aut(T, δ 0 )–orbit partition by supposing δ 0 (v) = δ 0 (w) and producing
χ ∈ Aut(T, δ 0 ) with χ(v) = w.
The automorphism χ is constructed inductively. Start by defining χ(v) := w. Since
0
δ is stable under neighbor refinement, for every o ∈ O0 we have # lk(v) ∩ (δ 0 )−1 (o) =
# lk(w)∩(δ 0 )−1 (o). Extend χ by choosing any bijection between these sets. This extends
χ to the 1–neighborhood of v.
Now suppose χ is defined on a subtree T 0 of T such that for every v 0 ∈ VT 0 , either v 0 is
a leaf or T 0 contains every edge incident to v 0 . Let v 0 be a leaf, and let u be the vertex of
T 0 adjacent to v 0 . For δ 0 (u)
χ via a bijection between lk(v) \ {u} ∩ (δ 0 )−1 (δ 0 (u))
extend
0
0
−1
and lk(χ(v )) \ {χ(u)} ∩ (δ ) (δ 0 (u)). For o 6= δ 0 (u) extend χ just as in the base
case.
4.2
Structure invariants
Let δs : T → Os be a neighbor refinement of δ as in Proposition 4.3. As defined, δs
actually encodes the history of the refinement process, not just the structure of T . We
define the structure invariant by forgetting this extraneous history. Let π0 : Os → O be
projection to the first coordinate. Choose an ordering of Im δ and let O[j] denote the
j–th ornament in the image. Similarly, for each 1 ≤ j ≤ # Im δ, choose an ordering of
π0−1 (O[j]) ∩ Im δs . Order Im δs lexicographically, and let Os [i] denote the i–th ornament.
Definition 4.6. The structure invariant S(T, δ, O) is the # Im δs ×# Im δs matrix whose
j, k–entry is the tuple consisting of the number of vertices in δs−1 (Os [j]) adjacent to each
vertex of δs−1 (Os [k]), the O–ornament π0 (Os [j]) and the O–ornament π0 (Os [k]). The
last two coordinates are called respectively the row and column ornament of the entry.
S(T, δ, O) can be seen as a block matrix, with blocks consisting of entries with the
same row ornaments and the same column ornaments. The structure invariant is well
defined up to permuting the O–blocks and permuting rows and columns within O–blocks,
ie, up to the choice of orderings of O and the π0−1 (O[j]).
26
Proposition 4.7. Let δ : T → O be a G–invariant decoration of a cocompact G–tree.
Let δ 0 : T 0 → O be a G0 –invariant decoration of a cocompact G0 –tree. There exists a
decoration-preserving isomorphism φ : T → T 0 if and only if S(T, δ, O) = S(T 0, δ 0, O), up
to permuting rows and columns within O–blocks.
In particular, with the above notations, T and T 0 must have the same sets of ornaments for a decoration-preserving isomorphism φ to exist.
Proof. It is clear that isomorphic decorated trees have the same structure invariants,
up to choosing the orderings of the ornaments. For the converse, assume that we have
reordered within O–blocks so that S(T, δ, O) = S(T 0, δ 0, O) = S. Construct a decorationpreserving tree isomorphism exactly as in the proof of Proposition 4.5.
Remark. When T is the universal cover of a finite graph Γ and the initial set of ornaments
is trivial then the structure invariant we have defined is just the well known degree
refinement of Γ. The lemma says that two graphs have the same degree refinement if
and only if they have isomorphic universal covers. A theorem of Leighton [27] says that
such graphs in fact have a common finite cover. There are also decorated versions of
Leighton’s Theorem, eg [34].
Observation. We get a quasi-isometry invariant of a group G by taking the structure
invariant of a cocompact QI(G)-tree with a QI(G)–invariant decoration.
This is a simple observation, but it does not seem to have appeared in the literature
in this generality.
Behrstock and Neumann [2, 3] have used special cases of this type of invariant, in a
different guise, to classify fundamental groups of some families of compact irreducible
3–manifolds of zero Euler characteristic. In both papers the tree is the Bass-Serre tree
for the geometric decomposition of such a 3–manifold along tori and Klein bottles, which
is the higher dimensional antecedent of the JSJ decompositions considered in this paper.
When the geometric decomposition has only Seifert fibered pieces the vertices are
decorated by the quasi-isometry type of the universal cover of the corresponding Seifert
fibered manifold. There are only two possible quasi-isometry types, according to whether
or not the Seifert fibered piece has boundary. Every vertex in the Bass-Serre tree has
infinite valence, so each entry of the structure invariant is either 0 or ∞.
Behrstock and Neumann [2] state their result in terms of ‘bi-similarity’4 classes of
bi-colored graphs. They show that each bi-similarity class is represented by a unique
minimal graph, and that two such 3–manifolds are quasi-isometric if and only if the bicolored Bass-Serre tree of their geometric decompositions have the same representative
minimal graph. Their minimal bi-colored graphs carry exactly the same information as
the structure invariant of the decorated Bass-Serre tree. One can construct their graph
by taking the vertex set to be the stable decoration set Os and connecting vertex Os [j]
to vertex Os [k] by an edge if and only if the j, k–entry of S is ∞. The vertices of the
graph are ‘bi-colored’ by the projection π0 : Os → O. Conversely, S can be recovered by
4
Their meaning of ‘similarity’ is different than in this paper.
27
replacing each edge in the graph by infinitely many edges, lifting the bi-coloring to the
universal covering tree, and calculating the structure invariant.
The second paper [3] extends their results to cases where the decomposition involves
some hyperbolic pieces. The decorations there are more complex.
4.3
Structure invariants for the JSJ tree of cylinders
Combining Proposition 4.7 with Theorem 2.9 and Corollary 2.10 proves:
Theorem 4.8. If G is a finitely presented one-ended group not commensurable to a
surface group, then the structure invariant for the JSJ tree of cylinders is a quasi-isometry
invariant of G, with respect to any of the following initial decorations:
1. Vertex type: rigid, hanging, or cylinder.
2. Vertex type and, if v is rigid, JGv K.
3. Vertex type and, if v is rigid, J(Gv , Pv )K.
Theorem 4.9. If G is hyperbolic and the JSJ decomposition of G has no rigid vertices
then the invariant of Theorem 4.8 is a complete quasi-isometry invariant.
Later we will prove a more general result, Theorem 8.5, that includes Theorem 4.9
as a special case. A brief sketch of a direct proof of Theorem 4.9 goes like this: Given
two hyperbolic groups as in Theorem 4.9, the structure invariants of Theorem 4.8 are
equivalent if and only if the groups have isomorphic decorated JSJ trees of cylinders.
Since all non-cylindrical vertices are hanging, it follows, using techniques of of Behrstock
and Neumann [2], that the groups are quasi-isometric if and only if they have isomorphic
decorated JSJ trees of cylinders. Details of the last claim can be found in Dani and
Thomas [15, Section 4]. The torsion-free case was previously written up in the thesis of
Malone [28].
4.4
4.4.1
Examples
An example with symmetry
Consider the graph of groups Γ in Figure 5, for which T (Γ) = Cyl(G(Γ)). Let G := G(Γ)
and T := T (Γ).
ha, bi
r0
a2 b2 āb̄
Z
c0
xy x̄z̄
hx, y, zi zy
h
Z
c
a2 b2 āb̄
ha, bi
r
Figure 5: Symmetric example
Take the initial decoration be vertex type, so δ0 (r), δ0 (r0 ) := ‘rigid’; δ0 (c), δ0 (c0 ) :=
‘cylindrical’; and δ0 (h) := ‘hanging’.
Let us compute the first neighbor refinement: The image of an edge injection into
a rigid vertex group is an infinite index subgroup, so rigid vertices of T are infinite
28
valence, and adjacent only to cylindrical vertices. Since δ0 does not distinguish between
cylindrical vertices, we have:
‘cylindrical’ 7→ ∞
fre,0 , fre0 ,0 := ‘rigid’
7→ 0
‘hanging’
7→ 0
In this example each edge group surjects onto its adjacent cylindrical vertex group,
so cylindrical vertices of T have valence 2. One neighbor is rigid and one is hanging, and
δ0 does not distinguish between the rigid vertices, so we have:
‘cylindrical’ 7→ 0
fec,0 , fce0 ,0 := ‘rigid’
7→ 1
‘hanging’
7→ 1
Finally, the image of an edge injection into a hanging vertex group is an infinite index
subgroup, so hanging vertices of T are infinite valence, and adjacent only to cylindrical
vertices. Since δ0 does not distinguish between cylindrical vertices, we have:
‘cylindrical’ 7→ ∞
:=
feh,0
‘rigid’
7→ 0
‘hanging’
7→ 0
The first refinement is therefore:
δ1 (e
r), δ1 (re0 ) := (‘rigid’, fre,0 )
δ1 (e
c), δ1 (ce0 ) := (‘cylindrical’, fec,0 )
δ1 (e
h) := (‘hanging’, feh,0 )
This is a trivial refinement, so δ0 is stable, and the structure invariant is given in Table 1.
‘cylindrical’ ‘rigid’ ‘hanging’
‘cylindrical’
0
1
1
‘rigid’
∞
0
0
‘hanging’
∞
0
0
Table 1: Structure invariant for symmetric example
Notice that we get the same structure invariant for the group G0 := G(Γ0 ) defined by
the graph of groups in Figure 6. This is expected, as the groups are quasi-isometric. In
fact, G is isomorphic to an index 2 subgroup of G0 .
29
hp, qi
Z
pq p̄q̄
h
Figure 6:
4.4.2
a2 b2 āb̄
c
ha, bi
r
Γ0
The example of Section 3.1
Let Γ be the graph of groups of Figure 2, for which T (Γ) = Cyl(G(Γ)). Let G := G(Γ)
and T := T (Γ). In this example we take the initial decoration to be by vertex type and
quasi-isometry type of vertex stabilizer. The vertices of T carry four possible ornaments,
(r, Jπ1 (M )K), (r, JF2 K), (h, JF2 K), and (c, JZK), where in the first coordinate r, h, and c
stand for ‘rigid’, hanging’, and ‘cylindrical’, respectively.
In the first refinement step, observe that each rigid and hanging vertex of T is adjacent
to infinitely many cylindrical vertices. The initial decoration does not distinguish any of
the cylindrical vertices, so δ1 does not give any new information about rigid and hanging
vertices.
On the other hand, we do distinguish some of the cylindrical vertices. This is because
δ0 does distinguish between three different kinds of rigid and hanging vertex, so we can
distinguish cylindrical vertices according to the number and kind of their neighbors.
re6
re1 , re2 , re3 , re4 , re5 , re7 , re8 , re9
e
h
ce1
ce2
ce3 , ce4 , ce6
ce5
(r, Jπ1 (M )K) (r, JF2 K) (h, JF2 K) (c, JZK)
(r, Jπ1 (M )K)
0
0
0
∞
(r, JF2 K)
0
0
0
∞
(h, JF2 K)
0
0
0
∞
0
5
0
0
1
1
0
0
(c, JZK)
0
2
0
0
0
1
1
0
Table 2: First refinement
Table 2 shows δ1 . In the first column, for convenience, we give representatives of
G–orbits of vertex v in T with a given ornament in O1 . The second column contains
δ0 (v). The remaining columns in each row give a row vector encoding fv,0 . For instance,
the last row says that a vertex v ∈ Gce5 has δ0 (v) = (c, JZK) and is adjacent to one vertex
in δ0−1 ((r, JF2 K)) and one vertex in δ0−1 ((h, JF2 K)).
Table 3 shows δ2 . The (c, JZK) columns are listed in the same order as the (c, JZK)
rows of Table 2. For instance, the fourth row says that for any given vertex in Gre2 , Gre3 ,
or Gre5 in T , it has infinitely many neighbors in Gce1 and infinitely many neighbors in
Gce3 ∪ Gce4 ∪ Gce6 and no other neighbors.
We can again count neighbors according to the new decoration, but we see in Table 4
that this does not distinguish any additional vertices. Thus, the refinement process has
stabilized, and Table 4 is the structure invariant. Up to permuting the ordering on O0
and the ordering of rows and columns within each O0 –block, this structure invariant
30
re6
re1
re4
re2 , re3 , re5
re7 , re8 , re9
e
h
ce1
ce2
ce3 , ce4 , ce6
ce5
(r, Jπ1 (M )K) (r, JF2 K) (h, JF2 K)
(c, JZK)
(r, Jπ1 (M )K)
0
0
0
0 ∞ 0 0
0
0
0
∞ ∞ 0 0
0
0
0
∞ 0 0 ∞
(r, JF2 K)
0
0
0
∞ 0 ∞ 0
0
0
0
0 0 ∞ 0
(h, JF2 K)
0
0
0
0 0 0 ∞
0
5
0
0 0 0 0
1
1
0
0 0 0 0
(c, JZK)
0
2
0
0 0 0 0
0
1
1
0 0 0 0
Table 3: Second refinement
re6
re1
re4
re2 , re3 , re5
re7 , re8 , re9
e
h
ce1
ce2
ce3 , ce4 , ce6
ce5
(r, Jπ1 (M )K)
(r, Jπ1 (M )K)
0
0
0
(r, JF2 K)
0
0
0
(h, JF2 K)
0
1
(c, JZK)
0
0
(r, JF2 K)
(h, JF2 K)
(c, JZK)
0 0 0 0
0
0 ∞ 0 0
0 0 0 0
0
∞ ∞ 0 0
0 0 0 0
0
∞ 0 0 ∞
0 0 0 0
0
∞ 0 ∞ 0
0 0 0 0
0
0 0 ∞ 0
0 0 0 0
0
0 0 0 ∞
1 1 3 0
0
0 0 0 0
1 0 0 0
0
0 0 0 0
0 0 1 1
0
0 0 0 0
0 1 0 0
1
0 0 0 0
Table 4: Structure invariant
completely determines T up to δ0 –preserving tree isomorphism.
5
5.1
A new decoration: Stretch factors
Relative quasi-isometric rigidity
In this section, we associate to the JSJ tree of a cylinders of a group new quasi-isometry
invariants that take into account the metric information carried by the various two-ended
edge groups. Indeed, consider an infinite order element of an edge stabilizer. Its image
in each of the adjacent vertex groups has some translation length, and the ratio of these
translation lengths gives a stretch factor that describes how the amalgamation distorts
distance as measured in the vertex groups. This stretch factor clearly depends on the
31
choice of metrics of the vertex groups, so it is not an intrinsic invariant of the group,
and, in general, it is not preserved by quasi-isometries. However, we show that when
the vertex groups satisfy an appropriate notion of quasi-isometric rigidity—a notion that
is satisfied by many interesting classes of groups—then such stretch factors are indeed
quasi-isometry invariants.
Definition 5.1. A finitely generated group G is quasi-isometrically rigid relative to the
peripheral structure P, or (G, P) is quasi-isometrically rigid, if there exists a proper
geodesic metric space X with peripheral structure P 0 and a quasi-isometry µ : (G, P) →
(X, P 0 ) such that
1. µ∗ (QI(G, P)) is a uniform subgroup of CI((X, µ(P))).
2. If g ∈ G is an infinite order element fixing an element of P then i 7→ µ(g i ) is a
coarse similitude.
The pair (X, P 0 ) is called a rigid model for (G, P).
Remark. In Proposition 5.7 we prove that 1 implies 2 if G is hyperbolic.
Lemma 5.2. If (X, P 0 ) is a rigid model for (G, P) then µ0 ◦ µ ∈ CIsom((X, P 0 )) for any
µ, µ0 ∈ QIsom((G, P), (X, P 0 )).
This fact motivates the terminology ‘rigid’.
Proof. For µ, µ0 ∈ QIsom((G, P), (X, P 0 )) we have µ0 ◦ µ ∈ QIsom((X, P 0 ), (X, P 0 )). For
any µ ∈ QIsom((G, P), (X, P 0 )) we have CI((X, P 0 )) < QI((X, P 0 )) = µ∗ (QI((G, P))),
so if µ∗ (QI((G, P))) < CI((X, P 0 )) then QI((X, P 0 )) = CI((X, P 0 )).
Definition 5.3. Let (X, P 0 ) be a rigid model for (G, P), and let g be an infinite order
element of G that fixes an element of P. Define the X–length of g, `X (g), to be the
multiplicative constant of the coarse similitude from Z to X defined by i 7→ µ(g i ), where
µ ∈ QIsom((G, P), (X, P 0 )).
If a positive power g k of an infinite order element g fixes an element of P define
`X (g) := k1 `X (g k ).
Lemma 5.2 implies that `X (g) is independent of the choice of quasi-isometry µ ∈
QIsom((G, P), (X, P 0 )).
We remark in the second case that if π : Z → kZ is a closest point projection then
i 7→ µ(g π(i) ) is a coarse similitude from Z to X with multiplicative constant k1 `X (g k ), so
this is a sensible definition for `X (g).
5.2
Examples of relative quasi-isometric rigidity
Let G be a finitely presented group. Let H be a finite collection of two-ended subgroups
of G. Let P be the peripheral structure consisting of distinct coarse equivalence classes
of conjugates of elements of H.
32
1. If G is quasi-isometric to a space X such that I(X) = QI(X) then (G, P) is rigid.
The peripheral structure plays no role in this case. Examples include:
(a) Irreducible symmetric spaces other than real or complex hyperbolic space;
thick Euclidean buildings; and products of such [26, 35].
(b) The ‘topologically rigid’ hyperbolic groups of Kapovich and Kleiner [25].
(c) Certain Fuchsian buildings [7, 46].
(d) Mapping class groups of non-sporadic hyperbolic surfaces [1].
2. If G is quasi-isometric to a space X such that CI(X) = QI(X) then (G, P) is rigid.
Again, the peripheral structure plays no role in this case. Xie gives an example of
a certain solvable Lie group with this property [47].
3. If X is a real or complex hyperbolic space of dimension at least 3 and G is quasiisometric to X then (G, P) is quasi-isometrically rigid whenever H is non-empty,
by a theorem of Schwartz [40].
4. If X 0 is the 3-valent tree, G is quasi-isometric to X 0 (so G is virtually free), and
G does not virtually split over 0 or 2–ended subgroups relative to H, then (G, P)
is quasi-isometrically rigid [12, 14]. In this case the model space X depends on P,
and is not necessarily isometric to X 0 .
5. If X = H2 , φ : G → X is a quasi-isometry, and G does not virtually split over 2–
ended subgroups relative to H, then (G, P) is quasi-isometrically rigid, as follows.
A result of Kapovich and Kleiner [25] shows that G has finite index in QI((G, P)).
Therefore, QI((G, P)) is a finitely generated group quasi-isometric to X. This
quasi-isometry induces a cobounded quasi-action of QI((G, P)) on X. Such a
quasi-action is quasi-isometrically conjugate to an isometric action on X, by a
theorem of Markovic [29].
The first four cases actually satisfy the following stronger version of quasi-isometric
rigidity:
Definition 5.4. We say G is strongly quasi-isometrically rigid relative to P, or (G, P)
is strongly quasi-isometrically rigid, if there is a proper geodesic space X such that if
(X, P 0 ) and (X, P 00 ) are rigid models for (G, P), then there is a coarse isometry φ of X
such that φ(P 0 ) = P 00 .
By contrast, the last case only satisfies the weaker version of rigidity.
For a non-example, consider G := Zn and X := Rn . For any H the group QI((G, P))
contains maps conjugate to homotheties of Rn . This implies that the multiplicative
constants in QI((G, P)) are unbounded, so QI((G, P)) cannot be conjugate into some
coarse isometry group.
It is also easy to find non-examples of relative rigidity via splittings: If G virtually
splits over a zero or two-ended group relative to H then (G, P) is not quasi-isometrically
33
rigid. In such an example there are generalized Dehn twist quasi-isometries preserving
P, powers of which again produce unbounded multiplicative constants in QI((G, P)).
The previous examples and non-examples naturally lead to the following question:
Question 1. If G is a hyperbolic group that is not quasi-isometric to H2 and does not virtually split over a zero or two-ended subgroup relative to H, is (G, P) quasi-isometrically
rigid? Strongly quasi-isometrically rigid?
5.3
Relative quasi-isometric rigidity for hyperbolic groups
In Theorem 5.6 we give a characterization of relative quasi-isometric rigidity for hyperbolic groups. This combines with Proposition 5.7 to show that the first condition of
Definition 5.1 implies the second for hyperbolic groups. Theorem 5.6 also provides an
alternative viewpoint that may be useful for resolving Question 1.
A space X is called visual if there exists an A ≥ 0 and x ∈ X such that for every
y ∈ X there exists an A–coarse-geodesic ray starting at x and passing within distance A
of y. It follows from [6, Proposition 5.2 and Proposition 5.6] that if X is quasi-isometric
to a visual hyperbolic space then X is a visual hyperbolic space.
Definition 5.5. A map φ : X → Y is an (α, M )–power-quasi-symmetric embedding if
for all distinct x, y, z ∈ X
dY (φ(x), φ(z))
dX (x, z)
≤η
,
dY (φ(x), φ(y))
dX (x, y)
where
(
M r1/α
η(r) =
M rα
for 0 < r < 1,
for 1 ≤ r.
Let PQS((X, dX ), (Y, dY )) denote the set of power-quasi-symmetric homeomorphisms,
and abbreviate PQS(X, dX ) := PQS((X, dX ), (X, dX )), which is a group. If Z is a collection of subsets of X, define:
PQS(X, d, Z) := {φ ∈ PQS((X, d)) | ∀Z ∈ Z, φ(Z) ∈ Z and ∃!Z 0 ∈ Z, φ(Z 0 ) = Z}
Theorem 5.6. Let G be a non-elementary hyperbolic group with a peripheral structure
P consisting of coarse equivalence classes of conjugates of finitely many two-ended subgroups. Fix, arbitrarily, a word metric d on G, a basepoint p ∈ G, and a visual metric
d∞ on ∂G. The following are equivalent:
1. (G, P) is quasi-isometrically rigid.
2. There exists a proper, geodesic, visual hyperbolic space X and a quasi-isometry
µ : G → X such that µ∗ (QI((G, P))) is a uniform subgroup of CI((X, µ(P))).
3. There exists a visual metric d0∞ on ∂G such that PQS(∂G, d∞ , ∂P) is power-quasisymmetrically conjugate to a uniform subgroup of BiLip(∂G, d0∞ ).
34
Proof. Since G is a visual hyperbolic space, so is X. Thus, equivalence of 1 and 2 is
immediate from the definition of relative rigidity.
Item 2 implies item 3 by taking d0∞ to be a visual metric on ∂X = ∂G and applying [6,
Theorem 6.5], which shows that µ extends to a power-quasi-symmetry of ∂G, and a
uniform subgroup of CI(X) extends to a uniform subgroup of BiLip(∂G, d0∞ ).
Conversely, there are several ‘hyperbolic cone’ constructions in the literature [6,8,11,
21] that take a metric space Z and produce a hyperbolic metric space Con(Z) such that
a visual metric on ∂ Con(Z) recovers Z with the given metric. We take Conr (Z) be the
‘truncated hyperbolic approximation with parameter r’ of Buyalo and Schroeder [11].
Item 3 implies item 2 by taking X to be the hyperbolic cone Conr (∂G, d0∞ ) for r
sufficiently small.
Let g be an infinite order element of a hyperbolic group G with a fixed word metric.
The isometry Lg defined by left multiplication by g has a well defined translation length,
which is positive. If µ : G → X is a quasi-isometry such that µ∗ (Lg ) is a coarse isometry,
it is not true in general that µ∗ (Lg ) has a well defined translation length. The next
proposition shows that we do still get a positive translation length for µ∗ (Lg ) in the
special case of relative rigidity.
Proposition 5.7. Let G be a hyperbolic group and P a peripheral structure such that
there exists a proper geodesic space X and a quasi-isometry µ : G → X such that
µ∗ (QI((G, P))) is a uniform subgroup of CI((X, µ(P))). For every infinite order element g ∈ G the map i 7→ µ(g i ) is a coarse similitude.
Before proving the proposition we need a few lemmas.
Lemma 5.8. Let X be proper geodesic space quasi-isometric to a non-elementary hyperbolic group. For i ∈ {0, 1}, let φi be a quasi-isometry of X, with [φ0 ] = [φ1 ] ∈ QI(X).
The distance between φ0 and φ1 is bounded in terms of the quasi-isometry constants of
φ0 and φ1 and the constants of X.
Proof. The following argument is standard, see for instance [44]. We briefly outline the
proof.
Since X is quasi-isometric to a visual hyperbolic space, it is a visual hyperbolic space.
Furthermore, every point x ∈ X can be realized as a quasi-center of an ideal geodesic
triangle ∆. Since φ0 and φ1 are coarsely equivalent, ∂φ0 = ∂φ1 , so φ0 (∆) and φ1 (∆) are
ideal quasi-geodesic triangles with the same ideal vertices, and quasi-geodesic constants
depending on those of φ0 and φ1 , respectively. The set of quasi-centers of uniformly
quasi-geodesic triangles with the same vertices is bounded in terms of the quasi-geodesic
constants and the hyperbolicity constant of X, and φ0 (x) and φ1 (x) both lie in this
set.
Corollary 5.9. If φ is an (M, A)–quasi-isometry and ψ is an A0 –coarse isometry with
[φ] = [ψ] ∈ QI(X) then there is an A00 depending only on M , A, A0 and X such that φ
is an A00 –coarse isometry.
35
Lemma 5.10. Let µ : Y → X be a quasi-isometry between visual hyperbolic spaces.
Suppose that φ is a loxodromic isometry of Y with translation length τ . Let y0 be a
point on an axis of φ, and set yi = φi (y0 ) and xi = µ(yi ). Suppose that {µ∗ (φi )}i∈Z are
uniform coarse isometries. Then
d(x0 , xi )
i→∞
i
L := lim
exists, and there exists an A such that i 7→ µ(φi (y0 )) is an (L, 2A)–coarse similitude.
Proof. The fact that {µ∗ (φi )}i∈Z are uniform coarse isometries implies that the difference
|d(xi , xj )−d(x0 , xj−i )| is uniformly bounded. Quasi-geodesic stability further implies that
|d(x0 , xi+j ) − d(x0 , xi ) − d(x0 , xj )| is uniformly bounded. Let A be the greater of these
two bounds.
Suppose L+ := lim sup d(x0 , xi )/i and L− := lim inf d(x0 , xi )/i are different. Note L− >
−
)/3. Choose some i such that
0 since i 7→ xi is a quasi-geodesic. Take := (L+ − Lq
d(x0 , xi )/i
< L− + and such that α :=
d(x0 ,xi )+2A
d(x0 ,xi )
q
+
<
2L+ +L−
.
L+ +2L−
Choose some j such
−
2L +L
, where q is the integer such that
that d(x0 , xi )/i > L+ − and such that q+1
q <
L+ +2L−
qi ≤ j < (q + 1)i.
The previous inequalities, together with the triangle inequality to decompose d(x0 , xj )
along x0 , xi , . . . , xqi , xj , yields:
L+ − <
d(x0 , xj )
d(x0 , xj )
(q + 1)(d(x0 , xi ) + 2A)
≤
≤
j
qi
qi
α(q + 1)d(x0 , xi )
α(q + 1)
=
≤
· (L− + )
qi
q
2L+ + L− L+ + 2L−
·
= L+ −
< +
L + 2L−
3
This is a contradiction, so L+ = L− = L.
Suppose there exists an i such that d(x0 , xi ) < Li − 2A. Then L = limj→∞
d(x0 ,xij )
ij
<
limj→∞ Lij
ij
= L, which is a contradiction.
If there exists an i such that d(x0 , xi ) > Li + 2A, then a similar computation leads
to a contradiction. Therefore, |d(x0 , xi ) − Li| ≤ 2A, which means i 7→ xi = µ(φi (y0 )) is
an (L, 2A)–coarse similitude.
Proof of Proposition 5.7. Let Lg be left multiplication on G by an infinite order element.
For all n ∈ Z the map Lng is an isometry, so µ∗ (Lng ) is a quasi-isometry whose constants
depend only on those of µ. By hypothesis, there exists an A such that for each n there
exists an A–coarse isometry cn ∈ CI(X) with [cn ] = [µ∗ (Lng )]. By Corollary 5.9, there
exists an A0 independent of n such that µ∗ (Lng ) is an A0 –coarse isometry. Now apply
Lemma 5.10.
36
5.4
Stretch factors
Let G be a finitely presented, one-ended group such that T := Cyl(G) has two-ended
edge stabilizers. Let Γ := G\Cyl(G). Let Y be an algebraic tree of spaces for G over T .
Vertices v0 and v1 of T that belong to a common cylinder have stabilizer groups that
intersect in a virtually cyclic subgroup.
ˆ denotes the modulus of Definition 2.5.
Recall that ∆
Definition 5.11. Let v0 and v1 be distinct quasi-isometrically rigid vertices of T contained in a common cylinder c. For i ∈ {0, 1}, choose a rigid model (Xvi , Pvi ) for
(Gvi , PvΓi ). Let hzi be an infinite cyclic subgroup of Gv0 ∩ Gv1 , and define the relative
stretch from v0 to v1 to be:
relStr(v0 , v1 , (Xv0 , Pv0 ), (Xv1 , Pv1 )) :=
`Xv1 (z)
`Xv0 (z)
Clearly, relStr(v0 , v1 , (Xv0 , Pv0 ), (Xv1 , Pv1 )) depends on the choices of (Xvi , Pvi ). Recall, by Lemma 5.2, it does not depend on the choice of quasi-isometries (Gv0 , PvΓ0 ) →
(Xv0 , Pv0 ) and (Gv1 , PvΓ1 ) → (Xv1 , Pv1 ).
Lemma 5.12. relStr(v0 , v1 , (Xv0 , Pv0 ), (Xv1 , Pv1 )) does not depend on the choice of hzi <
Gv0 ∩ Gv1 .
Proof. For i ∈ {0, 1}, let ei be an edge on the geodesic in T between v0 and v1 , with
ι(ei ) = vi . Let hz0 i < Ge0 and hz1 i < Ge1 be infinite cyclic subgroups of minimal index.
Since Ge0 and Ge1 are virtually cyclic, hzi has finite index in each of them.
`Xv1 (z1 ) ·
`Xv1 (z)
=
`Xv0 (z)
`Xv0 (z0 ) ·
[hz1 i:hz1 i∩hzi]
[hzi:hz1 i∩hzi]
[hz0 i:hz0 i∩hzi]
[hzi:hz0 i∩hzi]
=
`Xv1 (z1 ) [hz1 i : hz0 i ∩ hz1 i]
`Xv1 (z1 )
ˆ 0 , e1 )
·
=
· ∆(e
`Xv0 (z0 ) [hz0 i : hz0 i ∩ hz1 i]
`Xv0 (z0 )
The right-hand side is independent of the choice of z0 and z1 , since if, say, hz00 i is another
infinite cyclic subgroup of minimal index in Ge0 then:
`Xv0 (z00 ) = `Xv0 (z0 ) ·
[hz0 i : hz0 i ∩ hz00 i]
= `Xv0 (z0 )
[hz00 i : hz0 i ∩ hz00 i]
Corollary 5.13. If v0 , v1 , and v2 are quasi-isometrically rigid vertices in a common
cylinder then:
relStr(v0 , v1 , (Xv0 , Pv0 ), (Xv1 , Pv1 )) · relStr(v1 , v2 , (Xv1 , Pv1 ), (Xv2 , Pv2 ))
= relStr(v0 , v2 , (Xv0 , Pv0 ), (Xv2 , Pv2 ))
Proposition 5.14. Let φ be a quasi-isometry between finitely presented, one-ended groups
G and G0 whose JSJ trees of cylinders have two-ended edge stabilizers. Let Γ := G\Cyl(G)
and Γ0 := G0 \Cyl(G0 ). Suppose that v0 and v1 are distinct quasi-isometrically rigid vertices of T (Γ) contained in a cylinder c. Choose rigid models (Xv0 , Pv0 ) and (Xv1 , Pv1 )
for (Gv0 , PvΓ0 ) and (Gv1 , PvΓ1 ), respectively. Then:
relStr(v0 , v1 , (Xv0 , Pv0 ), (Xv1 , Pv1 )) = relStr(φ∗ (v0 ), φ∗ (v1 ), (Xv0 , Pv0 ), (Xv1 , Pv1 ))
37
Proof. Let Y be an algebraic tree of spaces for Γ, and let Y 0 be an algebraic tree of
spaces for Γ0 . For v ∈ {v0 , v1 } choose µv ∈ QIsom((Yv , PvΓ ), (Xv , Pv )). Note that
0
µv ◦ φv ∈ QIsom((Yφ0∗ (v) , PφΓ∗ (v) ), (Xv , Pv )). Define:
R := relStr(v0 , v1 , (Xv0 , Pv0 ), (Xv1 , Pv1 ))
R0 := relStr(φ∗ (v0 ), φ∗ (v1 ), (Xv0 , Pv0 ), (Xv1 , Pv1 ))
Choose two points x0 and x1 in µv0 (Ye0 ) such that dXv0 (x0 , x1 ) 0. The idea of the
proof is to approximate R by a quantity Q(x0 , x1 ) depending on x0 and x1 , and similarly
approximate R0 by Q0 (x0 , x1 ), and then show:
R=
lim
d(x0 ,x1 )→∞
Q(x0 , x1 ) =
lim
d(x0 ,x1 )→∞
Q0 (x0 , x1 ) = R0
In the following, quantities are ‘coarsely well defined’ if they are well defined up to
additive error independent of the choice of x0 and x1 .
By construction, Ye0 is a coset of Ge0 , and hze0 i is an infinite cyclic subgroup of
minimal index in Ge0 . Let g0 ∈ G such that Ye0 = g0 Ge0 . Since g0 hze0 i is coarsely dense
in Ye0 , there exist integers ki such that dYv0 (µ−1 (xi ), g0 zek0i ) is small.
Ye0 and Ye1 are coarsely equivalent, so closest point projection π : Ye0 → Ye1 is
coarsely well defined. Moreover, since Ge0 and Ge1 are commensurable, there exist
ˆ 0 ,e1 )+l
j ∆(e
∈ {±1} and l ∈ Z such that π(g0 zej 0 ) is bounded distance from g1 ze1
ˆ
j ∆(e
,e )+l
ze1 0 1
, where
is to be interpreted as ze1 raised to the greatest integer less than or equal to
ˆ
j ∆(e0 , e1 ) + l. Now we have the following string of relations, where ∼ indicates equality
up to additive error in the numerator and denominator, independent of x0 and x1 .
R=
`Xv1 (g1 ze1 g1−1 )
`Xv0 (g0 ze0 g0−1 )
ˆ 0 , e1 )
· ∆(e
ˆ
k ∆(e
,e )+l
ˆ
k ∆(e
,e1 )+l
0 1
0
),µv1 (g1 ze11
dXv (µv1 (g1 ze10
1
ˆ 0 ,e1 )+l)−(k0 ∆(e
ˆ 0 ,e1 )+l)|
|(k1 ∆(e
∼
k
k
dXv (µv0 (g0 ze00 ),µv0 (g0 ze01 ))
0
|k1 −k0 |
ˆ 0 ,e1 )+l
k ∆(e
!
ˆ 0 ,e1 )+l
k ∆(e
=
dXv1 (µv1 (g1 ze10
∼
dXv1 (µv1 (π(g0 zek00 )), µv1 (π(g0 zek01 )))
∼
))
), µv1 (g1 ze11
ˆ 0 , e1 )
· ∆(e
))
dXv0 (µv0 (g0 zek00 ), µv0 (g0 zek01 ))
dXv0 (µv0 (g0 zek00 ), µv0 (g0 zek01 ))
−1
dXv1 (µv1 ◦ π ◦ µ−1
v0 (x0 ), µv1 ◦ π ◦ µv0 (x1 ))
=: Q(x0 , x1 )
dXv0 (x0 , x1 )
We conclude R = limd(x0 ,x1 )→∞ Q(x0 , x1 ).
38
Similarly, if π 0 is closest point projection from φ(Ye0 ) to φ(Ye1 ) define:
Q0 (x0 , x1 ) :=
0
−1
dXv1 (µv1 ◦ φ ◦ π 0 ◦ φ ◦ µ−1
v0 (x0 ), µv1 ◦ φ ◦ π ◦ φ ◦ µv0 (x1 ))
dXv0 (x0 , x1 )
We have R0 = limd(x0 ,x1 )→∞ Q0 (x0 , x1 ). However, since Ye0 and Ye1 are coarsely equivalent, φ ◦ π 0 ◦ φ is coarsely equivalent to π, so Q(x0 , x1 ) ∼ Q0 (x0 , x1 ). We conclude:
R=
lim
d(x0 ,x1 )→∞
5.5
Q(x0 , x1 ) =
lim
d(x0 ,x1 )→∞
Q0 (x0 , x1 ) = R0
Uniformization
The stretch factors defined in the previous section depend on the choice of rigid model
for the vertex groups. We suppress this dependence by choosing models uniformly:
Definition 5.15. Let QItypes := {J(G, P)K} be the set of quasi-isometry classes of
finitely presented groups relative to peripheral structures. For each Q ∈ QItypes choose
a proper, geodesic space ZQ with peripheral structure PQ such that Q = J(ZQ , PQ )K.
Define Model(Q) := (ZQ , PQ ). If (ZQ , PQ ) is quasi-isometrically rigid then we choose
(ZQ , PQ ) to be a rigid model as in Section 5.1. We choose ZJR,RK = R.
Definition 5.16. If v0 and v1 are quasi-isometrically rigid vertices in a cylinder, define:
relStr(v0 , v1 ) = relStr(v0 , v1 , Model(Gv0 , PvΓ0 ), Model(Gv1 , PvΓ1 ))
5.6
Normalization for unimodular graphs of groups
Suppose that T := Cyl(G) has two-ended edge stabilizers and c is a unimodular cylinder
in T . Suppose that c contains some quasi-isometrically rigid vertices. Unimodularity
implies {relStr(v0 , v1 ) | v0 , v1 ∈ c are qi rigid} is bounded. Since stretch factors multiply
by Corollary 5.13, there exists a quasi-isometrically rigid v0 such that for every other
quasi-isometrically rigid vertex v1 in c we have relStr(v0 , v1 ) ≥ 1. Define relStr(c, v1 ) :=
relStr(v0 , v1 ).
Definition 5.17. Suppose that T := Cyl(G) has two-ended edge stabilizers. Let e be an
edge of T connecting a cylindrical vertex c to a quasi-isometrically rigid vertex v. Define
relStr(e) := relStr(c, v).
5.7
An example
Recall Example 1.1: Let Gi := a, b, t | tui t = vi , where ui and vi are words in ha, bi
given below. In each case Gi should be thought of as an HNN extension of ha, bi over Z
with stable letter t. Subdividing the edge gives a unimodular JSJ decomposition Γi of
Gi whose Bass-Serre Ti is equal to its tree of cylinders.
Let u0 := a, v0 := abab2 , u1 := ab, v1 := a2 b2 , u2 := ab2 , and v2 := a2 b.
39
The methods of [14] show that for each i, the pair {ui , vi } is Whitehead minimal,
with Whitehead graph equal to the complete graph on 4 vertices, so ha, bi is quasiisometrically rigid relative to the peripheral structure Pi coming from incident edge
groups, and the rigid model space is just the Cayley graph for ha, bi with respect to
{a, b}, ie, the 4–valent tree. Furthermore, QI((ha, bi , Pi )) is transitive on Pi , and
J(ha, bi , P1 )K = J(ha, bi , P2 )K = J(ha, bi , P3 )K. Therefore, Theorem 4.8 cannot distinguish G1 , G2 , and G3 .
In Ti , VC is a single orbit, and each vertex c ∈ VC has valence two. Fix some c ∈ VC ,
and suppose Rig(c) = {e, e0 }. Let us assume that e is an edge that attaches to Xτ (e)
along the image of a conjugate of hvi i and e0 is an edge that attaches to Xτ (e0 ) along
the image of a conjugate of hui i. The stabilizers of e and e0 are equal to the stabilizer
of c, which is infinite cyclic. This, together with the fact that the rigid model vertex
space is the Cayley graph (tree) for ha, bi with respect to {a, b}, and the fact that each ui
and vi is cyclically reduced with respect to {a, b}, means that Str(e) is the word length
of vi in ha, bi, and Str(e0 ) is the word length of ui in ha, bi. Thus, relStr(e0 ) = 1 and
relStr(e) = |vi |/|ui | = 5, 2, or 1, as i = 0, 1 or 2, respectively. By Proposition 5.14, no
one of these groups is quasi-isometric to the other.
6
Vertex constraints
In this section we assume that G is a one-ended, finitely presented group such that
T := Cyl(G) has two-ended cylinder stabilizers. Let Γ be the quotient graph of cylinders,
which is therefore a canonical JSJ decomposition of G over two-ended subgroups; recall
Section 2.3.4.
We suppose that X is a tree of spaces over T , quasi-isometric to G.
In Section 4 we saw how to decide if two vertices of T are in the same Aut(T, δ)–
orbit. In this section we would like to restrict further to subgroups of Aut(T, δ) induced
by QI(X), or, in the case that G is hyperbolic, by Homeo(∂X). We will actually do
something that is weaker in the quasi-isometry case, but has the advantage that the same
approach works for both quasi-isometries and boundary homeomorphisms. What we do
is restrict to elements of Aut(T, δ) that at each vertex look like they are induced by a
quasi-isometry or boundary homeomorphism of the appropriate vertex space. We also
add a compatibility condition below. First we explain the notation.
For [φ] ∈ QI(X) we can choose a representative φ that induces an automorphism φ∗
of T and splits as a tree of quasi-isometries φv := φ|Xv ∈ QI((Xv , Pv ), (Xφ∗ (v) , Pφ∗ (v) ))
over T .
Similarly, if G is hyperbolic, then X is hyperbolic and φ ∈ Homeo(∂X) induces an
automorphism φ∗ of T and splits as a tree of boundary homeomorphisms φv := φ|∂Xv ∈
Homeo((∂Xv , ∂Pv ), (∂Xφ∗ (v) , ∂Pφ∗ (v) )) over T .
Since cylinders are two-ended, each edge space is a quasi-line L in its respective
vertex space. Recall this means that there is a controlled embedding Ξ of R with image
L. In the hyperbolic case Ξ is actually a quasi-isometric embedding, and L has distinct
endpoints at infinity in the boundary of the vertex space containing it. In this case we
40
define an orientation of L to be a choice of one of these boundary points, and a boundary
homeomorphism of the vertex space that preserves ∂L is said to be orientation preserving
if it fixes ∂L and orientation reversing if it exchanges the two points of ∂L.
In the quasi-isometry case we know that Ξ([0, ∞)) and Ξ([0, −∞)) are not coarsely
equivalent. We define an orientation of L to be a choice of coarse equivalence class of
either Ξ([0, ∞)) or Ξ([0, −∞)). A quasi-isometry that coarsely preserves L is said to
be orientation preserving on L if it fixes the coarse equivalence classes of Ξ([0, ∞)) and
Ξ([0, −∞)), and orientation reversing if it exchanges them.
We seek χ ∈ Aut(T, δ) such that for every v ∈ VT there exists an element φv ∈
Map((Xv , Pv ), (Xχ(v) , Pχ(v) )) such that (φv )∗ = χ|lk(v) , subject to the following compatibility condition. In the quasi-isometry case we require that (αχ(e) ◦ φι(e) ) ◦ (φτ (e) ◦ αe )−1
is orientation preserving on Xχ(e) . In the boundary homeomorphism case we require that
(∂αχ(e) ◦ φι(e) ) ◦ (φτ (e) ◦ ∂αe )−1 is the identity on ∂Xχ(e) for every edge e. For brevity,
we say “(αχ(e) ◦ φι(e) ) ◦ (φτ (e) ◦ αe )−1 is orientation preserving on Xχ(e) ” in both cases.
In the boundary homeomorphism case we conclude, in Theorem 7.1, that such a
collection of φv patch together to give φ ∈ Homeo(∂X) with φ∗ = χ.
The analogous statement is not true for quasi-isometries. To patch together quasiisometries we need αχ(e) ◦φι(e) and φτ (e) ◦αe to be coarsely equivalent as maps, but we have
only assumed that they have coarsely equivalent image sets with the same orientations.
We also need to know that the φv have uniform quasi-isometry constants. These points
will be addressed in subsequent sections.
6.1
Partial Orientations
A partial orientation ζ of X assigns to each cylindrical vertex space and to each peripheral
set in each non-elementary vertex space either an orientation of that space or the value
‘NULL’.
A cylindrical vertex space or peripheral set is said to be ζ–oriented if its ζ value is
not ‘NULL’, and ζ–unoriented otherwise.
A cylindrical vertex is said to be ζ–oriented or ζ–unoriented if its vertex space is.
An edge e ∈ T is said to be ζ–oriented or ζ–unoriented if the corresponding edge
space in its incident non-elementary vertex is.
The sign of a map φ that takes an oriented space A to an oriented space B is 1
if the map is orientation preserving and −1 if it is orientation reversing. For a partial
orientation ζ we define signζ φ as usual when A and B are both ζ–oriented, and we define
signζ φ := 0 if either of them is ζ–unoriented.
One partial orientation, ζ 0 , extends another, ζ, if they agree on all ζ–oriented sets.
6.2
Cylindrical vertices
Definition 6.1. Let ζ be a partial orientation. Let c be a cylindrical vertex. The orientation imbalance at c with respect to a decoration δ : T → O and a partial orientation ζ is
O
the function Ωδ,ζ
c : O → Z /{−1, 1}, with the action by coordinate-wise multiplication,
defined as follows. Choose an orientation of Xc and for each e ∈ lk(c) let sign αe denote
41
the sign of αe with respect to the chosen orientation of Xc and the ζ–orientation of Xē ,
which we take to be 0 if e is ζ–unoriented. Define:
X
Ωδ,ζ
c (o) := [
sign αe ]
e∈lk(c)∩δ −1 (o)
If Ωδ,ζ
c is non-zero we call c an unbalanced cylinder.
Taking an equivalence class of function in the definition eliminates the dependence
on the arbitrary choice of orientation of Xc .
Proposition 6.2. Suppose δ and ζ are Map(X)–invariant. If there exists φ ∈ Map(X)
such that φ∗ fixes a cylindrical vertex c and reverses the orientation of Xc then Ωδ,ζ
c is
identically zero.
Proof. Suppose o ∈ O is such that there exists a ζ–oriented edge e ∈ lk(c) ∩ δ −1 (o). Let
v := τ (e). By Map(X)–invariance, ζ(Xφ∗ (e) ) = φv (ζ(Xē )) = αφ∗ (e) ◦ φc ◦ αe−1 (ζ(Xē )).
Since φc is orientation reversing, αe and αφ∗ (e) have opposite signs. Therefore, (φ)∗ |lk(c)
gives a bijection between edges in lk(c) ∩ δ −1 (o) whose attaching map have positive sign
and edges in lk(c) ∩ δ −1 (o) whose attaching map have negative sign. Since this is true
for every o ∈ O such that lk(c) ∩ δ −1 (o) is non-empty, Ωδ,ζ
c is identically zero.
Corollary 6.3. Suppose δ and ζ are Map(X)–invariant. If Gc contains an infinite
dihedral group then Ωδ,ζ
c is identically zero.
Proposition 6.4. Suppose δ and ζ are Map(X)–invariant. For every φ ∈ Map(X) we
δ,ζ
have Ωδ,ζ
c = Ωφ∗ (c) .
Proof. Choose some orientation on Xc and Xφ∗ (c) .
If no edge in lk(c)∩δ −1 (o) is ζ–oriented then Ωδ,ζ
c (o) = 0, and, by Map(X)–invariance,
the same are true for φ∗ (c).
Now consider o ∈ O such that there exists an edge e ∈ lk(c) ∩ δ −1 (o) such that e
is ζ–oriented. Let v := τ (e). By Map(X)–invariance, φ∗ (e) ∈ lk(φ∗ (c)) ∩ δ −1 (o) with
ζ(Xφ∗ (e) ) = φv (ζ(Xē )) = αφ∗ (e) ◦ φc ◦ αe−1 (ζ(Xē )). If φc is orientation reversing then αe
and αφ∗ (e) have opposite signs, so that:
X
X
sign αφ (e) = −
sign αe
∗
φ∗ (e)∈lk(φ∗ (c))∩δ −1 (o)
e∈lk(c)∩δ −1 (o)
If φc is orientation preserving then αe and αφ∗ (e) have the same signs, so that:
X
X
sign αφ∗ (e) =
sign αe
φ∗ (e)∈lk(φ∗ (c))∩δ −1 (o)
e∈lk(c)∩δ −1 (o)
The previous proposition shows we can use cylinder imbalances to distinguish different
cylinders. The following lemma shows this holds up under refinement of the decoration.
42
Lemma 6.5. Suppose δ 0 : T → O0 is a refinement of δ and ζ 0 is an extension of ζ.
Suppose that the δ–partition of edges of T is finer than the partition into ζ–oriented
δ 0 ,ζ 0
edges and ζ–unoriented edges. Let c be a cylindrical vertex. If Ωδ,ζ
c is non-zero then Ωc
is non-zero.
Let c0 be a cylindrical vertex distinct from c. If for every o ∈ O there exist ζ–oriented
edges in lk(c) ∩ δ −1 (o) if and only if there exist ζ–oriented edges in lk(c0 ) ∩ δ −1 (o) then
0 0
δ,ζ
δ 0 ,ζ 0
Ωδ,ζ
6= Ωδc0 ,ζ .
c 6= Ωc0 implies Ωc
Proof. If c is unbalanced then there exists an o ∈ O such that Ωδ,ζ
c (o) 6= 0, which implies
that there are ζ–oriented edges in lk(c) ∩ δ −1 (o). Since the δ–partition of edges of T
is finer than the partition into ζ–oriented edges and ζ–unoriented edges, all edges in
lk(c) ∩ δ −1 (o) are ζ–oriented. Since ζ 0 extends ζ, all edges in lk(c) ∩ δ −1 (o) are ζ 0 –
oriented, and since δ 0 refines δ, we have, with respect to ζ 0 and some fixed orientation of
Xc , that:
X
X
X
sign αe
sign αe =
o0 ∈δ 0 ◦δ −1 (o) e∈lk(c)∩δ −1 (o0 )
e∈lk(c)∩δ −1 (o)
The left hand side is non-zero, so one of the terms of the outer sum on the right hand
0 0
side must be non-zero. Thus, Ωδc ,ζ is not identically zero.
0 0
0 0
For the second statement, suppose, for contraposition, that Ωδc ,ζ = Ωδc0 ,ζ . Having
chosen orientations on Xc and Xc0 , there is an ∈ ±1 such that for all o0 ∈ O0 :
X
X
sign αe =
sign αe
e∈lk(c)∩(δ 0 )−1 (o0 )
e∈lk(c0 )∩(δ 0 )−1 (o0 )
If o ∈ O is such that there are no ζ–oriented edges in either lk(c) ∩ δ −1 (o) or lk(c0 ) ∩
δ −1 (o) then:
X
X
sign αe = 0 =
sign αe
e∈lk(c)∩δ −1 (o)
e∈lk(c0 )∩δ −1 (o)
Otherwise, by hypothesis, there are ζ–oriented edges in both lk(c)∩δ −1 (o) and lk(c0 )∩
δ,ζ
δ −1 (o). We conclude that Ωδ,ζ
c = Ωc0 from the following computation, in which the first
and third equalities are from the facts that the δ–partition of edges of T is finer than
the partition into ζ–oriented edges and ζ–unoriented edges and that δ 0 refines δ, and the
0 0
0 0
second equality is from the hypothesis that Ωδc ,ζ = Ωδc0 ,ζ .
X
X
X
o0 ∈δ 0 ◦δ −1 (o)
e∈lk(c)∩(δ 0 )−1 (o0 )
sign αe =
e∈lk(c)∩δ −1 (o)
sign αe
=
X
o0 ∈δ 0 ◦δ −1 (o)
X
sign αe =
e∈lk(c0 )∩(δ 0 )−1 (o0 )
X
sign αe
e∈lk(c0 )∩δ −1 (o)
Given δ and ζ that are both Map(X)–invariant we define the process of cylinder
refinement as follows.
43
1. By passing to the coarsest common refinement, we may assume that the δ–partition
of edges of T is finer than the partition into ζ–oriented edges and ζ–unoriented
edges. The refined δ is still Map(X)–invariant.
2. Consider the ζ–unoriented, unbalanced cylinders. It is possible to choose orientaδ,ζ
tions of their vertex spaces P
so that if c and c0 are twoP
such cylinders with Ωδ,ζ
c = Ωc0
then for all o ∈ O we have e∈lk(c)∩δ−1 (o) sign αe = e∈lk(c0 )∩δ−1 (o) sign αe . Extend
ζ to ζ 0 by taking these orientations of the unbalanced cylindrical vertex spaces.
3. If e is a ζ–unoriented edge such that c := ι(e) is cylindrical and ζ 0 –oriented, define
ζ 0 (Xē ) = αe (ζ 0 (Xc )).
4. Define O0 := O × {−1, 0, 1}. Define δ 0 (e) := (δ(e), signζ 0 (e)) for each edge and
δ 0 (v) := (δ(v), 0) for each vertex.
Lemma 6.6. Suppose that δ and ζ are both Map(X)–invariant and that the δ–partition
of edges of T is finer than the partition into ζ–oriented edges and ζ–unoriented edges.
Let δ 0 and ζ 0 be constructed via cylinder refinement, as above. Then δ 0 is a Map(X)–
invariant refinement of δ and ζ 0 is a Map(X)–invariant extension of ζ. Moreover, the
δ 0 –partition on edges is finer than the partition into ζ 0 –oriented and ζ 0 –unoriented edges.
Proof. Suppose e is a ζ–unoriented edge with ι(e) := c cylindrical and unbalanced. Then
e is ζ 0 –oriented and signζ 0 αe = 1, so δ 0 (e) = (δ(e), 1).
By invariance of δ and ζ, Proposition 6.4, and our choice of orientation on Xc and
Xφ∗ (c) , if φ ∈ Map(X) then φ∗ (c) is unbalanced, and φ∗ (e) is ζ–unoriented but ζ 0 oriented
with signζ 0 αφ∗ (e0 ) = 1. Moreover, φc is orientation preserving, so φ(ζ 0 (Xē )) = ζ 0 (Xφ∗ (e) ).
It also means that δ 0 (φ∗ (e)) = (δ(φ∗ (e)), 1) = (δ(e), 1) = δ 0 (e).
Now suppose e is ζ–oriented and ι(e) := c is cylindrical and unbalanced. Invariance
of ζ 0 on e is inherited from invariance of ζ. From the proof of Proposition 6.4, since φc
is orientation preserving, signζ 0 αe = signζ 0 αφ∗ (e) . Along with invariance of δ, this gives
us δ 0 (e) = δ 0 (φ∗ (e)).
For vertices and remaining edges, δ 0 (t) = (δ(t), 0) = (δ(φ∗ (t)), 0) = δ 0 (φ∗ (t)).
For the final claim, suppose e is ζ 0 –oriented and e0 is ζ 0 –unoriented. Since ζ 0 extends
ζ, e0 is also ζ–unoriented. If e is ζ–oriented then δ(e) 6= δ(e0 ) because the δ–partition
of edges of T is finer than the partition into ζ–oriented edges and ζ–unoriented edges.
Thus, δ 0 (e) 6= δ 0 (e0 ), since δ 0 refines δ.
If e is ζ–unoriented then δ 0 (e) = (δ(e), signζ 0 αe ) and δ 0 (e0 ) = (δ(e0 ), 0) differ in the
second coordinate.
Lemma 6.7. Suppose that δ and ζ are Map(X)–invariant and stable under cylinder
refinement. If c is an unbalanced cylindrical vertex and o ∈ O such that δ −1 (o)∩lk(c) 6= ∅
then either every edge in δ −1 (o) ∩ lk(c) has orientation preserving attaching map or every
edge in δ −1 (o) ∩ lk(c) has orientation reversing attaching map.
Proof. Cylinder refinement orients every edge in an unbalanced cylinder and distinguishes
edges with orientation preserving attaching map from those with orientation reversing
attaching map.
44
6.3
Non-elementary vertices
Given a tree of spaces whose underlying tree is decorated δ : T → O, we get a decoration
on the peripheral structure of each vertex space via mapping to the tree and composing
with δ.
Throughout this subsection we assume that δ : T → O is a Map(X)–invariant decoration and ζ is a Map(X)–invariant partial orientation.
Define:
a
PQ
PQ
O0 = O×
Map(ZQ , PQ )\(O × (∪x∈∂PQ x q {‘NULL’}) × (PQ q {‘NULL’}))
Q∈Maptypes
The left action of Map(ZQ , PQ ) is given by φ.(χ, ζ, e) := (χ ◦ φ−1 , ζ ◦ φ−1 , φ∗ (e)).
If e ∈ T is an edge with non-elementary terminus v := τ (e), Q = J(Xv , Pv )K, µv ∈
Map((Xv , Pv ), (ZQ , PQ )), and ζv is a partial orientation on Pv , define:
−1
δ 0 (v) := δ(v), Map(ZQ , PQ ).(δ|Pv ◦ µ−1
v , ζv ◦ µv , ‘NULL’)
−1
δ 0 (e) := δ(e), Map(ZQ , PQ ).(δ|Pv ◦ µ−1
v , ζv ◦ µv , (µv )∗ (e))
Note that the image is independent of the choice of µv ∈ Map((Xv , Pv ), (ZQ , PQ )).
Composition of δ 0 with projection to the first factor of O0 recovers δ, so δ 0 is a
refinement of δ.
Proposition 6.8. The refinement δ 0 of δ defined above is Map(X)–invariant.
Proof. Take φ ∈ Map(X). If e is an edge with v := τ (e) non-elementary, Q := JXv , Pv K,
and χ := µφ∗ (v) ◦ φv ◦ µ−1
v ∈ Map(ZQ , PQ ), then:
−1
δ 0 (φ∗ (e)) = δ(φ∗ (e)), Map(ZQ , PQ ).(δ|Pφ∗ (v) ◦ µ−1
,
ζ
◦
µ
,
(µ
)
(φ
(e)))
∗
∗
φ
(v)
φ
(v)
∗
∗
φ∗ (v)
φ∗ (v)
−1
−1
−1
= δ(e), Map(ZQ , PQ ).(δ|Pv ◦ φv ◦ µφ∗ (v) , ζv ◦ φv ◦ µ−1
,
(µ
)
(φ
(e)))
φ∗ (v) ∗ ∗
φ∗ (v)
−1
−1
−1
= δ(e), Map(ZQ , PQ ).(δ|Pv ◦ µ−1
v ◦ χ , ζv ◦ µv ◦ χ , (χ ◦ µv )∗ (e))
−1
= δ(e), Map(ZQ , PQ ).(δ|Pv ◦ µ−1
v , ζv ◦ µv , (µv )∗ (e))
= δ 0 (e)
Thus, δ 0 is Map(X)–invariant.
Proposition 6.9. For vertices v, w ∈ T , δ 0 (v) = δ 0 (w) if and only if there exists φ ∈
Map((Xv , Pv , δ, ζ), (Xw , Pw , δ, ζ)).
For edges e, f ∈ T with v := τ (e) and w := τ (f ) both non-elementary, δ 0 (e) = δ 0 (f )
if and only if there exists φ ∈ Map((Xv , Pv , δ, ζ), (Xw , Pw , δ, ζ)) with φ∗ (e) = f .
Proof. We give the proof for edges. The proof for vertices is similar.
Let Q = JXv , Pv K. By definition, δ 0 (e) = δ 0 (f ) if and only if δ(e) = δ(f ) and there
exists χ ∈ Map(ZQ , PQ ) such that
45
−1 = δ|
−1
1. δ|Pv ◦ µ−1
Pw ◦ µw
v ◦χ
−1 = ζ ◦ µ−1
2. ζv ◦ µ−1
w
v ◦χ
w
3. (χ ◦ µv )∗ (e) = (µw )∗ (f )
Define φ := µ−1
w ◦ χ ◦ µv ∈ Map((Xv , Pv ), (Xw , Pw )). Item 1 is equivalent to δ|Pv ◦
−1
φ = δ|Pw . Item 2 is equivalent to ζv ◦ φ−1 = ζw . Item 3 is equivalent to φ∗ (e) = f .
Corollary 6.10. There exists a Map(X)–invariant extension ζ 0 of ζ such that for any
edge e with v := τ (e) non-elementary, e is ζ 0 –unoriented if and only if the stabilizer of
Xē in Map(Xv , Pv , δ, ζ) contains an infinite dihedral group.
Proof. If e ζ–unoriented and the stabilizer of Xē in Map(Xv , Pv , δ, ζ) does not contain
an infinite dihedral group then define an extension ζ 0 of ζ on (δ 0 )−1 (δ 0 (e)) as follows.
Choose an orientation of Xē . If f is an edge with δ 0 (f ) = δ 0 (e) then, by Proposition 6.9,
there exists φ ∈ Map((Xv , Pv , δ, ζ), (Xw , Pw , δ, ζ)) with φ∗ (e) = f . This means that f is
ζ–unoriented, so we extend ζ by defining ζ 0 (Xf¯) := φ(ζ 0 (Xē )).
The orientation of Xf¯ is independent of the choice of φ because of the stabilizer
condition on Xē .
Definition 6.11. Given Map(X)–invariant decoration δ and partial orientation ζ, the
process of vertex refinement produces the Map(X)–invariant δ 0 and partial orientation
ζ 0 defined above.
6.4
Combining the local restrictions
In this section we have our main technical tools. Theorem 6.12 identifies Aut(T, δ)
orbits. Theorem 6.13 leverages this information to understand decoration preserving
isomorphisms between two different trees. Theorem 6.13 provides a blueprint for the
main classification theorems in the next two sections.
Theorem 6.12. Suppose δ : T → O is a Map(X)-invariant decoration and ζ is a
Map(X)-invariant partial orientation. Suppose δ and ζ are stable under neighbor, cylinder, and vertex refinement.
For edges e, f ∈ T we have δ(e) = δ(f ) if and only if there exists χ ∈ Aut(T, δ) such
that
• χ(e) = f
• For every u ∈ T there exists φu ∈ Map((Xu , Pu , δ, ζ), (Xχ(u) , Pχ(u) , δ, ζ)), such that
χ|lk(u) = (φu )∗ .
• For every edge e0 the map (αχ(e0 ) ◦ φι(e0 ) ) ◦ (φτ (e0 ) ◦ αe0 )−1 is orientation preserving
on Xχ(e0 ) .
46
Proof. If there exists χ ∈ Aut(T, δ) such that χ(e) = f then δ(e) = δ(f ). Conversely,
supposing δ(e) = δ(f ), we construct χ.
Define χ(e) := f .
By Proposition 6.9, there exists φτ (e) ∈ Map((Xτ (e) , Pτ (e) , δ, ζ), (Xτ (f ) , Pτ (f ) , δ, ζ))
with (φτ (e) )∗ (e) = f . Define χ|lk(τ (e)) := (φτ (e) )∗ .
Now, suppose that we have χ satisfying the desired properties defined on a subtree
0
T of T such that every leaf is non-elementary and T 0 contains every edge incident to
every non-leaf. Given an edge e0 with c := ι(e0 ) ∈
/ T 0 and τ (e0 ) ∈ T 0 , we show how to
extend χ to lk(c0 ), satisfying the desired properties. Then, by induction, we can extend
χ to all of T .
−1
Let χ(e0 ) := (φτ (e0 ) )∗ (e0 ). Define φc := αχ(e
◦ φτ (e0 ) ◦ αe0 so that (αχ(e0 ) ◦ φc ) ◦
0)
−1
(φτ (e0 ) ◦ αe0 ) is orientation preserving on Xχ(e0 ) .
Case 1: c is unbalanced. Extend χ to lk(c) by choosing a bijection between lk(c) \
{e0 }∩δ −1 (o) and lk(χ(c))\{χ(e0 )}∩δ −1 (o) for each o ∈ O. For each o these sets have the
same cardinality by neighbor stability. Since c is unbalanced, cylindrical stability implies
that c and all edges in lk(c) are ζ–oriented, and, for each o ∈ O, all edges in δ −1 (o) have
attaching maps with the same sign; recall Lemma 6.7. By Map(X)–invariance, the same
is true for χ(c), and for each e1 ∈ lk(c) we have signζ αe1 = signζ αχ(e1 ) .
By Proposition 6.9, there exists
φτ (e1 ) ∈ Map((Xτ (e1 ) , Pτ (e1 ) , δ, ζ), (Xτ (χ(e1 )) , Pτ (χ(e1 )) , δ, ζ))
with (φτ (e1 ) )∗ (e1 ) = χ(e1 ). Define φ|Xτ (e1 ) := φτ (e1 ) and χ|lk(τ (e1 )) := (φτ (e1 ) )∗ . By
construction, (αχ(e1 ) ◦ φc ) ◦ (φτ (e1 ) ◦ αe1 )−1 is orientation preserving on Xχ(e1 ) .
In the balanced cases, choose some orientation of Xc and Xχ(c) .
Case 2: c is balanced and o ∈ O is such that δ −1 (o) ∩ lk(c) 6= ∅ consists of ζ–
oriented edges. By neighbor stability, the total number, n, of edges in lk(c) ∩ δ −1 (o)
is equal to the total number of edges in lk(χ(c)) ∩ δ −1 (o). Since c is balanced, the
number of edges in lk(c) ∩ δ −1 (o) with orientation preserving attaching map is equal to
the number of edges in lk(c) ∩ δ −1 (o) with orientation reversing attaching map, so there
are n/2 of each. Cylinder stability implies χ(c) is also balanced, so there are n/2 edges
in lk(χ(c)) ∩ δ −1 (o) with orientation preserving attaching map and n/2 with orientation
reversing attaching map.
If sign αe0 = sign αχ(e0 ) then φc is orientation preserving. Define χ on δ −1 (o) ∩ lk(c) \
{e0 } by choosing any bijection with δ −1 (o) ∩ lk(χ(c)) \ {e0 } that preserves the signs of
the attaching maps.
If sign αe0 6= sign αχ(e0 ) then φc is orientation reversing. Define χ on δ −1 (o) ∩ lk(c) \
{e0 } by choosing any bijection with δ −1 (o) ∩ lk(χ(c)) \ {e0 } that exchanges the signs of
the attaching maps.
Extend φ and χ as in the previous case.
47
Case 3: c is balanced and o ∈ O is such that δ −1 (o) ∩ lk(c) 6= ∅ consists
of ζ–unoriented edges. By neighbor stability, lk(c) \ {e0 } ∩ δ −1 (o) and lk(χ(c)) \
{e0 } ∩ δ −1 (o) have the same cardinality, and we extend χ by an arbitrary bijection
between them. Take e1 ∈ δ −1 (o) ∩ lk(c) \ {e0 }. By Proposition 6.9, there exists
φ0τ (e1 ) ∈ Map((Xτ (e1 ) , Pτ (e1 ) , δ, ζ), (Xτ (χ(e1 )) , Pτ (χ(e1 )) , δ, ζ)) with (φ0τ (e1 ) )∗ (e1 ) = χ(e1 ).
Since e1 is ζ–unoriented and ζ is stable under vertex refinement, by Corollary 6.10 there
exists an element of Map(Xτ (e1 ) , Pτ (e1 ) , δ, ζ) reversing Xē1 . Define φτ (e1 ) := φ0τ (e1 ) if
(αχ(e1 ) ◦ φc ) ◦ (φ0τ (e1 ) ◦ αe1 )−1 is orientation preserving on Xχ(e1 ) , and define φτ (e1 ) to be
φ0τ (e1 ) precomposed with a Xτ (e1 ) –flip otherwise. Extend χ to lk(τ (e1 )) by (φτ (e1 ) )∗ .
Let ζ0 be the trivial partial orientation on X with constant value ‘NULL’. Let
δ0 : T → O0 be any Map(X)–invariant decoration of T . Perform neighbor, cylinder,
and vertex refinement repeatedly until all three stabilize, and let δ : T → O be the
resulting decoration and ζ the resulting partial orientation.
Now suppose X 0 is a tree of spaces over T 0 with finite cylinders and such that every
φ ∈ Map(X 0 ) splits as a tree of maps over T 0 . Let ζ00 be the trivial partial orientation,
and let δ00 : T 0 → O0 be a Map(X 0 )–invariant decoration of T 0 . (Note that δ0 and δ00
map to the same set of ornaments!) Let ζ 0 and δ 0 be the partial orientation extending ζ00
and the decoration refining δ00 that result from performing neighbor, cylinder, and vertex
refinement repeatedly until all three stabilize.
Recall that the process of cylinder refinement involved choosing Map(X)–invariant
orientations. We will need to account for the fact that these choices can be made differently in X and X 0 . Let ξ ∈ {−1, 1}O . Define ξ · ζ to be the partial orientation:
if ζ(Xt ) = ‘NULL’
‘NULL’
ξ · ζ(Xt ) = ζ(Xt )
if ξ(δ(t)) = 1
opposite of ζ(Xt ) if ξ(δ(t)) = −1
Theorem 6.13. With the above notation, the following are equivalent:
1. There exists χ ∈ Isom((T, δ0 ), (T 0 , δ00 )) such that:
0
0
(a) For every vertex v ∈ T there exists φv ∈ Map((Xv , Pv ), (Xχ(v)
, Pχ(v)
)), such
that χ|lk(v) = (φv )∗ .
(b) For every edge e ∈ T we have (αχ(e) ◦ φι(e) ) ◦ (φτ (e) ◦ αe )−1 is orientation
preserving on X 0 .
χ(e)
2. There exists a bijection β : Im δ → Im δ 0 and ξ ∈ {−1, 1}O such that:
(a) δ0 ◦ δ −1 = δ00 ◦ (δ 0 )−1 ◦ β
(b) When the rows and columns of S(T 0 , δ 0 , O0 ) are given the β–induced ordering
from S(T, δ, O), we have S(T, δ, O) = S(T 0 , δ 0 , O0 ).
(c) For every o ∈ Im δ such that δ −1 (o) consists of non-elementary vertices there
exists (equivalently, for every) v ∈ δ −1 (o) and v 0 ∈ (δ 0 )−1 (β(o)) so that
Map((Xv , Pv , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 )) is nonempty.
48
(d) For every o ∈ Im δ such that δ −1 (o) consists of cylindrical vertices there exists
(equivalently, for every) c ∈ δ −1 (o) and χ(c) ∈ (δ 0 )−1 (β(o)) so that Ωδ,ξ·ζ
=
c
δ 0 ,ζ 0
Ωχ(c) ◦ β.
Proof. If item 1 is true then δ0 = δ00 ◦ χ and ζ0 = ζ00 ◦ χ. Perform the same sequence of
refinements on δ0 and δ00 . Each time the partial orientation on X is extended by choosing
some orientation, push that choice forward to X 0 using the appropriate φc or φv . We get
the claims of item 2 with β the identity and ξ the constant map sending O to 1.
We complete the proof by showing that the hypotheses of item 2 allow us to build
a isomorphism χ ∈ Isom((T, β ◦ δ), (T 0 , δ 0 )) and a collection of maps φv satisfying the
conditions of item 1. Condition 2a implies χ ∈ Isom(T, δ0 ), (T 0 , δ00 )). The construction
is along the lines of that in the proof of Theorem 6.12: we inductively construct χ and
0
0
maps φv ∈ Map((Xv , Pv , β ◦ δ, ξ · ζ), (Xχ(v)
, Pχ(v)
, δ 0 , ζ)) with χ|lk(v) = (φv )∗ .
To begin, pick a non-elementary vertex v0 ∈ T and a vertex v00 ∈ (δ 0 )−1 (β ◦ δ(v0 )).
Define χ(v0 ) := v00 . By 2c and Theorem 6.12, there exists:
φv0 ∈ Map((Xv0 , Pv0 , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 ))
0
0
Define φ|Xv0 := φv0 and χ|lk(v0 ) := (φv0 )∗ .
Let e0 be an edge in lk(v0 ) with cylindrical initial vertex c := ι(e0 ). Define φc :=
(αe00 )−1 ◦ φv0 ◦ αe0 .
For the induction step we extend χ to lk(c). When c is balanced the construction is
virtually the same as that of Theorem 6.12, so we omit those cases. The remaining case
is that c is unbalanced.
Extend χ to lk(c) by choosing a bijection between lk(c) \ {e0 } ∩ δ −1 (o) and lk(χ(c)) \
{χ(e0 )} ∩ δ −1 (β(o)) for each o ∈ O. These sets have the same cardinality by condition
2b. Since c is unbalanced, cylindrical stability implies that c and all edges in lk(c) are
ζ–oriented, and, for each o ∈ O, all edges in δ −1 (o) have attaching maps with the same
sign; recall Lemma 6.7. This implies that for each o ∈ O, Ωδ,ξ·ζ
(o) = ±Ωδ,ζ
c
c (o), so, in
is
not
identically
zero.
Condition
2d
then
implies
χ(c)
is
unbalanced,
so
particular Ωδ,ξ·ζ
c
0
χ(c) and all of the edges in lk(χ(c)) are ζ –oriented, and edges with the same ornament
have attaching maps with the same sign.
0
We may choose the orientations on Xc and Xχ(c)
to be those given by ξ · ζ and ζ 0 ,
respectively. Together with condition 2d, this implies there exists ∈ ±1 such that for
all o ∈ O:
X
X
signξ·ζ αe =
signζ 0 αe0 0
e∈lk(c)∩δ −1 (o)
e0 ∈lk(χ(c))∩(δ 0 )−1 (β(o))
Since all the edges with a particular ornament have attaching maps of the same sign, this
0
means that for all e1 ∈ lk(c) we have signξ·ζ αe1 = signζ 0 αχ(e
. Therefore, the sign of
1)
0
0
0
αχ(e
◦ φc ◦ αe−1
= αχ(e
◦ (αχ(e
)−1 ◦ φv0 ◦ αe0 ◦ αe−1
1
1
1)
1)
0)
on Xē1 with respect to ξ · ζ and ζ 0 is ( · signξ·ζ αe0 · signξ·ζ αe1 )2 = +1.
49
By Proposition 6.9 and condition 2c, there exists
φτ (e1 ) ∈ Map((Xτ (e1 ) , Pτ (e1 ) , β ◦ δ, ξ · ζ), (Xτ0 (χ(e1 )) , Pτ0 (χ(e1 )) , δ 0 , ζ 0 ))
with (φτ (e1 ) )∗ (e1 ) = χ(e1 ). Define φ|Xτ (e1 ) := φτ (e1 ) and χ|lk(τ (e1 )) := (φτ (e1 ) )∗ . We know
(αχ(e1 ) ◦ φc ) ◦ (φτ (e1 ) ◦ αe1 )−1 is orientation preserving on X 0
because:
χ(e1 )
0
0
◦ φc ◦ αe−1
(ξ · ζ(Xē1 )) = ζ 0 (Xχ(e
αχ(e
) = φτ (e1 ) (ξ · ζ(Xē1 ))
1
1)
)
1
0
We remark that it is not required that φc (ξ · ζ(Xc )) = ζ 0 (Xχ(c)
), but this can easily
be arranged by redefining ξ(δ(c)) to be · ξ(δ(c)).
7
Classification of hyperbolic groups up to boundary homeomorphism from their two-ended JSJ splittings
We are now ready to prove our first classification theorem, characterizing the homeomorphism type of the Gromov boundary of a one-ended hyperbolic group from its JSJ tree
of cylinders.
Theorem 7.1. Let G be a one-ended hyperbolic group with non-trivial JSJ decomposition
over two-ended subgroups, with T := Cyl(G). Let X be an algebraic tree of spaces for
G over T . Let ζ0 be the trivial partial orientation on X. Take the initial decoration
δ0 on T to be by vertex type (‘cylindrical’, ‘rigid’, or ‘hanging’) and relative boundary
homeomorphism type. Perform neighbor, cylinder, and vertex refinement until all three
stabilize to give a decoration δ : T → O and a partial orientation ζ of X.
Let G0 be another one-ended hyperbolic group with non-trivial JSJ decomposition over
two-ended subgroups. Define T 0 , X 0 , δ00 , ζ00 , δ 0 : T 0 → O0 , and ζ 0 as we did for G. Then
∂G is homeomorphic to ∂G0 if and only if there exists a bijection β : Im δ → Im δ 0 and a
ξ ∈ {−1, 1}O such that:
1. δ0 ◦ δ −1 = δ00 ◦ (δ 0 )−1 ◦ β
2. When the rows and columns of S(T 0 , δ 0 , O0 ) are given the β–induced ordering from
S(T, δ, O), we have S(T, δ, O) = S(T 0 , δ 0 , O0 ).
3. For every o ∈ Im δ such that δ −1 (o) consists of non-elementary vertices there exists
(equivalently, for every) v ∈ δ −1 (o) and v 0 ∈ (δ 0 )−1 (β(o)) so that
Homeo((∂Xv , ∂Pv , β ◦ δ, ξ · ζ), (∂Xv0 0 , ∂Pv0 0 , δ 0 , ζ 0 ))
is nonempty.
4. For every o ∈ Im δ such that δ −1 (o) consists of cylindrical vertices there exists
0 0
(equivalently, for every) c ∈ δ −1 (o) and c0 ∈ (δ 0 )−1 (β(o)) such that Ωδ,ξ·ζ
= Ωδc0 ,ζ ◦β.
c
50
Proof. Since the initial decorations δ0 and δ00 are trivial, the given conditions are equivalent, by Theorem 6.13 for boundary homeomorphism, to the existence of χ ∈ Isom(T, T 0 )
such that:
0
0
1. For every vertex v ∈ T there exists φv ∈ Homeo((∂Xv , ∂Pv ), (∂Xχ(v)
, ∂Pχ(v)
)),
such that χ|lk(v) = (φv )∗ .
2. For every edge e ∈ T we have ∂αχ(e) ◦ φι(e) = φτ (e) ◦ ∂αe .
These conditions say there exists an isomorphism χ : T → T 0 and a tree of boundary
homeomorphisms over χ compatible with X and X 0 . By Theorem 2.28, this is equivalent
to the existence of a boundary homeomorphism between ∂X and ∂X 0 , hence between
∂G and ∂G0 .
Example 7.2. Recall the example of Section 5.7. The trees of cylinders of these groups
are isomorphic: they are bipartite with one orbit of valence 2 cylindrical vertex and one
orbit of infinite valence relatively rigid vertex. Recall also that the rigid vertices were discussed in Section 3, and the relative quasi-isometry (hence, boundary homeomorphism)
types of these three examples are the same.
There are two Gi –orbits of edge in Cyl(Gi ), so the only remaining question about
Homeo(∂Gi )–orbits is whether there are one or two orbits of edges. For boundary homeomorphism we do not care about stretch factors on the edges, so the edges have trivial
initial decoration.
The quasi-isometry group of the vertex stabilizer preserving the peripheral structure
coming from incident edge groups is transitive on peripheral sets, so the vertex refinement
is a trivial refinement of the initial decoration.
Moreover, the quasi-isometry group of the vertex stabilizer preserving the peripheral structure coming from incident edge groups contains and infinite dihedral group in
the stabilizer of each peripheral subset, so all cylinders are balanced, and the cylinder
refinement is a trivial refinement of the initial decoration.
We conclude there is only one Homeo(∂Gi ) orbit of edges in Cyl(Gi ), so the initial decoration and initial (trivial) partial orientation are stable. Since the structure
invariants are the same for the three Gi , they have homeomorphic boundaries.
8
Quasi-isometry classification of groups from their twoended JSJ splittings
We are now almost ready to prove our second main theorem, characterizing the quasiisometry type of a finitely presented one-ended group from its JSJ tree of cylinders.
Before doing so, we explain the extreme flexibity provided by the hanging vertices of the
tree.
51
8.1
Quasi-isometric flexibility of hanging spaces
Recall that the fixed model space for hanging vertices is the universal cover of a fixed
hyperbolic pair of pants Σ, with peripheral structure consisting of the coarse equivalence
classes of the boundary components of Σ̃.
Proposition 8.1 (cf [2, Theorem 1.2]). Let G be a finitely presented, one-ended group
admitting a JSJ decomposition over two-ended subgroups with two-ended cylinder stabilizers. Let Γ := G\Cyl(G) and T := Cyl(G) = T (Γ). Let X be an algebraic tree of spaces
for G over T . Let v be a hanging vertex in Γ. Let δ : T → O be a QI(X)–invariant decoration and let ζ be a QI(X)–invariant partial orientation. For each edge e ∈ Γ incident
to v, choose a positive real parameter σe . Let δ 0 : ∂ Σ̃ → O be a decoration of the peripheral structure of Σ̃ and let ζ 0 be a partial orientation of ∂Σ such that QI(Σ̃, ∂ Σ̃, δ 0 , ζ 0 )
acts coboundedly on Σ̃.
Suppose that for some e0 ∈ lk(e
v ) we are given a coarse similitude φ|Xe0 from Xe0 to a
component B0 of ∂ Σ̃ that respects the decoration and partial orientations. Suppose further
that there exists a φ0 ∈ QIsom((Xv , Pv , δ, ζ), (Σ̃, ∂ Σ̃, δ 0 , ζ 0 )) such that φ0 (Xe ) = B0 . Then
there exists φ ∈ QIsom((Xv , Pv , δ, ζ), (Σ̃, ∂ Σ̃, δ 0 , ζ)) extending φ|Xe0 that, for each edge
e ∈ lk(v) \ {e0 }, restricts to be a coarse similitude with multiplicative constant σe on Xe .
The quasi-isometry constants of φ can be bounded in terms of Gv , Pv , the coboundedness constant for QI(Σ̃, ∂ Σ̃, δ 0 , ζ 0 ) y Σ̃, the constants of φ|Xe0 , and the σv .
Sketch. The proof follows the same argument as [2, Theorem 1.2]. The idea is to build
φ inductively, peripheral set by peripheral set. We start with φXe0 . Let σ0 be the
multiplicative constant of φXe0 . Then we want to extend φ to peripheral sets that come
close to Xe0 in Xv , sending these to components of ∂ Σ̃ that come close to φ(Xe0 ).
For another peripheral set Xe , the number of peripheral sets in the QI(Xv , Pv , δ, ζ)–
orbit of Xe that come within some fixed distance K of a subsegment of Xe0 of length l
is coarsely dl for some d > 0.
Let d0r be such that there are coarsely d0r l peripheral sets in the QI(Σ̃, ∂ Σ̃, δ 0 , ζ 0 )–
orbit of φ0 (Xe ) that come within r of a subsegment of B0 of length l. The fact that
QI(Σ̃, ∂ Σ̃, δ 0 , ζ 0 ) y Σ̃ is C–cobounded for some C says that d0C > 0, and, in fact, d0r
grows exponentially in r. This means that there is a logarithmically growing function
whose value R at σd0 is such that d0R ≥ σd0 . Thus, for any l there is a way to send the
dl elements in the QI(Xv , Pv , δ, ζ)–orbit of Xe that come within distance K of a length
l subsegment S of Xe0 injectively to the peripheral sets in the QI(Σ̃, ∂ Σ̃, δ 0 , ζ 0 )–orbit of
φ0 (Xe ) that come within R of the length approximately σ0 l subsegment φ|Xe (S) of B0 .
In this way one builds a matching between the peripheral sets that come close to
Xe0 and the peripheral sets that come close to B0 , respecting decorations and partial
orientations. Then φ is defined along such a matched pair to be a coarse similarity with
the appropriate σe as multiplicative constant, and the neighbor-matching is repeated for
each such pair.
52
8.2
Geometric trees of spaces for groups with two-ended cylinders stabilizers
Let G be a finitely presented, one-ended group admitting a JSJ decomposition over twoended subgroups with two-ended cylinder stabilizers. Let Γ := G\Cyl(G) and T :=
Cyl(G) = T (Γ).
Recall that in Section 2.5 we built an algebraic tree of spaces Y quasi-isometric to
G, and gave conditions for a collection of quasi-isometries of the vertex spaces to patch
together to give a quasi-isometry of Y . Now we will construct a geometric tree of spaces
X by uniformizing the vertex spaces, that is, replacing each vertex space by its uniform
model from Section 5.5. The quasi-isometries between vertex spaces and their uniform
models will patch together to give a quasi-isometry from Y to X. Therefore, X will
be quasi-isometric to G. The price to pay for uniformizing the vertex spaces is that in
general G only admits a cobounded quasi-action on X, not a cocompact action, but this
will not affect us.
We use the same notation as in Section 2.5. Let Y be the algebraic tree of spaces
constructed there. For a relatively rigid vertex v ∈ Γ, fix a quasi-isometry νv : (Gv , Pv ) →
(ZJ(Gv ,Pv )K , PJ(Gv ,Pv )K ) from Gv to the chosen model space for the relative quasi-isometry
type of (Gv , Pv ).
If Gv is virtually cyclic choose a cyclic subgroup hzv i < Gv of minimal index. Define
νv by sending Gv onto hzv i by closest point projection and sending zvk to kσv ∈ R, where
σv is a positive real parameter chosen as follows. If v is not adjacent to any quasiisometrically rigid vertices then choose σv := 1. Otherwise, choose σv := min `Xw (zv ),
where the minimum is taken over quasi-isometrically rigid vertices w adjacent to v.
Remark 8.2. This choice of σv ’s is convenient because it will imply, for an edge e with
ι(e) cylindrical and τ (e) rigid, that the attaching map αeX constructed below is a coarse
similitude whose multiplicative constant is equal to the stretch factor relStr(e) defined
in Section 5.6.
For a hanging vertex v ∈ Γ we define νv as follows. For each edge e ∈ lk(v) choose a
cyclic subgroup hze i < Ge of minimal index. Define
νv : (Gv , Pv ) → (ZJ(Gv ,Pv )K , PJ(Gv ,Pv )K )
to be a quasi-isometry such that for each e ∈ lk(v), each coset of Ge , which is a peripheral
set in Pv , is sent to a peripheral set in PJ(Gv ,Pv )K by a coarse similitude with multiplicative
constant:
[hzτ (e) i:hzτ (e) i∩hze i]
[hze i:hz
i∩hze i] · στ (e)
τ (e)
`Gv (ze )
Here, `Gv (ze ) is the translation length of ze in the Cayley graph of Gv , which is non-zero
since Gv is hyperbolic, and στ (e) is the parameter for Gτ (e) chosen above. Such a quasiisometry can be constructed using Proposition 8.1. These particular values are chosen
to make Lemma 8.3, below, true.
53
Now, for each vertex v ∈ VT define Xv to be a copy of ZJ(Gv ,Pv )K with isometry
−1
µv : Xv → ZJ(Gv ,Pv )K . Define φv : Yv → Xv by x 7→ µ−1
v ◦ νv (h(v,i) x).
We define edge spaces and attaching maps in X to be compatible with those of Y ,
as follows. Consider an edge e with v := ι(e) and w := τ (e). There are h(v,i) and g(e,j)
e and e = h(v,i) g(e,j) ee. Define αeX := φw ◦ αeY ◦ πYe ◦ φ−1
such that v = h(v,i) v
v , where πYe
denotes closest point projection to Ye . The map αeX is coarsely well defined, since πYe
X
Y
−1
moves points of φ−1
v (Xe ) bounded distance. Define αē := φv ◦ αē ◦ πYē ◦ φw , where
πYē is closest point projection from Yw to the coarsely dense subset Yē . This map is well
defined, and is a quasi-isometry inverse to αeX , since πYē moves points bounded distance.
Chasing through these definitions on easily demonstrates:
Lemma 8.3. If c = ι(e) is cylindrical and v = τ (e) is hanging then αeX : Xc → Xv is a
coarse isometric embedding.
Proposition 8.4. With notation as above, G, X, and Y are quasi-isometric to one
another.
Proof. G is quasi-isometric to Y by Lemma 2.15. Proposition 2.16 implies X and Y are
quasi-isometric, since (φv ) is a tree of quasi-isometries over the identity on T compatible
with X and Y .
8.3
Quasi-isometries
Let G be a finitely presented, one-ended group with non-trivial JSJ decomposition over
two-ended subgroups such that:
• Every non-elementary vertex is either hanging or quasi-isometrically rigid relative
to the peripheral structure coming from incident edge groups.
• Cylinder stabilizers are two-ended.
If Question 1 has positive answer then every one-ended hyperbolic group with a nontrivial JSJ decomposition is of this form.
Theorem 8.5. Let G and G0 be finitely presented, one-ended groups with non-trivial JSJ
decompositions over two-ended subgroups with two-ended cylinder stabilizers.
Let T := Cyl(G). Let X be a geometric tree of spaces for G over T , as in Section 8.2.
Let ζ0 be the trivial partial orientation on X. Let δ0 the decoration on T that sends an
edge e incident to a rigid vertex to its relative stretch factor relStr(e) as in Definition 5.17,
sends other edges to ‘NULL’, and sends vertices to their vertex type (‘cylindrical’, ‘rigid’,
or ‘hanging’) and relative quasi-isometry type. Let ζ0 be the trivial partial orientation on
X. Perform neighbor, cylinder, and vertex refinement until all three stabilize to give a
decoration δ : T → O and a partial orientation ζ of X.
Define T 0 , X 0 , δ00 , ζ00 , δ 0 : T 0 → O0 , and ζ 0 for G0 as we did for G. In particular, X 0
is uniformized with respect to the same choice of model spaces from Definition 5.15.
Then G and G0 are quasi-isometric if and only if there exists a bijection β : Im δ →
Im δ 0 and ξ ∈ {−1, 1}O such that:
54
1. δ0 ◦ δ −1 = δ00 ◦ (δ 0 )−1 ◦ β
2. When the rows and columns of S(T 0 , δ 0 , O0 ) are given the β–induced ordering from
S(T, δ, O), we have S(T, δ, O) = S(T 0 , δ 0 , O0 ).
3. For every o ∈ Im δ such that δ −1 (o) consists of non-elementary vertices there exists
(equivalently, for every) v ∈ δ −1 (o) and v 0 ∈ (δ 0 )−1 (β(o)) so that
QIsom((Xv , Pv , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 ))
is nonempty.
4. For every o ∈ Im δ such that δ −1 (o) consists of cylindrical vertices, there exists
0 0
(equivalently, for every) c ∈ δ −1 (o) and c0 ∈ (δ 0 )−1 (β(o)) such that Ωδ,ξ·ζ
= Ωδc0 ,ζ ◦β.
c
The construction is a modification of the proof of Theorem 6.13. Recall in that case
we inductively built χ ∈ Isom((T, δ), (T, δ 0 )) and quasi-isometries
0
0
, δ 0 , ζ 0 ))
, Pχ(v)
φv ∈ QIsom((Xv , Pv , β ◦ δ, ξ · ζ), (Xχ(v)
such that (φv )∗ = χ|lk(v) . The proof of Theorem 6.13 mainly focuses on the inductive
step in the link of a cylindrical vertex, and chooses any φv as above such that (φv )∗
agrees with χ on the incoming edge to v.
In the present context we must be more careful about the choices of the φv . The proof
in Theorem 6.13 gives us a collection of quasi-isometries (φv ) such that for every edge e ∈
T with ι(e) cylindrical we have that (αχ(e) ◦ φι(e) ) ◦ (φτ (e) ◦ αe )−1 is orientation preserving
on X 0 , but now we require it to be coarsely the identity on X 0 . Furthermore, we
χ(e)
χ(e)
need the quasi-isometry constants of the φv to be uniformly bounded.
Here is how we achieve these requirements. Hanging vertices present no obstacles,
since by Proposition 8.1 they are so flexible. The real work is in dealing with the rigid
vertices. For these we choose in advance a finite number of quasi-isometries to use as
building blocks. Since the collection is finite, the constants are uniformly bounded. We
will choose the maps on cylinder spaces to be coarse isometries. It then remains to see that
if e ∈ ET is an edge with c := ι(e) cylindrical and v := τ (e) relatively rigid, that we can
make φc agree with a map φv constructed from the pre-chosen building blocks. We assume
we have chosen enough building blocks so that we can make (φv )∗ (e) = (φc )∗ (e), with the
correct orientation on Xē . This is handled by the same considerations as Theorem 6.13.
Additionally, we have set up the geometric tree of spaces so that the edge inclusion into
a rigid vertex is a coarse similitude whose multiplicative constant is the stretch factor
of the edge. Since we have incorporated the stretch factors into the decorations, we are
guaranteed that the stretch factor on e matches the stretch factor on χ(e). It follows
that (αχ(e) ◦ φc ) ◦ (φv ◦ αe )−1 is a coarse isometry that is orientation preserving. Finally,
we make it coarsely the identity by adjusting φv using the group action.
Proof of Theorem 8.5. By Theorem 2.9, Corollary 2.10, Proposition 2.16, and Proposition 5.14, X and X 0 are quasi-isometric if and only if there exists a tree of quasi-isometries
over an element of Isom((T, δ0 ), (T 0 , δ00 )) compatible with X and X 0 .
55
The existence of a tree of quasi-isometries over an element of Isom((T, δ0 ), (T 0 , δ00 ))
compatible with X and X 0 implies the above conditions. Our goal is to show the converse.
Suppose o ∈ O is an ornament such that δ −1 (o) consists of vertices that are relatively
quasi-isometrically rigid. Choose representatives vo,1 , . . . , vo,io of the G–orbits contained
in δ −1 (o). Suppose o0 ∈ O is an ornament such that δ −1 (o0 ) consists of edges incident
to v ∈ δ −1 (o). For each 1 ≤ i ≤ io choose representatives eo,i,o0 ,1 , . . . , eo,i,o0 ,jo0 of the
Gvo,i –orbits in δ −1 (o0 ) ∩ lk(vo,i ). For each i and j choose
Φo,i,o0 ,j ∈ QIsom((Xvo,i , Pvo,i , δ, ξ · ζ), (Xvo,1 , Pvo,1 , δ, ξ · ζ))
such that (Φo,i,o0 ,j )∗ (eo,i,o0 ,j ) = eo,1,o0 ,1 . Such quasi-isometries exist by Proposition 6.9.
0
0
Similarly choose representatives vβ(o),1
, . . . , vβ(o),i
of the G0 –orbits contained in
0
β(o)
(δ 0 )−1 (β(o)) and representatives e0β(o),i,β(o0 ),1 , . . . , e0β(o),i,β(o0 ),j 0
β(o0 )
of the G0vβ(o),i –orbits in
0
(δ 0 )−1 (β(o0 )) ∩ lk(vβ(o),i
) and quasi-isometries Φ0β(o),i,β(o0 ),j .
Choose a quasi-isometry Φo,o0 ∈ QIsom((Xvo,1 , Pvo,1 , δ, ξ · ζ), (Xv0 0
β(o),1
, Pv0 0
, δ 0 , ζ 0 ))
β(o),1
that takes eo,1,o0 ,1 to e0β(o),1,β(o0 ),1 . Such a quasi-isometry exists by condition 3 and Proposition 6.9. If eo,1,o0 ,1 is ζ–unoriented then we also choose
0
Φ−
o,o0 ∈ QIsom((Xvo,1 , Pvo,1 , δ, ξ · ζ), (Xv 0
β(o),1
, Pv0 0
, δ 0 , ζ 0 ))
β(o),1
−1 orientation reversing on
0
that takes eo,1,o0 ,1 to e0β(o),1,β(o0 ),1 such that Φ−
o,o0 ◦ (Φo,o )
Xē0 0
. Such a quasi-isometry exists by Corollary 6.10.
β(o),1,β(o0 ),1
We have chosen finitely many quasi-isometries Φ, so they have uniformly bounded
quasi-isometry constants.
Induction base case Begin the induction by choosing a cylindrical vertex c ∈ T and
a cylindrical vertex c0 ∈ (δ 0 )−1 (β(δ(c))). Define χ(c) := c0 . By construction Xc and Xc0 0
are copies of R. Define φc : Xc → Xc0 0 to be an isometry. If c is ζ–oriented we choose φc
so that φc (ξ · ζ(Xc )) = ζ 0 (Xc0 0 ). Extend χ to lk(c) as in Theorem 6.13.
Inductive steps for non-elementary vertices Suppose v = τ (e) is a non-elementary
0
vertex such that for c = ι(e) we have already defined a coarse isometry φc : Xc → Xχ(c)
0
0
and χ|lk(c) . Suppose further that if c is ζ–oriented then φc (ξ · ζ(Xc )) = ζ (Xχ(c) ). Let
c0 := χ(c) and e0 := χ(e).
Suppose v is rigid. Now gv = vδ(v0 ),i for some g ∈ G and some i, and hge =
eδ(v),i,δ(e),j for some j and some h ∈ Gvδ(v),i . Similarly, there are g 0 ∈ G0 and i0 such that
0
0
0
0
g 0 v 0 = vβ(δ(v)),i
such that h0 g 0 e0 = e0β(δ(v)),i0 ,β(δ(e)),j 0 .
0 , and j and h ∈ Gv 0
The map
β(δ(v)),i0
(h0 g 0 )−1 ◦ (Φ0β(δ(v0 )),i0 ,β(δ(e)),j 0 )−1 ◦ Φδ(v0 ),δ(e) ◦ Φδ(v0 ),i,δ(e),j ◦ hg
56
is an element of QIsom(Xv0 , Pv0 , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 )) taking e to e0 .
0
0
If e is ζ–unoriented then we also have that the map
(h0 g 0 )−1 ◦ (Φ0β(δ(v0 )),i0 ,β(δ(e)),j 0 )−1 ◦ Φ−
δ(v0 ),δ(e) ◦ Φδ(v0 ),i,δ(e),j ◦ hg
is an element of QIsom(Xv0 , Pv0 , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 )) taking e to e0 . For one of
0
0
these two, the composition with αe ◦ (φc )−1 ◦ (αe0 0 )−1 is orientation preserving on Xē .
Choose this one as φ0v .
Finally, choose a point x ∈ Xē . Since edge stabilizers act uniformly coboundedly on
their corresponding peripheral sets, we can choose an element k ∈ G0ē0 that is orientation
preserving on Xē0 0 and such that kφ0v (x) is boundedly close to αe0 0 ◦ φc ◦ (αe )−1 (x). Define
φv := kφ0v .
We have:
• φv ∈ QIsom((Xv0 , Pv0 , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 ))
0
0
• (φv )∗ (e) = e0
• φv (x) is boundedly close to αe0 0 ◦ φc ◦ (αe )−1 (x).
• φv ◦ (αe0 0 ◦ φc ◦ (αe )−1 )−1 is orientation preserving on Xē0 0 .
• φv is a composition of three of the pre-chosen Φ with multiplication by five group
elements, so the quasi-isometry constants of φv are bounded in terms of those of
the Φ and the constants for the group action.
By relative quasi-isometric rigidity, φv is a coarse isometry.
We also claim that αe0 0 ◦ φc ◦ (αe )−1 is a coarse isometry. This is because φc is a
coarse isometry, by the induction hypothesis, and αe and αe0 0 are, by construction (recall
Remark 8.2), coarse similitudes with multiplicative constants relStr(e) and relStr(e0 ),
which are equal, since:
relStr(e) = δ0 ◦ δ −1 (δ(e)) = δ00 ◦ (δ 0 )−1 ◦ β(δ(e)) = δ00 ◦ (δ 0 )−1 (δ 0 (e0 )) = δ00 (e0 ) = relStr(e0 )
Thus, φv ◦ (αe0 0 ◦ φc ◦ (αe )−1 )−1 is orientation preserving coarse isometry on Xē0 0 that
coarsely fixes a point. It follows that φv |X ē and αe0 0 ◦ φc ◦ (αe )−1 are coarsely equivalent.
Define χ|lk(v) := (φv )∗ .
0
−1 . Since φ is a
For each edge e00 ∈ lk(v) \ {ē} define φτ (e00 ) := αχ(e
00 ) ◦ φv ◦ (αe00 )
v
0
coarse isometry and αe00 and αχ(e00 ) are coarse similitudes with the same multiplicative
constant, as above, we have that φτ (e00 ) is a coarse isometry.
Suppose v is hanging. The map αe0 0 ◦ φc ◦ (αe )−1 : Xē → Xē0 0 is a coarse isometry,
since attaching maps to hanging vertex spaces are coarse isometries by Lemma 8.3 and
φc is a coarse isometry by the induction hypothesis.
Use condition 3 and Proposition 8.1 to produce a quasi-isometry
φv ∈ QIsom((Xv , Pv , β ◦ δ, ξ · ζ), (Xv0 0 , Pv0 0 , δ 0 , ζ 0 ))
57
that is a coarse isometry along each peripheral subset and that coarsely agrees with
αe0 0 ◦ φc ◦ (αe )−1 on Xē .
Define χ|lk(v) := (φv )∗ .
0
−1 is a coarse isometry,
For each e00 ∈ lk(v) \ {ē} the map φτ (e00 ) := αχ(e
00 ) ◦ φv ◦ (αe00 )
since attaching maps to hanging vertex spaces are coarse isometries by Lemma 8.3, and
φv is a coarse isometry along peripheral sets by construction.
Inductive step for cylindrical vertices Suppose c = ι(e) is cylindrical, χ is defined
on e, φτ (e) is defined, and φc is a coarse isometry such that φc is coarsely equivalent to
0
(αχ(e)
)−1 ◦ φτ (e) ◦ αe . Extend χ to lk(c) \ {e} as in Theorem 6.13.
This completes the induction. The result is χ ∈ Isom((T, δ), (T 0 , δ 0 )) and uniform
quasi-isometries (φv ) satisfying the conditions of Corollary 2.18, so (φv ) is a tree of
quasi-isometries over χ compatible with X and X 0 , as desired.
References
[1] Jason A. Behrstock, Bruce Kleiner, Yair N. Minsky, and Lee Mosher, Geometry and
rigidity of mapping class groups, Geom. Topol. 16 (2012), no. 2, 781–888.
[2] Jason A. Behrstock and Walter D. Neumann, Quasi-isometric classification of graph
manifold groups, Duke Math. J. 141 (2008), no. 2, 217–240.
[3] Jason A. Behrstock and Walter D. Neumann, Quasi-isometric classification of nongeometric 3-manifold groups, J. Reine Angew. Math. 669 (2012), 101–120.
[4] M. Bestvina and M. Feighn, A combination theorem for negatively curved groups, J.
Differential Geom. 35 (1992), no. 1, 85–101.
[5] Kingshook Biswas and Mahan Mj, Pattern rigidity in hyperbolic spaces: duality and
PD subgroups, Groups Geom. Dyn. 6 (2012), no. 1, 97–123.
[6] M. Bonk and O. Schramm, Embeddings of Gromov hyperbolic spaces, Geom. Funct.
Anal. 10 (2000), no. 2, 266–306.
[7] Marc Bourdon and Hervé Pajot, Rigidity of quasi-isometries for some hyperbolic
buildings, Comment. Math. Helv. 75 (2000), no. 4, 701–736.
[8] Marc Bourdon and Hervé Pajot, Cohomologie lp et espaces de Besov, J. Reine
Angew. Math. 558 (2003), 85–108.
[9] Brian H. Bowditch, Cut points and canonical splittings of hyperbolic groups, Acta
Math. 180 (1998), no. 2, 145–186.
[10] Martin R. Bridson and André Haefliger, Metric spaces of non-positive curvature,
Grundlehren der mathematischen Wissenschaften, vol. 319, Springer, Berlin, 1999.
58
[11] Sergei Buyalo and Viktor Schroeder, Elements of asymptotic geometry, EMS Monographs in Mathematics, European Mathematical Society (EMS), Zürich, 2007.
[12] Christopher H. Cashen, Splitting line patterns in free groups, Algebr. Geom. Topol.,
in press, arXiv:1009.2492.
[13] Christopher H. Cashen, Quasi-isometries between tubular groups, Groups Geom.
Dyn. 4 (2010), no. 3, 473–516.
[14] Christopher H. Cashen and Nataša Macura, Line patterns in free groups, Geom.
Topol. 15 (2011), no. 3, 1419–1475.
[15] Pallavi Dani and Anne Thomas, Quasi-isometry classification of certain right-angled
coxeter groups, preprint (2014), arXiv:1402.6224v4.
[16] Martin J. Dunwoody, The accessibility of finitely presented groups, Invent. Math. 81
(1985), no. 3, 449–457.
[17] Martin J. Dunwoody and Michah E. Sageev, JSJ-splittings for finitely presented
groups over slender groups, Invent. Math. 135 (1999), no. 1, 25–44.
[18] Max Forester, On uniqueness of JSJ decompositions of finitely generated groups,
Comment. Math. Helv. 78 (2003), no. 4, 740–751.
[19] Koji Fujiwara and Panos Papasoglu, JSJ-decompositions of finitely presented groups
and complexes of groups, Geom. Funct. Anal. 16 (2006), 70–125.
[20] Mikhael Gromov, Infinite groups as geometric objects, Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Warsaw, 1983) (Warsaw), PWN, 1984,
pp. 385–392.
[21] Mikhael Gromov, Hyperbolic groups, Essays in group theory, Math. Sci. Res. Inst.
Publ., vol. 8, Springer, New York, 1987, pp. 75–263.
[22] Vincent Guirardel and Gilbert Levitt, JSJ decompositions: definitions, existence,
uniqueness. I: The JSJ deformation space, preprint (2009), arXiv:0911.3173.
[23] Vincent Guirardel and Gilbert Levitt, Trees of cylinders and canonical splittings,
Geom. Topol. 15 (2011), no. 2, 977–1012.
[24] Ilya Kapovich, The combination theorem and quasiconvexity, Internat. J. Algebra
Comput. 11 (2001), no. 2, 185–216.
[25] Michael Kapovich and Bruce Kleiner, Hyperbolic groups with low-dimensional boundary, Ann. Sci. École Norm. Sup. (4) 33 (2000), no. 5, 647–669.
[26] Bruce Kleiner and Bernhard Leeb, Rigidity of quasi-isometries for symmetric spaces
and Euclidean buildings, Inst. Hautes Études Sci. Publ. Math. (1997), no. 86, 115–
197 (1998).
59
[27] Frank Thomson Leighton, Finite common coverings of graphs, J. Combin. Theory
Ser. B 33 (1982), no. 3, 231–238.
[28] William Malone, Topics in geometric group theory, Ph.D. thesis, University of Utah,
Salt Lake City, UT, 2010.
[29] Vladimir Markovic, Quasisymmetric groups, J. Amer. Math. Soc. 19 (2006), no. 3,
673–715.
[30] Alexandre Martin, Non-positively curved complexes of groups and boundaries, Geom.
Topol. 18 (2014), no. 1, 31–102.
[31] Alexandre Martin and Jacek Świątkowski, Infinitely-ended hyperbolic groups with
homeomorphic Gromov boundaries, J. Group Theory 18 (2015), no. 2, 273–289.
[32] Mahan Mj, Pattern rigidity and the Hilbert-Smith conjecture, Geom. Topol. 16
(2012), no. 2, 1205–1246.
[33] Lee Mosher, Michah Sageev, and Kevin Whyte, Quasi-actions on trees II: Finite
depth Bass-Serre trees, Mem. Amer. Math. Soc. 214 (2011), no. 1008, vi+105.
[34] Walter D. Neumann, On Leighton’s graph covering theorem, Groups Geom. Dyn. 4
(2010), no. 4, 863–872.
[35] Pierre Pansu, Métriques de Carnot-Carathéodory et quasiisométries des espaces
symétriques de rang un, Ann. of Math. (2) 129 (1989), no. 1, 1–60.
[36] Panos Papasoglu, Quasi-isometry invariance of group splittings, Ann. of Math. (2)
161 (2005), no. 2, 759–830.
[37] Panos Papasoglu and Kevin Whyte, Quasi-isometries between groups with infinitely
many ends, Comment. Math. Helv. 77 (2002), no. 1, 133–144.
[38] E. Rips and Z. Sela, Structure and rigidity in hyperbolic groups. I, Geom. Funct.
Anal. 4 (1994), no. 3, 337–371.
[39] E. Rips and Z. Sela, Cyclic splittings of finitely presented groups and the canonical
JSJ decomposition, Ann. of Math. (2) 146 (1997), no. 1, 53–109.
[40] Richard Evan Schwartz, Symmetric patterns of geodesics and automorphisms of surface groups, Invent. Math. 128 (1997), no. 1, 177–199.
[41] Jean-Pierre Serre, Trees, Springer Monographs in Mathematics, Springer-Verlag,
Berlin, 2003, Translated from the French original by John Stillwell, Corrected 2nd
printing of the 1980 English translation.
[42] John R. Stallings, On torsion-free groups with infinitely many ends, Ann. of Math.
(2) 88 (1968), 312–334.
60
[43] John R. Stallings, Group theory and three-dimensional manifolds, Yale University
Press, New Haven, Conn., 1971, A James K. Whittemore Lecture in Mathematics
given at Yale University, 1969, Yale Mathematical Monographs, 4.
[44] Pekka Tukia, Quasiconformal extension of quasisymmetric mappings compatible with
a Möbius group, Acta Math. 154 (1985), no. 3-4, 153–193.
[45] Diane M. Vavrichek, The quasi-isometry invariance of commensurizer subgroups,
Groups Geom. Dyn. 7 (2013), no. 1, 205–261.
[46] Xiangdong Xie, Quasi-isometric rigidity of Fuchsian buildings, Topology 45 (2006),
no. 1, 101–169.
[47] Xiangdong Xie, Quasisymmetric maps on the boundary of a negatively curved solvable Lie group, Math. Ann. 353 (2012), no. 3, 727–746.
Fakultät für Mathematik, Universität Wien, Oskar-Morgenstern-Platz 1, 1090 Wien,
Österreich
[email protected]
[email protected]
61
| 4math.GR
|
Neuronal Circuit Policies
Mathias Lechner * 1 Ramin M. Hasani * 1 Radu Grosu 1
arXiv:1803.08554v1 [q-bio.NC] 22 Mar 2018
Abstract
We propose an effective way to create interpretable control agents, by re-purposing the function of a biological neural circuit model, to govern simulated and real world reinforcement learning (RL) test-beds. We model the tap-withdrawal
(TW) neural circuit of the nematode, C. elegans,
a circuit responsible for the worm’s reflexive response to external mechanical touch stimulations,
and learn its synaptic and neuronal parameters as
a policy for controlling basic RL tasks. We also
autonomously park a real rover robot on a predefined trajectory, by deploying such neuronal
circuit policies learned in a simulated environment. For reconfiguration of the purpose of the
TW neural circuit, we adopt a search-based RL
algorithm. We show that our neuronal policies
perform as good as deep neural network policies
with the advantage of realizing interpretable dynamics at the cell level.
1. Introduction
Through natural evolution, the nervous system of the nematode, C. elegans, structured a near optimal wiring diagram
(White et al., 1986). Its stereotypic brain composed of 302
neurons connected through approximately 8000 chemical
and electrical synapses (Chen et al., 2006). C. elegans exhibits distinct behavioral mechanisms to process complex
chemical stimulations (Bargmann, 2006), avoid osmotic regions (Culotti & Russell, 1978), sleep (Nichols et al., 2017),
show adaptive behavior (Ardiel & Rankin, 2010), perform
mechanosensation (Chalfie et al., 1985b), and to control
muscles (Wen et al., 2012).
The functions of many neural circuits within its brain have
been identified (Wicks & Rankin, 1995; Chalfie et al.,
1985a; Li et al., 2012; Nichols et al., 2017). In particular, a neural circuit which is responsible for inducing a
forward/backward locomotion reflex when the worm is
*
Equal contribution 1 Cyber Physical Systems, TU Wien,
Austria.
Correspondence to: Mathias Lechner <[email protected]>, Ramin Hasani <[email protected]>.
mechanically exposed to touch stimulus on its body, has
been well-characterized (Chalfie et al., 1985a). The circuit
is called tap-withdrawal (TW) and it comprises 9 neuron
classes which are wired together by means of chemical and
electrical synapses. Synaptic polarities (either being excitatory or inhibitory) of the circuit have then been predicted,
suggesting that the circuit realizes a competitive behavior between forward and backward reflexes, in presence of touch
stimulations (Wicks & Rankin, 1995; Wicks et al., 1996).
Behavior of the tap-withdrawal (TW) reflexive response is
substantially similar to the control agent’s reaction in some
standard control settings such as the impulse response of
a controller operating on an Inverted Pendulum (Widrow,
1964; Doya, 2000; Russell & Norvig, 2010), a controller
acting on driving an under-powered car, to go up on a steep
hill, known as the Mountain Car (Moore, 1990; Singh &
Sutton, 1996), and a controller acting on the navigation of a
rover robot that plans to go from point A to B, on a planned
trajectory, with two control commands of angular and linear
velocity.
We intend to take advantage of the similarity and reconfigure the synaptic and neuronal parameters of a deterministic
dynamic model of the TW neural circuit, in each of the
mentioned control settings. We use publicly available reinforcement learning toolkits, to evaluate the performance
of our neuronal circuit policies. The environments include
the inverted pendulum (Schulman et al., 2017), the continuous mountain car of OpenAI Gym1 and rllab 2 , and the
Cart-pole of rllab (Duan et al., 2016). In a real robotic setting, We also determine a control task for a rover robot to
park autonomously in a specific parking spot, by a learned
TW neuronal policy. For all three control challenges, we
preserve the near-optimal wiring structure of the TW circuit
and adopt a search-based reinforcement learning (RL) algorithm for synaptic parametrization of the network. The
approach is named as neuronal circuit policies.
Our principle contribution in this work is to demonstrate the
performance of a compact neuronal circuit model from the
brain of the C. elegans worm, as an interpretable continuous
time recurrent neural network, in standard control and RL
settings. In our experimental evaluations, we demonstrate
1
2
https://github.com/openai/gym
https://github.com/rllab/rllab
Neuronal Circuit Policies
that our control agent can achieve the performance of the
conventional and the state-of-the-art artificial intelligence
(AI) control agents, by solving five RL tasks. We show how
a learned neuronal circuit policy in a simulated environment,
can be transferred to a real robotic environment. We also
demonstrate that the function of the neurons in a learned
neuronal network is interpretable.
2. Preliminaries
In this section, we first briefly describe the structure and
dynamics of the tap-withdrawal neural circuit. We then
introduce the mathematical neuron and synapse models utilized to build up the model of the circuit.
2.1. Tap-Withdrawal Neural Circuit Revisit
A mechanically exposed stimulus (i.e. tap) to the petri dish
in which the worm inhabits, results in the animal’s reflexive
response in the form of a forward or backward movement.
This response has been named as the tap-withdrawal reflex, and the circuit identified to underly such behavior is
known as the tap-withdrawal (TW) neural circuit (Rankin
et al., 1990). The circuit is shown in Figure 1. It is composed of four sensory neurons, PVD and PLM (posterior
touch sensors), AVM and ALM (anterior touch sensors),
four interneuron classes (AVD, PVC, AVA and AVB), and
two subgroup of motor neurons which are abstracted as a
forward locomotory neurons, FWD, and backward locomotory neurons, REV. Neurons recurrently synapse into each
other with excitatory and inhibitory synaptic links. It has
been shown that sensory neurons of the TW circuit get acA
B
tivated as a result of an input tap, and transfer the stimulus
through the modulatory interneurons PVC and AVD to the
command neurons AVA and AVB (Wicks et al., 1996). The
TW reflex is then modulated by a competition between these
two command neurons, either resulting in a forward scape
p ))
p ))
response (AVB’s activation
(( Tadominates AVA’s) or a reversal
(( Ta
scape response. Throughout the paper, we illustrate how
such recurrent neuronal network can be deployed in standard RL settings. We first sketch how we modeled neurons
and synapses to build up the TW circuit.
brane equation (Koch & Segev, 1998):
Cm
n
X
dvi
(i)
= GLeak VLeak − vi (t) +
Iin ,
dt
i=1
where Cm , GLeak and VLeak are parameters of the neuron
(i)
and Iin , stands for the external currents to the cell. We
adopted Eq. (1) to govern interneurons’ dynamics.
For interacting with the environment, We introduced sensory and motor neuron models, separately. A sensory component consists of two neurons Sp , Sn and a measurable
dynamic system variable, x. Sp gets activated when x has a
positive value, whereas Sn fires when x is negative. Mathematically, the potential of the neurons Sp , and Sn , as a
function of x, can be expressed as
if x ≤ 0
−70mV
50mV
Sp (x) := −70mV + xmax x if 0 < x ≤ xmax (2)
−20mV
if x > xmax
if x ≥ 0
−70mV
50mV
Sn (x) := −70mV + xmin x if 0 > x ≥ xmin (3)
−20mV
if x < xmin .
This maps the region [xmin , xmax ] of system variable x, to a
membrane potential range of [−70mV, −20mV ]. Note that
the potential range is selected to be close to the biophysics
of the nerve cells, where the resting potential is usually set
around -70 mV and a neuron can be considered to be active
when it has a potential around -20 mV (Hasani et al., 2017).
C
REV
PVD
PLM
AVA
AVD
PVC
AVB
AVM
ALM
DVA
2.2. Neuron Model
Most of the neurons in C. elegans are observed to exhibit
electrotonic dynamics (Kato et al., 2015), meaning that elecResponse
Response
tric charges spread passively inside a neuron creating graded
potentials. This implies that the neurons are non-spiking.
Dynamics of the neurons’ membrane potential therefore,
were modeled by the well-known, deterministic ordinary
differential equation (ODE), the single-compartment mem-
(1)
FWD
Legend:
Sensory neuron
Excitatory synapse
Inter neuron
Inhitory synapse
Motor neuron
Gap Junction
Figure 1. Tap-Withdrawal neural circuit schematic.
Neuronal Circuit Policies
Similar to sensory neurons, a motor component is composed of two neurons Mn , Mp and a controllable motor
variable y. Values of y is computed by y := yp + yn and
if Mp > −20mV
ymax ,
p +70mV )
yp (Mp ) := ymax (M
,
if Mp ∈ [−70, −20]mV
50mV
0,
if Mp < −70mV
(4)
if Mn > −20mV
ymin ,
n +70mV )
yn (Mn ) := ymin (M
,
if Mn ∈ [−70, −20]mV
50mV
0,
if Mn < −70mV
(5)
This maps the neuron potentials Mn and Mp , to the range
[ymin , ymax ]. FWD and REV motor classes in Figure 1, are
modeled in this fashion.
Chemical synapses are points at which two neurons trade
information by the release of neurotransmitters. The chemical synaptic current depends on a non-linear component
standing for their conductance strength, which is a function
of the presynaptic neurons’ potential, Vpre , and have a maximum weight of w,(standing for the maximum conductance
of the synapse) as, (Koch & Segev, 1998):
(6)
Moreover, the synaptic current linearly depends on the postsynaptic neuron’s membrane potential, Vpost , and therefore
can be formulated as, (Koch & Segev, 1998):
Is = g(Vpre )(E − Vpost ),
Input: A stochastic objective indicator f ,
a starting parameter θ
Input: Optimized parameter θ
fθ ← f (θ)
for k ← 1 to maximum iterations do
θ0 ← θ + rand()
fθ0 ← f (θ0 )
if fθ0 < fθ then
Set θ ← θ0
fθ ← fθ0
i←0
end if
i←i+1
if i > N then
fθ ← f (θ)
end if
end for
return θ
concrete discussion on the model implementation, and the
choice of parameters.
2.3. Synapse Model
g(Vpre ) = w/1 + eσ(Vpre +µ)
Algorithm 1 Random Search + Objective Indicator
(7)
where by varying E, the reversal potential of the synapse, it
realizes inhibitory or excitatory connection to their postsynaptic neurons.
An electrical synapse (gap-junction), which is a physical
junction between two neurons, was modeled by a constant
conductance, ω̂, where based on the Ohm’s law their bidirectional current between neurons j and i, can be computed
as
Iˆi,j = ω̂ vj (t) − vi (t) .
(8)
For simulating neural networks composed of such dynamic
models, we adopted a implicit numerical solver (Press et al.,
2007). Formally, we realized the ODE models in a hybrid
fashion which combine both implicit and explicit Euler’s
method. See supplementary materials, Section 2, for a
Note that one objective of the solver is to be employed
in a real-time control system. For reducing the complexity therefore, our method realizes a fixed-step solver. The
solver’s complexity for each time step ∆t is O(|# neurons|+
|# synapses|). The solver is implemented in C++ in which
we construct the TW circuit for performing specific control
tasks. We now need to formalize a learning platform to tune
the parameters of the circuit for the desired control problem.
3. Search-based Reinforcement Learning
In this section we formulate an RL setting for training the
parameters of the neural circuit to perform the balancing
of the inverted pendulum, control the mountain car, and to
park the rover robot.
The behavior of a neural circuit can be expressed as a policy
πθ (oi , si ) 7→ hai+1 , si+1 i, that maps an observation oi , and
an internal state si of the circuit, to an action ai+1 , and a
new internal state si+1 . This policy acts upon a possible
stochastic environment Env(ai+1 ), that provides an observation oi+1 , and a reward, ri+1 . The stochastic return is
PT
given by R(θ) := t=1 rt .
Objective of the Reinforcement
learning is to find a θ that
maximizes E R(θ) .
Approaches to find such optimal θ, can be categorized into
two major groups, based on how randomness is formulated
for the environment explorations (Schulman et al., 2015;
Salimans et al., 2017): I-Gradient-based and II-search-based
methods. The principle of gradient-based RL is to perform
random sampling for generating ai , and use the action’s
influence on the return value, to improve θ (Williams, 1992).
Search-based methods directly randomly sample parame-
Neuronal Circuit Policies
ters and estimate how good these random parameters are,
to update θ (Salimans et al., 2017; Szita & Lörincz, 2006).
Here, we adopted such search-based optimization which can
be applied in any control setting, regardless of the internal
structure of the policy, (black-box optimization). One major obstacle for search-based optimization is the stochastic
nature of the RL environment, which makes the objective
function, f (θ) = E(Rθ ), a probability distribution and the
underlying optimization problem an instance of Stochastic Optimization (Spall, 2003). Samples of the objective
distribution can be generated by running rollouts with πθ
on the environment. A possible solution to overcome a
high variance when estimating E(Rθ ), is to rely on a very
large number of samples (Salimans et al., 2017; Duan et al.,
2016), which nonetheless comes with high computational
costs. Our approach is based on a Random Search (RS)
(Rastrigin, 1963) optimization, combined with an Objective
Estimate (OE) as an objective function f : θ 7→ R+ .
The OE generates N rollouts with πθ on the environment
and computes an estimate of E(Rθ ) based on a filtering
mechanism on these N samples. We compared two filtering
strategies in this context; I-Taking the average of the N
samples, and II: taking the average of the worst k samples
out of N samples. The first strategy is equivalent to the
Sample Mean estimator, whereas the second strategy aims
to avoid getting mislead by outlying high samples of E(Rθ ).
Our objective for realizing the second strategy was the fact
that a suitable parameter θ makes the policy πθ , control the
environment in a decent way even in difficult situations (i.e.
rollouts with the lowest return).
In both strategies, to ensure that a single outlying high OE
for some θ does not hinder the algorithm to find a legitimately good parameter, the OE is reevaluated after it is
utilized M times, within the algorithm. The full algorithm
is outlined in Algorithm 1.
4. Experiments
The goal of our experiments is to answer the following
questions: 1) How would a neural circuit policy perform
in a basic standard control setting? 2) When possible, how
would the performance of our learned circuit compare to the
other methods? 3) Can we transfer a policy from a simulated
environment to a real environment? 4) Can we interpret the
behavior of the neural circuit policies?
We prepared five environmental settings for the TW sensory/motor neurons and then deployed our RL algorithm to
learn the parameters of the TW circuit to realize the given
control objective. Environments include I) Inverted pendulum of Roboschool (Schulman et al., 2017), II) Mountain
car of OpenAI Gym, III) Mountain car of rllab, IV) cart-pole
balancing of rllab and V) Parking a real rover robot from
A
- !̇
-"̇
+Control
- ! +!
B
+!̇
+"̇
-Control
Sensory Inputs
- ! +! +" -"
Tap-withdrawal
neural circuit
- Control + Control
Motor outputs
Figure 2. Inverted Pendulum setting for the TW circuit. A) observations and control variables in the inverted pendulum balance
problem. B) observation and control signals mapping of the pendulum to the TW circuit.
a learned policy in a simulated environment. The code is
available online 1 .
The major constraint that prevents the TW neural circuit
(shown in Figure 1) to be tested on arbitrary tasks, is its
limited number of sensory and motor neurons. As the TW
circuit allows us to incorporate only two input and one output variables, we selected tasks that can be solved with only
two observations. We chose environments from different
toolkits to evaluate our neural circuit policies on a variety of
dynamics, interactions and reward settings. For instance, the
two mountain car suites of OpenAI Gym and rllab realize
different positive rewards functions as explained in section
4.3.
4.1. Inverted pendulum with the TW neural circuit
The TW neural circuit shown in Figure 1, contains four
sensory neurons. It therefore, allows us to map the circuit
to only two input variables. Note that as we discussed in
section 2.2, a sensory component expressed in the form of
Eq. (2) and Eq. (3), consists of two neurons for incorporating positive and negative values of the dynamic system
variable. The inverted pendulum environment provides four
observation variables as shown in Figure 2A. The position
of the cart x, together with its velocity ẋ, the angle of the
pendulum ϕ2 along with its angular velocity ϕ̇.
Since the main objective of the controller is to balance the
pendulum in an upward position and make the car stay
within the horizontal borders, we fed ϕ, and x, as the inputs
to the sensors of the TW circuit, as illustrated in Figure 2B.
Control commands to the pendulum were originated by the
abstract motor neuron classes, FWD and REV components.
The activity of these components are governed by Eq. (4)
1
Code for all experiments is available online at:
https://github.com/mlech26l/neuronal_
circuit_policies
2
Remark: The environment further splits ϕ into sin(ϕ) and
cos(ϕ) to avoid the 2π → 0 discontinuity
Neuronal Circuit Policies
Table 1. Mapping the environmental variables to the sensory and motor neurons of the TW circuit, in different experiments
Experiment
Environment variable
Type
Positive neuron Negative neuron
ϕ
Sensor (pendulum angle)
PLM
AVM
Inverted Pendulum
x
Sensor (cart position)
ALM
PVD
a (Control)
Motor (move right/left)
FWD
REV
x
Sensor (car position)
PLM
AVM
ẋ
Sensor (car’s linear velocity)
ALM
PVD
Mountain Car
(OpenAI Gym)
a (Control)
Motor (move right/left)
FWD
REV
x
Sensor (car position)
PLM
AVM
Mountain Car
ẋ
Sensor (car’s linear velocity)
ALM
PVD
(rllab)
a (Control)
Motor (move right/left)
FWD
REV
ϕ
Sensor (pole angle)
PLM
AVM
ϕ̇
Sensor (pole angular velocity)
ALM
PVD
Cart-Pole
a (Control)
Motor (move right/left)
FWD
REV
x
Sensor (estimated x position)
PVD
y
Sensor (estimated y position)
PLM
Parking of a Rover
s
Sensor(start signal)
AVM
θ
Sensor (estimated angular pose)
ALM
a1 (Control)
Motor (angular velocity)
FWD
REV
a2 (Control)
Motor (linear velocity)
FWD/REV
We set up the search-based RL algorithm to
optimize the neurons and synapses parameters
ω, ω̂, σ, Cm , VLeak and GLeak in the Roboschool
RoboschoolInvertedPendulum-v1 environment
with an slight modification (Schulman et al., 2017) in
the reward calculation. It is desirable for the cart to
stay in the center of the horizontal space. Therefore,
we incorporated an additional axillary reward; if the
pendulum is in the up-right position and in the center, an
additional 20% reward is collected. This bonus linearly
decreased. For instance the bonus reward is 10% if the
cart is halfway to the end. Right at the borders the bonus
reward vanishes. A video of different stages of the learned
neuronal circuit policy for the inverted pendulum can be
viewed at https://youtu.be/iOHeQ7DhQv8
Note that we were also able the train the TW circuit to solve
original version of the inverted pendulum task when feeding
ϕ, and ϕ̇ into the circuit. However, we observed a slow drift
of the pendulum toward one of the two ends in final policy,
which we wished to eliminate.
Similar to configuration of the inverted pendulum, we set the
observational and motor variables for the tap withdrawal, as
illustrated in Figure 3B. The circuit was then learned by the
search-based RL algorithm. A video illustrating the control
of the car at various episodes during the learning process can
be viewed at https://youtu.be/lMrP1sXp3jk.
4.3. The mountain car (rllab) control with the TW
neural circuit
The environmental observations and motor actions for the
mountain car are the same in rllab and OpenAI Gym. Therefore, we set up the TW circuit the same way we did for the
mountain car in the Gym.
The major difference of the two environments are in the
way the reward is computed. The environment of rllab has
a graded continuous reward which is directly associated
with the position of the car, whereas in the OpenAI Gym
implementation, reward is sparse and gets allocated only
once the car reaches a certain altitude. Furthermore, in both
A
B
+!̇
4.2. The mountain car (OpenAI Gym) control with the
TW neural circuit
+C
on
tro
l
In this experiment we trained the TW circuit to drive the car
shown in Figure 3A uphill to the right-hand side, by generating gravitational momentum. The observation variables are
the car’s position (on the horizontal axis, x) together with
its linear velocity. The control signal applies force to the car
to move it to the right- and left-hand side, periodically, to
finally bring the car up on the hill.
-!̇
-C
on
tro
l
and Eq. (5), respectively, and represented in Figure 2B,
graphically.
0
!
Sensory Inputs
-! +! -!̇ +!̇
Tap-withdrawal
neural circuit
- Control + Control
Motor outputs
Figure 3. The Mountain car setup. A) observations and control
variables in the mountain car environment. B) mapping of the
observation and control signals to TW circuit.
Neuronal Circuit Policies
frameworks, reward is subjected to a penalty as follows;
in rllab the penalty is constant, whereas in the Gym the
amount of the penalty varies depending on the amplitude
of the performed action. Consequently, energy efficient
solutions achieve a higher score, in the Gym environment,
in contrast to the rllab version, where the highest scoring
solutions bring the car uphill as fast as possible.
4.4. Cart-pole (rllab) control with the TW circuit
The Cart-pole environmental setting is substantially similar
to that of the inverted pendulum. In contrast to the inverted
pendulum experiment, here we mapped the observation
variables, pole angle, ϕ and the pole’s angular velocity ϕ̇
to the sensor. As a result, the controller is not aware of
the position of the cart, x, and whether the boundary of the
movable space is nearby or not. An intuitive solution for
handling more control variables would be to add sensory
neurons to the TW circuit. However, in the present work
we intended to maintain the structure of the TW circuit
constant, and test its capabilities in control scenarios with
the resulting partial observability. Environmental variables
for the cart-pole suite were mapped to the circuit elements
as denoted in Table 1.
4.5. Autonomous parking of the rover with the TW
neural circuit
In this experiment, we generalized our TW neuronal circuit
policy to a real-world control setting. We let the TW circuit
learn to park a rover robot on a determined spot, given a set
of checkpoints which form a trajectory, in a deterministic
simulated environment. We then deployed the learned policy
on a mobile robot in a real environment shown in Figure
4A. The key objective here was to show the capability of the
method to perform well in a transformation from a simulated
environment to a real setting. For doing this, we developed
a custom deterministic simulated RL environment1 .
The rover robot provides four observational variables (Starting signal, position of the rover, x, y and its angular pose, θ),
together with two motor actions (linear and angular velocity,
v and w). We mapped all four observatory variables, as
illustrated in Figure 4B, to the sensors of the TW. Note that
here the geometric reference of the surrounding space is set
at the initial position of the robot. Therefore, observation
variables are only positive.
We mapped the linear velocity (which is a positive variable
throughout the parking task) to one motor neuron and the
same variable to another motor neuron. We determined two
motor neurons for the positive and negative angular velocity.
The mapping details are provided in Table 1. This configu1
https://github.com/mlech26l/neuronal_
circuit_policies
A
B
+𝑦
Sensory Inputs
+𝑥 +𝑦 𝜃
Start
I
+𝑥
IV
II
V
Tap-withdrawal
neural circuit
VI
+𝑤 -𝑤
+𝑣
Motor outputs
𝜃
III
Figure 4. Parking setup. A) Environmental setting for the parking
task. B) Mapping of the observations and control commands from
the environment to the TW circuit.
ration implies that the command neuron, AVA, controls two
motor neurons responsible for the turn-right and forward
motion-primitives, and AVB to control the turn-left and
forward motor neurons. Therefore, the TW circuit is able
to govern the robot’s locomotion in three different directions, forward, left- and right-turns, which are sufficient to
perform a parking trajectory. Furthermore, each command
neuron is able to move the rover forward and turn it to the
left or right.
4.5.1. RL SETTING FOR THE PARKING TASK
A set of checkpoints on a pre-defined parking trajectory
were determined in the custom simulated environment. For
every checkpoint, a deadline was assigned. At each deadline
a reward was given as the negative distance of the rover
to the current checkpoint. The checkpoints are placed to
resemble a real parking trajectory composed of a sequence
of motion primitives: Forward, turn left, forward, turn right,
forward and stop. We then learned the TW circuit, by the
RL algorithm. The learned policy has been mounted on
a Pioneer AT-3 mobile robot and performed a reasonable
parking performance. The Video of the performance of
the TW neuronal circuit policy on the parking task can be
viewed at https://youtu.be/Vwydc2ez9Wc.
5. Experimental Evaluation
In this section, we thoroughly assess the results of our experimentation. We qualitatively and quantitatively explain
the performance of our neuronal circuit policies. We then
benchmark our results where possible, with the existing
methods and describe the main attributes of our methodology. We finally illustrate how the activity of the learned
neuronal circuits can be interpretable.
5.1. Performance
The training algorithm was able to solve all five tasks, after
a reasonable number of iterations as shown in Figure 5. All
learning curves reached a stable state after the given number
of iterations.
Neuronal Circuit Policies
A
B
C
D
E
F
Figure 5. Performance as a function of the number of iterations. A) Inverted pendulum B) Mountain car (OpenAI Gym) C) Mountain car
(rllab) D) Cart-pole (rllab) E) parking trajectory of the rover robot F) Influence of filter size on training performance
Jumps in the learning curves of the mountain car in OpenAI
Gym (Figure 5B) are the consequence of the sparse reward.
The inverted pendulum of the Roboschool and the cartpole of rllab, realized substantially similar learning curves
(Figure 5A and 5D) . This indicates the robustness of the
policy and the learning algorithm to different frameworks for
a given control task. For the deterministic parking trajectory,
the learning curve approaches a stable state, exponentially
fast.
The final return values for the basic standard RL tasks (provided in Table 2), matches that of conventional policies
(Heidrich-Meisner & Igel, 2008), and the state-of-the-art
deep neural network policies learned by many RL algorithms (de Froissard de Broissia & Sigaud, 2016; Schulman
et al., 2017; Berkenkamp et al., 2017). Table 2, also depicts the average return over the entire training iterations.
The average return is not significant compared to the other
algorithms applied to the artificial neural network policies
(Duan et al., 2016), due to the smaller number of training
iterations. Training curves in our case however, reaches an
stable state reasonably fast in all tasks even with a fewer
number of epochs.
To take advantage of parallel hardware resources and to
further increase the performance of our training algorithm,
we deployed an ensemble of TW agents which were trained
independently in parallel. Depending on the difficulty of the
Table 2. Training results. Return value ± standard deviation. Performance of the learned neuronal circuit policies in terms of final
return and average return over all training iterations.
Environment
Final return
Average return
Inverted pendulum 1168.5 ± 21.7 394.1 ± 80.9
Mountain car (Gym) 91.5 ± 6.6
28.1 ± 9.6
Mountain car (rllab) −61.3 ± 1.4 −196.3 ± 78.0
Cart-pole balancing 3716.1 ± 240.1 1230.3 ± 272.2
Parking
−0.49 ± 0.63 −6.13 ± 1.41
task, between 100% (e.g. Roboschool’s inverted pendulum)
and 25% (e.g. in Gym’s Mountaincar) of the ensembles
were able to solve the task within the reported time frame.
The values reported in table 2, and Fig5, correspond to the
successful cases of the ensemble.
5.2. Effect of filter size on the training performance
Fig 2F shows how the choice of the objective-estimate, affects the training performance. The objective-estimate is
defined as the mean of the k (=filter size) returns out of 20
rollouts. When k = 20, the estimation is equal to the sample
mean. We tested two environments with different reward
settings: The mean return of Mountaincar (OpenAI Gym)
after 50,000 training iterations to represent a sparse reward
scenario and the mean return of our modified inverted pendulum task after 15,000 training iterations as an example of
a gradual reward setting. The results denote that in a sparse
reward setting, filtering of the high-outliers tend to degrade
the training performance while in a gradual reward setup,
filtering of the outliers improves the performance. The reported values in Fig2F, correspond to the average, when
running this experiment 10 times. Further discussions about
this can be found in the supplementary materials Section 4.
5.3. Interpretability of the neuronal circuit policies
Interpretability of neural network policies is still a challenge to be solved (Heess et al., 2016; Bacon et al., 2017).
Successful attempts on the development of interpretable
representation learning have been provided where the latent
space of the learned policies demonstrate interpretable skills
(Chen et al., 2016; Florensa et al., 2017). A neuronal circuit
policy has the distinct attribute of being interpretable at the
cell level. Here, we show how the activity of the learned TW
neuronal policies, in different domains is auditable. Figure
6A to 6C represent the normalized membrane potential of
neurons in learned policies for the inverted pendulum, the
Neuronal Circuit Policies
A
B
C
Figure 6. Neuronal activity of the Learned neuronal policies. A) The inverted pendulum circuit B) The mountain car (OpenAI Gym)
circuit C) The parking circuit. Individual neuron’s resting potential, Vleak , is mapped to 0, neuron’s maximum potential is mapped to 1
and neuron’s minimum potential is mapped to -1.
mountain car and the parking task. We describe the global
neuronal dynamics in these scenarios.
Inverted Pendulum - Activity of the learned Tapwithdrawal’s neurons in a successful episode of the inverted
pendulum control, is shown in Figure 6A. The learned circuit surprisingly realized a competitive behavior between
two sub-circuits within the network, similar to the reflexive
behavior observed in the worm (Wicks et al., 1996). Neurons AVM, AVD, AVA, and REV control the pendulum not
to fall on the left side, whereas a circuit composed of PLM,
PVC, AVB and FWD, controls the other side of the balance.
The antagonistic behavior between the command neurons
AVA and AVB balances the pendulum in the middle. Note
that DVA neuron acts as a mediator which couples the kinetics of the two sub-circuits. Note that Neurons’ dynamics
are realized with a variety of time-constants during the balancing control episodes. For instance PVC modulates fast
undulations while the motor neurons work with a slower
dynamic.
Mountain Car - As illustrated in Figure 6B, a gradual increase in the amplitude of the overall periodic dynamic of
the neurons is observed due to the alternation of the direction of movement to bring the car uphill. Sensory neurons
linearly transform the environmental observations to the
global activity of the system, forming a phase-shifted highly
synchronized behavior. In every episode, by the rising of
PDV (higher negative velocity) or PLM (be at a positive x),
all neurons approach their resting potential. This feedback
mechanism which prevents unstable behavior, originated
from inhibition of PVC and AVD by the ALM sensory neuron. At every period in which the car reaches the maximum
position on the left-hand side, a circuit composed of AVM,
AVD, AVB and FWD applies a force-pulse which pushes
the car to get closer to the target. AVA and AVB retained
their objective as in the TW actual circuit, as the command
neurons which enforce the network’s decision to the motor neurons REV and FWF. However their activity is not
fully antagonistic in this learned network. This is where the
competition "A more active command neuron, gives rise to
a more active downstream motor neuron, and as a results
winning the competition", is still present.
Parking of the rover - Figure 6C shows the activity of the
learned TW policy on the parking task. AVM receives the
starting signal and accordingly, a left turn together with
moving forward are initiated by the motor neuron FWD and
LFT. This is governed by the command neuron AVB. The
rover continues moving forward and initializes a right turn
by the motor neurons FWD and RGT. This is controlled by
AVA. AVD learned to control the turning-phases of the rover
by moderating its state, (a higher potential during right turns
and a lower membrane state during left turns). PVC and
DVA function as the pre-command neurons which incorporate the sensory and the network inputs, into a readable
action for the command neuron AVB.
6. Conclusions
In this work, we showed that a neuronal circuit policy can be
adopted to function as a controller for standard control tasks
expressing similar characteristics to the original purpose of
the circuit. We illustrated the performance of such policies
learned by a search-based RL algorithm, in standard basic
RL domains. We also generalized the use of such policies in
a real-domain robot control. More importantly, for various
domains we showed that our learned policies developed
interpretable skills at the neuron level.
Our neuronal policies are limited to the domains in which
the sensory observations and motor actions are as many
as the available sensory and motor neurons of the neural
circuit. Implementation of larger neural circuits to build
interpretable controllers for tasks with higher degrees of
freedom will be the focus of our continued effort. Moreover,
a policy gradient algorithm may significantly enhance the
performance of our policies. We open-sourced our policies, to encourage other researchers to further explore the
attributes of neuronal circuit policies and apply them to
other control and reinforcement learning domains.
Neuronal Circuit Policies
Acknowledgements
Authors would like to thank Jean V. De Carvalho for providing constructive feedbacks on the manuscript. This work
was supported with computation resources by Microsoft
Azure via the Microsoft Azure for Research Award program.
References
Ardiel, Evan L and Rankin, Catharine H. An elegant mind:
learning and memory in caenorhabditis elegans. Learning
& memory, 17(4):191–201, 2010.
Bacon, Pierre-Luc, Harb, Jean, and Precup, Doina. The
option-critic architecture. In AAAI, pp. 1726–1734, 2017.
Bargmann, Cornelia I. Chemosensation in c. elegans. WormBook, pp. 1–29, 2006.
Berkenkamp, Felix, Turchetta, Matteo, Schoellig, Angela,
and Krause, Andreas. Safe model-based reinforcement
learning with stability guarantees. In Advances in Neural
Information Processing Systems 30, pp. 908–919. Curran
Associates, Inc., 2017.
Chalfie, M, Sulston, JE, White, JG, Southgate, E, Thomson,
JN, and Brenner, S. The neural circuit for touch sensitivity
in Caenorhabditis elegans. Journal of Neuroscience, 5
(4):956–964, 1985a.
Chalfie, Martin, Sulston, John E, White, JOHN G, Southgate, Eileen, Thomson, J Nicol, and Brenner, Sydney.
The neural circuit for touch sensitivity in caenorhabditis
elegans. Journal of Neuroscience, 5(4):956–964, 1985b.
Chen, Beth L., Hall, David H., and Chklovskii, Dmitri B.
Wiring optimization can relate neuronal structure and
function. Proceedings of the National Academy of Sciences of the United States of America, 103(12):4723–
4728, 2006.
Chen, Xi, Duan, Yan, Houthooft, Rein, Schulman, John,
Sutskever, Ilya, and Abbeel, Pieter. Infogan: Interpretable
representation learning by information maximizing generative adversarial nets. In Advances in Neural Information
Processing Systems, pp. 2172–2180, 2016.
Culotti, Joseph G and Russell, Richard L. Osmotic avoidance defective mutants of the nematode caenorhabditis
elegans. Genetics, 90(2):243–256, 1978.
de Froissard de Broissia, Arnaud and Sigaud, Olivier. Actorcritic versus direct policy search: a comparison based on
sample complexity. CoRR, abs/1606.09152, 2016.
Doya, Kenji. Reinforcement learning in continuous time
and space. Neural computation, 12(1):219–245, 2000.
Duan, Yan, Chen, Xi, Houthooft, Rein, Schulman, John, and
Abbeel, Pieter. Benchmarking deep reinforcement learning for continuous control. In International Conference
on Machine Learning, pp. 1329–1338, 2016.
Florensa, Carlos, Duan, Yan, and Abbeel, Pieter. Stochastic
neural networks for hierarchical reinforcement learning.
arXiv preprint arXiv:1704.03012, 2017.
Hasani, Ramin M, Beneder, Victoria, Fuchs, Magdalena,
Lung, David, and Grosu, Radu. Sim-ce: An advanced
simulink platform for studying the brain of caenorhabditis
elegans. arXiv preprint arXiv:1703.06270, 2017.
Heess, Nicolas, Wayne, Greg, Tassa, Yuval, Lillicrap, Timothy, Riedmiller, Martin, and Silver, David. Learning
and transfer of modulated locomotor controllers. arXiv
preprint arXiv:1610.05182, 2016.
Heidrich-Meisner, Verena and Igel, Christian. Variable
metric reinforcement learning methods applied to the
noisy mountain car problem. In European Workshop on
Reinforcement Learning, pp. 136–150. Springer, 2008.
Kato, Saul, Kaplan, Harris S., Schrödel, Tina, Skora, Susanne, Lindsay, Theodore H., Yemini, Eviatar, Lockery,
Shawn, and Zimmer, Manuel. Global brain dynamics
embed the motor command sequence of Caenorhabditis
elegans. Cell, 163:656–669, October 2015.
Koch, Christof and Segev, Koch. Methods in Neuronal
Modeling - From Ions to Networks. MIT press, second
edition, 6 1998.
Li, Zhaoyu, Li, Yidong, Yi, Yalan, Huang, Wenming, Yang,
Song, Niu, Weipin, Zhang, Li, Xu, Zijing, Qu, Anlian,
Wu, Zhengxing, and Xu, Tao. Dissecting a central flipflop circuit that integrates contradictory sensory cues in
C. elegans feeding regulation. 3:776 EP –, Apr 2012.
Moore, Andrew William. Efficient memory-based learning
for robot control. 1990.
Nichols, Annika LA, Eichler, Tomáš, Latham, Richard, and
Zimmer, Manuel. A global brain state underlies c. elegans
sleep behavior. Science, 356(6344):eaam6851, 2017.
Press, William H., Teukolsky, Saul A., Vetterling,
William T., and Flannery, Brian P. Numerical Recipes
3rd Edition: The Art of Scientific Computing. Cambridge
University Press, New York, NY, USA, 3 edition, 2007.
Rankin, Catherine H, Beck, Christine DO, and Chiba,
Catherine M. Caenorhabditis elegans: a new model system for the study of learning and memory. Behavioural
brain research, 37(1):89–92, 1990.
Neuronal Circuit Policies
Rastrigin, L. A. About convergence of random search
method in extremal control of multi-parameter systems.
Avtomat. i Telemekh., 24:1467–1473, 1963.
Russell, Stuart J. and Norvig, Peter. Artificial Intelligence:
A Modern Approach. Pearson Education, 3 edition, 2010.
Salimans, Tim, Ho, Jonathan, Chen, Xi, and Sutskever,
Ilya. Evolution strategies as a scalable alternative to
reinforcement learning. 2017.
Schulman, John, Levine, Sergey, Moritz, Philipp, Jordan,
Michael I., and Abbeel, Pieter. Trust region policy optimization. CoRR, abs/1502.05477, 2015.
Schulman, John, Wolski, Filip, Dhariwal, Prafulla, Radford,
Alec, and Klimov, Oleg. Proximal policy optimization
algorithms. arXiv preprint arXiv:1707.06347, 2017.
Singh, Satinder P and Sutton, Richard S. Reinforcement
learning with replacing eligibility traces. Recent Advances in Reinforcement Learning, pp. 123–158, 1996.
Spall, James C. Introduction to Stochastic Search and Optimization. John Wiley & Sons, Inc., New York, NY, USA,
1 edition, 2003.
Szita, I. and Lörincz, A. Learning tetris using the noisy
cross-entropy method. Neural Computation, 18(12):2936–
2941, 2006. ISSN 0899-7667.
Wen, Quan, Po, Michelle D, Hulme, Elizabeth, Chen, Sway,
Liu, Xinyu, Kwok, Sen Wai, Gershow, Marc, Leifer, Andrew M, Butler, Victoria, Fang-Yen, Christopher, et al.
Proprioceptive coupling within motor neurons drives c.
elegans forward locomotion. Neuron, 76(4):750–761,
2012.
White, J. G., Southgate, E., Thomson, J. N., and Brenner,
S. The structure of the nervous system of the nematode
Caenorhabditis elegans. Philosophical Transactions of
the Royal Society of London B: Biological Sciences, 314
(1165):1–340, 1986.
Wicks, SR and Rankin, CH. Integration of mechanosensory
stimuli in Caenorhabditis elegans. Journal of Neuroscience, 15(3):2434–2444, 1995.
Wicks, Stephen R., Roehrig, Chris J., and Rankin,
Catharine H. A dynamic network simulation of the nematode tap withdrawal circuit: Predictions concerning
synaptic function using behavioral criteria. Journal of
Neuroscience, 16(12):4017–4031, 1996.
Widrow, Bernard. Pattern recognition and adaptive control.
IEEE Transactions on Applications and Industry, 83(74):
269–277, 1964.
Williams, Ronald J. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Mach.
Learn., 8(3-4):229–256, May 1992.
000
001
002
003
004
005
006
007
008
009
010
011
012
013
014
015
016
017
018
019
020
021
022
023
024
025
026
027
028
029
030
031
032
033
034
035
036
037
038
039
040
041
042
043
044
045
046
047
048
049
050
051
052
053
054
Supplementary Materials
1. Videos of the performance of the learned neuronal circuit policies
Description
TW circuit controls an inverted pendulum
at different stages of the training process
TW circuit controls a mountain car
at different stages of the training process
TW circuit performs the parking task
Parking task with four
degrees-of-freedom
URL
https://youtu.be/iOHeQ7DhQv8
https://youtu.be/lMrP1sXp3jk
https://youtu.be/Vwydc2ez9Wc
https://youtu.be/jVQqKoHopTU
Table S1. Videos
2. Neural Circuit Implementation and Setup
In this section we describe how to integrate the neuron and synapse equations into a computational framework to build up
the TW circuit. Due to non-linearity of the sigmoid function in Eq. (7)Main-text, the neuron’s differential equation, Eq.
(1)Main-text, becomes non-linear. Unfortunately, there are no complete theory of solving problems of this class, explicitly
(Gerald, 2012). Thus, for simulating neural networks composed of such dynamic neuron models, we adopted a numerical
implicit solver.
When a network structure is dense and full of synaptic pathways, the set of ODEs (neuron potentials), defined in Eq.
(1)Main-text, becomes stiff (Press et al., 2007). Therefore, in order to overcome stability issues we used an implicit
derivation approximation method as follows (Press et al., 2007):
dv
v(t) − v(t − ∆t )
≈
dt
∆t
for some small ∆t .
(1)
In this way, we discretize the time variable and transform the set of ODEs into a set of iterative equations.
For each neuron, Eq. (1)Main-text, exposed to chemical synaptic currents in the form of Eq. (7)Main-text, and gap
junction currents in the form of Eq. (8)Main-text, if we apply approximation of the Eq. (1)supplementary and assume
vpre (t) ≈ vpre (t − ∆t ), we can show that the membrane potential of that neuron at time t, is computable by:
v(t) =
[
Cm
v(t − ∆t ) + GLeak VLeak +
∆
Xt
X
ωex,i ERev,ex +
ωinh,i ERev,inh +
i∈Ex
X
i∈GJ
[
i∈Inh
ωgj,i vpre (t − ∆t )]/
X
Cm
+ GLeak +
ωex,i +
∆t
i∈Ex
X
X
ωinh,i +
ωgj,i ]
i∈Inh
i∈GJ
(2)
Supplementary Materials
055
056
057
058
059
060
061
062
063
064
065
066
067
068
069
070
071
072
073
074
075
076
077
078
079
080
081
082
083
084
085
086
087
088
089
090
091
092
093
094
095
096
097
098
099
100
101
102
103
104
105
106
107
108
109
In Eq. (2)Supplementary, ωex,i , ωinh,i , ωgj,i , respectively stand for the overall conductance of the excitatory synapse,
inhibitory synapse and the gap junction, where ωex,i = gex,i (vpre ), ωinh,i = ginh,i (vpre ), and ωgj,i = ω̂. Variables together
with their boundaries, and constants in Eq. (2)Supplementary, are summarized in the Table 2.
Parameter
Cm
GLeak
Erev excitatory
Erev inhibitory
VLeak
µ
σ
ω
ω̂
Value
Variable
Variable
0mV
-90mV
Variable
-40mV
Variable
Variable
Variable
Lower bound
1mF
50mS
Upper bound
1F
5S
-90mV
0mV
0.05
0S
0S
0.5
3S
3S
Table S2. Parameters and their bounds of a neural circuit
Formally, Eq. (2)Supplementary, was realized in a hybrid fashion which combine both implicit and explicit Euler’s method;
The overall neuron equation, Eq. (1)Main-text is approximated by the implicit Euler’s method while the parts substituted
from Eq. (7)Main-text and Eq. (8)Main-text, were estimated by an explicit Euler’s method.
The motivation for implementing such hybrid solver, was to make the resulting algorithm of simulating the neuronal network
be separable into the following steps:
(i)
1. Compute all the incoming currents, Iin , form all synapses to the cell using the most recent values of v(t)
2. Update all v(t) by Eq. (2)Supplementary
This is significantly effective when implementing a neural network on a real-time controller.
3. Experimental Setup Parameters
3.1. Experimental setup table
Iterations
Horizon
Sample size
Filter size
Inverted Pendulum
25,000
1000
20
10
Mountaincar (OpenAI Gym)
50,000
1000
20
20
Mountaincar (rllab)
50,000
500
20
20
Cart-pole
50,000
500
20
10
Parking
20,000
320
1
1
Table S3. Experiment Parameters
With the aim of gaining a better performance, and utilizing parallel hardware resources and to increase the statistical
confidence, all experiments are performed by an ensemble of 12 agents. Due to the stochasticity of the training algorithm,
not all agents were able to solve the tasks in the given time frame. The observed success-rates are: Mountaincar (OpenAI
gym) 25%, Mountaincar (rllab) 42%, Cart-pole 42%, Inverted Pendulum 100% and Parking 100%.
3.2. Neuron’s and synapse’s parameter-boundaries in the optimization setting
Type
ω
σ
Cm
GLeak
VLeak
Lower bound
0
0.05
0.001
0.05
-90
Upper bound
3
0.5
1
5
0
Table S4. Types of parameters that are optimized and range of valid values
Supplementary Materials
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
3.3. Sensory neuron mappings
As introduced in section 2.2, input and output values are mapped to the potential of sensory respectively motor neurons by
an affine mapping. This affine mapping is defined by the minimum and maximum value of the particular input or output
value. For each of the five RL environments we set these boundary values separately, according to the following table:
Environment
Inverted pendulum
Mountaincar (OpenAI Gym)
Mountaincar (rllab)
Cart-pole
Parking
Variable
x
ϕ
a
x
ẋ
a
x
ẋ
a
ϕ
ϕ̇
a
x
y
θ
Start signal
Linear velocity
Angular velocity
Minimum
-1
-0.12
-0.3
-0.02
-0.12
-1
-0.8
-1.5
-1
-0.15
-1
-1
-1
Maximum
+1
+0.12
+0.3
+0.02
+0.12
+1
+0.8
+1.5
+1
+0.15
+1
+1
+1
+1
+1
+1
+1
+1
Table S5. Input and output boundary values used to define the affine sensory and motor mappings
4. Discussion on the influence of the filtering on the training performance
We experienced that a large filter size, i.e an objective estimate, which filters out only few outlying samples, performs better
when the reward is sparse. We observed that learning curve of such environments usually has the shape of a series of step
functions. One example is the Mountaincar (OpenAI Gym) environment, where a positive reward is only given once at the
end, when the entire task has been solved. A large filter size performs better, in such tasks, because the episodes that have
been solved by luck, are not filtered out and instead, have a large effect on the estimate. This is crucial for the training in a
sparse reward setting, since during the learning phase, we observed that the agent first is unable to solve the task, then is
able to solve only a few cases (e.g. when the car starts somewhere uphill) and so on, until the final agent solves the task even
in the difficult scenarios.
Furthermore, we observed that a small filter size, i.e. an objective-estimate which is computed only from the lowest samples,
performs slightly better in tasks with gradually increasing reward. Learning curve of such environments are usually shaped
smoother. An example of such task is the inverted pendulum (Roboschool), where a positive reward is given all the time
when the pendulum is still facing upwards. Our filtering strategy performs slightly better here, because, as originally
intended, the estimate is not influenced by outlying high samples. In the inverted pendulum task, such samples occur when
the pendulum start in an almost straight pose and a high return can be collected without any specific action.
5. TW circuit can realize more degrees of freedom
In our first parking experiment, the TW circuit is able to make a turn left and move the rover forward with only one command
neuron being active. This means that the circuit is able to solve the task with having binary activation states (active, not
active) of the command neuron. To test the flexibility of the TW circuit and underlying neuron model, we set up a second
parking experiment. In this experimental setup, we connected the command neuron AVA to two motor neurons responsible
for turning right and moving backwards, and AVB to two motor neurons responsible for turning left and moving forward.
In this setup, the controller is principally able to move the robot to 4 different directions: Forward, Backward, turn left
Supplementary Materials
and turn right. Furthermore, the TW circuit is not able to move the rover forward and turn right with only command
neuron being active. If the TW circuit tends to make a right turn and move the rover forward at the same time, (which is
necessary to solve this task), the circuit must be able to do this via a synchronized cooperation of the two command neurons.
With this configuration, our goal was to test whether the TW circuit can express multiple output primitives with only two
command neurons, by operating them in more than two potential states. We conclude that the training algorithm was able to
parametrize the TW circuit, such that the agent can keep the trajectory checkpoints, correctly. A video on this scenario can
be viewed at https://youtu.be/jVQqKoHopTU.
6. Neuron traces of figure 6
Time (s)
20
25
PVD (mV)
15
Time (s)
PLM (mV)
10
AVA (mV)
5
AVD (mV)
0
PVC (mV)
50
AVB (mV)
25
25
50
75
AVM (mV)
25
50
25
50
75
25
50
75
ALM (mV)
70
80
25
50
75
25
50
75
DVA (mV)
50
0
50
25
50
50
75
LFT (mV)
10 15 20 25 30 35 40 45
0
25
50
75
80
50
60
70
25
50
75
RGT (mV)
5
25
50
60
25
50
25
50
FWD (mV)
0
PVD (mV)
40
PLM (mV)
20
AVA (mV)
25
50
75
AVD (mV)
40
50
PVC (mV)
70
60
AVB (mV)
60
40
60
AVM (mV)
AVM (mV)
40
60
ALM (mV)
40
60
DVA (mV)
60
FWD (mV)
AVB (mV)
40
40
60
ALM (mV)
50
60
70
40
DVA (mV)
60
50
0
50
FWD (mV)
40
C
REV (mV)
PVD (mV)
PLM (mV)
50
60
70
PVC (mV)
60
70
AVA (mV)
B
AVD (mV)
A
REV (mV)
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
25
50
75
0
5
10
Time (s)
15
Figure S1. Neuron activity represented in figure 6. A) Inverted pendulum, B) Mountain car C) Parking
References
Gerald, Teschl. Ordinary Differential Equations and Dynamical Systems, volume 140 of Graduate Studies in Mathematics.
American Mathematical Society, 2012.
Press, William H., Teukolsky, Saul A., Vetterling, William T., and Flannery, Brian P. Numerical Recipes 3rd Edition: The
Art of Scientific Computing. Cambridge University Press, New York, NY, USA, 3 edition, 2007.
| 9cs.NE
|
Sensor Array Design Through Submodular
Optimization
Gal Shulkind, Student Member, IEEE, Stefanie Jegelka, Gregory W. Wornell, Fellow, IEEE
arXiv:1705.06616v2 [cs.IT] 28 Dec 2017
Abstract
We consider the problem of far-field sensing by means of a sensor array. Traditional array geometry design techniques are
agnostic to prior information about the far-field scene. However, in many applications such priors are available and may be utilized
to design more efficient array topologies.
We formulate the problem of array geometry design with scene prior as one of finding a sampling configuration that enables
efficient inference, which turns out to be a combinatorial optimization problem. While generic combinatorial optimization problems
are NP-hard and resist efficient solvers, we show how for array design problems the theory of submodular optimization may be
utilized to obtain efficient algorithms that are guaranteed to achieve solutions within a constant approximation factor from the
optimum. We leverage the connection between array design problems and submodular optimization and port several results of
interest. We demonstrate efficient methods for designing arrays with constraints on the sensing aperture, as well as arrays respecting
combinatorial placement constraints.
This novel connection between array design and submodularity suggests the possibility for utilizing other insights and
techniques from the growing body of literature on submodular optimization in the field of array design.
Index Terms
Array design, Submodular optimization, Far field sensing
I. I NTRODUCTION
Sensor arrays for spatial sensing are widely deployed in a wide range of applications including radar, sonar, medical imaging
and radio astronomy and there is a vast literature on the topics of array design and array processing from the last century [1],
[2].
A major goal in designing arrays is efficiently meeting sensing specifications with a limited budget of sensing elements,
which is often a main determinant of system cost in terms of dimensions, weight, complexity and manufacturing cost. However,
traditionally most studies assume a uniform and linear design, often a truncated half-wavelength array, or restrictions thereof.
Indeed, the problem of designing non-uniform arrays hints at combinatorial optimization and is computationally intractable as
we discuss later.
In beamforming arrays attaining a desired resolution level is often a primary concern, achieved by means of manipulating
beam pattern parameters such as lobe widths and positions. The problem of designing efficient array geometries that facilitate
desired beam patterns has been widely studied in the past. Techniques involving array thinning start with a dense uniform
geometry, removing elements while maintaining performance within specified bounds [3]. Other approaches consider various
methods such as swarm optimization [4], dynamic programming [5], genetic algorithms [6], inversion [7] and Bayesian
compressive sampling [8].
In direction-of-arrival estimation applications other specialized techniques have emerged for finding efficient array designs,
such as optimizing the corresponding Cramer-Rao error bound [9], or designing according to the nested array methodology
for increasing the available number of degrees of freedom [10], and theoretical results for estimation performance in these
settings have been developed [24], [25], [26]. Alternatively, other studies have addressed array design from the point of view
of performing efficient compressive sampling of a Wide Sense Stationary (WSS) process, e.g. for a setting with a parametrized
covariance matrix [11], for wide-band power spectrum estimation [12] and evaluating performance in the presence of modeling
errors [13]. However, these samplers are optimized for reconstructing the second order statistics of the scene and not the actual
scene realization.
In virtually all design methodologies some assumptions are made on the scene of interest. These represent beliefs, constraints
or knowledge that hold over the unobserved scene. A favorable design is one meeting requirements while taking into account
these assumptions. For example, in direction of arrival applications we may assume some limit on the number of point targets
present in the scene [9]. In beamforming we assume some separation level between objects of interest that implies a certain
resolution level is required [1], or some scene sparsity structure [14], [15].
In this paper we study the problem of inference on a scene of interest through measurements collected by an array of sensors.
The approach we take for designing array geometries is novel in that we consider settings where some Bayesian prior on the
This work was supported, in part, by NSF CAREER Award 1553284.
This work was supported, in part, by DARPA under Contract No. HR0011-16-C-0030, and by NSF under Grant No. CCF-1319828.
The authors are with the Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA 02139 USA
(e-mail: {shulkind,stefje,gww}@mit.edu).
1
scene is available at the time of design and propose exploiting this knowledge by adapting the array geometry accordingly to
achieve efficient inference with a limited budget for sensors.
Prior knowledge is in fact not a rarity at all. Frequently, the same device is used to sense multiple similar scenes, where past
examples are indicative. A medical imaging device, for instance, is typically used to image the same organ across different
patients. In other cases, we may have prior knowledge in the form of scene properties such as smoothness or adherence to
spatial constraints. We incorporate such knowledge as a prior in a Bayesian model, such that sensing the scene is equivalent
to performing inference in this model, and the problem of array geometry design asks to select a geometry that optimizes the
quality of inference. This naturally turns out to be a combinatorial optimization problem, which can be extremely hard to solve
even approximately without exploiting additional structure.
We show how to choose a cost function for the array geometry design problem that holds the property of submodularity
[16]. A submodular set function is one that exhibits diminishing marginal gains, i.e. adding additional elements results in
diminishing benefit. Recently, there has been significant progress on the theory of optimizing submodular functions [17], [18],
[19], [20]. In particular, these results state that, while NP-hard, submodular maximization admits variants of greedy algorithms
that are guaranteed to achieve near-optimal solutions, i.e. within a constant factor. Submodularity has been used in connection
with sensor placement problems, for example in [21], [22], [23], however these works are not tailored to the far field scenes
and models that we focus on here.
Our connections and formulations open avenues for leveraging those results for efficient array design with strong guarantees.
We demonstrate this by showcasing the design of array geometries in settings with arbitrary apertures and settings where sensor
placement respects matroid combinatorial constraints by porting results from the growing body of literature on submodular
optimization. We demonstrate that together with our new adaptive formulations, the exploitation of prior knowledge leads to
higher quality inference at lower cost in terms of the number of sensing elements.
The paper is structured as follows: In Section II, we review far-field sensing and introduce a statistical framework for
modeling a-priori scene distributions. In Section III, we define a cost-function to quantify the inference loss associated with
array configurations and derive a corresponding array design optimization problem. We show that this optimization problem
belongs to the class of submodular optimization problems and adapt established algorithms and guarantees to our setting. In
Section IV, we review the theory of matroids and showcase its use in designing array geometries with combinatorial placement
constraints. Finally, in Section V we summarize numerical experiments exemplifying and validating our theory. The appendix
provides proofs for some of our claims.
II. P ROBLEM FORMULATION
In this section we formulate the array design problem. We focus on far-field sensing applications, although, as will become
apparent, the same techniques could also be generalized to other settings where the measurement process is linear. For simplicity
we consider a one dimensional setting as the extension to multiple dimensions is straightforward. We begin with a review of
the far-field sensing model to establish notation, and then pose the sensing problem as a Bayesian inference problem.
A. Far-Field Sensing
The far-field sensing setup is depicted in Fig. 1. At a distance from an observation axis x we have a scene of interest,
characterized by an illumination function β̃(θ) that depends on the angle θ∈[− π2 , + π2 ] between the direction of observation
and the axis of measurement broadside. We are to place N sensors along the observation axis at positions S ≡ {x1 , . . . , xN }.
Fig. 1. Far-field sensing. Antenna elements are depicted as blue dots.
The passive sensors pick up the radiation emitted from the scene, which is assumed to be narrow-band around a fixed
wavelength λ. Recall that in an ideal noiseless setting, r(x), the reading collected at position x, is given according to [1]:
1
π
Z+ 2
r(x)=
β̃(θ)ej
2π
λ x sin(θ)
Z+ 2
cos(θ)dθ=
−π
2
β(ψ)ej
4π
λ xψ
dψ,
(1)
− 21
where we have introduced ψ ≡ sin2 θ such that ψ ∈ [− 12 , + 21 ] and β(ψ) ≡ β̃(sin−1 (2ψ)).
Eq. (1) bears close resemblance to the definition of the Fourier transform. Namely, defining f (t) to be the Fourier transform
pair of β(ψ)
Z
Z
F
j2πtψ
f (t) = β(ψ)e
dψ β(ψ)= f (t)e−j2πtψ dt
(2)
F −1
2
2
we immediately identify, comparing (1)
2 and (2) that r(x) = f ( λ x). That is, collecting measurements at x ∈ 1{xn }1 is equivalent
to sampling the function f (t) at t ∈ λ xn . Notice that since the support of β(ψ) is restricted to ψ ∈ [− 2 , + 2 ] we observe
that f (t) is band-limited with bandwidth 1. Thus, as a consequence of the Whittaker-Kotelnikov-Shannon sampling theorem
we have that perfect reconstruction of f (t), and subsequently the scene β(ψ), is possible from an infinite set of samples taken
at t ∈ {. . . , −1, 0, +1, . . .} which are exactly the samples collected by an infinite λ2 -spaced array.
Finally, we consider the effect of noise by introducing the vector of N noisy measurements f˜ = [f˜1 , . . . , f˜N ]> modelled
according to:
2
˜
xn + wn
n = 1, . . . , N
(3)
fn ≡ r(xn ) + wn = f
λ
where wn are additive noise components. Stacked into a vector w = [w1 , . . . , wN ]> , we assume throughout that the noise is
a complex, circular, Gaussian random vector w ∼ CN (0, Σww ) [27]. We assume the noise to be independent across different
2
sensors, i.e., Σww = σw
IN , where IN is an N ×N identity matrix such that the noise is i.i.d. across different sensors.
B. Bayesian Formulation
The sensing problem we consider in this paper entails the estimation of the illumination function β(ψ) or equivalently its
inverse Fourier transform f (t) from the set of noisy measurements f˜. Even in the noiseless setting this problem is gravely
ill-posed as infinitely many wildly varying scenes map to any given finite set of observations1 .
To cope with this ill-posedness, some prior belief or knowledge pertaining to β(ψ) (or equivalently f (t)) must hence be
incorporated into the model, and this could be achieved in several ways. Wingham [29] proposed selecting one specific f (t) of
the multiple such functions consistent with the samples, namely the minimum norm solution. Alternatively, some constraints
or other preferences may be imposed on the solution by penalizing the inversion. For example one may require the solution
to satisfy some constraints, e.g. lie in some pre-specified sub-space of the function space, or regularize the inversion, e.g. to
promote smoothness or low total variation [30].
In this study we take a Bayesian approach and impose a prior distribution on the scene β(ψ). Subsequently, sensing is
equivalent to performing inference in this model. The prior may be assigned based on past observations over the distribution
of scenes or based on a-priori knowledge of scene properties.
We conceptually think of our scene as a natural image, and as such, our Bayesian formulation entails posing a prior that is
consistent with its expected characteristics, tailored to the specific application that is of interest to the user. Utilizing Gaussian
distributions as priors for natural scenes and images has been successfully applied in the past [31], [32], [33] and our work
uses similar notions, with the Gaussian prior specifically expressed in the frequency domain (e.g., as in [9]). We note that this
Gaussian description naturally captures the statistics of spread-out scenes and is not immediately suitable for serving as a prior
for point-source target models. The formulations and methods that we develop in subsequent sections can be adapted for use
with other scene priors in lieu of the Gaussian one we focus on for this study. In particular, our formulations may account for
settings with point-source scenes by posing a point-process prior (e.g. a determinantal point process, DPP [34]) that is better
suited for capturing priors of point-targets in specific environments. With this alternative prior we could follow similar steps to
develop efficient array designs for those applications. Studying such alternative priors and resulting formulations is a possible
interesting extension to our work.
In what follows, to be concrete we detail one specific choice for the form of the scene prior, expressed as a Gaussian
distribution in the frequency domain:
1) Discrete Representation: Assigning a prior on a continuous function β(ψ) may seem like a daunting task. However, in
many scenarios this can be considerably simplified if the scene may be efficiently expanded in a countable basis of functions
such that the prior may be imposed in thediscrete domain of expansion coefficients. In what follows, we consider expanding
β(ψ) by means of Fourier basis functions ej2πmψ |m = . . . , −1, 0, +1, . . . in ψ ∈ [− 21 , + 12 ]. Alternatively, other expansions
may be used with similar results. We write:
X
1 1
β(ψ) =
βm ej2πmψ ,
ψ ∈ [− , + ]
(4)
2 2
m
1
Z+ 2
βm ≡
β(ψ)e−j2πmψ dψ
(5)
− 12
where {βm } are the Fourier expansion coefficients. The usual Parseval relation holds:
Z
X
|β(ψ)|2 dψ =
|βm |2
m
1 The
mapping between a band-limited function f (t) and a finite set of its (generally non-uniform) samples {f (tn )} is not bijective [28].
(6)
3
In lieu of the prior on β(ψ) we will impose a prior distribution over {βm }. As will become evident such a description is
especially suited for applications involving real world smooth scenes β(ψ) as suggested by the following properties of the
Fourier series expansion [35]:
|m|→∞
(a) (Riemann-Lebesgue lemma) Let β(ψ) be any integrable function. Then |βm | −→ 0.
α
(b) Let β(ψ) ∈ C r where C r is the space of r-times continuously differentiable functions over some domain. Then |βm | ≤ |m|
r
∂r
with α = supψ | ∂ψr β(ψ)| [35].
Thus, for any nicely behaved scene β(ψ) the high frequency Fourier expansion coefficients diminish to zero with an
asymptotically polynomial rate determined by the level of smoothness.
2) Prior: In the sequel we use Gaussian priors on the coefficients {βm }:
2
βm ∼ CN (0, σm
)
(7)
2
where βm are independent, complex, circular and Gaussian, and σm
are the corresponding variances, which are assumed known
in advance. Using (6) we define the expected scene power:
Z
X
2
P ≡ E |β(ψ)|2 dψ =
σm
(8)
m
2
The σm
can be set following some initial measurements of sample functions β(ψ) or taking into account prior knowledge.
For example if we have a-priori knowledge that β(ψ) ∈ C r we may use
1
m=0
2
(9)
σm
=
1
m
6= 0
2r
m
which is a very simple distribution respecting the polynomial variance decay. For the rest of the paper we adopt the prior in
(9) and take r = 1 to promote continuously differentiable functions. Also notice that the resulting prior distribution on β(ψ) is
2
Wide Sense Stationary (WSS), with the coefficients σm
determining the shape of the power spectrum, based on the available
prior information we have on the scene.
C. Observation Model
With the prior stated in the discrete {βm } domain as described above, our next goal is to circumvent β(ψ) by directly
stating the problem in terms of the measurements f˜n and the coefficients βm , replacing the continuous representation with
manageable discrete counterparts. Substituting the sum (4) into Equation (1) we have:
1
r(xn ) =
Z+ 2 X
− 12
m
βm ej2πmψ ej
4π
λ xn ψ
dψ =
X
Knm βm
(10)
m
where
1
Z+ 2
Knm ≡
2
2
ej2π( λ xn +m)ψ dψ = sinc(m+ xn )
λ
(11)
− 21
with sinc(x) ≡
sin(πx)
πx .
Plugging this into (3) we finally have the full observation model
X
n = 0, . . . , N − 1
f˜n =
Knm βm + wn
(12)
m
and the sensing problem amounts to estimating the coefficients {βm } given the observation vector f˜. As we have assumed a
Gaussian distribution for the noise vector w as well as for the prior over {βm } the posterior distribution p({βm } |f˜) turns out
to be Gaussian with a convenient analytic expression, as we detail in Section III-D.
III. A RRAY D ESIGN W ITH AN A PERTURE C ONSTRAINT
In the previous section we formulated the problem of far-field sensing in a Bayesian setting with a prior on the distribution
of the underlying scene. In this section we design corresponding array geometries to facilitate efficient sensing that exploit the
prior knowledge and model. Initially we consider simple constraints on the total number of sensors we can place and on the
aperture where these can be situated. Specifically, we assume that up to N sensors are free to be placed over some aperture
A on the real line, e.g. for simplicity we can consider A ≡ {x| − a ≤ x ≤ a} , a ∈ R+ . In Section IV we consider more
sophisticated combinatorial placement constraints and show that the same formulations we develop here may be adapted to
those more challenging use cases.
4
A. Formulating a Cost Function
In order to make the problem of optimal array design well-posed in our Bayesian setting we need to specify a performance
measure that will be used as a cost function to compare different designs and choose the best one. Revisiting our observation
model (12) we notice that the array design determines the coefficients Kmn through the set of sensor positions S, as defined in
Section II-A. With any given design the sensing problem amounts to inferring the posterior p({βm } |f˜). A natural performance
measure in this setting quantifies the quality of inference, i.e., the information gained by performing the sensing experiments
which results in updating our beliefs about the coefficients {βm } from the prior p({βm }) to the posterior p({βm } |f˜). This
problem has been extensively studied in the context of statistical inference and experimental design [36], [37]. If we are
interested in making general inference to learn the state of the world (represented by the posterior distribution of the Fourier
coefficients) then the natural performance measure, which is referred to as Bayes D-optimality is the mutual information
between the expansion coefficients and the collected measurements [36], [37]:
I(f˜S ; {βm }) = H({βm }) − H({βm } |f˜S )
(13)
where I(·; ·) is the mutual information and H(·) the Shannon entropy. The subscript S explicitly emphasizes the dependence
of the measurements on the set of sensor positions S. Notice that this cost function measures the reduction in uncertainty of
the scene expansion coefficients before and after measurements are made. The larger the mutual information the more we trust
the values of the coefficients {βm } |f˜S .
With the cost function in place the array design problem becomes
S ? = argmax I(f˜S ; {βm }),
(14)
S⊆A,|S|≤N
which is an NP-hard combinatorial optimization problem. However, we will show later that we can obtain a constant-factor
approximation for (14) using efficient computational techniques.
B. Finite Dimensional Approximation
Solving (14) under our model (12) involves manipulations of the infinite sequence {βm } which may not be amenable to
computer representation. To make our formulation tractable we approximate the infinite sequence {βm } with a finite truncated
set of coefficients {βm |m ∈ M}, where M is some finite set. Stacked in vector form β, we consider the simplified finite
dimensional approximation of (12):
fˆS = KS β + w
(15)
where KS is an N ×|M| matrix formed by restricting Knm on m ∈ M and the dependency on the sampling set S again
made explicit. Notice that hat notation is replacing the previous tilde.
We show that the approximate finite-dimensional model (15) is a good proxy for the original infinite-dimensional model
(12) for a suitably selected M. Indeed, if M is chosen to only exclude those βm coefficients that in expectation contribute a
marginally small part of the energy of the full infinite sequence {βm } then the mutual information derived from the approximate
model (15) will closely track that of the infinite dimensional model in (12). More precisely we have the following result:
P 2
3
2
2
Lemma III.1. Let the prior on {βm } be i.i.d. according to βm ∼ CN (0, σm
), M fixed, and ≡
σm satisfies < σw
N−2 .
m∈M
/
We have:
3
−N log(1+
3
N 2
N 2
)≤I(f˜S ; {βm })−I(fˆS ; β)≤−N log(1− 2 )
2
σw
σw
Proof. See Appendix A.
3
→0
2
ˆ
˜
By virtue of the last lemma and N log(1± N
2 ) −→ 0 we have that I(fS ; β) is an arbitrarily accurate proxy for I(fS ; {βm })
σw
for small enough, such that in lieu of problem (14) we now continue with the simplified finite dimensional approximation:
S ? = argmax I(fˆS ; β)
S⊆A,|S|≤N
and the results will be accurate to within the approximation bounds from Lemma III.1.
(16)
5
C. Grid Discretization
Next we turn to discretizing the aperture A to cast the array design problem in the form of a generic finite selection problem.
We thus restrict the choice of sampling positions to the finite set
o
n
(17)
V ≡ x1 , . . . , x|V| ⊂ A
For the sequel we take V to be a uniform δ-spaced grid of positions in A. We adapt the array design problem (16) accordingly
as:
Sd? = argmax I(fˆS ; β)
(18)
S⊆V,|S|≤N
with the subscript d implying discretization. The next result can be used to quantify the level of discretization δ necessary to
guarantee performance close to optimum within some specified error bound:
Lemma III.2. With V a uniform grid of sampling positions with distance δ between adjacent positions we have:
3
I(fˆSd? ; β) ≤ I(fˆS ? ; β) ≤ I(fˆSd? ; β)+N log(1+
4δP (1+δ)N 2
)
2
λσw
Proof. See Appendix B.
With this last lemma in place the array design problem (16) may be further approximated in the more convenient finite
combinatorial problem form of (18) with guarantees on the accuracy of the resulting designs. In the sequel we assume that
δ is chosen such as to meet desired accuracy levels, as prescribed in Lemma III.2, and work with the simplified formulation
(18).
D. Sensing
With the observation model of (15), coupled with a Gaussian distribution for the noise vector w and a Gaussian prior for the
coefficients vector β, calculation of the posterior distribution β|fˆS is particularly simple and can be performed analytically.
Concretely, we have as a result of all random variables being Gaussian β|fˆS ∼ CN (µ̂, Σ̂) and the parameters are given
according to the conventional Gaussian conditional parameters:
fˆ
µ̂ = Σβ fˆΣ−1
fˆfˆ S
Σ†
Σ̂ = Σββ − Σβ fˆΣ−1
fˆfˆ β fˆ
where Σβ fˆ = Σββ KS† ,
(19)
2
Σfˆfˆ = KS Σββ KS† + Σww and Σββ = diag[σ12 , . . . , σM
].
E. Submodular Optimization
The next step is to prescribe an efficient algorithm for the solution of (18). As we will show shortly (18) is an instance of
a known NP-hard problem such that it is widely believed that no efficient algorithm for its solution exists. However, due to
the structure of the cost function an efficient approximation algorithm is known to exist with strong theoretical guarantees. In
this subsection we survey the relevant results and adapt them to our needs.
1) Submodularity: We begin by invoking the submodularity property of set functions [16]:
Definition III.1. Let G : 2V → R be a set function.
(a) G is submodular if it satisfies the property of decreasing marginals: ∀S, T ⊆ V such that S ⊆ T and x ∈ V\T it holds
that G(S ∪ {x}) − G(S) ≥ G(T ∪ {x}) − G(T ).
(b) G is monotonic (increasing) if ∀S, T ⊆ V s.t. S ⊆ T we have G(S) ≤ G(T ).
As it turns out, our cost function is monotonic and submodular as the next result shows (this is almost identical to corollary
4 in [38]. We reproduce it here with adaptations to our setting for completeness):
Lemma III.3. Let V be defined as before, and define the set function G : 2V → R according to G(S) = I(fˆS ; β). Then G is
submodular and monotonic (increasing).
Proof. Expanding the mutual information according to I(x; y) = H(x)−H(x|y) we have:
G(S∪ {x})−G(S) = H(fˆS∪{x} )−H(fˆS )
−[H(fˆS∪{x} |β)−H(fˆS |β)] = H(fˆ{x} |fˆS )−H(fˆ{x} |β)
(20)
6
where in the last equality we used the conditional independence of the components of fˆS∪{x} given β. Substituting T for S
we immediately get:
[G(S∪ {x})−G(S)]−[G(T ∪ {x})−G(T )]
= H(fˆ{x} |fˆS )−H(fˆ{x} |fˆT )
(21)
Using S ⊆ T we have H(fˆ{x} |fˆS ) ≥ H(fˆ{x} |fˆT ) such that G(S∪ {x})−G(S) ≥ G(T ∪ {x})−G(T ) and G is submodular.
To prove monotonicity it is enough to show G(S∪ {x})−G(S) ≥ 0. This time expand the mutual information according to
I(x; y) = H(y)−H(y|x):
G(S∪ {x})−G(S) = H(β|fˆS )−H(β|fˆS∪{x} )
(22)
Conditioning can never increase entropy so H(β|fˆS )≥H(β|fˆS∪{x} ) and the result follows.
2) Efficient Solvers: Our optimization problem (18) is the maximization of a monotonic submodular function. A greedy
algorithm, shown as Algorithm 1, solves this problem to within the best possible approximation factor, as the following
fundamental theorem states.
Algorithm 1 Greedy Submodular Maximization
1: function G REEDY M AX (G(·), V, N )
2:
S←∅
3:
for i = 1 to N do
4:
x? = argmaxx∈V\S G(S ∪ {x})
5:
S ← S ∪ {x? }
6:
Return S
The following theorem guarantees that the greedy algorithm achieves approximately optimal performance:
Theorem III.1. (Nemhauser [16]) Let G be a monotonic, submodular set function and S ? = argmax G(S).
S⊆V,|S|≤N
Let S gr be the set retrieved by the greedy maximization Algorithm 1. We have the following guarantee for the performance of
the greedy algorithm:
1
1
G(S gr ) ≥ (1 − (1 − )N )G(S ? ) ≥ (1 − )G(S ? )
N
e
Moreover, no polynomial time algorithm can provide a better approximation guarantee unless P=NP [39].
Combining the guarantees of Theorem III.1, Lemma III.1 and III.2 we derive an approximation bound on the original
problem (14):
Corollary III.1.
"
#
3
2
2
1
λσ
+4δP
(1+δ)N
w
(1− ) I(f˜S ? ; {βm })−N log
≤I(fˆS gr ; β)
2 −λN 23
e
λσw
(23)
We hence apply Algorithm 1 to solve our optimization problem (18). The algorithm runs in time O(|V|N ), linear in the
size of the set V and the number of selected elements N [40] such that it is easily implementable for problems of large size.
However, as it turns out even more efficient variants of the algorithm have been introduced and studied. One such variant
commonly referred to as the ’lazy greedy’ algorithm was studied in [40] and it was shown to offer substantial running-time
improvements in practice (with an unlikely worst-case theoretical performance upper bounded by that of the conventional
greedy algorithm). Our numerical experiments described in Section V implement this more efficient variant to reduce running
time.
3) Improved Approximation Bounds: While Theorem III.1 guarantees an approximation bound of (1 − 1e ) ≈ 63% for the
efficient greedy algorithm this guarantee is not tight. It is possible to derive a tighter data-dependent online bound on the gap
between the cost of the greedy solution G(S gr ) and that of the optimal solution G(S ? ) [41]. Specifically, we have:
X
G(S gr )≥G(S ? )− max
(G(S gr ∪ {e})−G(S gr ))
(24)
B:|B|≤N
e∈B
which takes O(|V| log |V|) evaluations of G(S) to compute and sort. We use (24) to improve the distance from optimality
bound in some of our numerical solutions in Section V.
7
F. Design Example: A Simple Ideal Setting
In the previous subsections we formulated the array design problem in a setting with constraints on the aperture A and
the number of sensors N and showed how a greedy algorithm (Algorithm 1) is guaranteed to efficiently find an approximate
solution.
Here, we study a particular instance of that problem, where the Signal to Noise Ratio (SNR) is high, and the aperture is
effectively unconstrained (the N sensors may be placed anywhere on the real line). Under these conditions a truncated λ2 -spaced
array is traditionally considered the design of choice in the conventional non-Bayesian setting (this is a truncated version of
the infinite λ2 -spaced design mentioned in Section II-A). We show next that the truncated λ2 -spaced design naturally emerges as
the approximately optimal solution as retrieved by our schemes in Bayesian settings where the a-priori β distribution satisfies
some conditions. Specifically, we have the following result:
Theorem III.2. Consider the high SNR regime σP2 → ∞ and assume the prior from (7) takes a symmetric, monotonically
w
2
2
2
2
decreasing form, i.e. σm
= σ−m
and σm
≥ σm
whenever 0 ≤ m1 < m2 . In addition, take V as an arbitrarily dense set of
1
2
sampling points on R, and M = −M, . . . , M with M → ∞.
We then have that a greedy solver on (18) will return a length N , λ2 -spaced truncated uniform array centered around x = 0.
Proof. See Appendix C.
The last theorem studies one class of simple idealized problems where the greedy solution is reminiscent of generic nonBayesian array designs. However, notice that our formalism is also useful in more challenging design problems such as when
the aperture A takes on arbitrary forms, and the effects of noise and application-tailored priors are considered. Furthermore, in
Section IV we demonstrate that additional combinatorial constraints may be naturally incorporated into our formulation such
that even more challenging problems become manageable.
IV. A RRAY D ESIGN WITH M ATROID C ONSTRAINTS
In Section III we formalized the array design problem in a setting where we imposed constraints on the aperture and the
number of sensors. Specifically, the constraints were S ⊆ V, |S| ≤ N . In many practical scenarios these may be too simplistic
to accurately represent real world design constraints. For example in applications where sensors are heavy and mounted on
support beams we may want to restrict the number of sensors in specific sections of the aperture.
In this section we briefly review key elements from matroid theory which is a branch in combinatorics [42] and survey
results from submodular optimization with matroid constraints guaranteeing the existence of efficient approximate solvers for
this class of problems. We continue to show that these mathematical structures may be utilized to impose constraints of interest
in array design enriching the set of problems our Bayesian formulation can describe and solve.
A. Submodular Optimization with Matroid Constraints
We begin by defining matroids and their corresponding independent sets [16]:
Definition IV.1. A finite matroid M is a pair (V, I) where V is a ground set and I is a collection of subsets of V (the
independent sets) that satisfies the following properties:
1) The empty set is independent: ∅ ∈ I
2) A subset of an independent set is independent:
X ⊂ Y, Y ∈ I ⇒ X ∈ I
3) If X is an independent set and Y is a larger indepedent set, X can be augmented to a larger independent set by adding
an element from Y \X:
X, Y ∈I, |X|<|Y | ⇒ ∃e ∈ Y \X s.t. X∪ {e} ∈ I
A matroid structure may be used to classify subsets of a ground set V into permissible subsets which belong to I and
non permissible subsets which do not belong to I. In the next subsection we show that using this formalism we can express
interesting array design constraints.
From the theory of submodular optimization we have the following results for submodular optimization with matroid
constraints [43], [44]. Let M = (V, I) be a matroid and G(S) a monotonic, submodular set function. There exists an efficient
approximate solver for the problem argmaxS∈I G(S). Specifically, a greedy solver (maximizing the immediate marginal benefit
at each step) taking the form
(
)
S gr ←S gr ∪
argmax
[G(S gr ∪ {e})−G(S gr )]
(25)
e:e∈S
/ gr ,S gr ∪{e}∈I
and stopping when no more elements e can be added is guaranteed to achieve a half-approximation bound:
1
G(S gr ) ≥ max G(S).
2 S∈I
1
The constant factor may be tightened to 1 − e by utilizing specialized randomized algorithms [44].
(26)
8
B. Combinatorial Constraints in Array Design
Here we invoke one specific well known matroid structure and show its application in expressing useful array design
constraints. Let V be a ground
S set of gridTpoints where sensors are allowed to be placed as before. Let V1 , . . . , VK be a
partition of the set V, i.e.
Vk = V, Vi Vj = ∅, ∀i 6= j, and let N, n1 , . . . , nK be a set of integers.
k
We define the (cardinality constrained) partition matroid [44] M = (V, I) with the following definition for the collection
of independent sets: a subset S⊆V is an independent subset S∈I if it holds |S∩V| ≤ N, |S∩Vj | ≤ nj , ∀j.
In the context of array design the partition matroid may be useful in expressing practical constraints over sensor placement
configurations. For example if the subsets Vi represent closed line sections, e.g. a physical partitioning of the aperture into
zones, and I represents the collection of all permissible designs then the structure of the matroid limits the number of sensors
that may be placed in the ith zone to ni which may be an important engineering constraint coupled with some specific
application. We solve:
S ? = argmaxS∈I I(fˆS ; β)
(27)
Applying the results from the previous subsection we immediately have an efficient approximate solver for the array design
problem coupled with a partition matroid constraint. In Section V we detail such a design for one numerical example.
V. N UMERICAL E XPERIMENTS
In this section we perform numerical experiments validating our theoretical results and exemplifying them. We showcase
an array design with cardinality and aperture constraints as prescribed in Section III and design arrays adhering to matroid
constraints as prescribed in Section IV.
Our initial setting is as follow. We fix λ = 1 throughout as the wavelength only serves to scale the x axis. The aperture is set
as A = {x| − 3.5 ≤ x ≤ 3.5} and the selection set V is chosen as a uniform grid of 113 positions from A spaced δ = 0.0625
apart. We set out to design an array consisting of N = 11 sensor locations. The prior for {βm } is set as per (9) with r = 1
and normalized to sum to P = 1. For the simulations we consider the truncated vector β formed when restricting the set of m
coefficients to 901 consecutive elements centered around the origin, i.e., we set M = {−450, . . . , +450}. For the preliminary
design we implement the lazy greedy algorithm of Section III and plot the results in the left column of Fig. 2 as a function
of the SNR which we define here as NPσ2 . Blue markers denote the full selection set V and red markers delineate the active
w
S selected by the algorithm.
In the high SNR regime (SNR = 10dB or higher values) the resulting design is a truncated λ2 uniform array as predicted
according to Theorem III.2. As the SNR decreases the reliability of the measurements deteriorates and the algorithm prefers
locating samplers right next to each other on expense of widening the array as this serves to average out the noise.
The performance in terms of mutual information I(fˆS ; β) for the selected locations S appears in the title of the plots (in
natural units). Notice for example that for the 5dB SNR design the achieved mutual information is 12.54. Using Theorem
III.1 we have that the optimal design cannot achieve mutual information better than 1−1 1 12.54 = 19.83. This bound can be
e
improved using the improved bounding method briefly described following Theorem III.1 to show that the optimal performance
is not greater than 17.45.
The truncation level dictated by our choice of M translates to ∼
= 1e−4 and the truncation bounds from Lemma III.1 read
(for the lower, extreme SNR case): −0.45≤I(f˜S ; {βm })−I(fˆS ; β)≤0.47. We find empirically that these bounds are extremely
loose and M can be shrunk considerably without substantially compromising accuracy. To achieve a similar upper bound on
I(fˆS ? ; β) − I(fˆSd? ; β) as per Lemma III.2 a discretization level of δ ∼
= 2e−4 is needed. However, we empirically find that
our choice of δ = 0.0625 is accurate enough as further refining the grid does not significantly change the design. Our lemmas
prove to be extremely pessimistic as is expected given that the proofs take into account worst-case scenarios.
The array geometries above, derived according to the formulations of Section III, are designed to optimize the quality
of inference between the measurements and the scene expansion coefficients β. Many sensing applications of interest
specifically involve imaging the scene, that is reconstructing β(ψ) from the measurements. Our next experiment was designed
to empirically evaluate the Mean Square Error (MSE) performance in scene reconstruction from measurements collected using
the prescribed designs. First, we designed five array geometries as described above, optimized for several target SNR levels
{30dB, 12dB, 10dB, 5dB, 0dB}. We set up a Monte-Carlo experiment where 1000 scenes were randomly drawn from the
distribution of Section II-B2. For each scene, noisy measurements were collected by each of the five optimized arrays. The
measurements were repeated with five different synthetic noise levels corresponding to the five target SNR levels.
We repeatedly performed maximum likelihood estimation [45] of the expansion coefficients β, and synthesized an estimated
scene β̂(ψ) according to (4). The MSE discrepancy between β̂(ψ) and the true scene is depicted in Fig. 3. It is evident that
the quality of inference criterion is indicative of MSE performance, as each of the five geometries yielded the best MSE
performance at its specified target SNR level.
Next we solve a corresponding set of design problems with matroid constraints installed to limit the number of sensors in given
aperture segments. Specifically, we use the (cardinality constrained) partition matroid from Section IV-B with N = 11, ni =
1, ∀i and Vi spanning consecutive line segments of length 0.5: Vi =[−0.25+0.5·i, +0.25+0.5·i). The matroid constraints limits
9
Fig. 2. (left) Array designs with an aperture constraint and various SNR levels. (right) Array designs with combinatorial placement constraints and various
SNR levels.
Fig. 3. Reconstruction MSE for various arrays designed for fixed SNR levels, with actual SNR levels swept.
the proximity between sensor elements, which may be useful in real world use cases. We plot the results for the matroid
constrained designs in the right column of Fig. 2. Notice that while the theoretical guarantees pertaining to the greedy matroid
optimization scheme of Section IV is 0.5 compared to 1− 1e with the cardinality constraints of Section III, the actual performance
achieved in the constrained design instances is not far from those achieved with the simple cardinality constraints.
VI. C ONCLUDING REMARKS
We introduced a novel framework for designing sensor arrays. Our setting is Bayesian in the sense that our model incorporates
the notion of prior beliefs about the scene of interest. Our goal is designing arrays that are adapted to the scene and perform
efficient inference. We showed that this NP-hard combinatorial selection problem may be efficiently approximated by porting
results from the theory of submodular set function optimization.
10
Initially, we showed how to apply our formalism in design problems with straightforward cardinality and aperture constraints.
Later we showed that more challenging combinatorial constraints may be enforced by utilizing results and concepts from the
field of matroids and submodular optimization with matroid constraints.
Our formulation connecting the array design problem and submodularity may further be extended to reap additional benefits,
namely tackling problems such as robust array design or adaptive design that evolve as the scene is learned. We leave the
treatment of some of these and other problems to separate study [46].
A PPENDIX A
P ROOF OF L EMMA III.1
For simplicity of notation, we suppress the subscript S throughout. Begin by expanding the mutual information expressions:
I(f˜; {βm }) = H(f˜) − H(f˜| {βm })
I(fˆ; β) = H(fˆ) − H(fˆ|β)
(28)
Examining (12) and (15) we have H(f˜| {βm }) = H(fˆ|β) = H(w) and so:
I(f˜; {βm }) − I(fˆ; β) = H(f˜) − H(fˆ)
(29)
Next, notice that f˜ and fˆ are both circular, complex, Gaussian random N -length vectors, such that their entropies are given
according to [27]:
H(f˜) = log((πe)N det(Σ̃))
H(fˆ) = log((πe)N det(Σ̂))
Σ̃ij = E[f˜i f˜j∗ ]
Σ̂ij = E[fˆi fˆ∗ ]
j
(30)
∗
2
Further expand using the independence between w and βm and E[βm βm
0 ] = δmm0 σm :
X
X
∗
∗
∗
Σ̃ij = E[(
Kim βm + wi )(
Kjm
0 βm0 + wj )]
m
= [Σww ]ij +
X
m0
∗
2
Kim Kjm σm
(31)
m
Σ̂ij = E[(
X
Kim βm + wi )(
X
∗
∗
∗
Kjm
0 βm0 + wj )]
m0 ∈M
m∈M
X
= [Σww ]ij +
∗
2
Kim Kjm
σm
(32)
m∈M
Comparing (31) and (32) and using:
X
∗
2
Kim Kjm
σm
≤
m∈M
/
X
∗
2
Kim Kjm
σm
≤
X
2
σm
=
(33)
m∈M
/
m∈M
/
we have |Σ̃ij − Σ̂ij | ≤ such that we may write:
Σ̂ = Σ̃ + X
(34)
for some N × N matrix X satisfying |Xij | ≤ 1. We use (34) to bound the determinants. Σ̃ is positive-definite and invertible
such that we can write:
det(Σ̂) = det(Σ̃ + X) = det(Σ̃)det(IN + Σ̃−1 X)
(35)
H(fˆ)−H(f˜) = log(det(IN +Σ̃−1 X))= log(det(X̃))
(36)
Substituting (35) in (30) we have:
with X̃ ≡ IN + Σ̃−1 X. We turn next to bounding the term log(det(X̃)). First notice:
[Σ̃−1 X]ij =
X
Σ̃−1
im Xmj ≤
m
≤
X
Σ̃−1
im Xmj
m
√
−1
Σ̃−1
k∞ ≤ N kΣ̃−1 k2
im ≤ kΣ̃
m
√
X
√
= N σmax (Σ̃−1 ) = N
√
1
N
≤ 2
σw
σmin (Σ̃)
(37)
11
√
Where we have used the matrix norm equivalence kAk∞ ≤ N kAk2 (for N ×N matrices) and σmax (·) (σmin (·)) is the
2
maximal (minimal)
singular value such that σmin (Σ̃) ≥ σw
. Thus we have that X̃ has
√ diagonal elements centered around 1:
√
P
−1)
N
X̃im ≤ Nσ(N
.
X̃ii − 1 ≤ σ2 and the row-sums over non-diagonal entries satisfy
2
w
w
m6=i
Applying the Gershgorin circle theorem we have for the eigenvalues of X̃:
√
√
NN
NN
1−
≤
|λ
|
≤
1
+
i
2
2
σw
σw
det(X̃) is a positive real number as the quotient of the determinants of two positive definite matrices det(X̃) =
Q
Q
that we can write det(X̃) = i λi = i |λi | and consequently:
3
N log(1 −
det(Σ̂)
det(Σ̃)
such
3
N 2
N 2
) ≤ log(det(X̃)) ≤ N log(1 + 2 )
2
σw
σw
(38)
This finally leads to:
3
−N log(1+
3
N 2
N 2
) ≤ I(f˜; {βm })−I(fˆ; β) ≤ −N log(1− 2 )
2
σw
σw
(39)
A PPENDIX B
P ROOF OF L EMMA III.2
The left inequality is trivial. We have I(fˆSd? ; β) ≤ I(fˆS ? ; β) as the second optimization is over a larger set.
To prove the right inequality we show that for every S ⊆ A there is Sd ⊆ V such that
3
I(fˆS ; β) ≤ I(fˆSd ; β)+N log(1+
4δP (1+δ)N 2
)
2
λσw
we will show this for |S|=N but the proof for other cardinalities is identical.
With distance δ between adjacent elements of V, for every S= {x1 , . . . , xN } ⊂ A there is a set Sd = xd1 , . . . , xdN
that Sd ⊆ V and |xi − xdi | ≤ 2δ for all i. We have, similarly to Appendix A:
(40)
such
I(fˆS ; β) − I(fˆSd ; β) = H(fˆS ) − H(fˆSd )
(41)
H(fˆS ) = log((πe)N det(Σ̂S ))
H(fˆS ) = log((πe)N det(Σ̂Sd ))
(42)
Using the model (15):
d
where:
Σ̂S ≡ E[fˆS fˆS† ] = KS Σββ KS† + Σww
Σ̂Sd ≡ E[fˆS fˆ† ] = KS Σββ K † + Σww
d
Sd
d
Sd
†
Σββ ≡ E[ββ ]
(43)
and H(fˆS ) − H(fˆSd ) = log(det(Σ̂S )) − log(det(Σ̂Sd )).
Both KS and KSd are size N × |M| matrices defined as per the definition in (11), such that:
2
2
|[KS ]nm −[KSd ]nm | = |sinc(m+ xn )−sinc(m+ xdn )|
λ
λ
2
2
≤ 2 |xq −xdq | ≤ δ
λ
λ
(44)
where for the first inequality we have used the fact that sinc(·) is Lipschitz with constant smaller than 2. We can thus define
∆ ≡ KS − KSd and we have |∆nm | ≤ λ2 δ. Substitution in (43) yields:
Σ̂S = Σ̂Sd + ∆Σββ ∆† + ∆Σββ KS† d + KSd Σββ ∆†
(45)
12
We bound the perturbation terms by noticing:
X
[∆Σββ ∆† ]ij =
∆im [Σββ ]mm ∆jm
m
X
2
4
≤ ( δ)2
[Σββ ]mm ≤ 2 δ 2 P
λ
λ
m
[∆Σββ KS† d ]ij =
X
∆im [Σββ ]mm [KSd ]jm
m
≤
X
2
2
δ·1·
[Σββ ]mm ≤ δP
λ
λ
m
(46)
and overall we have:
|Σ̂Sij − Σ̂Sijd | ≤
4
4
4
δP + 2 δ 2 P = δP (1 + δ)
λ
λ
λ
(47)
define: 0 ≡ λ4 δP (1 + δ) and we have
Σ̂S = Σ̂Sd + 0 X
(48)
with N × N matrix X satisfying |Xij | ≤ 1 which is akin to (34). We thus port the results from Appendix A here (we only
need the lower bound):
3
0 N 2
−N log(1 + 2 ) ≤ I(fˆSd ; β) − I(fˆS ; β)
σw
(49)
which, upon substitution of 0 is equivalent to (40).
A PPENDIX C
P ROOF OF T HEOREM III.2
The greedy algorithm sequentially selects elements according to the rule x? = argmaxx∈V\S I(fˆS∪{x} ; β) where S is the
set of elements selected so far. We recursively show that the added elements can be selected on a λ2 -spaced grid centered
around x = 0. Expanding the mutual information as in Appendix A we have:
argmaxx∈V\S I(fˆS∪{x} ; β)=argmaxx∈V\S H(fˆS∪{x} )
(50)
We begin by showing that the first selected element is x1 = 0. Indeed, using the results from Appendix A:
argmax H(fˆ{x} ) = argmax det(Σ̂11 ) = argmax Σ̂11
x∈V
x∈V
where again using Appendix A and under the assumptions of the theorem (high SNR):
X
X
2
∗
2
2
Σ̂11 =
K1m K1m
σm
=
sinc2 (m+ x)σm
λ
m
m
X
2
2
2
≤ σ0
sinc (m+ x) = σ02
λ
m
2
where we used σm
≤ σ02 ,
(51)
x∈V
∀m 6= 0 and the identity:
X
sinc(m+a)sinc(m+b) = sinc(b−a)
(52)
(53)
m
It easy to see that in (52) equality is achieved for the choice x1 = 0 which
is the claim.
Next, assume that the greedy algorithm has already picked a set S = x1 , . . . , x|S| of adjacent elements on a λ2 -spaced
grid centered around x = 0 and show that the next element to be added is an adjacent location on the same λ2 -spaced grid.
We have using H(x, y)=H(x)+H(y|x):
2
argmax H(fˆS∪{x} )=argmax H(fˆ{x} |fˆS )=argmax σx|S
x∈V\S
x∈V\S
(54)
x∈V\S
2
where σx|S
is the conditional variance of the Gaussian observation collected at x given the Gaussian observations made at the
set S:
†
2
σx|S
= σx2 − ΣxS Σ−1
SS ΣxS
(55)
13
with the usual definitions for the covariance matrices (as in Appendix A):
X
2
2
2
2
2
[ΣxS ]1i =
sinc(m+ x)sinc(m+ xi )σm
=sinc(m(i)+ x)σm(i)
λ
λ
λ
m
X
2
2
2
2
[ΣSS ]ij =
sinc(m+ xi )sinc(m+ xj )σm
=δii σm(i)
λ
λ
m
where in the last equations almost all sinc(·) functions nulled out as the xi ’s are situated on a
m(i)≡− λ2 xi , such that I ≡ {m(i)} is a set of consecutive integers. Additionally, we have:
X
2
2
σx2 =
sinc2 (m+ x)σm
λ
m
(56)
λ
2
grid, and we have defined
(57)
Substituting back into (55) we have:
2
σx|S
=
X
=
X
m
m∈I
/
X
2
2
2
2
sinc2 (m+ x)σm
−
sinc2 (m(i)+ x)σm(i)
λ
λ
i∈I
2
2
sinc (m+ x)σm
λ
2
(58)
and this is similar to the optimization over the selection of the first location in (52) with the optimum achieved by selecting
/ I which is the next adjacent location on the λ2 grid which completes the proof.
x such that λ2 x = m for the first m ∈
R EFERENCES
[1] H. L. Van Trees, Detection, estimation, and modulation theory, optimum array processing. John Wiley & Sons, 2004.
[2] H. Krim and M. Viberg, “Two decades of array signal processing research: the parametric approach,” Signal Processing Magazine, IEEE, vol. 13, no. 4,
pp. 67–94, 1996.
[3] R. L. Haupt, “Thinned arrays using genetic algorithms,” Antennas and Propagation, IEEE Transactions on, vol. 42, no. 7, pp. 993–999, 1994.
[4] M. M. Khodier and C. G. Christodoulou, “Linear array geometry synthesis with minimum sidelobe level and null control using particle swarm
optimization,” Antennas and Propagation, IEEE Transactions on, vol. 53, no. 8, pp. 2674–2679, 2005.
[5] M. Skolnik, G. Nemhauser, and J. Sherman, “Dynamic programming applied to unequally spaced arrays,” IEEE Transactions on Antennas and
Propagation, vol. 12, no. 1, pp. 35–43, 1964.
[6] R. L. Haupt and D. H. Werner, Genetic algorithms in electromagnetics. John Wiley & Sons, 2007.
[7] B. P. Kumar and G. Branner, “Design of unequally spaced arrays for performance improvement,” IEEE Transactions on Antennas and Propagation,
vol. 47, no. 3, pp. 511–523, 1999.
[8] G. Oliveri, M. Carlin, and A. Massa, “Complex-weight sparse linear array synthesis by bayesian compressive sampling,” Antennas and Propagation,
IEEE Transactions on, vol. 60, no. 5, pp. 2309–2326, 2012.
[9] H. Gazzah and S. Marcos, “Cramer-rao bounds for antenna array design,” IEEE Transactions on Signal Processing, vol. 54, no. 1, pp. 336–345, 2006.
[10] P. P. Vaidyanathan and P. Pal, “Sparse sensing with co-prime samplers and arrays,” IEEE Transactions on Signal Processing, vol. 59, no. 2, pp. 573–586,
2011.
[11] D. Romero, R. Lopez-Valcarce, and G. Leus, “Compression limits for random vectors with linearly parameterized second-order statistics,” IEEE
Transactions on Information Theory, vol. 61, no. 3, pp. 1410–1425, 2015.
[12] D. D. Ariananda and G. Leus, “Compressive wideband power spectrum estimation,” IEEE Transactions on signal processing, vol. 60, no. 9, pp.
4775–4789, 2012.
[13] A. Koochakzadeh and P. Pal, “Performance of uniform and sparse non-uniform samplers in presence of modeling errors: A cramér-rao bound based
study,” IEEE Transactions on Signal Processing, vol. 65, no. 6, pp. 1607–1621, 2017.
[14] J. D. Krieger, Y. Kochman, and G. W. Wornell, “Design and analysis of multi-coset arrays,” in 2013 IEEE International Conference on Acoustics, Speech
and Signal Processing. IEEE, 2013, pp. 3781–3785.
[15] ——, “Multi-coset sparse imaging arrays,” IEEE Transactions on Antennas and Propagation, vol. 62, no. 4, pp. 1701–1715, 2014.
[16] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, “An analysis of approximations for maximizing submodular set functions,” Mathematical Programming,
vol. 14, no. 1, pp. 265–294, 1978.
[17] S. Fujishige, Submodular functions and optimization. Elsevier, 2005, vol. 58.
[18] N. Buchbinder, M. Feldman, J. Seffi, and R. Schwartz, “A tight linear time (1/2)-approximation for unconstrained submodular maximization,” SIAM
Journal on Computing, vol. 44, no. 5, pp. 1384–1402, 2015.
[19] J. Vondrák, “Symmetry and approximability of submodular maximization problems,” SIAM Journal on Computing, vol. 42, no. 1, pp. 265–304, 2013.
[20] B. Mirzasoleiman, A. Karbasi, R. Sarkar, and A. Krause, “Distributed submodular maximization: Identifying representative elements in massive data,”
in Advances in Neural Information Processing Systems, 2013, pp. 2049–2057.
[21] A. Krause, A. Singh, and C. Guestrin, “Near-optimal sensor placements in gaussian processes: Theory, efficient algorithms and empirical studies,” Journal
of Machine Learning Research, vol. 9, no. Feb, pp. 235–284, 2008.
[22] M. Shamaiah, S. Banerjee, and H. Vikalo, “Greedy sensor selection: Leveraging submodularity,” in 49th IEEE conference on decision and control (CDC).
IEEE, 2010, pp. 2572–2577.
[23] A. Krause and C. Guestrin, “Submodularity and its applications in optimized information gathering,” ACM Transactions on Intelligent Systems and
Technology (TIST), vol. 2, no. 4, p. 32, 2011.
[24] P. Stoica and A. Nehorai, “Performance study of conditional and unconditional direction-of-arrival estimation,” IEEE Transactions on Acoustics, Speech,
and Signal Processing, vol. 38, no. 10, pp. 1783–1795, 1990.
[25] M. Wang and A. Nehorai, “Coarrays, music, and the cramér–rao bound,” IEEE Transactions on Signal Processing, vol. 65, no. 4, pp. 933–946, 2017.
[26] A. Koochakzadeh and P. Pal, “Cramér–rao bounds for underdetermined source localization,” IEEE Signal Processing Letters, vol. 23, no. 7, pp. 919–923,
2016.
[27] F. D. Neeser and J. L. Massey, “Proper complex random processes with applications to information theory,” Information Theory, IEEE Transactions on,
vol. 39, no. 4, pp. 1293–1302, 1993.
[28] J. L. Yen, “On nonuniform sampling of bandwidth-limited signals,” Circuit Theory, IRE Transactions on, vol. 3, no. 4, pp. 251–257, 1956.
14
[29] D. J. Wingham, “The reconstruction of a band-limited function and its fourier transform from a finite number of samples at arbitrary locations by singular
value decomposition,” Signal Processing, IEEE Transactions on, vol. 40, no. 3, pp. 559–570, 1992.
[30] Y. C. Eldar, Sampling Theory: Beyond Bandlimited Systems. Cambridge University Press, 2015.
[31] M. K. Mihcak, I. Kozintsev, K. Ramchandran, and P. Moulin, “Low-complexity image denoising based on statistical modeling of wavelet coefficients,”
IEEE Signal Processing Letters, vol. 6, no. 12, pp. 300–303, 1999.
[32] A. Levin, R. Fergus, F. Durand, and W. T. Freeman, “Image and depth from a conventional camera with a coded aperture,” ACM transactions on graphics
(TOG), vol. 26, no. 3, p. 70, 2007.
[33] A. Levin, Y. Weiss, F. Durand, and W. T. Freeman, “Understanding and evaluating blind deconvolution algorithms,” in Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009, pp. 1964–1971.
[34] A. Kulesza, B. Taskar et al., “Determinantal point processes for machine learning,” Foundations and Trends R in Machine Learning, vol. 5, no. 2–3,
pp. 123–286, 2012.
[35] G. Friesecke, “Lecture notes in fourier analysis [MA4064],” 2013.
[36] K. Chaloner and I. Verdinelli, “Bayesian experimental design: A review,” Statistical Science, pp. 273–304, 1995.
[37] J. M. Bernardo, “Expected information as expected utility,” The Annals of Statistics, pp. 686–690, 1979.
[38] A. Krause and C. E. Guestrin, “Near-optimal nonmyopic value of information in graphical models,” Proceedings of the Twenty-First Conference on
Uncertainty in Artificial Intelligence, 2012.
[39] U. Feige, “A threshold of ln n for approximating set cover,” Journal of the ACM (JACM), vol. 45, no. 4, pp. 634–652, 1998.
[40] M. Minoux, “Accelerated greedy algorithms for maximizing submodular set functions,” in Optimization Techniques. Springer, 1978, pp. 234–243.
[41] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen, and N. Glance, “Cost-effective outbreak detection in networks,” in Proceedings of the
13th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2007, pp. 420–429.
[42] J. G. Oxley, Matroid theory. Oxford University Press, USA, 2006, vol. 3.
[43] A. Krause and D. Golovin, “Submodular function maximization,” Tractability: Practical Approaches to Hard Problems, vol. 3, p. 19, 2012.
[44] G. Calinescu, C. Chekuri, M. Pál, and J. Vondrák, “Maximizing a monotone submodular function subject to a matroid constraint,” SIAM Journal on
Computing, vol. 40, no. 6, pp. 1740–1766, 2011.
[45] T. Adali, P. J. Schreier, and L. L. Scharf, “Complex-valued signal processing: The proper way to deal with impropriety,” IEEE Transactions on Signal
Processing, vol. 59, no. 11, pp. 5101–5125, 2011.
[46] G. Shulkind, S. Jegelka, and G. W. Wornell, “Multiple wavelength sensing array design,” in Acoustics, Speech and Signal Processing (ICASSP), 2017
IEEE International Conference on. IEEE, 2017, pp. 3424–3428.
| 7cs.IT
|
Mining Process Model Descriptions of Daily Life
through Event Abstraction
arXiv:1705.10202v1 [cs.LG] 25 May 2017
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
Abstract Process mining techniques focus on extracting insight in processes from
event logs. Process mining has the potential to provide valuable insights in (un)healthy
habits and to contribute to ambient assisted living solutions when applied on data
from smart home environments. However, events recorded in smart home environments are on the level of sensor triggers, at which process discovery algorithms
produce overgeneralizing process models that allow for too much behavior and that
are difficult to interpret for human experts. We show that abstracting the events to
a higher-level interpretation can enable discovery of more precise and more comprehensible models. We present a framework for the extraction of features that can
be used for abstraction with supervised learning methods that is based on the XES
IEEE standard for event logs. This framework can automatically abstract sensorlevel events to their interpretation at the human activity level, after training it on
training data for which both the sensor and human activity events are known. We
demonstrate our abstraction framework on three real-life smart home event logs and
show that the process models that can be discovered after abstraction are more precise indeed.
Niek Tax
Eindhoven University of Technology, e-mail: [email protected]
Natalia Sidorova
Eindhoven University of Technology e-mail: [email protected]
Reinder Haakma
Philips Research e-mail: [email protected]
Wil M.P. van der Aalst
Eindhoven University of Technology e-mail: [email protected]
1
2
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
1 Introduction
Process mining is a fast growing discipline that combines methods from computational intelligence, data mining, process modeling and process analysis [1]. Process discovery, the task of extracting process models from logs, plays an important role in process mining. There are many different process discovery algorithms
([2, 7, 37, 36, 18]), which can discover many different types of process models, including BPMN models, Petri nets, process trees, UML activity diagrams, and statecharts.
While originally the scope of process mining has been on business processes, it
has broadened in recent years towards other application areas, including the analysis
of human behavior [29, 31, 32]. Process model descriptions of human behavior can
be used amongst others to aid lifestyle coaching for healthy living, or to asses the
ability of independent living of elderly or people with illness.
Events in the event log are generated by e.g. motion sensors placed in the home,
power sensors placed on appliances, open/close sensors placed on closets and cabinets, etc. This clearly distinguishes process mining for smart homes from the traditional application domain of business processes, where events in the log are logged
by IT systems when an business tasks are performed.
In event logs from business processes the event labels generally have a clear
semantic meaning, like register mortgage request. In the smart home domain the
events are on the sensor level, while the human expert is interested in analyzing the
behavior in terms of activities of daily life. Additionally, simply using the sensor that
generated the event as the event label has been shown to result in non-informative
process models that overgeneralize the event log and allow for too much behavior [34]. In the field of process mining such overgeneralizing process models are
generally referred to as being imprecise.
In our earlier work [31] we showed how to discover more precise process models
by taking the name of the sensor as a starting point for the event label and then refine
the labels using the time of the day at which the event occurred. However, labels in
such process models still represent sensors, and they have no direct interpretation
on the human activity level. In this paper we leverage diary style annotations of
the activities performed on a human activity level and use them learn a mapping
from sensor-level events to human activity events. This enables discovery of process
models that describe the human activities directly, leading to more comprehensible
and more precise descriptions of human behavior. Often it is infeasible or simply too
expensive to obtain such diaries for periods of time longer than a couple of weeks.
To mine a process model of human behavior more than a couple of weeks of data is
needed. Therefore, there is a need to infer human level interpretations of behavior
from sensors.
With supervised learning techniques the mapping from sensor-level events to
human activity level events can be learned through examples, without requiring a
hand-made ontology of how human activities relate to sensors. Similar approaches
have been explored in the activity recognition field, where continuous-valued time
series from sensors are mapped to time series of human activity. Change points in
Mining Process Model Descriptions of Daily Life through Event Abstraction
3
these time series are triggered by sensor-level events like opening/closing the fridge
door, and the annotations of the higher level events (e.g. cooking) are often obtained
through manual activity diaries. However, in contrast to techniques from the activity recognition field, we operate on discrete events on the sensor-level instead of
continuous time series.
In this paper we extend the work started in [33]. We describe a framework for
supervised abstraction of events that enables the discovery of more precise process
models from smart home event logs. Additionally, the process models obtained represent human activity directly, thereby enabling direct analysis of human behavior
itself, instead of indirect analysis through sensor-level models. In Section 2 we give
an overview of the related work from the activity recognition field. Basic concepts,
notations, and definitions that we use throughout the rest of the paper are introduced
in Section 3. In Section 4 we explain conceptually why abstraction from sensor-level
to human activity level events can help to the process discovery step to find more
precise process models. In Section 5 we describe a framework for retrieving useful
features for abstraction from event logs using specific concepts of the IEEE XES
standard for event logs [12]. Section 6 demonstrates the added value of supervised
event abstraction for process mining in the smart home domain and show that it
enables discovery of more precise models on three real life smart home event logs.
Section 7 concludes the paper and identifies some areas of future work.
2 Related Work
Event abstraction based on supervised learning is an unexplored problem in process
mining. Most related work for abstracting from sensor-level to human activity level
events can be found in the field of activity recognition, which focuses on the task
of detecting different types of human activity from either passive sensors [14, 30],
wearable sensors [3, 16], or cameras [22].
Activity recognition methods generally operate on discrete time windows over
the time series of continuous-valued sensor values and aim to map each time window onto the correct type of human activity, e.g. eating or sleeping. Activity recognition methods can be classified into probabilistic approaches [14, 30, 3, 16] and ontological reasoning approaches [5, 25]. The advantage of probabilistic approaches
over ontological reasoning approaches is their ability to handle noisy, uncertain and
incomplete sensor data [5].
Tapia [30] was the first to explore supervised learning for inference of human
activities from passive sensors, using a naive Bayes classifier. Many more recent
activity recognition approaches use probabilistic graphical models [14, 13]: Van
Kasteren et al. [14] explored Conditional Random Fields [17] and Hidden Markov
Models [23], and Van Kasteren and Kröse [13] applied Bayesian Networks [10]
on the activity recognition task. Kim et al. [15] found Hidden Markov Models to
be unable to capture long-range or transitive dependencies between observations,
4
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
which results in difficulties recognizing multiple interacting activities (concurrent
or interwoven). Conditional Random Fields do not possess these limitations.
Our work differentiates itself from existing activity recognition work in the form
of the input data on which they operate and in the goal that it aims to achieve. On the
input side, activity recognition techniques consider the data to be a multidimensional
time series of the sensor values over time, based on which time windows are mapped
onto human activities. An appropriate time window size is determined using domain
knowledge of the data set. Instead, we aim for a generic method that does not require
this domain knowledge, and that works in general for any event log. An approach
based on time windows contrasts with our aim for generality, as no single time
window size exists that is suitable for all event logs. The durations of the events
within a single event log might differ drastically (e.g. one event might take seconds,
while another event takes months), in which case time window based approaches
will either miss short events in case of larger time windows or resort to very large
numbers of time windows resulting in very long computational time. Therefore, we
map each individual sensor-level event to a human activity level event and do not
use time windows. In a smart home environment context with passive sensors, each
change in a binary sensor value can be considered to be a low-level event. A second
difference with existing activity recognition techniques is that our framework aims
to find an abstraction of the data that enables discovery of more precise process
models, where classical activity recognition methods do not have a link with the
application of process mining.
Other related work can be found in the area of process mining, where several
techniques address the challenge of abstracting low-level (e.g. sensor-level) events
to high level (e.g. human activity level) events ([4, 11, 9]). Most existing event
abstraction methods rely on clustering methods, where each cluster of low-level
events is interpreted as one single high-level event. However, using unsupervised
learning introduces two new problems. First, it is unclear how to label high-level
events that are obtained by clustering low-level events. Current techniques require
the user / process analyst to provide high-level event labels themselves based on
domain knowledge. Secondly, unsupervised learning gives no guidance with respect
to the desired level of abstraction. Many existing event abstraction methods contain
one or more parameters to control the size the clusters, and finding the right level of
abstraction providing meaningful results is often a matter of trial-and-error.
One abstraction technique from the process mining field that does not rely on
unsupervised learning is proposed by Mannhardt et al. [20]. This approach relies on
domain knowledge to abstract low-level events into high-level events, requiring the
user to specify a low-level process model for each high-level activity. However, in
the context of human behavior it is unreasonable to expect the user to provide the
process model in sensor terms for each human activity from domain knowledge.
Mining Process Model Descriptions of Daily Life through Event Abstraction
5
3 Preliminaries
In this section we introduce basic concepts and notation used throughout the paper.
We use the standard sequence definition, and denote a sequence by listing its
elements, e.g. we write ha1 , a2 , . . . , an i for a (finite) sequence s : {1, . . . , n} → S of
elements from some alphabet S, where s(i) = ai for any i ∈ {1, . . . , n}.
3.1 Petri nets
A process modeling notation that is commonly used in process mining is the Petri
net. Petri nets are directed bipartite graphs consisting of transitions and places, connected by arcs. Transitions represent activities, while places represent the state of
the system before and after execution of a transition. Labels are assigned to transitions to indicate the type of activity that they represent. A special label τ is used to
represent invisible transitions, which are only used for routing purposes and do not
represent any real activity.
Definition 1 (Labeled Petri net). A labeled Petri net is a tuple N = (P, T, F, R, ℓ)
where P is a finite set of places, T is a finite set of transitions such that P ∩ T = 0,
/
and F ⊆ (P × T ) ∪ (T × P) is a set of directed arcs, called the flow relation, R is
/ R is a label representing an
a finite set of labels representing event types, with τ ∈
invisible action, and ℓ : T → R ∪ {τ } is a labeling function that assigns a label to
each transition.
The state of a Petri net is defined w.r.t. the state that a process instance can be in
during its execution. A state of a Petri net is captured by the marking of its places
with tokens. In a given state, each place is either empty, or it contains a certain
number of tokens. A transition is enabled in a given marking if all places with an
outgoing arc to this transition contain at least one token. Once a transition fires (i.e.
is executed), a token is removed from all places with outgoing arcs to the firing transition and a token is put to all places with incoming arcs from the firing transition,
leading to a new state.
Definition 2 (Marking, Enabled transitions, and Firing). A marked Petri net is
a pair (N, M), where N = (P, T, F, L, ℓ) is a labeled Petri net and M ∈ NP denotes
the marking of N. For n ∈ (P ∪ T ) we use •n and n• to denote the set of inputs and
outputs of n respectively. Let C(s, e) indicate the number of occurrences (count) of
element e in multiset s. A transition t ∈ T is enabled in a marking M of net N if
∀p ∈ •t : C(M, p) > 0. An enabled transition t may fire, removing one token from
each of the input places •t and producing one token for each of the output places t•.
Figure 1 shows four Petri nets, with the circles representing places, the squares
representing transitions. The gray rectangles represent (invisible) τ -transitions.
belong to the final marking, indicating that the process exPlaces depicted as
ecution can terminate in this marking.
6
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
The Petri net shown in Figure 1c initially has one token in the place p1, indicated
by the dot. Firing the enabled silent transition takes the token from p1 and puts a
token in both p2 and p3, enabling both MC and DCC. When MC fires, it takes the
token from p2 and puts a token in p4. When DCC fires, it takes the token from p3
and puts a token in p5. After MC and DCC have both fired, resulting in a token in
both p4 and p5, W is enabled. Executing W takes the token from both p4 and p5,
and puts a token in p6, which is a place that belongs to the final marking, indicates
that the process execution can stop here. Alternatively, it can fire the silent transition, taking the token from p6 and placing a token in p2 and p5, which allows for
execution of MC and W to reach the marking consisting of p6 again. We refer the
interested reader to [24] for an extensive review of Petri nets.
3.2 Conditional Random Field
We consider the recognition of human activity level events as a sequence labeling
task in which each sensor-level event is classified into one of the human activity
level events. Linear-chain Conditional Random Fields (CRFs) [17] are a type of
probabilistic graphical model which has shown to perform well on many sequence
labeling tasks in the fields of language processing and computer vision. Conceptually CRFs can be regarded as a sequential version of multiclass logistic regression,
i.e., the predictions in the prediction sequence are dependent on each other. A CRF
models the conditional probability distribution of the label sequence given an observation sequence using a log-linear model. Linear-chain CRFs take the following
form:
p(y|x) =
1
exp( ∑ ∑ λk fk (t, yt−1 , yt , x))
Z(x)
t=1 k
(1)
where Z(x) is the normalization factor which makes sure that the values of the probability distribution range from zero to one. X = hx1 , . . . , xn i is an observation sequence (the sensor-level events), Y = hy1 , . . . , yn i is the associated label sequence
(the human activity level events), fk and λk respectively are feature functions and
their weights. Feature functions, which can be binary or real valued, are defined
on the observations and are used to compute label probabilities. In contrast to Hidden Markov Models [23], CRFs do not assume the feature functions to be mutually
independent.
4 Motivating Example
Figure 1 shows a simplistic example demonstrating how a process can seem unstructured at the sensor level of events, while being structured at a human behavior level.
Petri net 1c shows the process at the human activity level. The Taking medicine hu-
Mining Process Model Descriptions of Daily Life through Event Abstraction
DCC
MC
p2
7
CD
p4
W
p1
p6
DCC
p3
p5
DCC = Dishes & Cups Cabinet
MC = Medicine Cabinet
W = Water
CD = Cutlery Drawer
D = Dishwasher
DCC = Dishes & Cups Cabinet
D
(b)
(a)
Taking Medicine
Eating
(c)
(d)
Fig. 1 A human activity level process model (c) where the two transitions themselves are defined
as process models (shown in a and b), and the Inductive Miner result on the sensor-level traces
generated from this model (d).
man activity level activity is itself defined as a process, which is shown in Figures
1a to 1c. Eating is also defined as a process, which is shown in Figure 1b. When we
apply the Inductive Miner process discovery algorithm [18] to sensor-level traces
generated by the hierarchical process of Figure 1c, we obtain the process model
shown in Figure 1d. This process model allows for almost all possible sequences
over the alphabet {CD, D, DCC, MC,W }, with the only constraint introduced by the
model being that if a W occurs, then it has to be preceded by a DCC event. Firing
of all other transitions in the model can be skipped. Behaviorally this model is very
close to the so called ”flower” model [1], the model that allows for all behavior over
its alphabet. The alternating structure between Taking medicine and Eating that was
present in the human activity level process in Figure 1c cannot be observed in the
process model in Figure 1d. This is caused by high variance in start and end events
of the sensor-level subprocesses of Taking medicine and Eating as well as by the
overlap in types of activities between these two subprocesses. Both subprocesses
contain DCC, and the miner cannot see that there are actually two different contexts
for the DCC activity to split the label in the model. Abstracting the sensor-level
events to their respective human activity level events before applying process discovery to the resulting human activity log unveils the alternating structure between
Eating and Taking medicine as shown in Figure 1c.
8
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
Fig. 2 XES event log meta-model, as defined in [12].
5 Event Abstraction as a Sequence Labeling Task
In this section we describe the framework for supervised abstraction of events based
on Conditional Random Fields (CRFs). Additionally, we describe feature functions
for event logs in a general way by using the IEEE XES standard [12]. XES, which is
an abbreviation for eXtensible Event Stream, is the IEEE standard for process mining event logs. An overview of the XES file structure which is shown in Figure 2.
An event log is defined as a set of traces, which in itself are a sequences of events.
The log, traces and events can all contain one or more attributes, which consist of
a key and a value of a certain type. Event or trace attributes may be global, which
indicates that the attribute needs to be defined for each event or trace respectively.
A log contains one or more classifiers, which can be seen as labeling functions on
the events of a log, defined on global event attributes. Extensions define a set of
attributes on log, trace, or event level, in such a way that the semantics of these
attributes are clearly defined. One can view XES extensions as a specification of attributes that events, traces, or event logs themselves frequently contain. XES defines
the following standard extensions:
Mining Process Model Descriptions of Daily Life through Event Abstraction
9
Concept Specifies the generally understood name of the event/trace/log (attribute
’Concept:name’).
Lifecycle Specifies the lifecycle phase (attribute ’Lifecycle:transition’) that the
event represents in a transactional model of their generating activity. The Lifecycle extension also specifies a standard transactional model for activities.
Organizational Specifies three attributes for events, which identify the actor having caused the event (attribute ’Organizational:resource’), his role in the organization (attribute ’Organizational:role’), and the group or department within the
organization where he is located (attribute ’Organizational:group’).
Time Specifies the date and time at which an event occurred (attribute ’Time:timestamp’).
We introduce a special attribute of type String with key label, which represents
the human activity level activity. The concept name of an event is then considered to
be a sensor-level name of an event. The label attribute specifies the human activity
level label for each event individually, allowing for example one sensor-level event
of type Dishes & cups cabinet to be of human activity level type Taking medicine,
and another sensor-level event of the same type to be of human level activity type
Eating. Note that for some traces human level activity annotations might be available, in which case its events contain the label attribute, while other traces might not
be annotated. Human activity level interpretations of unannotated traces, obtained
by inferring the label attribute based on information that is present in the annotated
traces, allow the use of unannotated traces for process discovery and conformance
checking on a human activity level.
Figure 3 provides a conceptual overview of the supervised event abstraction
method. The approach takes two inputs: 1) a set of annotated traces, which are traces
where the human activity level event that each sensor level event belongs to (the label attribute of the sensor-level event) is known, and 2) a set of unannotated traces,
which are traces where only the sensor-level events known. Conditional Random
Fields (CRFs) are trained of the annotated traces to create a probabilistic mapping
from sensor-level events to human activity level events. This mapping, once obtained, can be applied to the unannotated traces in order to estimate the corresponding human activity level event for each sensor-level event (its label attribute). Often
multiple consecutive sensor-level events will have the same label attribute. We assume that multiple human activity level events cannot occur in parallel. This enables
us to interpret a sequence of events with identical label values as a single human activity level event. To obtain a final human activity level log, we collapse sequences
of events with the same value for the label attribute into two events with this value
as concept name, where the first event has a lifecycle ’start’ and the second has the
lifecycle ’complete’. Table 1 and Table 2 illustrate this collapsing procedure through
an input and output event log.
The method described in this section is implemented and available for use as
package AbstractEventsSupervised in the ProM 6 [35] process mining toolkit and
makes use of the GRMM [28] implementation of CRFs.
We now show for each XES extension how it can be translated into useful feature
functions for event abstraction. Note that we do not limit ourselves to XES logs that
10
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
Table 1 A trace with predicted human activity level annotations (label)
Case Time:timestamp
Concept:name
label
1
1
1
1
1
1
1
1
Medicine cabinet
Dishes & cups cabinet
Water
Dishes & cups cabinet
Dishwasher
Dishes & cups cabinet
Medicine cabinet
Water
Taking medicine
Taking medicine
Taking medicine
Eating
Eating
Taking medicine
Taking medicine
Taking medicine
03/11/2015 08:45:23
03/11/2015 08:46:11
03/11/2015 08:46:45
03/11/2015 08:47:59
03/11/2015 08:47:89
03/11/2015 17:10:58
03/11/2015 17:10:69
03/11/2015 17:11:18
Table 2 The resulting human activity level log after collapsing the consecutive identical label
values of the trace in Table 1.
Case Time:timestamp
Concept:name
Lifecycle:transition
1
1
1
1
1
1
Taking medicine
Taking medicine
Eating
Eating
Taking medicine
Taking medicine
Start
Complete
Start
Complete
Start
Complete
03/11/2015 08:45:23
03/11/2015 08:46:45
03/11/2015 08:47:59
03/11/2015 08:47:89
03/11/2015 17:10:58
03/11/2015 17:11:18
Fig. 3 Conceptual overview
of Supervised Event Abstraction.
Annotated traces
Learn
Sensor-level to
Activity-level
Mapping
Activity-level
log
Unannotated traces
Apply
Traces with
predicted
annotations
contain all XES extensions; when a log contains a subset of the extensions, a subset of the feature functions will be available for the supervised learning step. This
approach leads to a feature space of unknown size, potentially causing problems related to the curse of dimensionality. To address this we use L1-regularized CRFs. In
the training phase we search for values of weight vector λ that minimize the cross
entropy between the ground truth target and the predicted label on the training data.
L1-regularization adds a λ penalty terms to this minimization function that is proportionate to the size of the weight vector, giving the model an incentive not to use
all of the available features (i.e., setting some features to zero weight). This results
in prediction models that are sparse and therefore simpler, which helps to prevent
overfitting.
Mining Process Model Descriptions of Daily Life through Event Abstraction
11
Fig. 4 The histogram
representation and a Gaussian
Mixture Model fitted to
timestamps values of the
plates cupboard sensor in the
Van Kasteren data set.
5.1 From a XES Log to a Feature Space
We now discuss per XES extension how feature functions can be obtained.
Concept extension The sensor-level labels (concept names) of the preceding
events in a trace can contain useful contextual information for classification into
the correct human activity level event type. Based on the n-gram ha1 , a2 , . . . , an i
consisting of the sensor-level labels of the n last-seen events in a trace, we can
estimate a categorical probability distribution over the classes of human activity
level activities from the training log, such that the probability of class l is equal to the
number of times that the n-gram was observed while the n-th event was annotated
with class l, divided by the total number of times that the n-gram was observed.
The CRF model requires feature functions with numerical range. A feature function
based on the concept extension has two parameters, n and l, and is valued with the
estimated categorical probability density of the current sensor-level event having
human activity level label l given the n-gram with the last n sensor-level event labels.
It can be useful to combine multiple features that are based on the concept extension,
where the features have different values for n and l.
Organizational extension Similar to the concept extension feature functions,
categorical probability distributions can be estimated on the training set for ngrams of resource, role, or group attributes of the last n events. Likewise, an organizational extension based feature function with three parameters, n-gram size
n, o ∈ {resource, role, group}, and label l, is valued with the probability density
according to the estimated categorical probability distribution of label l given the
n-gram of the last n event resources/roles/groups.
Time extension In terms of time, there are several potentially existing patterns.
A certain type of human activity might for example be concentrated in a certain
parts of a day, of a week, or of a month. This concentration can however not be
modeled with a single Gaussian distribution, as it might be the case that a type of
human activity has high probability to occur in the morning or in the evening, but
low probability to occur in the afternoon in-between. A mixture distribution consisting of multiple components is therefore needed to model the probability distri-
12
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
bution over timestamps. The most well-known mixture distribution is the Gaussian
Mixture Model (GMM), where each component of the mixture is defined by a normal distribution. The circular, non-Euclidean, nature of the data space of time-ofthe-day, time-of-the-week, or time-of-the-month however introduces problems for
the GMM, as, using time-of-the-day as an example, 00:00 is actually very close
to 23:59. Figure 4 illustrates this problem. The Gaussian component with a mean
around 10 o’clock has a standard deviation that is much higher than what one would
expect when looking at the histogram, as the GMM tries to explain the data points
just after midnight with this component. These data points just after midnight would
however have been much better explained with the Gaussian component with the
mean around 20 o’clock, which is much closer in time. Alternatively, we use a
mixture model with components of the von Mises distribution, which is a close
approximation of a normal distribution wrapped around the circle. To determine the
correct number of components of such a von Mises Mixture Model (VMMM) we
use Bayesian Information Criterion (BIC) [27], choses the number of components
which explains the data with the highest likelihood, while adding a penalty for the
number of model parameters. A VMMM is estimated on training data, modeling the
probabilities of each type of human activity based on the time passed since the start
of the day, week or month. A time extension feature function with two parameters,
t ∈ {day, week, month, . . .} and label l, is valued with the VMMM-estimated probability of label l given the t view on the event timestamp. An alternative approach to
estimate the probability density on data that lies on a manifold, such as a circle, is
described by Cohen and Welling [6].
Lifecycle extension & Time extension The XES standard [12] defines several
lifecycle stages of process activities, which represent the transactional model of their
generating activity. Lifecycle values that are commonly found in real life logs are
start and complete which respectively represent when this activity started and ended
However, a larger set of lifecycle values is defined in the XES standard, including
schedule, suspend, and resume. The time differences between different stages of an
activity lifecycle can be calculated when an event log possesses both the lifecycle
extension and the time extension. For example, when observing the complete of an
activity, the time between this complete and the corresponding start of this activity
can contain useful information for predicting the correct human activity label. Finding the associated start event becomes nontrivial when multiple instances of the
same activity are in parallel, as it is then unknown which complete event belongs
to which start event. We assume consecutive lifecycle steps of activities running
in parallel to occur in the same order as the preceding lifecycle step. For example,
when we observe two start events of an activity of type A in a row, followed by two
complete events of type A, we assume the first complete to belong to the first start,
and the second complete to belong to the second start.
The XES standard defines an ordering over the lifecycle values. For each type
of human activity, we fit a Gaussian Mixture Model (GMM) to the set of time differences between each two consecutive lifecycle steps. A feature based on both the
combination of the lifecycle and the time extension with activity label parameter
l and lifecycle c is valued the probability density of activity l as estimated by the
Mining Process Model Descriptions of Daily Life through Event Abstraction
13
GMM given the time between the current event and lifecycle value c. Bayesian Information Criterion (BIC) [27] is again used to determine the number of components
of the GMM. Note that while these features are time-based, regular GMMs can be
used instead of VMMMs since time duration is a Euclidean, non-circular, space.
5.2 Evaluating Human Activity Level Event Predictions for
Process Mining Applications
Existing approaches in the field of activity recognition take as input time windows
where each time window is represented by a feature vector that describes the sensor
activity or status during that time window. Hence, evaluation methods in the activity
recognition field are window-based, using evaluation metrics like the percentage of
correctly classified time slices [30, 13, 14]. There are two reasons to deviate from
this evaluation methodology in a process mining setting. First, our method operates
on events instead of time windows. Second, the accuracy of the resulting high level
sequences is much more important for many process mining techniques (e.g. process
discovery, conformance checking) than the accuracy of predicting each individual
minute of the day.
A well-known metric for the distance of two sequences is the Levenshtein distance [19]. However, Levenshtein distance is not suitable to compare sequences of
human actions, as human behavior sometimes includes branches in which it does
not matter in which order two activities are performed. For example, most people
shower and have breakfast after waking up, but people do not necessarily always
perform the two in the same order. Indeed, when ha, bi is the sequence of predicted
human activities, and hb, ai is the actual sequence of human activities, we consider
this to be only a minor error, since it is often not relevant in which order two parallel
activities are executed. Levenshtein distance would assign a cost of 2 to this abstraction, as transforming the predicted sequence into the ground truth sequence would
require one deletion and one insertion operation. For example, most people shower
and have breakfast after waking up, but people do not necessarily always perform
the two in the same order. An evaluation measure that better reflects the prediction
quality of event abstraction is the Damerau-Levenstein distance [8], which adds a
swapping operation to the set of operations used by Levenshtein distance. DamerauLevenshtein distance would assign a cost of 1 to transform ha, bi into hb, ai. To obtain comparable numbers for different numbers of predicted events we normalize the
Damerau-Levenshtein distance by the maximum of the length of the ground truth
trace and the length of the predicted trace and subtract the normalized DamerauLevenshtein distance from 1 to obtain Damerau-Levenshtein Similarity (DLS).
14
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
6 Case Studies
In this section we evaluate the supervised event abstraction framework on three case
studies on real life smart home data sets.
6.1 Experimental setup
We include three real life smart home event logs in the evaluation: the Van Kasteren
event log [14], and two event logs from a smart home experiment conducted by MIT
[30]. All three event logs used in for the evaluation consist of multidimensional time
series data with all dimensions binary, where each binary dimension represents the
state of one in-home sensor. These sensors include motion sensors, open/close sensors, and power sensors (discretized to 0/1 states). We transform the multidimensional time series data from sensors into events by regarding each sensor change
point as an event. Cases are created by grouping events together that occurred in the
same day, with a cut-off point at midnight. High-level labels are provided for the
event logs.
The following XES extensions can be used for these event logs:
Concept The sensor that generated the event.
Time The time stamp of the sensor change point.
Lifecycle Start when the event represents a sensor value change from 0 to 1 and
Complete when it represents a sensor value change from 1 to 0.
Note that human activity level annotations are provided for all traces in the three
event logs. To evaluate how well the supervised event abstraction method generalized to unannotated traces, we artificially use a part of the traces to train the abstraction model and apply them on a test set where we regard the annotations to
be non-existent. We evaluate the obtained human activity labels against the ground
truth labels in a Leave-One-Trace-Out-Cross-Validation setup where we iteratively
leave out one trace to evaluate how well this mapping generalizes to unseen events
and cases. We measure the accuracy of the human activity level traces compared
to the ground truth human activity level traces in terms of Damerau-Levenshtein
similarity [8].
Additionally, we evaluate the quality of the process model that can be discovered from the human activity level traces. To discover a process model from the
human activity level event log we use the Inductive Miner [18]. There are several
criteria to express the fit between a process model and an event log in the area of
process mining. Two of those criteria are fitness [26], which measures the degree to
which the behavior that is observed in the event log can be replayed on the process
model, and precision [21], which measures the degree to which the behavior that
was never observed in the event log cannot be replayed on the process model. Low
precision typically is indicates an overly general process model, that allows for too
Mining Process Model Descriptions of Daily Life through Event Abstraction
15
much behavior. We compare the fitness and precision of the models produced by the
Inductive Miner on the sensor-level log and the human activity level log.
6.2 Case Study 1: Van Kasteren Event Log
For the first case study we use the smart home environment log described in Van
Kasteren et al. [14]. The Van Kasteren log contains 1285 events divided over fourteen different sensors. The log contains 23 days of data.
The average Damerau-Levenshtein similarity between the predicted human activity level traces in the Leave-One-Trace-Out-Cross-Validation experimental setup
and the ground truth human activity level traces is 0.7516, which shows that the
supervised event abstraction method produces traces which are fairly similar to the
ground truth.
Figure 5 shows the result of the Inductive Miner [18] for the sensor-level events
in the Van Kasteren data set. The resulting process model starts with a choice between four activities: hall-toilet door, hall-bedroom door, hall-bathroom door, and
frontdoor. After this choice the model branches into three parallel blocks, where the
upper block consists of a large choice between eight different activities. The other
two parallel blocks respectively contain a loop of the cups cupboard and the fridge.
This model closely resembles the flower model, which allows for all behavior in any
arbitrary order. There seems to be very little structure on this sensor level of event
granularity.
Figure 6 shows the result of the Inductive Miner on the aggregated set of predicted test traces. We can see that the main daily routine starts with breakfast, after
which the subject leaves the house to go to work. After work the subject prepares
dinner and goes to bed. The activities use toilet and take shower are put in parallel
to this sequence of activities, indicating that they occur at different places in the
sequences of activities.
Table 3 shows the effect of the abstraction on the fitness and precision of the
models discovered by the Inductive Miner. It shows that the precision of the model
discovered on the abstracted log is much higher than the precision of the model
discovered on the sensor data, indicating that the abstraction helps discovering a
model that is more behaviorally constrained and more specific.
6.3 Case Study 2: MIT Household A Event Log
For the second case study we use the data of household A of a smart home experiment conducted by MIT [30]. Household A contains data of 16 days of living, 2701
sensor-level events registered by 26 different sensors. The human level activities are
provided in the form of a taxonomy of activities on three levels, called heading, category and subcategory. On the heading level the human activities are very general in
16
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
Fig. 5 Inductive Miner result on the sensor-level events of the Van Kasteren event log.
Fig. 6 Inductive Miner result on the human activity level events discovered from the Van Kasteren
event log.
Table 3 Effect of abstraction on fitness and precision of the process model discovered by the
Inductive Miner.
Event log
Abstraction
Fitness
precision
Van Kasteren
Van Kasteren
No (Figure 5)
Yes (Figure 6)
0.9111
0.7918
0.3308
0.7804
MIT household A
MIT household A
No (Figure 7)
Yes (Figure 8)
0.9916
0.9880
0.2289
0.3711
MIT household B
MIT household B
No (Figure 9)
Yes (Figure 10)
1.0
0.9305
0.2389
0.4319
nature, such as the activity personal needs. The eight different activities on the heading level branch into 19 different activities on the category level, where personal
needs branches into e.g. eating, sleeping, and personal hygiene. The 19 categories
are divided over 34 subcategories, which contain very specific human activities. At
the subcategory level the category meal cleanup is for example divided into washing dishes and putting away dishes. At the subcategory level there are more types of
Mining Process Model Descriptions of Daily Life through Event Abstraction
17
Fig. 7 Inductive Miner result on the sensor-level MIT household A event log.
Fig. 8 Inductive Miner result on the discovered human activity level events on the MIT household
A log.
human activities than there are sensors-level activities, which makes the abstraction
task very hard. Therefore, we set the target label to the category level.
Figure 7 shows the model discovered with the Inductive Miner on the sensor-level
events of the MIT household A log. The model obtained allows for too much behavior, as it contains two large choice blocks. We found a Damerau-Levenshtein similarity of 0.6348 in the Leave-One-Trace-Out-Cross-Validation experiment. Note
that the abstraction accuracy on this log is lower than the abstraction accuracy on
the Van Kasteren event log. However, the MIT household A log contains more different types of human activity, resulting in a more difficult prediction task with a
higher number of possible target classes. Figure 10 shows the process model discovered from the human activity level traces that we predicted from the sensor-level
events. Even though the model is too large to print in a readable way, from its shape
it is clear that the abstracted model is much more behaviorally constrained than the
sensor-level model. The precision and fitness values in Table 3 show that indeed the
process model after abstraction has become behaviorally more specific while the
portion of behavior of the data that fits the process model remains more or less the
same.
18
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
Fig. 9 Inductive Miner result on the sensor-level MIT household B event log.
Fig. 10 Inductive Miner result on the discovered human activity level events on the MIT household
B log
6.4 Case Study 3: MIT Household B Event Log
For the third case study we use the data of household B of the MIT smart home
experiment [30]. Household B contains data of 17 days of living, 1962 sensorlevel events registered by 20 different sensors. Identically to MIT household A the
human-level activities are provided as a three-level taxonomy. Again, we use the
subcategory level of this taxonomy as target activity label.
The model discovered with the Inductive Miner [18] from the sensor-level events
is shown in Figure 9. The model obtained allows for too much behavior, as it
contains two large choice blocks. We found a Damerau-Levenshtein similarity of
0.5865 in the Leave-One-Trace-Out-Cross-Validation experiment, which is lower
than the similarity found on the MIT A data set while the target classes of the abstraction are the same for the two data sets. This can be explained by the fact that
there is less training data for this event log, as household B contains 1932 sensorlevel events where household A contains 2701 sensor-level events. Figure 10 shows
the process model discovered from abstracted log. Again this model is not readable
due to its size, but its shape shows it to be behaviorally quite specific. The precision
and fitness values in Table 3 also that process model after abstraction has indeed
Mining Process Model Descriptions of Daily Life through Event Abstraction
19
become behaviorally more specific while the portion of behavior of the data that fits
the process model decreased only slightly.
7 Conclusion
In this paper we presented a framework to abstract events using supervised learning
which has been implemented in the ProM process mining toolkit. An important
part of the framework is a generic way to extract useful features for abstraction
from the extensions defined in the XES IEEE standard for event logs. We propose
the Damerau-Levenshtein Similarity for evaluation of the abstraction results, and
motivate why it fits the application of process mining. Finally, we showed on three
real life smart home data sets that application of the supervised event abstraction
framework enables us to mine more precise process model description of human
life compared to what could be mined from the original data on the sensor-level.
Additionally, these process models contain interpretable event labels on the human
behavior activity level.
References
[1] van der Aalst WMP (2016) Process Mining: Data Science in Action. Springer
[2] van der Aalst WMP, Weijters T, Maruster L (2004) Workflow mining: Discovering process models from event logs. IEEE Transactions on Knowledge and
Data Engineering 16(9):1128–1142
[3] Bao L, Intille SS (2004) Activity recognition from user-annotated acceleration
data. In: Pervasive Computing, LNCS, Springer, pp 1–17
[4] Bose RPJC, van der Aalst WMP (2009) Abstractions in process mining: A
taxonomy of patterns. In: Business Process Management, LNCS, Springer, pp
159–175
[5] Chen L, Nugent C (2009) Ontology-based activity recognition in intelligent
pervasive environments. International Journal of Web Information Systems
5(4):410–430
[6] Cohen T, Welling M (2015) Harmonic exponential families on manifolds.
In: Proceedings of The 32nd International Conference on Machine Learning,
JMLR Workshop & Conference Proceedings, pp 1757–1765
[7] Conforti R, Dumas M, Garcı́a-Bañuelos L, La Rosa M (2016) BPMN miner:
Automated discovery of bpmn process models with hierarchical structure. Information Systems 56:284–303
[8] Damerau FJ (1964) A technique for computer detection and correction of
spelling errors. Communications of the ACM 7(3):171–176
20
N. Tax, N. Sidorova, R. Haakma, W.M.P. van der Aalst
[9] van Dongen BF, Adriansyah A (2010) Process mining: Fuzzy clustering and
performance visualization. In: Business Process Management Workshops,
Springer, LNBIP, pp 158–169
[10] Friedman N, Geiger D, Goldszmidt M (1997) Bayesian network classifiers.
Machine Learning 29(2-3):131–163
[11] Günther CW, Rozinat A, van der Aalst WMP (2010) Activity mining by global
trace segmentation. In: Business Process Management Workshops, Springer,
LNBIP, pp 128–139
[12] IEEE (2016) IEEE standard for eXtensible Event Stream (XES) for achieving
interoperability in event logs and event streams. IEEE Std 1849-2016 pp 1–50,
DOI 10.1109/IEEESTD.2016.7740858
[13] van Kasteren T, Kröse B (2007) Bayesian activity recognition in residence for
elders. In: Proceedings of the 3rd IET International Conference on Intelligent
Environments, IEEE, pp 209–212
[14] van Kasteren T, Noulas A, Englebienne G, Kröse B (2008) Accurate activity
recognition in a home setting. In: Proceedings of the 10th International Conference on Ubiquitous Computing, ACM, pp 1–9
[15] Kim E, Helal S, Cook D (2010) Human activity recognition and pattern discovery. Pervasive Computing 9(1):48–53
[16] Kwapisz JR, Weiss GM, Moore SA (2011) Activity recognition using cell
phone accelerometers. ACM SIGKDD Explorations Newsletter 12(2):74–82
[17] Lafferty J, McCallum A, Pereira FCN (2001) Conditional random fields: probabilistic models for segmenting and labeling sequence data. In: Proceedings of
the 18th International Conference on Machine Learning, Morgan Kaufmann
[18] Leemans SJJ, Fahland D, van der Aalst WMP (2013) Discovering blockstructured process models from event logs - a constructive approach. In: Application and Theory of Petri Nets and Concurrency, LNCS, Springer, pp 311–
329
[19] Levenshtein VI (1966) Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady 10:707–710
[20] Mannhardt F, de Leoni M, Reijers HA, van der Aalst WMP, Toussaint PJ
(2016) From low-level events to activities - a pattern-based approach. In: International Conference on Business Process Management, Springer, pp 125–141
[21] Munoz-Gama J, Carmona J (2010) A fresh look at precision in process conformance. In: International Conference on Business Process Management,
Springer, pp 211–226
[22] Poppe R (2010) A survey on vision-based human action recognition. Image
and Vision Computing 28(6):976–990
[23] Rabiner LR, Juang BH (1986) An introduction to hidden Markov models.
ASSP Magazine 3(1):4–16
[24] Reisig W (2012) Petri nets: an introduction, vol 4. Springer Science & Business Media
[25] Riboni D, Bettini C (2011) OWL 2 modeling and reasoning with complex
human activities. Pervasive and Mobile Computing 7(3):379–395
Mining Process Model Descriptions of Daily Life through Event Abstraction
21
[26] Rozinat A, van der Aalst WMP (2008) Conformance checking of processes
based on monitoring real behavior. Information Systems 33(1):64–95
[27] Schwarz G (1978) Estimating the dimension of a model. The Annals of Statistics 6(2):461–464
[28] Sutton C (2006) GRMM: Graphical models in Mallet. Implementation available at http://mallet{}cs{}umass{}edu/grmm
[29] Sztyler T, Carmona J, Völker J, Stuckenschmidt H (2016) Self-tracking
reloaded: applying process mining to personalized health care from labeled
sensor data. In: Transactions on Petri Nets and Other Models of Concurrency
XI, Springer Berlin Heidelberg, pp 160–180
[30] Tapia EM, Intille SS, Larson K (2004) Activity recognition in the home using
simple and ubiquitous sensors. In: Pervasive Computing, LNCS, Springer, pp
158–175
[31] Tax N, Alasgarov E, Sidorova N, Haakma R (2016) On generation of timebased label refinements. In: Proceedings of the 25th International Workshop
on Concurrency, Specification and Programming, CEUR-WS.org, pp 25–36
[32] Tax N, Sidorova N, van der Aalst WMP, Haakma R (2016) Heuristic approaches for generating local process models through log projections. In: Proceedings of IEEE Symposium on Computational Intelligence and Data Mining, To appear.
[33] Tax N, Sidorova N, Haakma R, van der Aalst WMP (2016) Event abstraction
for process mining using supervised learning techniques. In: Proceedings of
the SAI Intelligent Systems Conference, IEEE, pp 161–170
[34] Tax N, Sidorova N, Haakma R, van der Aalst WMP (2016) Log-based evaluation of label splits for process models. Procedia Computer Science 96:63–72
[35] Verbeek HMW, Buijs JCAM, Van Dongen BF, van der Aalst WMP (2010)
ProM 6: The process mining toolkit. In: Proceedings of the Business Process
Management Demonstration Track, CEUR-WS.org, pp 34–39
[36] Weijters AJMM, Ribeiro JTS (2011) Flexible heuristics miner (FHM). In: Proceedings of the IEEE Symposium on Computational Intelligence and Data
Mining, IEEE, pp 310–317
[37] van Zelst SJ, van Dongen BF, van der Aalst WMP (2015) Avoiding over-fitting
in ILP-based process discovery. In: International Conference on Business Process Management, Springer International Publishing, pp 163–171
| 2cs.AI
|
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
1
Skeleton Based Human Action Recognition with
Global Context-Aware Attention LSTM Networks
arXiv:1707.05740v5 [cs.CV] 11 Jan 2018
Jun Liu, Student Member, IEEE, Gang Wang, Senior Member, IEEE,
Ling-Yu Duan, Member, IEEE, Kamila Abdiyeva, Student Member, IEEE, and Alex C. Kot, Fellow, IEEE
Abstract—Human action recognition in 3D skeleton sequences
has attracted a lot of research attention. Recently, Long ShortTerm Memory (LSTM) networks have shown promising performance in this task due to their strengths in modeling the
dependencies and dynamics in sequential data. As not all skeletal
joints are informative for action recognition, and the irrelevant
joints often bring noise which can degrade the performance, we
need to pay more attention to the informative ones. However,
the original LSTM network does not have explicit attention
ability. In this paper, we propose a new class of LSTM network, Global Context-Aware Attention LSTM (GCA-LSTM), for
skeleton based action recognition, which is capable of selectively
focusing on the informative joints in each frame by using a
global context memory cell. To further improve the attention
capability, we also introduce a recurrent attention mechanism,
with which the attention performance of our network can be
enhanced progressively. Besides, a two-stream framework, which
leverages coarse-grained attention and fine-grained attention, is
also introduced. The proposed method achieves state-of-the-art
performance on five challenging datasets for skeleton based action
recognition.
Index Terms—Action Recognition, Long Short-Term Memory,
Global Context Memory, Attention, Skeleton Sequence.
I. I NTRODUCTION
CTION recognition is a very important research problem
owing to its relevance to a wide range of applications, such as video surveillance, patient monitoring, robotics,
human-machine interaction, etc [1], [2], [3]. With the development of depth sensors, such as RealSense and Kinect [4], [5],
[6], 3D skeleton based human action recognition has received
much attention, and a lot of advanced methods have been
proposed during the past few years [7], [8], [9], [10].
Human actions can be represented by a combination of the
motions of skeletal joints in 3D space [11], [12]. However,
this does not indicate all joints in the skeleton sequence
are informative for action recognition. For instance, the hand
joints’ motions are quite informative for the action clapping,
while the movements of the foot joints are not. Different action
sequences often have different informative joints, and in the
same sequence, the informativeness degree of a body joint
may also change over the frames. Thus, it is beneficial to
selectively focus on the informative joints in each frame of the
A
J. Liu, K. Abdiyeva, and A. C. Kot are with School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, 639798.
E-mail: {jliu029, abdi0001, eackot}@ntu.edu.sg.
G. Wang is with Alibaba Group, Hangzhou, 310052, China. Email: [email protected].
L.-Y. Duan is with National Engineering Lab for Video Technology, Peking
University, Beijing, 100871, China. Email: [email protected].
Manuscript received July 1, 2017.
sequence, and try to ignore the features of the irrelevant ones,
as the latter contribute very little for action recognition, and
even bring noise which corrupts the performance [13]. This
selectively focusing scheme can also be called attention, which
has been demonstrated to be quite useful for various tasks,
such as speech recognition [14], image caption generation
[15], machine translation [16], and so on.
Long Short-Term Memory (LSTM) networks have strong
power in handling sequential data [17]. They have been
successfully applied to language modeling [18], RGB based
video analysis [19], [20], [21], [22], [23], [24], [25], [26], [27],
and also skeleton based action recognition [12], [28], [29].
However, the original LSTM does not have strong attention
capability for action recognition. This limitation is mainly
owing to LSTM’s restriction in perceiving the global context
information of the video sequence, which is, however, often
very important for the global classification problem – skeleton
based action recognition.
In order to perform reliable attention over the skeletal joints,
we need to assess the informativeness degree of each joint in
each frame with regarding to the global action sequence. This
indicates that we need to have global contextual knowledge
first. However, the available context information at each evolution step of LSTM is relatively local. In LSTM, the sequential
data is fed to the network as input step by step. Accordingly,
the context information (hidden representation) of each step
is fed to the next one. This implies the available context at
each step is the hidden representation from the previous step,
which is quite local when compared to the global information
1
.
In this paper, we extend the original LSTM model and
propose a Global Context-Aware Attention LSTM (GCALSTM) network which has strong attention capability for
skeleton based action recognition. In our method, the global
context information is fed to all evolution steps of the GCALSTM. Therefore, the network can use it to measure the
informativeness scores of the new inputs at all steps, and adjust
the attention weights for them accordingly, i.e., if a new input
is informative regarding to the global action, then the network
takes advantage of more information of it at this step, on the
contrary, if it is irrelevant, then the network blocks the input
at this step.
1 Though in LSTM, the hidden representations of the latter steps contain
wider range of context information than that of the initial steps, their context
is still relatively local, as LSTM has trouble in remembering information too
far in the past [30].
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
2
Second LSTM Layer
Refine
Attention Iteration #1
Attention Iteration #2
Global
Context
Memory
Informativeness Gate
Softmax
Classifier
First LSTM Layer
Initialize
Fig. 1: Skeleton based human action recognition with the Global Context-Aware Attention LSTM network. The first LSTM
layer encodes the skeleton sequence and generates an initial global context representation for the action sequence. The second
layer performs attention over the inputs by using the global context memory cell to achieve an attention representation for the
sequence. Then the attention representation is used back to refine the global context. Multiple attention iterations are performed
to refine the global context memory progressively. Finally, the refined global context information is utilized for classification.
Our proposed GCA-LSTM network for skeleton based
action recognition includes a global context memory cell and
two LSTM layers, as illustrated in Fig. 1. The first LSTM
layer is used to encode the skeleton sequence and initialize
the global context memory cell. And the representation of the
global context memory is then fed to the second LSTM layer to
assist the network to selectively focus on the informative joints
in each frame, and further generate an attention representation
for the action sequence. Then the attention representation is
fed back to the global context memory cell in order to refine
it. Moreover, we propose a recurrent attention mechanism
for our GCA-LSTM network. As a refined global context
memory is produced after the attention procedure, the global
context memory can be fed to the second LSTM layer again
to perform attention more reliably. We carry out multiple
attention iterations to optimize the global context memory
progressively. Finally, the refined global context is fed to the
softmax classifier to predict the action class.
In addition, we also extend the aforementioned design of our
GCA-LSTM network in this paper, and further propose a twostream GCA-LSTM, which incorporates fine-grained (jointlevel) attention and coarse-grained (body part-level) attention,
in order to achieve more accurate action recognition results.
The contributions of this paper are summarized as follows:
• A GCA-LSTM model is proposed, which retains the
sequential modeling ability of the original LSTM, meanwhile promoting its selective attention capability by introducing a global context memory cell.
• A recurrent attention mechanism is proposed, with which
the attention performance of our network can be improved
progressively.
• A stepwise training scheme is proposed to more effectively train the network.
• We further extend the design of our GCA-LSTM model,
and propose a more powerful two-stream GCA-LSTM
network.
The proposed end-to-end network yields state-of-the-art
performance on the evaluated benchmark datasets.
This work is an extension of our preliminary conference
paper [31]. Based on the previous version, we further propose
a stepwise training scheme to train our network effectively and
efficiently. Moreover, we extend our GCA-LSTM model and
further propose a two-stream GCA-LSTM by leveraging finegrained attention and coarse-grained attention. Besides, we
extensively evaluate our method on more benchmark datasets.
More empirical analysis of the proposed approach is also
provided.
The rest of this paper is organized as follows. In Section
II, we review the related works on skeleton based action
recognition. In Section III, we introduce the proposed GCALSTM network. In Section IV, we introduce the two-stream
attention framework. We provide the experimental results in
Section V. Finally, we conclude the paper in Section VI.
•
II. R ELATED W ORK
In this section, we first briefly review the skeleton based
action recognition methods which mainly focus on extracting
hand-crafted features. We then introduce the RNN and LSTM
based methods. Finally, we review the recent works on attention mechanism.
A. Skeleton Based Action Recognition with Hand-crafted Features
In the past few years, different feature extractors and classifier learning methods for skeleton based action recognition
have been proposed [32], [33], [34], [35], [36], [37], [38], [39],
[40], [41], [42], [43], [44].
Chaudhry et al. [45] proposed to encode the skeleton
sequences to spatial-temporal hierarchical models, and then
use linear dynamical systems (LDSs) to learn the dynamic
structures. Vemulapalli et al. [46] represented each action as
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
a curve in a Lie group, and then utlized a support vector
machine (SVM) to classify the actions. Xia et al. [47] proposed
to model the temporal dynamics in action sequences with
the Hidden Markov models (HMMs). Wang et al. [48], [49]
introduced an actionlet ensemble representation to model the
actions meanwhile capturing the intra-class variances. Chen
et al. [50] designed a part-based 5D feature vector to explore
the relevant joints of body parts in skeleton sequences. Koniusz
et al. [51] introduced tensor representations for capturing the
high-order relationships among body joints. Wang et al. [52]
proposed a graph-based motion representation in conjunction
with a SPGK-kernel SVM for skeleton based activity recognition. Zanfir et al. [53] developed a moving pose framework
together with a modified k-NN classifier for low-latency action
recognition.
B. Skeleton Based Action Recognition with RNN and LSTM
Models
Very recently, deep learning, especially recurrent neural
network (RNN), based approaches have shown their strength in
skeleton based action recognition. Our proposed GCA-LSTM
network is based on the LSTM model which is an extension
of RNN. In this part, we review the RNN and LSTM based
methods as below, since they are relevant to our method.
Du et al. [12] introduced a hierarchical RNN model to
represent the human body structure and temporal dynamics
of the joints. Veeriah et al. [54] proposed a differential
gating scheme to make the LSTM network emphasize on the
change of information. Zhu et al. [28] proposed a mixednorm regularization method for the LSTM network in order to
drive the model towards learning co-occurrence features of the
skeletal joints. They also designed an in-depth dropout mechanism to effectively train the network. Shahroudy et al. [55]
introduced a part-aware LSTM model to push the network
towards learning long-term context representations of different
body parts separately. Liu et al. [29], [56] designed a 2D
Spatio-Temporal LSTM framework to concurrently explore the
hidden sources of action related context information in both
temporal and spatial domains. They also introduced a trust gate
mechanism [29] to deal with the inaccurate 3D coordinates of
skeletal joints provided by the depth sensors.
Beside action recognition, RNN and LSTM models have
also been applied to skeleton based action forecasting [57]
and detection [58], [57].
Different from the aforementioned RNN/LSTM based approaches, which do not explicitly consider the informativeness
of each skeletal joint with regarding to the global action sequence, our proposed GCA-LSTM network utilizes the global
context information to perform attention over all the evolution
steps of LSTM to selectively emphasize the informative joints
in each frame, and thereby generates an attention representation for the sequence, which can be used to improve the
classification performance. Furthermore, a recurrent attention
mechanism is proposed to iteratively optimize the attention
performance.
3
C. Attention Mechanism
Our approach is also related to the attention mechanism
[14], [16], [59], [60], [61], [62], [63] which allows the
networks to selectively focus on specific information. Luong
et al. [62] proposed a network with attention mechanism for
neural machine translation. Stollenga et al. [64] designed a
deep attention selective network for image classification. Xu
et al. [15] proposed to incorporate hard attention and soft attention for image caption generation. Yao et al. [65] introduced
a temporal attention model for video caption generation.
Though a series of deep learning based models have been
proposed for video analysis in existing works [66], [67], most
of them did not consider the attention mechanism. There are
several works which explored attention, such as the methods in
[60], [68], [69]. However, our method is significantly different
from them in the following aspects: These works use the
hidden state of the previous time step of LSTM, whose context
information is quite local, to measure the attention scores for
the next time step. For the global classification problem action recognition, the global information is crucial for reliably
evaluating the importance (informativeness) of each input to
achieve a reliable attention. Therefore, we propose a global
context memory cell for LSTM, which is utilized to measure
the informativeness score of the input at each step. Then the
informativeness score is used as a gate (informativeness gate,
similar to the input gate and forget gate) inside the LSTM
unit to adjust the contribution of the input data at each step
for updating the memory cell. To the best of our knowledge,
we are the first to introduce a global memory cell for LSTM
network to handle global classification problems. Moreover,
a recurrent attention mechanism is proposed to iteratively
promote the attention capability of our network, while the
methods in [60], [68], [69] performed attention only once.
In addition, a two-stream attention framework incorporating
fine-grained attention and coarse-grained attention is also
introduced. Owing to the new contributions, our proposed
network yields state-of-the-art performance on the evaluated
benchmark datasets.
III. GCA-LSTM N ETWORK
In this section, we first briefly review the 2D SpatioTemporal LSTM (ST-LSTM) as our base network. We then introduce our proposed Global Context-Aware Attention LSTM
(GCA-LSTM) network in detail, which is able to selectively
focus on the informative joints in each frame of the skeleton
sequence by using global context information. Finally, we
describe our approach to training our network effectively.
A. Spatio-Temporal LSTM
In a generic skeleton based human action recognition problem, the 3D coordinates of the major body joints in each
frame are provided. The spatial dependence of different joints
in the same frame and the temporal dependence of the same
joint among different frames are both crucial cues for skeleton
based action analysis. Very recently, Liu et al. [29] proposed a
2D ST-LSTM network for skeleton based action recognition,
which is capable of modeling the dependency structure and
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
4
(J,T)
Second ST-LSTM Layer
Temporal (frame)
(J,T)
(n)
0.0
hj,t
(j-1,t) (j,t)
0.3
0.0
r j,t(n)
0.1
0.4
0.0
(j,t-1)
Refine
Global Context
Memory (n)
Initialize
(0)
Spatial (joint)
j,t
Fig. 2: Illustration of the ST-LSTM network [29]. In the spatial
direction, the body joints in each frame are arranged as a
chain and fed to the network as a sequence. In the temporal
dimension, the body joints are fed over the frames.
context information in both spatial and temporal domains
simultaneously.
As depicted in Fig. 2, in ST-LSTM model, the skeletal
joints in a frame are arranged and fed as a chain (the spatial
direction), and the corresponding joints over different frames
are also fed in a sequence (the temporal direction).
Specifically, each ST-LSTM unit is fed with a new input
(xj,t , the 3D location of joint j in frame t), the hidden
representation of the same joint at the previous time step
(hj,t−1 ), and also the hidden representation of the previous
joint in the same frame (hj−1,t ), where j ∈ {1, ..., J} and
t ∈ {1, ..., T } denote the indices of joints and frames, respectively. The ST-LSTM unit has an input gate (ij,t ), two forget
gates corresponding to the two sources of context information
(T )
(S)
(fj,t for the temporal dimension, and fj,t for the spatial
domain), together with an output gate (oj,t ).
The transition equations of ST-LSTM are formulated as
presented in [29]:
ij,t
σ
f (S)
σ
xj,t
j,t
(T )
hj−1,t (1)
fj,t =
σ W
σ
hj,t−1
oj,t
tanh
uj,t
cj,t
= ij,t
+
+
hj,t
uj,t
(S)
fj,t
(T )
fj,t
= oj,t
cj−1,t
(2)
cj,t−1
tanh(cj,t )
(3)
where cj,t and hj,t denote the cell state and hidden representation of the unit at the spatio-temporal step (j, t), respectively,
uj,t is the modulated input,
denotes the element-wise
product, and W is an affine transformation consisting of model
parameters. Readers are referred to [29] for more details about
the mechanism of ST-LSTM.
B. Global Context-Aware Attention LSTM
Several previous works [13], [50] have shown that in each
action sequence, there is often a subset of informative joints
which are important as they contribute much more to action
analysis, while the remaining ones may be irrelevant (or even
First ST-LSTM Layer
Fig. 3: Illustration of our GCA-LSTM network. Some arrows
are omitted for clarity.
noisy) for this action. As a result, to obtain a high accuracy of
action recognition, we need to identify the informative skeletal
joints and concentrate more on their features, meanwhile trying
to ignore the features of the irrelevant ones, i.e., selectively
focusing (attention) on the informative joints is useful for
human action recognition.
Human action can be represented by a combination of
skeletal joints’ movements. In order to reliably identify the
informative joints in an action instance, we can evaluate the informativeness score of each joint in each frame with regarding
to the global action sequence. To achieve this purpose, we need
to obtain the global context information first. However, the
available context at each evolution step of LSTM is the hidden
representation from the previous step, which is relatively local
when compared to the global action.
To mitigate the aforementioned limitation, we propose to
introduce a global context memory cell for the LSTM model,
which keeps the global context information of the action
sequence, and can be fed to each step of LSTM to assist the
attention procedure, as illustrated in Fig. 3. We call this new
LSTM architecture as Global Context-Aware Attention LSTM
(GCA-LSTM).
1) Overview of the GCA-LSTM network: We illustrate
the proposed GCA-LSTM network for skeleton based action
recognition in Fig. 3. Our GCA-LSTM network contains three
major modules. The global context memory cell maintains an
overall representation of the whole action sequence. The first
ST-LSTM layer encodes the skeleton sequence, and initializes
the global context memory cell. The second ST-LSTM layer
performs attention over the inputs at all spatio-temporal steps
to generate an attention representation of the action sequence,
which is then used to refine the global context memory.
The input at the spatio-temporal step (j, t) of the first STLSTM layer is the 3D coordinates of the joint j in frame t.
The inputs of the second layer are the hidden representations
from the first layer.
Multiple attention iterations (recurrent attention) are performed in our network to refine the global context memory
iteratively. Finally, the refined global context memory can be
used for classification.
To facilitate our explanation, we use hj,t instead of hj,t to
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
5
denote the hidden representation at the step (j, t) in the first
ST-LSTM layer, while the symbols, including hj,t , cj,t , ij,t ,
and oj,t , which are defined in Section III-A, are utilized to
represent the components in the second layer only.
2) Initializing the Global Context Memory Cell: Our GCALSTM network performs attention by using the global context
information, therefore, we need to obtain an initial global
context memory first.
A feasible scheme is utilizing the outputs of the first layer
to generate a global context representation. We can average
the hidden representations at all spatio-temporal steps of the
first layer to compute an initial global context memory cell
(IF(0) ) as follows:
IF(0) =
J
T
1 XX
hj,t
JT j=1 t=1
(4)
We may also concatenate the hidden representations of the
first layer and feed them to a feed-forward neural network, then
use the resultant activation as IF(0) . We empirically observe
these two initialization schemes perform similarly.
3) Performing Attention in the Second ST-LSTM Layer:
By using the global context information, we evaluate the
informativeness degree of the input at each spatio-temporal
step in the second ST-LSTM layer.
In the n-th attention iteration, our network learns an in(n)
formativeness score (rj,t ) for each input (hj,t ) by feeding
the input itself, together with the global context memory cell
(IF(n−1) ) generated by the previous attention iteration to a
network as follows:
hj,t
(n)
ej,t = We1 tanh We2
(5)
IF(n−1)
(n)
rj,t =
J
P
(n)
exp(ej,t )
T
P
(6)
(n)
exp(eu,v )
u=1 v=1
(n)
where rj,t ∈ (0, 1) denotes the normalized informativeness
score of the input at the step (j, t) in the n-th attention
iteration, with regarding to the global context information.
(n)
The informativeness score rj,t is then used as a gate of the
ST-LSTM unit, and we call it informativeness gate. With the
assistance of the learned informativeness gate, the cell state of
the unit in the second ST-LSTM layer can be updated as:
cj,t
=
(n)
rj,t
ij,t
+ (1 −
+ (1 −
(n)
rj,t )
(n)
rj,t )
uj,t
fj,t
(S)
cj−1,t
(T )
fj,t
cj,t−1
(7)
The cell state updating scheme in Eq. (7) can be explained
as follows: (1) if the input (hj,t ) is informative (important)
with regarding to the global context representation, then we
let the learning algorithm update the cell state of the second
ST-LSTM layer by importing more information of it; (2) on
the contrary, if the input is irrelevant, then we need to block
the input gate at this step, meanwhile relying more on the
history information of the cell state.
4) Refining the Global Context Memory Cell: We perform
attention by adopting the cell state updating scheme in Eq. (7),
and thereby obtain an attention representation of the action
sequence. Concretely, the output of the last spatio-temporal
step in the second layer is used as the attention representation
(F (n) ) for the action. Finally, the attention representation F (n)
is fed to the global context memory cell to refine it, as
illustrated in Fig. 3. The refinement is formulated as follows:
F (n)
(n)
IF(n) = ReLu WF
(8)
IF(n−1)
(n)
where IF(n) is the refined version of IF(n−1) . Note that WF
is not shared over different iterations.
Multiple attention iterations (recurrent attention) are carried
out in our GCA-LSTM network. Our motivation is that after
we obtain a refined global context memory cell, we can use
it to perform the attention again to more reliably identify
the informative joints, and thus achieve a better attention
representation, which can then be utilized to further refine the
global context. After multiple iterations, the global context can
be more discriminative for action classification.
5) Classifier: The last refined global context memory cell
IF(N ) is fed to a softmax classifier to predict the class label:
ŷ = softmax Wc IF(N )
(9)
The negative log-likelihood loss function [70] is adopted to
measure the difference between the true class label y and the
prediction result ŷ. The back-propagation algorithm is used to
minimize the loss function. The details of the back-propagation
process are described in Section III-C.
C. Training the Network
In this part, we first briefly describe the basic training
method which directly optimizes the parameters of the whole
network, we then propose a more advanced stepwise training
scheme for our GCA-LSTM network.
1) Directly Train the Whole Network: Since the classification is performed by using the last refined global context,
to train such a network, it is natural and intuitive to feed the
action label as the training output at the last attention iteration,
and back-propagate the errors from the last step, i.e., directly
optimize the whole network as shown in Fig. 4(a).
2) Stepwise Training: Owing to the recurrent attention
mechanism, there are frequent mutual interactions among
different modules (the two ST-LSTM layers and the global
context memory cell, see Fig. 3) in our network. Moreover,
during the progress of multiple attention iterations, new parameters are also introduced. Due to these facts, it is rather
difficult to simply optimize all parameters and all attention
iterations of the whole network directly as mentioned above.
Therefore, we propose a stepwise training scheme for our
GCA-LSTM network, which optimizes the model parameters
incrementally. The details of this scheme are depicted in
Fig. 4(b) and Algorithm 1.
The proposed stepwise training scheme is effective and
efficient in optimizing the parameters and ensuring the convergence of the GCA-LSTM network. Specifically, at each
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
(N)
6
(N)
(N)
(N)
Softmax
(N)
Softmax
(N)
(1)
(1)
(1)
(1)
(0)
(0)
(0)
(1)
(0)
Softmax
(0)
(0)
(0)
Training Step #0
(a)
(1)
Softmax
(1)
(0)
Training Step #1
Training Step #N
(b)
Fig. 4: Illustration of the two network training methods. (a) Directly train the whole network. (b) Stepwise optimize the network
parameters. In this figure, the global context memory cell IF(n) is unfolded over the attention iterations. The training step #n
corresponds to the n-th attention iteration. The black and red arrows denote the forward and backward passes, respectively.
Some passes, such as those between the two ST-LSTM layers, are omitted for clarity. Better viewed in colour.
Algorithm 1 Stepwise train the GCA-LSTM network.
1:
2:
3:
4:
5:
6:
7:
Randomly initialize the parameters of the whole network
with zero-mean Gaussian.
for n = 0 to N do // n is the training step
Feed the action label as the training output at the
attention iteration n.
do
Training an epoch: optimizing the parameters used
in the iterations 0 to n via back-propagation.
while Validation error is decreasing
end for
training step n, we only need to optimize a subset of parameters and modules which are used by the attention iterations 0
to n. 2 Training this shrunken network is more effective and
efficient than directly training the whole network. At the step
n + 1, a larger scale network needs to be optimized. However,
the training at step n + 1 is also very efficient, as most of
the parameters and passes have already been optimized (pretrained well) by its previous training steps.
IV. T WO - STREAM GCA-LSTM N ETWORK
In the aforementioned design (Section III), the GCA-LSTM
network performs action recognition by selectively focusing on
the informative joints in each frame, i.e., the attention is carried
out at joint level (fine-grained attention). Beside fine-grained
attention, coarse-grained attention can also contribute to action
analysis. This is because some actions are often performed
at body part level. For these actions, all the joints from the
same informative body part tend to have similar importance
degrees. For example, the postures and motions of all the
joints (elbow, wrist, palm, and finger) from the right hand
are all important for recognizing the action salute in the NTU
RGB+D dataset [55], i.e., we need to identify the informative
2 Note
that #0 is not an attention iteration, but the process of initializing
the global context memory cell (IF(0) ). To facilitate the explantation of the
stepwise training, we here temporally describe it as an attention iteration.
body part “right hand” here. This implies coarse-grained (body
part-level) attention is also useful for action recognition.
As suggested by Du et al. [12], the human skeleton can
be divided into five body parts (torso, left hand, right hand,
left leg, and right leg) based on the human physical structure.
These five parts are illustrated as the right part of Fig. 5.
Therefore, we can measure the informativeness degree of each
body part with regarding to the action sequence, and then
perform coarse-grained attention.
Specifically, we extend the design of out GCA-LSTM
model, and introduce a two-stream GCA-LSTM network here,
which jointly takes advantage of a fine-grained (joint-level) attention stream and a coarse-grained (body part-level) attention
stream.
The architecture of the two-stream GCA-LSTM is illustrated in Fig. 5. In each attention stream, there is a global
context memory cell to maintain the global attention representation of the action sequence, and also a second STLSTM layer to perform attention. This indicates we have
two separated global context memory cells in the whole
architecture, which are respectively the fine-grained attention
(n)
memory cell (IF(F ) ) and the coarse-grained attention memory
(n)
cell (IF(C) ). The first ST-LSTM layer, which is used to encode
the skeleton sequence and initialize the global context memory
cells, is shared by the two attention streams.
The process flow (including initialization, attention, and
refinement) in the fine-grained attention stream is the same
as the GCA-LSTM model introduced in Section III. The
operation in the coarse-grained attention stream is also similar.
The main difference is that, in the second layer, the coarsegrained attention stream performs attention by selectively
focusing on the informative body parts in each frame.
Concretely, in the attention iteration n, the network learns
(n)
an informativeness score (rP,t ) for each body part P (P ∈
{1, 2, 3, 4, 5}) as:
(n)
eP,t
= W e3
tanh We4
h̄P,t
(n−1)
IF(C)
!!!
(10)
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
7
Fine-Grained Attention Stream
Second ST-LSTM Layer (Fine-Grained)
hj,t
(F)
r j,t(n)
0.0
0.3
0.0
0.1
Classifier
(n)
(F)
(J,T)
0.4
Coarse-Grained Attention Stream
(n)
(C)
Refine
0.0
Second ST-LSTM Layer (Coarse-Grained)
(J,T)
Refine
hj,t
(C)
Fine-Grained
Attention (n)
Memory (F)
Coarse-Grained
Attention
(n)
Memory (C)
Initialize
Initialize
0.1
0.2
3
2
4
3
7
8
5
1
6
10
0.1
0.3
2
4
7
8
5
9
1
6
13
10
9
13
j,t
14
11
11
First ST-LSTM Layer
15
12
12
14
15
Fig. 5: Illustration of the two-stream GCA-LSTM network, which incorporates fine-grained (joint-level) attention and coarsegrained (body part-level) attention. To perform coarse-grained attention, the joints in a skeleton are divided into five body parts,
and all the joints from the same body part share a same informative score. In the second ST-LSTM layer for coarse-grained
attention, we only show two body parts at each frame, and other body parts are omitted for clarity.
(n)
(n)
rP,t =
exp(eP,t )
5 P
T
P
(n)
exp(eu,v )
(11)
u=1 v=1
where h̄P,t is the representation of the body part P at frame
t, which is calculated based on the hidden representations of
all the joints that belong to P , with average pooling as:
1 X
h̄P,t =
hj,t
(12)
JP
and Berkeley MHAD [73] datasets. To investigate the effectiveness of our approach, we conduct extensive experiments
with the following different network structures:
•
j∈P
where JP denotes the number of joints in body part P .
To perform coarse-grained attention, we allow each joint j
in body part P to share the informativeness degree of P , i.e.,
at frame t, all the joints in P use the same informativeness
(n)
score rP,t , as illustrated in Fig. 5. Hence, in the coarse-grained
attention stream, if j ∈ P , then the cell state of the second
ST-LSTM layer is updated at the spatio-temporal step (j, t)
as:
cj,t
=
(n)
rP,t
ij,t
+ (1 −
+ (1 −
(n)
rP,t )
(n)
rP,t )
uj,t
•
•
(S)
cj−1,t
(T )
cj,t−1
fj,t
fj,t
(13)
Multiple attention iterations are also performed in the proposed two-stream GCA-LSTM network. Finally, the refined
(N )
fine-grained attention memory IF(F ) and coarse-grained at-
“ST-LSTM + Global (1)”. This network architecture
is similar to the original two-layer ST-LSTM network
in [29], but the hidden representations at all spatiotemporal steps of the second layer are concatenated and
fed to a one-layer feed-forward network to generate a
global representation of the skeleton sequence, and the
classification is performed on the global representation;
while in [29], the classification is performed on single
hidden representation at each spatio-temporal step (local
representation).
“ST-LSTM + Global (2)”. This network structure is
similar to the above “ST-LSTM + Global (1)”, except
that the global representation is obtained by averaging
the hidden representations of all spatio-temporal steps.
“GCA-LSTM”. This is the proposed Global ContextAware Attention LSTM network. Two attention iterations
are performed by this network. The classification is
performed on the last refined global context memory cell.
The two training methods (direct training and stepwise
training) described in Section III-C are also evaluated for
this network structure.
(N )
tention memory IF(C) are both fed to the softmax classifier,
and the prediction scores of these two streams are averaged
for action recognition.
The proposed step-wise training scheme can also be applied
to this two-stream GCA-LSTM network, and at the training step #n, we simultaneously optimize the two attention
streams, both of which correspond to the n-th attention iteration.
V. E XPERIMENTS
We evaluate our proposed method on the NTU RGB+D [55],
SYSU-3D [71], UT-Kinect [47], SBU-Kinect Interaction [72],
In addition, we also adopt the large scale NTU RGB+D and
the challenging SYSU-3D as two major benchmark datasets
to evaluate the proposed “two-stream GCA-LSTM” network.
We use Torch7 framework [74] to perform our experiments.
Stochastic gradient descent (SGD) algorithm is adopted to
train our end-to-end network. We set the learning rate, decay
rate, and momentum to 1.5×10−3 , 0.95, and 0.9, respectively.
The applied dropout probability [75] in our network is set to
0.5. The dimensions of the global context memory representation and the cell state of ST-LSTM are both 128.
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
8
TABLE II: Performance of the two-stream GCA-LSTM network on the NTU RGB+D dataset.
A. Experiments on the NTU RGB+D Dataset
The NTU RGB+D dataset [55] was collected with Kinect
(V2). It contains more than 56 thousand video samples. A
total of 60 action classes were performed by 40 different
subjects. To the best of our knowledge, this is the largest
publicly available dataset for RGB+D based human action
recognition. The large variations in subjects and viewpoints
make this dataset quite challenging. There are two standard
evaluation protocols for this dataset: (1) Cross subject (CS):
20 subjects are used for training, and the remaining subjects
are used for testing; (2) Cross view (CV): two camera views
are used for training, and one camera view is used for testing.
To extensively evaluate the proposed method, both protocols
are tested in our experiment.
We compare the proposed GCA-LSTM network with stateof-the-art approaches, as shown in TABLE I. We can observe
that our proposed GCA-LSTM model outperforms the other
skeleton-based methods. Specifically, our GCA-LSTM network outperforms the original ST-LSTM network in [29] by
6.9% with the cross subject protocol, and 6.3% with the cross
view protocol. This demonstrates that the attention mechanism
in our network brings significant performance improvement.
Both “ST-LSTM + Global (1)” and “ST-LSTM + Global
(2)” perform classification on the global representations, thus
they achieve slightly better performance than the original STLSTM [29] which performs classification on local representations. We also observe “ST-LSTM + Global (1)” and “STLSTM + Global (2)” perform similarly.
The results in TABLE I also show that using the stepwise
training method can improve the performance of our network
in contrast to using the direct training method.
Method
GCA-LSTM (coarse-grained only)
GCA-LSTM (fine-grained only)
Two-stream GCA-LSTM
Two-stream GCA-LSTM with stepwise training
CS
74.1%
74.3%
76.2%
77.1%
CV
81.6%
82.8%
84.7%
85.1%
B. Experiments on the SYSU-3D Dataset
The SYSU-3D dataset [71], which contains 480 skeleton
sequences, was collected with Kinect. This dataset includes
12 action classes which were performed by 40 subjects. The
SYSU-3D dataset is very challenging as the motion patterns
are quite similar among different action classes, and there are
lots of viewpoint variations in this dataset.
We follow the standard cross-validation protocol in [71]
on this dataset, in which 20 subjects are adopted for training
the network, and the remaining subjects are kept for testing.
We report the experimental results in TABLE III. We can
observe that our GCA-LSTM network surpasses the stateof-the-art skeleton-based methods in [80], [71], [29], which
demonstrates the effectiveness of our approach in handling the
task of action recognition in skeleton sequences. The results
also show that our proposed stepwise training scheme is useful
for our network.
TABLE III: Experimental results on the SYSU-3D dataset.
Method
LAFF (SKL) [80]
Dynamic Skeletons [71]
ST-LSTM [56]
ST-LSTM + Global (1)
ST-LSTM + Global (2)
GCA-LSTM (direct training)
GCA-LSTM (stepwise training)
Accuracy
54.2%
75.5%
76.5%
76.8%
76.6%
77.8%
78.6%
TABLE I: Experimental results on the NTU RGB+D dataset.
Method
Skeletal Quads [76]
Lie Group [46]
Dynamic Skeletons [71]
HBRNN [12]
Deep RNN [55]
Deep LSTM [55]
Part-aware LSTM [55]
JTM CNN [77]
STA Model [68]
SkeletonNet [78]
Visualization CNN [79]
ST-LSTM [29]
ST-LSTM + Global (1)
ST-LSTM + Global (2)
GCA-LSTM (direct training)
GCA-LSTM (stepwise training)
CS
38.6%
50.1%
60.2%
59.1%
56.3%
60.7%
62.9%
73.4%
73.4%
75.9%
76.0%
69.2%
70.5%
70.7%
74.3%
76.1%
CV
41.4%
52.8%
65.2%
64.0%
64.1%
67.3%
70.3%
75.2%
81.2%
81.2%
82.6%
77.7%
79.5%
79.4%
82.8%
84.0%
Using this challenging dataset, we also evaluate the performance of the two-stream attention model. The results in
TABLE IV show that the two-stream GCA-LSTM network is
effective for action recognition.
TABLE IV: Performance of the two-stream GCA-LSTM network on the SYSU-3D dataset.
Method
GCA-LSTM (coarse-grained only)
GCA-LSTM (fine-grained only)
Two-stream GCA-LSTM
Two-stream GCA-LSTM with stepwise training
Accuracy
76.9%
77.8%
78.8%
79.1%
C. Experiments on the UT-Kinect Dataset
We also evaluate the performance of the two-stream GCALSTM network, and report the results in TABLE II. The results
show that by incorporating fine-grained attention and coarsegrained attention, the proposed two-stream GCA-LSTM network achieves better performance than the GCA-LSTM with
fine-grained attention only. We also observe the performance
of two-stream GCA-LSTM can be improved with the stepwise
training method.
The UT-Kinect dataset [47] was recorded with a stationary
Kinect. The skeleton sequences in this dataset are quite noisy.
A total of 10 action classes were performed by 10 subjects,
and each action was performed by the same subject twice.
We follow the standard leave-one-out-cross-validation protocol in [47] to evaluate our method on this dataset. Our
approach yields state-of-the-art performance on this dataset,
as shown in TABLE V.
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
TABLE V: Experimental results on the UT-Kinect dataset.
Method
Grassmann Manifold [81]
Histogram of 3D Joints [47]
Riemannian Manifold [82]
Key-Pose-Motifs Mining [83]
Action-Snippets and Activated Simplices [84]
ST-LSTM [29]
ST-LSTM + Global (1)
ST-LSTM + Global (2)
GCA-LSTM (direct training)
GCA-LSTM (stepwise training)
Accuracy
88.5%
90.9%
91.5%
93.5%
96.5%
97.0%
97.0%
97.5%
98.5%
99.0%
D. Experiments on the SBU-Kinect Interaction Dataset
The SBU-Kinect Interaction dataset [72] includes 8 action
classes for two-person interaction recognition. This dataset
contains 282 sequences corresponding to 6822 frames. The
SBU-Kinect Interaction dataset is challenging because of (1)
the relatively low accuracies of the coordinates of skeletal
joints recorded by Kinect, and (2) complicated interactions
between two persons in many action sequences.
We perform 5-fold cross-validation evaluation on this
dataset by following the standard protocol in [72]. The experimental results are depicted in TABLE IX. In this table,
HBRNN [12], Deep LSTM [28], Co-occurrence LSTM [28],
and ST-LSTM [29] are all LSTM based models for action
recognition in skeleton sequences, and are very relevant to our
network. We can see that the proposed GCA-LSTM network
achieves the best performance among all of these methods.
9
the recurrent attention mechanism proposed by us is useful
for the GCA-LSTM network.
Specifically, we also evaluate the performance of 3 attention
iterations by using the large scale NTU RGB+D dataset, and
the results are shown in TABLE VIII. We find the performance
of 3 attention iterations is slightly better than 2 iterations
if we share the parameters over different attention iterations
(see columns (a) and (b) in TABLE VIII). This consistently
shows using multiple attention iterations can improve the
performance of our network progressively. We do not try more
iterations due to the GPU’s memory limitation.
We also find that if we do not share the parameters over
different attention iterations (see columns (c) and (d) in
TABLE VIII), then too many iterations can bring performance
degradation (the performance of using 3 iterations is worse
than that of using 2 iterations). In our experiment, we observe
the performance degradation is caused by over-fitting (increasing iteration number will introduce new parameters if we do
not share parameters). But the performance of two iterations is
still significantly better than one iteration in this case. We will
also give the experimental analysis of the parameter sharing
schemes detailed in Section V-G.
G. Evaluation of Parameter Sharing Schemes
The Berkeley MHAD dataset was recorded by using a
motion capture system. It contains 659 sequences and 11
action classes, which were performed by 12 different subjects.
We adopt the standard experimental protocol on this dataset,
in which 7 subjects are used for training and the remaining
5 subjects are held out for testing. The results in TABLE X
show that our method achieves very high accuracy (100%) on
this dataset.
As the Berkeley MHAD dataset was collected with a motion
capture system rather than a Kinect, thus the coordinates of
the skeletal joints are relatively accurate. To evaluate the robustness with regarding to the input noise, we also investigate
the performance of our GCA-LSTM network on this dataset
by adding zero mean input noise to the skeleton sequences,
and show the results in TABLE VI. We can see that even if we
add noise with the standard deviation (σ) set to 12cm (which
is significant noise in the scale of human body), the accuracy
of our method is still very high (92.7%). This demonstrates
that our method is quite robust against the input noise.
As formulated in Eq. (5), the model parameters We1 and
We2 are introduced for calculating the informativeness score
at each spatio-temporal step in the second layer. Also multiple
attention iterations are carried out in this layer. To regularize
the parameter number inside our network and improve the
generalization capability, we investigate two parameter sharing
strategies for our network: (1) Sharing within iteration: We1
and We2 are shared by all spatio-temporal steps in the same
attention iteration; (2) Sharing cross iterations: We1 and We2
are shared over different attention iterations. We investigate
the effect of these two parameter sharing strategies on our
GCA-LSTM network, and report the results in TABLE VIII.
In TABLE VIII, we can observe that: (1) Sharing parameters
within iteration is useful for enhancing the generalization
capability of our network, as the performance in columns (b)
and (d) of TABLE VIII is better than (a) and (c), respectively.
(2) Sharing parameters over different iterations is also helpful
for handling the over-fitting issues, but it may limit the
representation capacity, as the network with two attention
iterations which shares parameters within iteration but does
not share parameters over iterations achieves the best result
(see column (d) of TABLE VIII). As a result, in our GCALSTM network, we only share the parameters within iteration,
and two attention iterations are used.
F. Evaluation of Attention Iteration Numbers
H. Evaluation of Training Methods
We also test the effect of different attention iteration numbers on our GCA-LSTM network, and show the results in
TABLE VII. We can observe that increasing the iteration
number can help to strength the classification performance
of our network (using 2 iterations obtains higher accuracies
compared to using only 1 iteration). This demonstrates that
The previous experiments showed that using the stepwise
training method can improve the performance of our network
in contrast to using direct training (see TABLE I, V, III, IX).
To further investigate the performance of these two training
methods, we plot the convergence curves of our GCA-LSTM
network in Fig. 6.
E. Experiments on the Berkeley MHAD Dataset
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
10
TABLE VI: Evaluation of robustness against the input noise. Gaussian noise N (0, σ 2 ) is added to the 3D coordinates of the
skeletal joints.
Standard deviation (σ) of noise
Accuracy
0.1cm
100%
1cm
99.3%
2cm
98.5%
4cm
97.5%
8cm
95.6%
12cm
92.7%
16cm
80.4%
32cm
61.5%
TABLE VII: Performance comparison of different attention iteration numbers (N ).
#Attention Iteration
1
2
NTU RGB+D (CS)
72.9%
76.1%
NTU RGB+D (CV)
81.8%
84.0%
UT-Kinect
98.0%
99.0%
SYSU-3D
77.8%
78.6%
Berkeley MHAD
100%
100%
TABLE VIII: Performance comparison of different parameter sharing schemes.
#Attention Iteration
1
2
3
(a)
w/o sharing within iteration
w/ sharing cross iterations
71.0%
73.0%
73.1%
(b)
w/ sharing within iteration
w/ sharing cross iterations
72.9%
74.3%
74.4%
(c)
w/o sharing within iteration
w/o sharing cross iterations
71.0%
73.4%
69.3%
(d)
w/ sharing within iteration
w/o sharing cross iterations
72.9%
76.1%
73.2%
We analyze the convergence curves (Fig. 6) of the stepwise
training method as follows. By using the proposed stepwise
training method, at the training step #0, we only need to train
the subnetwork for initializing the global context (IF(0) ), i.e.,
only a subset of parameters and modules need to be optimized,
thus the training is very efficient and the loss curve converges
very fast. When the validation loss stops decreasing, we start
the next training step #1. Step #1 contains new parameters
and modules for the first attention iteration, which have not
been optimized yet, therefore, loss increases immediately at
this epoch. However, most of the parameters involved at this
step have already been pre-trained well by the previous step
#0, thus the network training is quite effective, and the loss
drops to a very low value after only one training epoch.
By comparing the convergence curves of the two training
methods, we can find (1) the network converges much faster if
we use stepwise training, compared to directly train the whole
network. We can also observe that (2) the network is easier
to get over-fitted by using direct training method, as the gap
between the train loss and validation loss starts to rise after the
20th epoch. These observations demonstrate that the proposed
stepwise training scheme is quite useful for effectively and
efficiently training our GCA-LSTM network.
TABLE X: Experimental results on the Berkeley MHAD
dataset
I. Evaluation of Initialization Methods and Attention Designs
Fig. 6: Convergence curves of the GCA-LSTM network with
two attention iterations by respectively using stepwise training
(in red) and direct training (in green) on the NTU RGB+D
dataset. Better viewed in colour.
In Section III-B2, we introduce two methods to initialize
the global context memory cell (IF(0) ). The first is averaging
TABLE IX: Experimental results on the SBU-Kinect Interaction dataset.
Method
Yun et al. [72]
CHARM [85]
Ji et al. [86]
HBRNN [12]
Deep LSTM [28]
Co-occurrence LSTM [28]
SkeletonNet [78]
ST-LSTM [29]
GCA-LSTM (direct training)
GCA-LSTM (stepwise training)
Accuracy
80.3%
83.9%
86.9%
80.4%
86.0%
90.4%
93.5%
93.3%
94.1%
94.9%
Method
Ofli et al. [43]
Vantigodi et al. [87]
Vantigodi et al. [88]
Kapsouras et al. [89]
ST-LSTM [29]
GCA-LSTM (direct training)
GCA-LSTM (stepwise training)
Accuracy
95.4%
96.1%
97.6%
98.2%
100%
100%
100%
Training step #0
5
Training step #1
4.5
Training step #2
4
3.5
3
2.5
2
1.5
1
0.5
Epoch
0
1
8
15
22
29
36
43
50
Stepwise training (train loss)
Stepwise training (validation loss)
Direct training (train loss)
Direct training (validation loss)
the hidden representations of the first layer (see Eq. (4)),
and the second is using a one-layer feed-forward network to
obtain IF(0) . We compare these two initialization methods in
TABLE XI. The results show that these two methods perform
similarly. In our experiment, we also find that by using feedforward network, the model converges faster, thus the scheme
of feed-forward network is used to initialize the global context
memory cell in our GCA-LSTM network.
(n)
In the GCA-LSTM network, the informativeness score rj,t
is used as a gate within LSTM neuron, as formulated in Eq.
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
TABLE XI: Performance comparison of different methods of
initializing the global context memory cell.
Method
Averaging
Feed-forward network
NTU RGB+D (CS)
73.8%
74.3%
NTU RGB+D (CV)
83.1%
82.8%
(7). We also explore to replace this scheme with soft attention
method [15], [62], i.e., the attention representation F (n) is
PJ PT
(n)
calculated as
j=1
t=1 rj,t hj,t . Using soft attention, the
accuracy drops about one percentage point on the NTU
RGB+D dataset. This can be explained as equipping LSTM
(n)
neuron with gate rj,t provides LSTM better insight about
when to update, forget or remember. In addition, it can keep
the sequential ordering information of the inputs hj,t , while
soft attention loses ordering and positional information.
J. Visualizations
To better understand our network, we analyze and visualize
the informativeness score evaluated by using the global context
information on the large scale NTU RGB+D dataset in this
section.
We analyze the variations of the informativeness scores
over the two attention iterations to verify the effectiveness
of the recurrent attention mechanism in our method, and
show the qualitative results of three actions (taking a selfie,
pointing to something, and kicking other person) in Fig. 7.
The informativeness scores are computed with soft attention
for visualization. In this figure, we can see that the attention
performance increases between the two attention iterations. In
the first iteration, the network tries to identify the potential
informative joints over the frames. After this attention, the
network achieves a good understanding of the global action.
Then in the second iteration, the network can more accurately
focus on the informative joints in each frame of the skeleton
sequence. We can also find that the informativeness score of
the same joint can vary in different frames. This indicates that
our network performs attention not only in spatial domain, but
also in temporal domain.
In order to further quantitatively evaluate the effectiveness
of the attention mechanism, we analyze the classification
accuracies of the three action classes in Fig. 7 among all the
actions. We observe if the attention mechanism is not used, the
accuracies of these three classes are 67.7%, 71.7%, and 81.5%,
respectively. However, if we use one attention iteration, the
accuracies rise to 67.8%, 72.4%, and 83.4%, respectively. If
two attention iterations are performed, the accuracies become
67.9%, 73.6%, and 86.6%, respectively.
To roughly explore which joints are more informative for
the activities in the NTU RGB+D dataset, we also average
the informativeness scores of the same joint in all the testing
sequences, and visualize it in Fig. 8. We can observe that
averagely, more attention is assigned to the hand and foot
joints. This is because in the NTU RGB+D dataset, most of the
actions are related to the hand and foot postures and motions.
We can also find that the average informativeness score of the
right hand joint is higher than that of left hand joint. This
indicates most of the subjects are right-handed.
11
VI. C ONCLUSION
In this paper, we have extended the original LSTM network
to construct a Global Context-Aware Attention LSTM (GCALSTM) network for skeleton based action recognition, which
has strong ability in selectively focusing on the informative joints in each frame of the skeleton sequence with the
assistance of global context information. Furthermore, we
have proposed a recurrent attention mechanism for our GCALSTM network, in which the selectively focusing capability
is improved iteratively. In addition, a two-stream attention
framework is also introduced. The experimental results validate the contributions of our approach by achieving state-ofthe-art performance on five challenging datasets.
ACKNOWLEDGEMENT
This work was carried out at the Rapid-Rich Object Search
(ROSE) Lab at Nanyang Technological University (NTU), Singapore. The ROSE Lab is supported by the National Research
Foundation, Singapore, under its Interactive Digital Media
(IDM) Strategic Research Programme. We acknowledge the
support of NVIDIA AI Technology Centre (NVAITC) for the
donation of the Tesla K40 and K80 GPUs used for our research
at the ROSE Lab. Jun Liu would like to thank Qiuhong Ke
from University of Western Australia for helpful discussions.
R EFERENCES
[1] J. Zheng, Z. Jiang, and R. Chellappa, “Cross-view action recognition
via transferable dictionary learning,” IEEE Transactions on Image Processing, 2016.
[2] F. Liu, X. Xu, S. Qiu, C. Qing, and D. Tao, “Simple to complex
transfer learning for action recognition,” IEEE Transactions on Image
Processing, 2016.
[3] Y.-G. Jiang, Q. Dai, W. Liu, X. Xue, and C.-W. Ngo, “Human action
recognition in unconstrained videos by explicit motion modeling,” IEEE
Transactions on Image Processing, 2015.
[4] J. Han, L. Shao, D. Xu, and J. Shotton, “Enhanced computer vision with
microsoft kinect sensor: A review,” IEEE transactions on cybernetics,
2013.
[5] G. Zhang, J. Liu, H. Li, Y. Q. Chen, and L. S. Davis, “Joint human
detection and head pose estimation via multistream networks for rgb-d
videos,” IEEE Signal Processing Letters, 2017.
[6] G. Zhang, L. Tian, Y. Liu et al., “Robust real-time human perception
with depth camera,” in ECAI, 2016.
[7] L. L. Presti and M. La Cascia, “3d skeleton-based human action
classification: A survey,” Pattern Recognition, 2016.
[8] F. Han, B. Reily, W. Hoff, and H. Zhang, “Space-time representation of
people based on 3d skeletal data: a review,” arXiv, 2016.
[9] J. K. Aggarwal and L. Xia, “Human activity recognition from 3d data:
A review,” Pattern Recognition Letters, 2014.
[10] J. Zhang, W. Li, P. O. Ogunbona, P. Wang, and C. Tang, “Rgb-d-based
action recognition datasets: A survey,” Pattern Recognition, 2016.
[11] M. Ye, Q. Zhang, L. Wang, J. Zhu, R. Yang, and J. Gall, “A survey
on human motion analysis from depth data,” in Time-of-flight and depth
imaging. sensors, algorithms, and applications, 2013.
[12] Y. Du, W. Wang, and L. Wang, “Hierarchical recurrent neural network
for skeleton based action recognition,” in CVPR, 2015.
[13] M. Jiang, J. Kong, G. Bebis, and H. Huo, “Informative joints based
human action recognition using skeleton contexts,” Signal Processing:
Image Communication, 2015.
[14] J. K. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y. Bengio,
“Attention-based models for speech recognition,” in NIPS, 2015.
[15] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel,
and Y. Bengio, “Show, attend and tell: Neural image caption generation
with visual attention,” in ICML, 2015.
[16] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by
jointly learning to align and translate,” in ICLR, 2015.
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
Action
12
Attention iteration #1
Attention iteration #2
(1)
Taking
a
selfie
(2)
Pointing
to
something
(3)
Kicking
other
person
Fig. 7: Examples of qualitative results on the NTU RGB+D dataset. Three actions (taking a selfie, pointing to something, and
kicking other person) are illustrated. The informativeness scores of two attention iterations are visualized. Four frames are
shown for each iteration. The circle size indicates the magnitude of the informativeness score for the corresponding joint in a
frame. For clarity, the joints with tiny informativeness scores are not shown.
[26]
[27]
Right hand
Left hand
[28]
[29]
Fig. 8: Visualization of the average informativeness gates for
all testing samples. The size of the circle around each joint
indicates the magnitude of the corresponding informativeness
score.
[30]
[31]
[32]
[33]
[17] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, 1997.
[18] M. Sundermeyer, R. Schlüter, and H. Ney, “Lstm neural networks for
language modeling.” in INTERSPEECH, 2012.
[19] M. Ibrahim, S. Muralidharan, Z. Deng, A. Vahdat, and G. Mori, “A
hierarchical deep temporal model for group activity recognition,” in
CVPR, 2016.
[20] S. Yeung, O. Russakovsky, G. Mori, and L. Fei-Fei, “End-to-end learning
of action detection from frame glimpses in videos,” in CVPR, 2016.
[21] J. Yue-Hei Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals,
R. Monga, and G. Toderici, “Beyond short snippets: Deep networks
for video classification,” in CVPR, 2015.
[22] Z. Wu, X. Wang, Y.-G. Jiang, H. Ye, and X. Xue, “Modeling spatialtemporal clues in a hybrid deep learning framework for video classification,” in ACM MM, 2015.
[23] J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, K. Saenko, and T. Darrell, “Long-term recurrent convolutional
networks for visual recognition and description,” in CVPR, 2015.
[24] Q. Ke, M. Bennamoun, S. An, F. Sohel, and F. Boussaid, “Leveraging
structural context models and ranking score fusion for human interaction
prediction,” arXiv, 2017.
[25] Q. Ke, M. Bennamoun, S. An, F. Bossaid, and F. Sohel, “Spatial, struc-
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
tural and temporal feature learning for human interaction prediction,”
arXiv, 2016.
N. Srivastava, E. Mansimov, and R. Salakhutdinov, “Unsupervised
learning of video representations using lstms,” in ICML, 2015.
S. Ma, L. Sigal, and S. Sclaroff, “Learning activity progression in lstms
for activity detection and early detection,” in CVPR, 2016.
W. Zhu, C. Lan, J. Xing, W. Zeng, Y. Li, L. Shen, and X. Xie, “Cooccurrence feature learning for skeleton based action recognition using
regularized deep lstm networks,” in AAAI, 2016.
J. Liu, A. Shahroudy, D. Xu, and G. Wang, “Spatio-temporal lstm with
trust gates for 3d human action recognition,” in ECCV, 2016.
J. Weston, S. Chopra, and A. Bordes, “Memory networks,” in ICLR,
2015.
J. Liu, G. Wang, P. Hu, L.-Y. Duan, and A. C. Kot, “Global contextaware attention lstm networks for 3d action recognition,” in CVPR, 2017.
J. Luo, W. Wang, and H. Qi, “Group sparsity and geometry constrained
dictionary learning for action recognition from depth maps,” in ICCV,
2013.
A. Shahroudy, T.-T. Ng, Q. Yang, and G. Wang, “Multimodal multipart
learning for action recognition in depth videos,” IEEE Transactions on
Pattern Analysis and Machine Intelligence, 2016.
M. Meng, H. Drira, M. Daoudi, and J. Boonaert, “Human-object
interaction recognition by learning the distances between the object and
the skeleton joints,” in FG, 2015.
X. Yang and Y. Tian, “Effective 3d action recognition using eigenjoints,”
Journal of Visual Communication and Image Representation, 2014.
J. Wang and Y. Wu, “Learning maximum margin temporal warping for
action recognition,” in ICCV, 2013.
I. Lillo, J. Carlos Niebles, and A. Soto, “A hierarchical pose-based approach to complex action understanding using dictionaries of actionlets
and motion poselets,” in CVPR, 2016.
H. Rahmani, A. Mahmood, D. Q. Huynh, and A. Mian, “Real time action
recognition using histograms of depth gradients and random decision
forests,” in WACV, 2014.
C. Chen, R. Jafari, and N. Kehtarnavaz, “Fusion of depth, skeleton, and
inertial data for human action recognition,” in ICASSP, 2016.
L. Tao and R. Vidal, “Moving poselets: A discriminative and interpretable skeletal motion representation for action recognition,” in
ICCVW, 2015.
A. Shahroudy, G. Wang, and T.-T. Ng, “Multi-modal feature fusion for
action recognition in rgb-d sequences,” in ISCCSP, 2014.
ACCEPTED TO IEEE TRANSACTIONS ON IMAGE PROCESSING
[42] P. Wang, W. Li, P. Ogunbona, Z. Gao, and H. Zhang, “Mining mid-level
features for action recognition based on effective skeleton representation,” in DICTA, 2014.
[43] F. Ofli, R. Chaudhry, G. Kurillo, R. Vidal, and R. Bajcsy, “Sequence
of the most informative joints (smij): A new representation for human
skeletal action recognition,” Journal of Visual Communication and
Image Representation, 2014.
[44] R. Anirudh, P. Turaga, J. Su, and A. Srivastava, “Elastic functional
coding of human actions: from vector-fields to latent variables,” in
CVPR, 2015.
[45] R. Chaudhry, F. Ofli, G. Kurillo, R. Bajcsy, and R. Vidal, “Bioinspired dynamic 3d discriminative skeletal features for human action
recognition,” in CVPRW, 2013.
[46] R. Vemulapalli, F. Arrate, and R. Chellappa, “Human action recognition
by representing 3d skeletons as points in a lie group,” in CVPR, 2014.
[47] L. Xia, C.-C. Chen, and J. Aggarwal, “View invariant human action
recognition using histograms of 3d joints,” in CVPRW, 2012.
[48] J. Wang, Z. Liu, Y. Wu, and J. Yuan, “Mining actionlet ensemble for
action recognition with depth cameras,” in CVPR, 2012.
[49] J. Wang, Z. Liu, Y. Wu, and J. S. Yuan, “Learning actionlet ensemble for
3d human action recognition,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, 2014.
[50] H. Chen, G. Wang, J.-H. Xue, and L. He, “A novel hierarchical
framework for human action recognition,” Pattern Recognition, 2016.
[51] P. Koniusz, A. Cherian, and F. Porikli, “Tensor representations via kernel
linearization for action recognition from 3d skeletons,” in ECCV, 2016.
[52] P. Wang, C. Yuan, W. Hu, B. Li, and Y. Zhang, “Graph based skeleton
motion representation and similarity measurement for action recognition,” in ECCV, 2016.
[53] M. Zanfir, M. Leordeanu, and C. Sminchisescu, “The moving pose: An
efficient 3d kinematics descriptor for low-latency action recognition and
detection,” in ICCV, 2013.
[54] V. Veeriah, N. Zhuang, and G.-J. Qi, “Differential recurrent neural
networks for action recognition,” in ICCV, 2015.
[55] A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu rgb+d: A large scale
dataset for 3d human activity analysis,” in CVPR, 2016.
[56] J. Liu, A. Shahroudy, D. Xu, A. C. Kot, and G. Wang, “Skeleton-based
action recognition using spatio-temporal lstm network with trust gates,”
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
[57] A. Jain, A. R. Zamir, S. Savarese, and A. Saxena, “Structural-rnn: Deep
learning on spatio-temporal graphs,” in CVPR, 2016.
[58] Y. Li, C. Lan, J. Xing, W. Zeng, C. Yuan, and J. Liu, “Online human
action detection using joint classification-regression recurrent neural
networks,” in ECCV, 2016.
[59] C. Xiong, S. Merity, and R. Socher, “Dynamic memory networks for
visual and textual question answering,” in ICML, 2016.
[60] S. Sharma, R. Kiros, and R. Salakhutdinov, “Action recognition using
visual attention,” in ICLRW, 2016.
[61] A. Kumar, O. Irsoy, P. Ondruska, M. Iyyer, J. Bradbury, I. Gulrajani,
V. Zhong, R. Paulus, and R. Socher, “Ask me anything: Dynamic
memory networks for natural language processing,” in ICML, 2016.
[62] M.-T. Luong, H. Pham, and C. D. Manning, “Effective approaches to
attention-based neural machine translation,” in EMNLP, 2015.
[63] S. Sukhbaatar, A. Szlam, J. Weston, and R. Fergus, “End-to-end memory
networks,” in NIPS, 2015.
[64] M. F. Stollenga, J. Masci, F. Gomez, and J. Schmidhuber, “Deep
networks with internal selective attention through feedback connections,”
in NIPS, 2014.
[65] L. Yao, A. Torabi, K. Cho, N. Ballas, C. Pal, H. Larochelle, and
A. Courville, “Describing videos by exploiting temporal structure,” in
ICCV, 2015.
[66] K. Simonyan and A. Zisserman, “Two-stream convolutional networks
for action recognition in videos,” in NIPS, 2014.
[67] A. Shahroudy, T.-T. Ng, Y. Gong, and G. Wang, “Deep multimodal feature analysis for action recognition in rgb+ d videos,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, 2017.
[68] S. Song, C. Lan, J. Xing, W. Zeng, and J. Liu, “An end-to-end spatiotemporal attention model for human action recognition from skeleton
data.” in AAAI, 2017.
[69] Y. Wang, S. Wang, J. Tang, N. O’Hare, Y. Chang, and B. Li, “Hierarchical attention network for action recognition in videos,” arXiv, 2016.
[70] A. Graves, “Supervised sequence labelling,” in Supervised Sequence
Labelling with Recurrent Neural Networks, 2012.
[71] J.-F. Hu, W.-S. Zheng, J. Lai, and J. Zhang, “Jointly learning heterogeneous features for rgb-d activity recognition,” in CVPR, 2015.
13
[72] K. Yun, J. Honorio, D. Chattopadhyay, T. L. Berg, and D. Samaras,
“Two-person interaction detection using body-pose features and multiple
instance learning,” in CVPRW, 2012.
[73] F. Ofli, R. Chaudhry, G. Kurillo, R. Vidal, and R. Bajcsy, “Berkeley
mhad: A comprehensive multimodal human action database,” in WACV,
2013.
[74] R. Collobert, K. Kavukcuoglu, and C. Farabet, “Torch7: A matlab-like
environment for machine learning,” in NIPSW, 2011.
[75] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from overfitting,” Journal of Machine Learning Research, 2014.
[76] G. Evangelidis, G. Singh, and R. Horaud, “Skeletal quads: Human action
recognition using joint quadruples,” in ICPR, 2014.
[77] P. Wang, Z. Li, Y. Hou, and W. Li, “Action recognition based on joint
trajectory maps using convolutional neural networks,” in ACM MM,
2016.
[78] Q. Ke, S. An, M. Bennamoun, F. Sohel, and F. Boussaid, “Skeletonnet:
Mining deep part features for 3-d action recognition,” IEEE Signal
Processing Letters, 2017.
[79] M. Liu, H. Liu, and C. Chen, “Enhanced skeleton visualization for view
invariant human action recognition,” Pattern Recognition, 2017.
[80] J.-F. Hu, W.-S. Zheng, L. Ma, G. Wang, and J. Lai, “Real-time rgb-d
activity prediction by soft regression,” in ECCV, 2016.
[81] R. Slama, H. Wannous, M. Daoudi, and A. Srivastava, “Accurate 3d
action recognition using learning on the grassmann manifold,” Pattern
Recognition, 2015.
[82] M. Devanne, H. Wannous, S. Berretti, P. Pala, M. Daoudi, and
A. Del Bimbo, “3-d human action recognition by shape analysis of
motion trajectories on riemannian manifold,” IEEE Transactions on
Cybernetics, 2015.
[83] C. Wang, Y. Wang, and A. L. Yuille, “Mining 3d key-pose-motifs for
action recognition,” in CVPR, 2016.
[84] C. Wang, J. Flynn, Y. Wang, and A. L. Yuille, “Recognizing actions in
3d using action-snippets and activated simplices,” in AAAI, 2016.
[85] W. Li, L. Wen, M. Choo Chuah, and S. Lyu, “Category-blind human
action recognition: A practical recognition system,” in ICCV, 2015.
[86] Y. Ji, G. Ye, and H. Cheng, “Interactive body part contrast mining for
human interaction recognition,” in ICMEW, 2014.
[87] S. Vantigodi and R. V. Babu, “Real-time human action recognition from
motion capture data,” in NCVPRIPG, 2013.
[88] S. Vantigodi and V. B. Radhakrishnan, “Action recognition from motion
capture data using meta-cognitive rbf network classifier,” in ISSNIP,
2014.
[89] I. Kapsouras and N. Nikolaidis, “Action recognition on motion capture
data using a dynemes and forward differences representation,” Journal
of Visual Communication and Image Representation, 2014.
| 1cs.CV
|
Ising Processing Units:
Potential and Challenges for
Discrete Optimization
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
arXiv:1707.00355v1 [math.OC] 2 Jul 2017
Los Alamos National Laboratory, Los Alamos, New Mexico, USA
Abstract. The recent emergence of novel computational devices, such
as adiabatic quantum computers, CMOS annealers, and optical parametric oscillators, presents new opportunities for hybrid-optimization algorithms that leverage these kinds of specialized hardware. In this work, we
propose the idea of an Ising processing unit as a computational abstraction for these emerging tools. Challenges involved in using and benchmarking these devices are presented, and open-source software tools are
proposed to address some of these challenges. The proposed benchmarking tools and methodology are demonstrated by conducting a baseline
study of established solution methods to a D-Wave 2X adiabatic quantum computer, one example of a commercially available Ising processing
unit.
1
Introduction
As the challenge of scaling traditional transistor-based Central Processing Unit
(CPU) technology continues to increase, experimental physicists and high-tech
companies have begun to explore radically different computational technologies,
such as adiabatic quantum computers (AQCs) [29], gate-based quantum computers [28,43], CMOS annealers [52,53], neuromorphic computers [42], and optical
parametric oscillators [40,27,30]. The goal of all of these technologies is to leverage the dynamical evolution of a physical system to perform a computation that
is challenging to emulate using traditional CPU technology (e.g., the simulation
of quantum physics) [19]. Despite their entirely disparate physical implementations, AQCs, CMOS annealers, and optical parametric oscillators are unified by
a common mathematical abstraction known as the Ising model, which has been
widely adopted by the physics community for the study of naturally occurring
discrete optimization [12]. Furthermore, this kind of “Ising machine” [40,27] is
already commercially available with more than 2000 decision variables in the
form of AQCs developed by D-Wave Systems [16].
The emergence of physical devices that can quickly solve Ising models is
particularly relevant to the optimization and operations research communities,
because the core purpose of these devices is to perform discrete optimization.
As this technology matures, it may be possible for this specialized hardware to
2
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
rapidly solve classical combinatorial problems, such as Max-Cut [24] or MaxClique [37]. Preliminary studies have even suggested that some classes of Constraint Satisfaction Problems may be effectively encoded in such devices because
of their combinatorial structure [6,5,47,50]. Furthermore, an Ising model coprocessor could have significant impacts on solution methods for a variety of fundamental combinatorial problem classes, such as MAX-SAT [21,44] and integer
programming [46]. At this time, however, it remains unclear how established
optimization algorithms should leverage this emerging technology.
Similar to an arithmetic logic unit or a graphics processing unit (GPU), we
propose the idea of an Ising processing unit (IPU) as the computational abstraction for physical devices that perform optimization of Ising models. This
work first provides a brief introduction to the IPU abstraction and develops the
mathematical foundations and real-world considerations of this novel hardware.
Second, it conducts a baseline benchmarking study of an IPU to demonstrate
the current capabilities of such a device. Hence, this work makes both literature
survey and technical contributions. The survey contributions include (1) presenting foundational Ising model results from other disciplines in terminology
that is familiar to the optimization community (Section 2), and (2) highlighting key features of IPUs to provide context for future algorithmic developments
utilizing these analog devices (Section 3). The technical contributions include
(1) developing a benchmarking methodology for IPUs, which is enabled by proposed open-source software tools (Section 4); and (2) demonstrating the proposed benchmarking tools by conducting a baseline evaluation of a D-Wave 2X
IPU to support future work in the area (Section 5). To the best of our knowledge,
this is the first benchmarking study to use an integer programming solver for
identifying challenging IPU test cases. Because of the maturity and commercial
availability of the D-Wave IPU, this work often refers to that architecture as an
illustrative example. However, based on our understanding of other IPU technologies, the methods and tools proposed herein are applicable to all emerging
IPU hardware realizations.
2
A Brief Introduction to Ising Models
This section introduces the notations of the paper and provides a brief introduction to Ising models because they are the core mathematical abstraction of
IPUs. The Ising model refers to the class of graphical models where the nodes,
N , represent spin variables (i.e., σi ∈ {−1, 1} ∀i ∈ N ) and the edges, E, represent interactions of spin variables (i.e., σi σj ∀i, j ∈ E). A local field hi ∀i ∈ N
can be specified for each node, and an interaction strength Jij ∀i, j ∈ E can be
specified for each edge. Given these data, the energy of the Ising model is defined
as
X
X
E(σ) =
Jij σi σj +
hi σi
(1)
i,j∈E
i∈N
Applications of the Ising model typically consider one of two tasks. Some applications focus on finding the lowest possible energy of the Ising model, namely
Ising Processing Units: Potential and Challenges for Discrete Optimization
3
finding the globally optimal solution of the following binary quadratic optimization problem:
min : E(σ)
(2)
s.t.: σi ∈ {−1, 1} ∀i ∈ N
Other applications are interested in sampling from the Boltzmann distribution
of the Ising model’s states:
P r(σ) ∝ e
−E(σ)
τ
(3)
where τ is a parameter representing the effective temperature of the Boltzmann
distribution [54]. It is valuable to observe that in the Boltzmann distribution, the
lowest energy states have the highest probability. Hence, the task of sampling
from a Boltzmann distribution is similar to the task of finding the lowest energy
of the Ising model. Indeed, as τ approaches 0, the sampling task smoothly transforms into the aforementioned optimization task. This paper focuses exclusively
on the mathematical program presented in (2), the optimization task.
Frustration The notion of frustration is common in the study of Ising models
and refers to any instance of (2) where the optimal solution, σ ∗ , satisfies the
property
X
X
E(σ ∗ ) >
−|Jij | −
|hi |
(4)
i,j∈E
i∈N
A canonical example is the following three node problem:
h1 = 0, h2 = 0, h3 = 0
(5a)
J12 = −1, J23 = −1, J13 = 1
(5b)
Observe that, in this case, there are a number
of optimal solutions such that
P
E(σ ∗ ) = −2 but none such that E(σ) = i,j∈E −|Jij | = −3.
Gauge Transformations A valuable property of the Ising model is the gauge
transformation, which characterizes an equivalence class of Ising models. For
illustration, consider the optimal solution of Ising model S, σ s∗ . One can construct a new Ising model T where the optimal solution is the same, except that
σit∗ = −σis∗ for a particular node i ∈ N is as follows:
t
s
Jij
= −Jij
∀i, j ∈ E(i)
hti
=
−hsi
(6a)
(6b)
where E(i) indicates the neighboring edges of node i. This S-to-T manipulation
is referred to as a gauge transformation. Given a complete source state σ s and
a complete target state σ t , this transformation is generalized to all of σ by
t
s s s t t
Jij
= Jij
σi σj σi σj ∀i, j ∈ E
hti
=
hsi σis σit
∀i ∈ N
(7a)
(7b)
4
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
It is valuable to observe that by using this gauge transformation property, one
can consider the class of Ising models where the optimal solution is σi∗ = −1 ∀i ∈
N or any arbitrary vector of −1, 1 values without loss of generality.
Bijection of Ising and Boolean Optimization It is also useful to observe that
there is a bijection between Ising optimization (i.e., σ ∈ {−1, 1}) and Boolean
optimization (i.e., x ∈ {0, 1}). The transformation of σ ⇒ x is given by
σi + 1
∀i ∈ N
2
σi σj + σi + σj + 1
xi xj =
∀i, j ∈ E
4
xi =
(8a)
(8b)
and the inverse x ⇒ σ is given by
σi = 2xi − 1 ∀i ∈ N
(9a)
σi σj = 4xi xj − 2xi − 2xj + 1 ∀i, j ∈ E
(9b)
Consequently, any results from solving Ising models are also immediately applicable to the following class of Boolean optimization problems:
X
X
min :
cij xi xj +
ci x i
(10)
i,j∈E
i∈N
s.t.: xi ∈ {0, 1} ∀i ∈ N
The Ising model provides a clean mathematical abstraction for understanding
the computation of IPUs. However, in practice, a number of hardware implementation factors present additional challenges for computing with IPUs.
3
Features of Ising Processing Units
The core inspiration for developing IPUs is to take advantage of the natural
evolution of a discrete physical system to find high-quality solutions to an Ising
model (2) [29,40,53]. Consequently, to the best of our knowledge, all IPUs developed to date are analog machines, which present a number of challenges that
the optimization community is not accustomed to considering.
Effective Temperature The ultimate goal of current IPUs is to solve the optimization problem (2) and determine the globally optimal solution to the input Ising
model. In practice, however, a variety of analog factors preclude IPUs from reliably finding globally optimal solutions. As a first-order approximation, current
IPUs behave like a Boltzmann sampler (3) with some hardware-specific effective
temperature, τ [7]. It has also been observed that the effective temperature of
an IPU can vary around a nominal value based on the Ising model that is being
executed [4]. This suggests that the IPU’s performance can change based on the
structure of the problem input.
Ising Processing Units: Potential and Challenges for Discrete Optimization
σ4
σ5
σ6
σ7
σ12
σ13
σ14
σ15
σ0
σ1
σ2
σ3
σ8
σ9
σ10
σ11
σ20
σ21
σ22
σ23
σ28
σ29
σ30
σ31
σ16
σ17
σ18
σ19
σ24
σ25
σ26
σ27
5
Fig. 1. A 2-by-2 Chimera Graph Illustrating the Variable Product Limitations of a
D-Wave 2X IPU.
Environmental Noise One of the primary contributors to the sampling nature
of IPUs is environmental factors. All analog machines are subject to faults due
to environmental noise; for example, even classical computers can be affected by
cosmic rays. However, given the relative novelty of IPUs, the effects of environmental noise are noticeable in current hardware. The effects of environmental
noise contribute to the perceived effective temperature τ of the IPU.
Problem Coefficients In traditional optimization applications, the problem coefficients are often rescaled to best suit floating-point arithmetic. Similarly, IPUs
have digital-to-analog converters that can encode a limited number of values;
typically these values are represented as numbers in the range of -1 to 1. Some
IPUs allow for hundreds of steps within this range, [29,53] whereas others support only the discrete set of {-1, 0, 1} [40]. In either case, the Ising model must
be rescaled into the IPU’s operating range. The values of the coefficients used
in the IPU are critically important because these values are often perturbed by
environmental noise and hardware biases.
Coefficient Biases Once an Ising model is input into an IPU, its coefficients
are subject to at least two sources of bias. The first source of bias is a model
programming error that occurs independently each time the IPU is configured for
a computation. This bias is often mitigated by programming the IPU multiple
times with an identical input and combining the results from all executions.
The second source of bias is a persistent coefficient error, which is an artifact of
the IPU manufacturing and calibration process. Because this bias is consistent
across multiple IPU executions, this source bias is often mitigated by performing
multiple gauge transformations on the input and combining the results from all
executions.
Topological Limitations Another significant feature of IPUs is a restricted topology for variable products. In classical optimization (e.g., (2)), it is assumed that
every variable can interact with every other variable, that is, an Ising model
6
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
where an edge connects every pair of variables. However, because of the hardware implementation of an IPU, it may not be possible for some variables to
interact. For example, the current D-Wave IPUs are restricted to the chimera
topology, which is a two-dimensional lattice of unit cells, each of which consist
of a 4-by-4 bipartite graph (e.g., see Figure 1). In addition to these restrictions,
fabrication errors can also lead to random failures of nodes and edges in the IPU
hardware. Indeed, as a result of these minor imperfections, every D-Wave IPU
developed to date has a unique topology [9,18,32]. Research and development
of algorithms for embedding various kinds of Ising models into a specific IPU
topology is still an active area of research [6,11,13,34].
3.1
Challenges of Benchmarking Ising Processing Units
These novel IPU features present unique challenges for benchmarking these devices. These challenges fall into roughly two categories: finding interesting Ising
models for testing and comparing with classical methods.
Standardized Benchmark Libraries Research and development in optimization
algorithms has benefited greatly from standardized benchmark libraries [35,20,26].
However, direct application of these libraries to IPUs is out of scope for the near
term for the following reasons: (1) the Ising model is a binary quadratic program,
which is sufficiently restrictive to preclude the use of many standard problem
libraries; (2) even in cases where the problems of interest can be mapped directly
to the Ising model (e.g., Max-Cut, Max-Clique), the task of embedding the given
problems onto the IPU’s hardware graph can be prohibitive; and (3) even if an
embedding can be found, it is not obvious that the problem’s coefficients will be
amenable to the IPU’s operating range [14].
Another option is to build a new standardized benchmark library specifically
for the evaluation of IPUs. However, given that every IPU is unique and within
an IPU each variable also has individual properties (e.g., persistent bias), it is
not immediately clear that a new static benchmark library could be applied to
multiple IPUs.
Comparison with Classical Algorithms Because of the radically different hardware of CPUs vs IPUs and the sampling nature of the IPUs, how to conduct a
fair comparison of these two technologies is not immediately clear [38,39,33]. For
example, is the comparison more fair or less fair if classical algorithms utilize
multiple cores or GPUs? Indeed, comparisons of D-Wave’s IPU with classical
algorithms have resulted in vigorous discussion about what algorithms and metrics should be used to make such comparisons [1,32,2]. It is widely accepted that
IPUs do not provide optimality guarantees and should be compared to classical
heuristic methods rather than methods providing optimality proofs. However,
it is less clear whether heuristics should be specialized to solve problems for a
specific IPU architecture (e.g., [49,48]) or whether classical methods should be
designed for the most general form of the Ising model (2). This debate will most
likely continue for several years. In this work, our goal is not to answer these
Ising Processing Units: Potential and Challenges for Discrete Optimization
7
challenging questions but rather to develop benchmarking tools that will assist
researchers in exploring these questions and replicating previous results as the
subject of IPUs evolves.
4
Tools for Benchmarking Ising Processing Units
To help address a number of the previously discussed challenges, in this work
we propose three open-source tools for assisting in benchmarking IPUs:
– bqpjson: a language-independent json-based test case exchange format designed with IPUs in mind1
– dwig: a collection of algorithms for IPU test case generation2
– bqpsolvers: a collection of tools for encoding bqpjson data into various
optimization solvers3
We now review the design motivations and core features of each of these tools.
4.1
Data Management
The primary design goals of bqpjson are to (1) encode all of the information
required to replicate the exact execution that was performed on a specific IPU
hardware; (2) allow the IPU test case to be transformed back to the original problem domain, if one exists; and (3) not be prescriptive about the mathematical
representation of the test case (i.e., (2) or (10)). To that end, bqpjson encodes
the following class of binary quadratic programs, which is a generalization of (2)
and (10):
X
X
min : s
cij bi bj +
ci bi + o
(11)
i,j∈E
i∈N
s.t.: bi ∈ {0, 1} ∀i ∈ N or bi ∈ {−1, 1} ∀i ∈ N
The key features of this model are (1) the decision variables can be represented as
either spin (i.e., {−1, 1}) or Boolean (i.e., {0, 1}) values; (2) the objective offset
term, o, is required in order to ensure an idempotent transformation between
the spin and Boolean variable domains; (3) the explicit scaling term, s, allows
for the problem coefficients (i.e., c) to be specified within an IPU’s operating
range while preserving the units of the problem’s original objective function;
and (4) the collection of decision variables N allows the variable identifiers to
have arbitrary values, such as the names used by specific IPU hardware (unlike
traditional test case formats, where the variables are numbered from 1 to n).
1
2
3
The source code is available at https://github.com/lanl-ansi/bqpjson
The source code is available at https://github.com/lanl-ansi/dwig
The source code is available at https://github.com/lanl-ansi/bqpsolvers
8
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
In addition to these problem specification features, bqpjson also has extensive support for documenting test cases. A description section provides humanreadable information about the test case, whereas a metadata section is used
to encode useful machine-readable information. bqpjson also supports encoding solutions so that test cases can include details about target or best-known
solutions to specific cases.
Implementation and Tools The current implementation of bqpjson is provided
as a free and open-source Python package that can be installed from the source
code repository or from the Python Package Index (PyPI). This package includes
(1) detailed documentation of the bqpjson file format; (2) tools for validation of
JSON data to ensure they conform to the bqpjson specification; (3) bqpjsonto-bqpjson idempotent transformations between the spin and Boolean variable
domains; (4) translation of bqpjson data into other established problem formats, such as QUBO [10] and MiniZinc [45]; and (5) a collection of command
line tools for bash script processing of bqpjson data.
4.2
Case Generation
Because of the challenges associated with mapping established optimization test
cases to the IPU hardware, the IPU benchmarking community has adopted the
practice of generating tests on a case-by-case basis for specific IPUs [32,25,33,18].
This practice amounts to finding interesting values for h and J in (1). In some
cases the procedures for generating these values are elaborate [18,31] and are
designed to leverage theoretical results about Ising models [25]. To ensure the
experimental reproducibility in the context of IPU-specific case generation, in
this paper we developed the D-Wave Instance Generator (dwig). dwig represents a collection of algorithms that generate well-studied random Ising models.
Our hope is that as long as the same test case generation algorithm is used, the
average values of previous experimental results will be replicable across multiple
IPUs. The dwig script currently implements five problem classes proposed in
the literature [32,31,33,18], each of which we will briefly introduce. For a detailed
description, please refer to the source publication of the problem class.
Random (RAN-k and RANF-k) To the best of our knowledge, the RAN-k problem was first proposed in [32] and consists simply of assigning each value of h
to 0 and each value of J uniformly at random from the set
{−k, −k + 1, . . . , −2, −1, 1, 2, . . . , k − 1, k}
(12)
The RANF-k problem is a simple variant of RAN-k where the values of h are
also selected uniformly at random from (12). As we will later see, RAN-1 and
RANF-1, where h, J ∈ {−1, 1}, are an interesting subclass of this problem.
Frustrated Loops (FL-k and FCL-k) The frustrated loop problem was originally
proposed in [25] and then later refined to the FL-k problem in [31]. It consists
Ising Processing Units: Potential and Challenges for Discrete Optimization
9
of generating a collection of random cycles in the IPU graph. In each cycle,
all of the edges are set to −1 except one random edge, which is set to 1 to
produce frustration. A scaling factor α is used to control how many random cycles
should be generated, and the parameter k determines how many cycles each edge
can participate in. A key property of the FL-k generation procedure is that a
globally optimal solution is maintained at σi = −1 ∀i ∈ N and σi = 1 ∀i ∈ N
[31]. However, by default, dwig uses a gauge transformation to obfuscate this
solution pair to a random assignment of σ.
A variant of the frustrated loop problem is the frustrated cluster loop problem, FCL-k [33]. The FCL-k problem is inspired by the chimera network topology
(i.e., Figure 1). The core idea is that tightly coupled variables (e.g., σ1 ..σ7 in
Figure 1) should form a cluster where all of the variables take the same value.
This is achieved by setting all of the values of J within the cluster to −1. For
the remaining edges between clusters, the previously described frustrated cycles generation scheme is used. It is important to note that a polynomial time
algorithm is known for solving the FCL-k problem class on chimera graphs [39].
It is worthwhile to mention that the FL-k and FCL-k case generators are
effectively solving a cycle packing problem on the IPU graph. Hence, the randomized algorithm implemented in dwig is not guaranteed to find a solution if
one exists, and in practice it fails for the highly constrained settings of α and k.
Weak-Strong Cluster Networks (WSCNs) The WSCN problem was proposed in
[18] and is highly specialized to the chimera network topology. The basic building
block of a WSCN is a pair of spin clusters in the chimera graph (e.g., σ1 ..σ7 and
σ8 ..σ15 in Figure 1). In the strong cluster the values of h are set to the strong
force parameter sf and in the weak cluster the values of h are set to the weak
force parameter wf. All of the values of J within and between this cluster pair
are set to −1. Once a number of weak-strong cluster pairs have been placed, the
strong clusters are connected to each other using random values of J ∈ {−1, 1}.
The values of sf = −1.0 and wf = 0.44 are recommended by [18]. The motivation
for the WSCN design is that the clusters create many deep local minima that
are difficult for local search methods to escape.
Implementation The current implementation of dwig is provided as a Python
script. The implementation takes an IPU hardware specification as input and
produces bqpjson test cases for each of the problem classes described above.
4.3
Baseline Solution Methods
Because of the novelty of bqpjson, solvers do not currently support loading
bqpjson cases directly. Hence, to make bqpjson usable we provide the bqpsolvers code base, which includes a collection of simple tools for running
bqpjson data on a variety of established Ising model optimization algorithms.
bqpsolvers is designed as a collaborative space where state-of-the-art bqpjson solvers can be community-curated and leveraged for future IPU evaluations.
10
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
At this time, bqpsolvers includes connections to the following solution
methods: (1) mixed integer programming (MIP) [17,8], (2) large neighborhood
search (LNS) via the Hamze-Freitas-Selby (HFS) algorithm [23,49], and (3) adiabatic quantum computation (AQC) via the D-Wave IPU [29]. Each of these is
briefly introduced.
Mixed Integer Programming Modern MIP solvers are highly optimized for Boolean
variables, so we choose to develop a MIP model based on the Boolean optimization variant of the Ising model, (10), as follows:
X
X
min : s
cij xij +
ci xi + o
(13a)
i,j∈E
i∈N
s.t.:
xij ≥ xi + xj − 1 ∀i, j ∈ E
(13b)
xij ≤ xi ∀i, j ∈ E
(13c)
xij ≤ xj ∀i, j ∈ E
(13d)
xij ∈ {0, 1} ∀i, j ∈ E
xi ∈ {0, 1} ∀i ∈ N
In this formulation we convert the binary quadratic program defined in (10) to
a binary linear program by lifting the variable products xi xj into a new variable
xij and adding linear constraints to capture the xij = xi ∧xj ∀i, j ∈ E constraints
[8]. This formulation was implemented using Gurobi [22].
Large Neighborhood Search The state-of-the-art heuristic for solving Ising models on the chimera graphs is an LNS-based method called HFS [23,49]. The core
idea of this algorithm is to extract low treewidth subgraphs of the given Ising
model and then use dynamic programming to compute the optimal configuration
of these subgraphs. This extract and optimize process is repeated until a specified time limit is reached. This solver utilizes a highly optimized open-source C
implementation of HFS [48] and leverages the bqpjson tools to convert bqpjson data into the HFS’s proprietary data format.
Adiabatic Quantum Computation The AQC solver runs the provided bqpjson model natively on a D-Wave IPU. This solver assumes that the bqpjson model was generated on the specified IPU and that the variable identifiers
used in the bqpjson model will map directly to the IPU hardware. No attempt
is made to transform a general bqpjson model into the specified IPU hardware.
Recalling that the IPU behaves like a Boltzmann sampler rather than a classical
optimization solver, this code takes a number of samples from the IPU (typically
10,000) and reports the best solution found among all of the samples, making
the IPU’s output consistent with the other optimization tools considered here.
Implementation All of the bqpsolvers are provided as independent Python
scripts with consistent input and output interfaces.
Ising Processing Units: Potential and Challenges for Discrete Optimization
5
11
A Study of Established Methods
In this section we conduct an in-depth computational study of the IPU instance
generation procedures and possible solution methods. The first goal of this study
is to understand what classes of problems are most useful for the study of IPUs.
The second goal is to provide a baseline of comparison for future works in this
area. This computational study is divided into two phases. First, we focus on a
broad parameter sweep of all possible instance generation procedures and use a
MIP solver to determine which of the problem classes are most challenging. Second, after the most challenging problem classes have been identified, a detailed
study is conducted to compare and contrast the MIP, LNS, and AQC solution
methods.
Throughout this section, the following notations are used to describe the
runtime analysis of the algorithms: U B denotes the objective value of the best
feasible solution produced by the algorithm within the time limit, LB denotes the
value of the best lower bound produced by the algorithm within the time limit,
T denotes the algorithm runtime in seconds, T O denotes that the algorithm hit
a time limit of 600 seconds, µ(·) denotes the mean of a collection of values, sd(·)
denotes the standard deviation of a collection of values, and max(·) denotes the
maximum of a collection of values.
Computation Environment The IPU computation is conducted on a D-Wave 2X
AQC [15]. This computer is a 12-by-12 chimera topology with random omissions;
in total, the IPU has 1095 qubits and 3061 couplers and an effective temperature
of τ ∈ (0.091, 0.053) depending on the problem being solved [51,36]. Unless
otherwise noted, the AQC is configured to produce 10,000 samples using a 5microsecond annealing time per sample and a random gauge transformation
every 100 samples. The reported runtime of the AQC reflects the amount of time
used on the IPU hardware; it does not include the overhead of communication
or scheduling of the computation.
The classical computing algorithms are run on HPE ProLiant XL170r servers
with dual Intel 2.10GHz CPUs and 128GB memory. Gurobi 7.0.2 [22] is used as
the MIP solver and is configured to use one thread. The highly specialized and
optimized HFS algorithm [48] is used as an LNS-based heuristic and also uses
one thread.
5.1
Identifying Challenging Cases
Broad Parameter Sweep In this first experiment, we conduct a parameter sweep
of all the inputs to the problem generation algorithms described in Section 4.2.
Table 1 provides a summary of the input parameters for each problem class.
The values of each parameter are encoded with the following triple: (start..stop :
step size). When two parameters are required for a given problem class, the cross
product of all parameters is used. For each problem class and each combination
of parameter settings, 250 random problems are generated in order to produce a
reasonable estimate of the average difficulty of that configuration. Each problem
12
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
Problem
RAN-k
RANF-k
FL-k
FCL-k
WSCN
First Param.
k ∈ (1..5 : 1)
k ∈ (1..5 : 1)
k ∈ (1..5 : 1)
k ∈ (1..5 : 1)
wf ∈ (−1..1 : 0.2)
Second Param.
NA
NA
α ∈ (0..1 : 0.1)
α ∈ (0..1 : 0.1)
sf ∈ (−1..1 : 0.2)
Table 1. Parameter Settings of Various Problems.
Problem # Cases µ(|N |)
RAN
1250 1095
RANF
1250 1095
FL
6944 1008
FCL
8347
888
WSCN
30250
949
µ(|E|)
3061
3061
2126
2282
2313
µ(T ) sd(T ) max(T )
TO
—
TO
TO
—
TO
1.82 1.06
16.80
4.19 2.81
41.40
0.25 0.87
17.90
Table 2. MIP Runtime on Various IPU Benchmark Problems (seconds).
is generated using all of the decision variables available on the IPU. The results
of this parameter sweep are summarized in Table 2. It is important to note
that the FL and FCL generation methods are randomized algorithms and are
not guaranteed to succeed. This explains why the number of cases for these
problems is not a multiple of 250.
The results presented in Table 2 indicate that, at this problem size, all variants of the FL, FCL, and WSCN problems are easy for modern MIP solvers. This
suggests that these problems are not ideal candidates for benchmarking IPUs.
In contrast, the RAN and RANF cases consistently hit the runtime limit of the
MIP solver, suggesting that these problems are more challenging optimization
tasks. This result is consistent with a similar observation in the SAT community,
where random SAT problems are known to be especially challenging [41,3]. To
get a better understanding of these RAN problem classes, we next perform a
detailed study of these problems for various values of the parameter k.
The RAN and RANF Problems In this second experiment, we focus on the
RAN-k and RANF-k problems and conduct a detailed parameter sweep of k ∈
(1..10 : 1). To accurately measure the runtime difficulty of the problem, we also
reduce the size of the problem from 1095 variables to 194 variables so that the
MIP solver can reliably terminate within a 600-second time limit. The results of
this parameter sweep are summarized in Table 3.
The results presented in Table 3 indicate that (1) as the value of k increases,
both the RAN and RANF problems become easier; and (2) the RANF problem is
easier than the RAN problem. The latter is not surprising because the additional
linear coefficients in the RANF problem break many of the symmetries that exist
in the RAN problem. These results suggest that it is sufficient to focus on the
RAN-1 and RANF-1 cases for a more detailed study of IPU performance. This
is a serendipitous outcome for IPU benchmarking because restricting the problem coefficients to {−1, 0, 1} reduces artifacts caused by noise and the numeral
precision of the analog hardware.
Ising Processing Units: Potential and Challenges for Discrete Optimization
13
k # Cases µ(|N |) µ(|E|) µ(T ) sd(T ) max(T ) µ(T ) sd(T ) max(T )
Problems of Increasing k
RAN-k
RANF-k
1
250
194 528 340.0 195.0
TO 14.10 15.20
82.70
2
250
194 528 89.3 64.3
481 2.97 3.41
22.70
3
250
194 528 64.8 28.3
207 1.67 1.48
10.70
4
250
194 528 58.0 29.5
250 1.25 0.83
6.10
5
250
194 528 49.0 23.0
131 1.12 0.77
6.98
6
250
194 528 49.0 22.4
119 1.05 0.59
4.47
7
250
194 528 45.0 22.8
128 1.04 0.75
7.60
8
250
194 528 44.8 23.7
121 1.01 0.62
5.43
9
250
194 528 42.3 22.3
110 0.98 0.60
5.08
10
250
194 528 39.8 22.1
107 0.91 0.43
3.09
Table 3. MIP Runtime on RAN-k and RANF-k IPU Benchmark Problems (seconds)
5.2
An IPU Evaluation using RAN-1 and RANF-1
Now that the RAN-1 and RANF-1 problem classes have been identified as the
most challenging IPU test cases, we perform two detailed studies on these problems using all three algorithmic approaches (i.e., AQC, LNS, and MIP). The first
study focuses on the scalability trends of these solution methods as the problem
size increases, whereas the second study focuses on a runtime analysis of the
largest cases that can be evaluated on our IPU hardware.
Scalability Analysis In this experiment, we increase the problem size gradually
to understand the scalability profile of each of the solution methods (AQC,
LNS, and MIP). The results are summarized in Table 4. Focusing on the smaller
problems, where the MIP solver provides an optimality proof, we observe that
both the AQC and the LNS methods find the optimal solution in all of the
sampled test cases, suggesting that both heuristic solution methods are of high
quality. Focusing on the larger problems, we observe that, in just a few seconds,
both AQC and LNS find feasible solutions that are of higher quality than what
the MIP solver can find in 600 seconds. This suggests that both methods are
producing high-quality solutions at this scale. As the problem size grows, a slight
quality discrepancy emerges favoring LNS over AQC; however, this discrepancy
in average solution quality is less than 1% of the best known value.
Detailed Runtime Analysis Given that both the AQC and the LNS solution
methods have very similar solution qualities, it is prudent to perform a detailed
runtime study to understand the quality vs runtime tradeoff. To develop a runtime profile of the LNS algorithm, the solver’s runtime limit is set to values
ranging from 0.01 to 10.00 seconds. In the case of the AQC algorithm, the number of requested samples is set to values ranging from 10 to 10,000, which has
the effect of scaling the runtime of the IPU process.4 The results of this study
4
It is important to emphasize that the IPU runtime does not include any communication or scheduling overheads.
14
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
AQC
LNS
# Cases µ(|N |) µ(|E|) µ(U B) µ(T ) µ(U B) µ(T ) µ(U B)
RAN-1 Problems of Increasing Size
250
30
70
-44 3.53
-44 10
-44
250
69 176
-110 3.57
-110 10
-110
250
122 321
-199 3.60
-199 10
-199
250
194 528
-325 3.64
-325 10
-325
250
275 751
-462 3.68
-462 10
-461
250
375 1030
-633 3.73
-633 10
-629
250
486 1337
-821 3.77
-822 10
-814
250
613 1689 -1038 3.77 -1039 10 -1021
250
761 2114 -1296 3.76 -1297 10 -1262
250
923 2578 -1574 3.77 -1576 10 -1525
250 1095 3061 -1870 3.80 -1873 10 -1806
RANF-1 Problems of Increasing Size
250
30
70
-53 3.53
-53 10
-53
250
69 176
-127 3.56
-127 10
-127
250
122 321
-229 3.61
-229 10
-229
250
194 528
-370 3.66
-370 10
-370
250
275 751
-526 3.71
-526 10
-526
250
375 1030
-719 3.76
-719 10
-719
250
486 1337
-934 3.81
-934 10
-933
250
613 1689 -1179 3.82 -1179 10 -1178
250
761 2114 -1472 3.82 -1472 10 -1470
250
923 2578 -1786 3.82 -1787 10 -1778
250 1095 3061 -2121 3.86 -2122 10 -2110
MIP
µ(LB)
µ(T )
-44 0.05
-110 0.48
-199 15.90
-327 340.00
-483
TO
-673
TO
-881
TO
-1116
TO
-1401
TO
-1713
TO
-2045
TO
-53
-127
-229
-370
-527
-727
-954
-1211
-1520
-1856
-2212
0.02
0.13
0.67
14.10
128.00
471.00
588.00
TO
TO
TO
TO
Table 4. A Comparison of Solution Quality and Runtime as Problem Size Increases
on RAN-1 and RANF-1.
are summarized in Figure 2. Note that the stochastic sampling nature of the
IPU results in some noise for small numbers of samples. However, the overall
trend is still clear.
The results presented in Figure 2 further illustrate that (1) the RAN problem class is more challenging than the RANF problem class, and (2) regardless
of the runtime configuration used, the LNS heuristic slightly outperforms the
AQC; however, the average solution quality is always within 1% of each other.
Combining all of the results from this section suggests that the D-Wave 2X IPU
is comparable to state-of-the-art optimization methods on commodity classical
computing hardware.
6
Conclusion
In this work we have introduced the idea of an IPU as a computational abstraction for emerging physical devices that optimize Ising models. We have
demonstrated a number of unexpected challenges in benchmarking such devices
and have proposed bqpjson , dwig , and bqpsolvers to help address these
Ising Processing Units: Potential and Challenges for Discrete Optimization
1e−02
1e−01
1e+00
Runtime (seconds, log)
1e+01
−2119
−2117
AQC
LNS
−2121
−1870
−1865
−1860
AQC
LNS
Average Objective Value (n = 200)
RANF−1 Runtime Trend
−1875
Average Objective Value (n = 200)
RAN−1 Runtime Trend
15
1e−02
1e−01
1e+00
1e+01
Runtime (seconds, log)
Fig. 2. Detailed Runtime Analysis of the AQC (D-Wave 2X) and LNS Heuristic (HFS)
on the RAN-1 (left) and RANF-1 (right) Problem Classes.
challenges. We hope that these tools will assist the research community in developing novel algorithms that leverage IPUs.
The baseline study of the D-Wave 2X IPU suggests that averaging over collections of randomly generated test cases is a reasonable strategy for benchmarking
IPUs. However, finding a class of challenging randomly generated test cases is
not trivial. The study verified that at least one commercially available IPU is
comparable to current state-of-the-art classical methods. However, a detailed
runtime analysis did not demonstrate any time vs quality configuration where
the IPU outperformed a state-of-the-art classical method. That said, the IPU
hardware seems to be largely invariant in quality and runtime across a wide
variety of problem classes and instance sizes. If this trend continues as the IPU’s
hardware increases in size, one would expect the IPU to overtake the current
state-of-the-art classical methods because of its parallel computational nature.
Overall, we conclude that the emergence of IPUs is an interesting development for the optimization community and warrants continued study. Considerable work remains to determine the best classes of test cases for benchmarking
IPUs and to develop a better understanding of how to fairly compare this heterogeneous hardware to established classical computing methods.
Acknowledgments
We gratefully acknowledge Denny Dahl for his feedback on the tools proposed
herein and the U.S. Department of Energy for supporting this work through Los
Alamos National Laboratory’s LDRD Program.
16
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
References
1. Aaronson, S.: D-wave: Truth finally starts to emerge. Published online at http:
//www.scottaaronson.com/blog/?p=1400 (May 2013), accessed: 04/28/2017
2. Aaronson, S.: Insert d-wave post here. Published online at http://www.
scottaaronson.com/blog/?p=3192 (Mar 2017), accessed: 04/28/2017
3. Balyo, T., Heule, M.J.H., Jarvisalo, M.: Sat competition 2016: Recent developments. In: Proceedings of the Thirty-First National Conference on Artificial Intelligence. pp. 5061–5063. AAAI’17, AAAI Press (2017)
4. Benedetti, M., Realpe-Gómez, J., Biswas, R., Perdomo-Ortiz, A.: Estimation of
effective temperatures in quantum annealers for sampling applications: A case
study with possible applications in deep learning. Phys. Rev. A 94, 022308 (Aug
2016), https://link.aps.org/doi/10.1103/PhysRevA.94.022308
5. Bian, Z., Chudak, F., Israel, R., Lackey, B., Macready, W.G., Roy, A.: Discrete
optimization using quantum annealing on sparse ising models. Frontiers in Physics
2, 56 (2014), http://journal.frontiersin.org/article/10.3389/fphy.2014.
00056
6. Bian, Z., Chudak, F., Israel, R.B., Lackey, B., Macready, W.G., Roy, A.: Mapping constrained optimization problems to quantum annealing with application
to fault diagnosis. Frontiers in ICT 3, 14 (2016), http://journal.frontiersin.
org/article/10.3389/fict.2016.00014
7. Bian, Z., Chudak, F., Macready, W.G., Rose, G.: The ising model: teaching an
old problem new tricks. Published online at https://www.dwavesys.com/sites/
default/files/weightedmaxsat_v2.pdf (2010), accessed: 04/28/2017
8. Billionnet, A., Elloumi, S.: Using a mixed integer quadratic programming solver
for the unconstrained quadratic 0-1 problem. Mathematical Programming 109(1),
55–68 (2007), http://dx.doi.org/10.1007/s10107-005-0637-9
9. Boixo, S., Ronnow, T.F., Isakov, S.V., Wang, Z., Wecker, D., Lidar, D.A., Martinis, J.M., Troyer, M.: Evidence for quantum annealing with more than one hundred qubits. Nat Phys 10(3), 218–224 (Mar 2014), http://dx.doi.org/10.1038/
nphys2900, article
10. Booth, M.: qbsolv. https://github.com/dwavesystems/qbsolv (2017)
11. Boothby, T., King, A.D., Roy, A.: Fast clique minor generation in chimera qubit
connectivity graphs. Quantum Information Processing 15(1), 495–508 (Jan 2016),
http://dx.doi.org/10.1007/s11128-015-1150-6
12. Brush, S.G.: History of the lenz-ising model. Rev. Modern Phys. 39, 883–893 (Oct
1967), https://link.aps.org/doi/10.1103/RevModPhys.39.883
13. Cai, J., Macready, W.G., Roy, A.: A practical heuristic for finding graph minors
(2014), https://arxiv.org/abs/1406.2741
14. Coffrin, C., Nagarajan, H., Bent, R.: Challenges and Successes of Solving Binary
Quadratic Programming Benchmarks on the DW2X QPU. Tech. rep., Los Alamos
National Laboratory (LANL) (2016)
15. D-Wave Systems Inc.: The d-wave 2x quantum computer technology overview.
Published online at https://www.dwavesys.com/sites/default/files/D-Wave%
202X%20Tech%20Collateral_0915F.pdf (2015), accessed: 04/28/2017
16. D-Wave Systems Inc.: Customers. Published online at https://www.dwavesys.
com/our-company/customers (2017), accessed: 04/28/2017
17. Dash, S.: A note on qubo instances defined on chimera graphs. arXiv preprint
arXiv:1306.1202 (2013), https://arxiv.org/abs/1306.1202
Ising Processing Units: Potential and Challenges for Discrete Optimization
17
18. Denchev, V.S., Boixo, S., Isakov, S.V., Ding, N., Babbush, R., Smelyanskiy, V.,
Martinis, J., Neven, H.: What is the computational value of finite-range tunneling? Phys. Rev. X 6, 031015 (Aug 2016), https://link.aps.org/doi/10.1103/
PhysRevX.6.031015
19. Feynman, R.P.: Simulating physics with computers. International Journal of Theoretical Physics 21(6), 467–488 (1982)
20. Gent, I.P., Walsh, T.: CSPlib: A Benchmark Library for Constraints, pp. 480–
481. Springer Berlin Heidelberg, Berlin, Heidelberg (1999), http://dx.doi.org/
10.1007/978-3-540-48085-3_36
21. de Givry, S., Larrosa, J., Meseguer, P., Schiex, T.: Solving Max-SAT as Weighted
CSP, pp. 363–376. Springer Berlin Heidelberg, Berlin, Heidelberg (2003), http:
//dx.doi.org/10.1007/978-3-540-45193-8_25
22. Gurobi Optimization, Inc.: Gurobi optimizer reference manual. Published online
at http://www.gurobi.com (2014)
23. Hamze, F., de Freitas, N.: From fields to trees. In: Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence. pp. 243–250. UAI ’04, AUAI Press,
Arlington, Virginia, United States (2004), http://dl.acm.org/citation.cfm?id=
1036843.1036873
24. Haribara, Y., Utsunomiya, S., Yamamoto, Y.: A Coherent Ising Machine for MAXCUT Problems: Performance Evaluation against Semidefinite Programming and
Simulated Annealing, pp. 251–262. Springer Japan, Tokyo (2016), http://dx.doi.
org/10.1007/978-4-431-55756-2_12
25. Hen, I., Job, J., Albash, T., Rønnow, T.F., Troyer, M., Lidar, D.A.: Probing for
quantum speedup in spin-glass problems with planted solutions. Phys. Rev. A 92,
042325 (Oct 2015), https://link.aps.org/doi/10.1103/PhysRevA.92.042325
26. Hoos, H.H., Stutzle, T.: Satlib: An online resource for research on sat (2000)
27. Inagaki, T., Haribara, Y., Igarashi, K., Sonobe, T., Tamate, S., Honjo, T., Marandi,
A., McMahon, P.L., Umeki, T., Enbutsu, K., Tadanaga, O., Takenouchi, H., Aihara, K., Kawarabayashi, K.i., Inoue, K., Utsunomiya, S., Takesue, H.: A coherent
ising machine for 2000-node optimization problems. Science 354(6312), 603–606
(2016), http://science.sciencemag.org/content/354/6312/603
28. International Business Machines Corporation: Ibm building first universal quantum
computers for business and science. Published online at https://www-03.ibm.com/
press/us/en/pressrelease/51740.wss (2017), accessed: 04/28/2017
29. Johnson, M.W., Amin, M.H.S., Gildert, S., Lanting, T., Hamze, F., Dickson, N.,
Harris, R., Berkley, A.J., Johansson, J., Bunyk, P., Chapple, E.M., Enderud, C.,
Hilton, J.P., Karimi, K., Ladizinsky, E., Ladizinsky, N., Oh, T., Perminov, I., Rich,
C., Thom, M.C., Tolkacheva, E., Truncik, C.J.S., Uchaikin, S., Wang, J., Wilson,
B., Rose, G.: Quantum annealing with manufactured spins. Nature 473(7346),
194–198 (May 2011), http://dx.doi.org/10.1038/nature10012
30. Kielpinski, D., Bose, R., Pelc, J., Vaerenbergh, T.V., Mendoza, G., Tezak, N.,
Beausoleil, R.G.: Information processing with large-scale optical integrated circuits. In: 2016 IEEE International Conference on Rebooting Computing (ICRC).
pp. 1–4 (Oct 2016)
31. King, A.D., Lanting, T., Harris, R.: Performance of a quantum annealer on rangelimited constraint satisfaction problems. arXiv preprint arXiv:1502.02098 (2015)
32. King, J., Yarkoni, S., Nevisi, M.M., Hilton, J.P., McGeoch, C.C.: Benchmarking a quantum annealing processor with the time-to-target metric. arXiv preprint
arXiv:1508.05087 (2015)
18
Carleton Coffrin, Harsha Nagarajan, and Russell Bent
33. King, J., Yarkoni, S., Raymond, J., Ozfidan, I., King, A.D., Nevisi, M.M., Hilton,
J.P., McGeoch, C.C.: Quantum annealing amid local ruggedness and global frustration (2017), https://arxiv.org/abs/1701.04579
34. Klymko, C., Sullivan, B.D., Humble, T.S.: Adiabatic quantum programming: minor embedding with hard faults. Quantum Information Processing 13(3), 709–729
(2014), http://dx.doi.org/10.1007/s11128-013-0683-9
35. Koch, T., Achterberg, T., Andersen, E., Bastert, O., Berthold, T., Bixby, R.,
Danna, E., Gamrath, G., Gleixner, A., Heinz, S., Lodi, A., Mittelmann, H., Ralphs,
T., Salvagnin, D., Steffy, D., Wolter, K.: Miplib 2010: Mixed integer programming
library version 5. Mathematical Programming Computation 3(2), 103–163 (2011)
36. Lokhov, A.Y., Vuffray, M., Misra, S., Chertkov, M.: Optimal structure and parameter learning of ising models (2016), https://arxiv.org/abs/1612.05024
37. Lucas, A.: Ising formulations of many np problems. Frontiers in Physics 2, 5
(2014), http://journal.frontiersin.org/article/10.3389/fphy.2014.00005
38. Mandrà, S., Zhu, Z., Wang, W., Perdomo-Ortiz, A., Katzgraber, H.G.: Strengths
and weaknesses of weak-strong cluster problems: A detailed overview of state-ofthe-art classical heuristics versus quantum approaches. Phys. Rev. A 94, 022337
(Aug 2016), https://link.aps.org/doi/10.1103/PhysRevA.94.022337
39. Mandr, S., Katzgraber, H.G., Thomas, C.: The pitfalls of planar spin-glass benchmarks: Raising the bar for quantum annealers (again) (2017), https://arxiv.org/
abs/1703.00622
40. McMahon, P.L., Marandi, A., Haribara, Y., Hamerly, R., Langrock, C., Tamate, S.,
Inagaki, T., Takesue, H., Utsunomiya, S., Aihara, K., et al.: A fully-programmable
100-spin coherent ising machine with all-to-all connections. Science p. aah5178
(2016)
41. Mitchell, D., Selman, B., Levesque, H.: Hard and easy distributions of sat problems.
In: Proceedings of the Tenth National Conference on Artificial Intelligence. pp.
459–465. AAAI’92, AAAI Press (1992), http://dl.acm.org/citation.cfm?id=
1867135.1867206
42. Modha, D.S.: Introducing a brain-inspired computer. Published online at
http://www.research.ibm.com/articles/brain-chip.shtml (2017), accessed:
04/28/2017
43. Mohseni, M., Read, P., Neven, H., Boixo, S., Denchev, V., Babbush, R.,
Fowler, A., Smelyanskiy, V., Martinis, J.: Commercialize quantum technologies in five years. Nature 543, 171174 (2017), http://www.nature.com/news/
commercialize-quantum-technologies-in-five-years-1.21583
44. Morgado, A., Heras, F., Liffiton, M., Planes, J., Marques-Silva, J.: Iterative and
core-guided maxsat solving: A survey and assessment. Constraints 18(4), 478–534
(2013), http://dx.doi.org/10.1007/s10601-013-9146-2
45. Nethercote, N., Stuckey, P.J., Becket, R., Brand, S., Duck, G.J., Tack, G.:
MiniZinc: Towards a Standard CP Modelling Language, pp. 529–543. Springer
Berlin Heidelberg, Berlin, Heidelberg (2007), http://dx.doi.org/10.1007/
978-3-540-74970-7_38
46. Nieuwenhuis, R.: The IntSat Method for Integer Linear Programming, pp. 574–589.
Springer International Publishing, Cham (2014), http://dx.doi.org/10.1007/
978-3-319-10428-7_42
47. Rieffel, E.G., Venturelli, D., O’Gorman, B., Do, M.B., Prystay, E.M., Smelyanskiy, V.N.: A case study in programming a quantum annealer for hard operational planning problems. Quantum Information Processing 14(1), 1–36 (2015),
http://dx.doi.org/10.1007/s11128-014-0892-x
Ising Processing Units: Potential and Challenges for Discrete Optimization
19
48. Selby, A.: Qubo-chimera. https://github.com/alex1770/QUBO-Chimera (2013)
49. Selby, A.: Efficient subgraph-based sampling of ising-type models with frustration
(2014), https://arxiv.org/abs/1409.3934
50. Venturelli, D., Marchand, D.J.J., Rojo, G.: Quantum annealing implementation of
job-shop scheduling (2015), https://arxiv.org/abs/1506.08479
51. Vuffray, M., Misra, S., Lokhov, A., Chertkov, M.: Interaction screening: Efficient
and sample-optimal learning of ising models. In: Lee, D.D., Sugiyama, M., Luxburg,
U.V., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing
Systems 29, pp. 2595–2603. Curran Associates, Inc. (2016)
52. Yamaoka, M., Yoshimura, C., Hayashi, M., Okuyama, T., Aoki, H., Mizuno, H.:
24.3 20k-spin ising chip for combinational optimization problem with cmos annealing. In: 2015 IEEE International Solid-State Circuits Conference - (ISSCC) Digest
of Technical Papers. pp. 1–3 (Feb 2015)
53. Yoshimura, C., Yamaoka, M., Aoki, H., Mizuno, H.: Spatial computing architecture
using randomness of memory cell stability under voltage control. In: 2013 European
Conference on Circuit Theory and Design (ECCTD). pp. 1–4 (Sept 2013)
54. Zdeborova, L., Krzakala, F.: Statistical physics of inference: thresholds and algorithms. Advances in Physics 65(5), 453–552 (2016), http://dx.doi.org/10.1080/
00018732.2016.1211393
LA-UR-17-23494
| 2cs.AI
|
arXiv:1408.1675v3 [cs.PL] 12 Aug 2014
Database Queries that Explain their Work
James Cheney
Amal Ahmed
Umut A. Acar
University of Edinburgh
[email protected]
Northeastern University
[email protected]
Carnegie Mellon University &
INRIA-Rocquencourt
[email protected]
Abstract
promising results have been identified, a scientist may want to
extract just that information that is needed to explain the result,
without showing all of the intermediate search steps or uninteresting results. Conversely, if the results are counterintuitive, the
scientist may want to identify the underlying data that contributed
to the anomalous result. Missier et al. [24] introduced the idea of
a “golden trail”, or a subset of the provenance trace that explains,
or allows reproduction of, a high-value part of the output. They
proposed techniques for extracting “golden trails” using recursive
Datalog queries over provenance graphs; however, they did not
propose definitions of correctness or reproducibility.
It is a natural question to ask how we know when a proposed
solution, such as Missier et al.’s “golden trail” queries, correctly
explains or can be used to correctly reproduce the behavior of the
original computation. It seems to have been taken for granted that
simple graph traversals suffice to at least overapproximate the desired subset of the graph. As far as we know, it is still an open
question how to define and prove such correctness properties for
most provenance techniques. In fact, these properties might be defined and formalized in a number of ways, reflecting different modeling choices or requirements. In any case, in the absence of clear
statements and proofs of correctness, claims that different forms
of provenance “explain” or allow “reproducibility” are difficult to
evaluate objectively.
The main contribution of this paper is to formalize and prove the
correctness of an approach to fine-grained provenance for database
queries. We build on our approach developed in prior work, which
we briefly recapitulate. Our approach is based on analogies between the goals of provenance tracking for databases and workflows, and those of classical techniques for program comprehension and analysis, particularly program slicing [31] and information flow [28]. Both program slicing and information flow rely critically on notions of dependence, such as the familiar control-flow
and data-flow dependences in programming languages.
We previously introduced a provenance model for NRC (including difference and aggregation operations) called dependency
provenance [12], and showed how it can be used to compute data
slices, that is, subsets of the input to the query that include all of
the information relevant to a selected part of the output. Some other
forms of provenance for database query languages, such as howprovenance [20], satisfy similar formal guarantees that can be used
to predict how the output would change under certain classes of input changes, specifically those expressible by semiring homomorphisms. For example, Amsterdamer et al.’s system [3] is based on
the semiring provenance model, so the effects of deletions on parts
of the output can be predicted by inspecting their provenance, but
other kinds of changes are not supported.
More recently, we proposed an approach to provenance called
self-explaining computation [11] and explored it in the context
of a general-purpose functional programming language [1, 27].
In this approach, detailed execution traces are used as a form of
Provenance for database queries or scientific workflows is often
motivated as providing explanation, increasing understanding of
the underlying data sources and processes used to compute the
query, and reproducibility, the capability to recompute the results
on different inputs, possibly specialized to a part of the output.
Many provenance systems claim to provide such capabilities; however, most lack formal definitions or guarantees of these properties, while others provide formal guarantees only for relatively limited classes of changes. Building on recent work on provenance
traces and slicing for functional programming languages, we introduce a detailed tracing model of provenance for multiset-valued
Nested Relational Calculus, define trace slicing algorithms that extract subtraces needed to explain or recompute specific parts of the
output, and define query slicing and differencing techniques that
support explanation. We state and prove correctness properties for
these techniques and present a proof-of-concept implementation in
Haskell.
Keywords
provenance, database queries, slicing
1. Introduction
Over the past decade, the use of complex computer systems in science has increased dramatically: databases, scientific workflow systems, clusters, and cloud computing based on frameworks such as
MapReduce [16] or PigLatin [26] are now routinely used for scientific data analysis. With this shift to computational science based
on (often) unreliable components and noisy data comes decreased
transparency, and an increased need to understand the results of
complex computations by auditing the underlying processes.
This need has motivated work on provenance in databases, scientific workflow systems, and many other settings [7, 25]. There is
now a great deal of research on extending such systems with rich
provenance-tracking features. Generally, these systems aim to provide high-level explanations intended to aid the user in understanding how a computation was performed, by recording and presenting
additional “trace” information.
Over time, two distinct approaches to provenance have emerged:
(1) the use of annotations propagated through database queries to
illustrate where-provenance linking results to source data [7], lineage or why-provenance linking result records to sets of witnessing input records [15], or how-provenance describing how results
were produced via algebraic expressions [20], and (2) the use of
graphical provenance traces to illustrate how workflow computations construct final results from inputs and configuration parameters [5, 21, 29]. However, to date few systems formally specify the
semantics of provenance or give formal guarantees characterizing
how provenance “explains” results.
For example, scientists often conduct parameter sweeps to
search for interesting results. The provenance trace of such a computation may be large and difficult to navigate. Once the most
1
In general, we use sequences of natural numbers [i1 , . . . , in ] ∈ N∗
as indices, and we maintain a stronger invariant: the set of indexes
used in a multiset must form a prefix code. We define a semantics
for NRC expressions over such collections that is fully deterministic and does not resort to generation of fresh intermediate labels;
in particular, the union operation adjusts the labels to maintain distinctness.
The second technical challenge involves extending patterns for
partial collections. In our previous work, any subexpression of a
value can be replaced by a hole. This works well in a conventional functional language, where typical values (such as lists and
trees) are essentially initial algebras built up by structural induction.
However, when we consider unordered collections such as bags,
our previous approach becomes awkward.
For example, if we want to use a pattern to focus on only the
B field of the second record in query result Q(R), we can only
do this by deleting the other record values, and the smallest such
pattern is {[1, r1 ].✷, [2, r2 ].hA:✷, B:3, C:✷i, [2, r3 ].✷}. This is
tolerable if there are only a few elements, but if there are hundreds
or millions of elements and we are only interested in one, this is
a significant overhead. Therefore, we introduce enriched patterns
that allow us to replace entire subsets with holes, for example,
p′ = {[2, r2 ].hB:3; ✷i} ∪˙ ✷. Enriched patterns are not just a
convenience; we show experimentally that they allow traces and
slices to be smaller (and computed faster) by an order of magnitude
or more.
provenance. Traces explain results in the sense that they can be
replayed to recompute the results, and they can be sliced to obtain
smaller traces that provide more concise explanations of parts of
the output. Trace slicing also produces a slice of the input showing
what was needed by the trace to compute the output. Moreover,
other forms of provenance can be extracted from traces (or slices),
and we also showed that traces can be used to compute program
slices efficiently through lazy evaluation. Finally, we showed how
traces support differential slicing techniques that can highlight the
differences between program runs in order to explain and precisely
localize bugs in the program or errors in the input data.
Our long-term vision is to develop self-explaining computation
techniques covering all components used in day-to-day scientific
practice. Databases are probably the single most important such
component. Since our previous work already applies to a generalpurpose programming language, one way to proceed would be to
simply implement an interpreter for NRC in this language, and inherit the slicing behavior from that. However, without some further
inlining or optimization, this naive strategy would yield traces that
record both the behavior of the NRC query and its interpreter, along
with internal data structures and representation choices whose details are (intuitively) irrelevant to understanding the high-level behavior of the query.
1.1 Technical overview
In this paper, we develop a tracing semantics and trace slicing techniques tailored to NRC (over a multiset semantics). This semantics evaluates a query Q over an input database (i.e. environment γ
mapping relation names to table values), yielding the usual result
value v as well as a trace T . Traces are typically large and difficult
to decipher, so we consider a scenario where a user has run Q, inspected the results v, and requests an explanation for a part of the
result, such as a field value of a single record. As in our previous
work for functional programs, we use partial values with “holes” ✷
to describe parts of the output that are to be explained. For example, if the result of a program is just a pair (1, 2) then the pattern
(1, ✷) can be used to request an explanation for just the first component. Given a partial value p matching the output, our approach
computes a “slice” consisting of a partial trace and a partial input
environment, where components not necessary for recomputing the
explained output part p have been deleted.
The main technical contribution of this paper over our previous
work [1, 27] is its treatment of tracing and slicing for collections.
There are two underlying technical challenges; we illustrate both
(and our solutions) via a simple example query Q = σA<B (R) ∪
ρA7→B,B7→A (σA≥B (R)) over a table R with attributes A, B. Here,
σφ is relational selection of all tuples satisfying a predicate φ
and ρA7→B,B7→A is renaming. Thus, Q simply swaps the fields of
records where A ≥ B, and leaves other records alone.
The first challenge is how to address elements of multisets
reliably across different executions and support propagation of
addresses in the output backwards towards the input. Our solution
is to use a mildly enriched semantics in which multiset elements
carry explicit labels; that is, we view multisets of elements from X
as functions I → X from some index set to X. For example, if R
is labeled as follows and we use this enriched semantics to evaluate
the above query Q on R, we get a result:
id
[r1 ]
R=
[r2 ]
[r3 ]
A
1
2
4
B
2
3
3
C
7
8
9
id
[1, r1 ]
Q(R) =
[1, r2 ]
[2, r3 ]
A
1
2
3
B
2
3
4
1.2 Outline
The rest of this paper is structured as follows. Section 2 reviews the
(multiset-valued) Nested Relational Calculus and presents our tracing semantics and deterministic labeling scheme. Section 3 presents
trace slicing, including a simple form of patterns. Section 4 shows
how to enrich our pattern language to allow for partial record and
partial set patterns, which considerably increase the expressiveness
of the pattern language, leading to smaller slices. Section 5 presents
the query slicing algorithm and shows how to compute differential
slices. Section 6 presents additional examples and discussion. Section 7 presents our implementation demonstrating the benefits of
laziness and enriched patterns for trace slicing. Section 8 discusses
related and future work and Section 9 concludes.
Due to space limitations, and in order to make room for examples and high-level discussion, some (mostly routine) formal details
and proofs are relegated to the appendix.
2. Traced Evaluation for NRC
The nested relational calculus [9] is a simply-typed core language
with collection types that can express queries on nested data similar
to those of SQL on flat relations, but has simpler syntax and cleaner
semantics. In this section, we show how to extend the ideas and
machinery developed in our previous work on traces and slicing for
functional languages [27] to NRC. Developing formal foundations
for tracking provenance in the presence of unordered collections
presents a number of challenges not encountered in the functional
programming setting, as we will explain.
2.1 Syntax and Dynamic Semantics
Figure 1 presents the abstract syntax of NRC expressions, values,
and traces. The expression ∅ denotes the empty collection, {e} constructs a singleton collection, and e1 ∪ e2 takes the (multiset) union
of two collections. The operation sum e computes the sum of a collection of integers, while the predicate empty e tests whether the
collection denoted by e is empty. Additional aggregation operations
such as count, maximum and average can
S easily be accommodated.
Finally, the comprehension operation {e′ | x ∈ e} iterates over
C
7
8
9
where in each case the id column contains a distinct index ri . In
Q(R), the first ‘1’ or ‘2’ in each index indicates whether the row
was generated by the left or right subexpression in the union (’∪’).
2
γ, e ⇓ v, T
γ, c ⇓ c, c
γ, e1 ⇓ c1 , T1
···
γ, en ⇓ cn , Tn
γ, f(e1 , . . . , en ) ⇓ f̂(c1 , . . . , cn ), f(T1 , . . . , Tn )
γ, e1 ⇓ v1 , T1
···
γ, x ⇓ γ(x), x
γ, e ⇓ hA1 :v1 , . . . , An :vn i, T
γ, en ⇓ vn , Tn
γ, hA1 :e1 , . . . , An :en i ⇓ hA1 :v1 , . . . , An :vn i, hA1 :T1 , . . . , An :Tn i
γ, e ⇓ true, T
γ, e1 ⇓ v1 , T1
γ, if(e, e1 , e2 ) ⇓ v1 , if(T, e1 , e2 ) ⊲true T1
γ, e1 ⇓ v1 , T1
γ, e2 ⇓ v2 , T2
γ, e1 ∪ e2 ⇓ 1 · v1 ⊎ 2 · v2 , T1 ∪ T2
γ, e ⇓ false, T
γ, e2 ⇓ v2 , T2
γ, if(e, e1 , e2 ) ⇓ v2 , if(T, e1 , e2 ) ⊲false T2
γ, e ⇓ v, T
γ, x ∈ v, e′ ⇓∗ v′ , Θ
[
[
γ, {e′ | x ∈ e} ⇓ v′ , {e′ | x ∈ T } ⊲ Θ
γ, e ⇓ v, T
v=∅
γ, empty e ⇓ true, empty T
γ, x ∈ v, e ⇓∗ v, Θ
∗
γ, x ∈ ∅, e ⇓ ∅, ∅
γ, e1 ⇓ v1 , T1
γ[x 7→ v1 ], e2 ⇓ v2 , T2
γ, let x = e1 in e2 ⇓ v2 , let x = T1 in T2
γ, e.Ai ⇓ vi , T.Ai
γ, ∅ ⇓ ∅, ∅
γ, e ⇓ v, T
γ, {e} ⇓ {ǫ.v}, {T }
γ, e ⇓ {ℓ1 .v1 , . . . , ℓn .vn }, T
γ, sum e ⇓ v1 +̂ . . . +̂vn , sum T
γ, e ⇓ v, T
v 6= ∅
γ, empty e ⇓ false, empty T
γ, x ∈ v1 , e ⇓∗ v1′ , Θ1
γ, x ∈ v2 , e ⇓∗ v2′ , Θ2
∗ ′
γ, x ∈ v1 ⊎ v2 , e ⇓ v1 ⊎ v2′ , Θ1 ⊎ Θ2
γ[x 7→ v], e ⇓ v′ , T
γ, x ∈ {ℓ.v}, e ⇓∗ ℓ · v′ , {ℓ.T }
Figure 2. Traced evaluation.
Other operations on labels and labeled collections will be introduced in due course.
The labels on the elements of a collection provide us with a persistent address for a particular element of the collection. This capability is essential when asking and answering provenance queries
about parts of the source or output data, and when tracking finegrained dependencies.
Both expressions and traces are subject to a type system. NRC
types include collection types {τ } which are often taken to be sets,
bags (multisets), or lists, though in this paper, we consider multiset
collections only. However, types do not play a significant role in
this paper so the typing rules are omitted. For expressions, the
typing judgment Γ ⊢ e : τ is standard and the typing rules for
trace well-formedness Γ ⊢ T : τ are presented in Appendix A.
f ::= + | − | ∗ | / | = | < | ≤ | · · ·
e ::= c | f(e1 , . . . , en ) | x | let e = x in e′
| hA1 : e1 , . . . , An : en i | e.A | if(e, e′ , e′′ )
S
| ∅ | {e} | e1 ∪ e2 | {e′ | x ∈ e}
| empty e | sum e | · · ·
Labels
ℓ ::= ℓ.ℓ′ | ǫ | n
Values
v ::= c | hA1 : v1 , . . . An : vn i
| ∅ | {ℓ1 .v1 , . . . , ℓn .vn }
Environments γ ::= [x1 7→ v1 , . . . , xn 7→ vn ]
Traces
T ::= · · · | if(T, e′ , e′′ ) ⊲true T | if(T, e′ , e′′ ) ⊲false T
S
|
{e | x ∈ T } ⊲ Θ
Trace Sets
Θ ::= {ℓ1 .T1 , . . . , ℓn .Tn }
Types
τ ::= int | bool | hA1 : τ1 , . . . , An : τn i | {τ }
Type Contexts Γ ::= x1 : τ1 , . . . , xn : τn
Operations
Expressions
Traced evaluation NRC traces include a trace form corresponding to each of the expressions described above. The structure of the
traces is best understood by inspecting the typing rules (Figure 12)
and the traced evaluation rules (Figure 2), which define a judgment
γ, e ⇓ v, T indicating that evaluating an expression e in environment γ yields a value v and a trace T . We assume an environment
Σ associating constants and function symbols with their types, and
write f̂ or +̂ for the semantic operations corresponding to f or +,
and so on. In most cases, the trace form is similar to the expression
form; for example the trace of a constant or variable is a constant
trace c, the trace of a primitive operation f(e1 , . . . , en ) is a primitive operation trace f(T1 , . . . , Tn ) applied to the traces Ti of the
arguments ei , the trace of a record expression is a trace record constructor hA1 : T1 , . . . , An : Tn i, and the trace of a field projection
e.A is a trace T.A. Also, the trace of a let-binding is a let-binding
trace let x = T1 in T2 , where x is bound in T2 . In these cases,
the traces mimic the expression structure.
The traced evaluation rules for conditionals illustrate that traces
differ from expressions in recording control flow decisions. The
trace of a conditional is a conditional trace if(T, e1 , e2 ) ⊲b T ′
where T is the trace of the conditional test, b is the Boolean
value of the test e1 , and T ′ is the trace of the taken branch. The
expressions e1 and e2 are not strictly necessary but retained to
preserve structural similarity to the original expression.
The trace of ∅ is a constant trace ∅. To evaluate a singletoncollection constructor {e}, we evaluate e to obtain a value v and
return the singleton {ǫ.v} with empty label ǫ. We return the sin-
Figure 1. NRC expressions, values, traces, and types.
the collection obtained by evaluating e, evaluating e′ (x) with x
bound to each element of the collection in turn, and returning a
collection containing the union of all of the results. We sometimes
consider pairs (e1 , e2 ), a special case of records h#1 : e1 , #2 : e2 i
using two designated field names #1 and #2 . Many trace forms are
similar to those for expressions; only the differences are shown.
Labels are sequences ℓ = [ii , . . . , in ] ∈ N∗ , possibly empty.
The empty sequence is written ǫ, and labels can be concatenated
ℓ · ℓ′ ; concatenation is associative. Record field names are written
A, B, A1 , A2 , . . ..
Values in NRC include constants c, which we assume include at least booleans and integers. Record values are essentially partial functions from field names to values, written hA1 :
v1 , . . . , An : vn i. Collection values are essentially partial, finitedomain functions from labels in N∗ to values, which we write
{ℓ1 .v1 , . . . , ℓn .vn }. Since they denote functions, collections and
records are identified up to reordering of their elements, and their
field names or labels are always distinct. We write ℓ · v for the
operation that prepends ℓ to each of the labels in a set v, that is,
ℓ · {ℓ1 .v1 , . . . , ℓn .vn } = {ℓ · ℓ1 .v1 , . . . , ℓ · ℓn .vn } .
3
gleton trace {T } recording the trace for the evaluation of the element. To evaluate the union of two expressions, we evaluate each
one and take the semantic union (written ⊎) of the resulting collections, with a ‘1’ or ‘2’ concatenated onto the beginning of each
label to reflect whether each element came from the first or second
part of the union; the union trace T1 ∪ T2 records the traces for the
evaluation of the two subexpressions. For sum e, evaluating e yields
a collection of numbers whose sum we return, together with a sum
trace sum T recording the trace for evaluation of e. Evaluation of
emptiness tests empty e is analogous,
S yielding a trace empty T .
To evaluate a comprehension {e′ | x ∈ e}, we first evaluate e,
which yields a collection v and trace T , and then (using auxiliary
judgment γ, x ∈ v, e′ ⇓∗ v ′ , Θ) evaluate e′ repeatedly with
x bound to each element vi of the collection v to get resulting
values vi′ and corresponding traces Ti′ . We return a new collection
v ′ = {ℓ1 · v1′ , . . . , ℓn · vn′ }; similarly we return a labeled set of
traces Θ = {ℓ1 .T1 , . . . , ℓn .Tn }. (Analogously to values, trace sets
are essentially finite partial functions from labels to traces). For
each of these collections, we prepend the appropriate label ℓi of the
corresponding input element.
A technical point of note is that the resulting trace Ti may contain free occurrences of x. As in our trace semantics for functional
programs, these variables serve as markers in Ti that will be critical
for the trace replaySsemantics. The comprehension trace records,
using the notation {e′ | x ∈ T } ⊲ {ℓ1 .T1 , . . . , ℓn .Tn }, that the
trace T was used to compute a multiset v, and x was bound to
each element ℓi .vi in v in turn, with trace Ti showing how the corresponding subset of the result was computed. The comprehension
trace also records the expression e′ and bound variable x, which are
again not strictly necessary but preserve the structural similarity to
the original expression.
At this point it is useful to provide some informal motivation
for the labeling semantics, compared for example to other semantics that use annotations or labels as a form of provenance. We do
not view the labels themselves as provenance; instead, they provide a useful infrastructure for traces, which do capture a form of
provenance. Moreover, by calculating the label of each part of an
intermediate or final result deterministically (given the labels on
the input), we provide a way to reliably refer to parts of the output,
which otherwise may be unaddressable in a multiset-valued semantics. This is essential for supporting compositional slicing for operations such as let-binding or comprehension, where the output of
one subexpression becomes a part of the input for another.
A central point of our semantics is that evaluation preserves the
property that labels uniquely identify the elements of each multiset.
This naturally assumes that the labels on the input collections
are distinct. In fact, a stronger property is required: evaluation
preserves the property that set labels form a prefix code. In the
following, we write x ≤ y to indicate that sequence x is a prefix of
sequence y.
γ, T y v
···
γ, T1 y c1
γ, c y c
γ, Tn y cn
γ, f(T1 , . . . , Tn ) y f̂(c1 , . . . , cn )
γ, T1 y v1
γ, x y γ(x)
γ[x 7→ v1 ], T2 y v2
γ, let x = T1 in T2 y v2
γ, T1 y v1
···
γ, Tn y vn
γ, hA1 :T1 , . . . , An :Tn i y hA1 :v1 , . . . , An :vn i
γ, T y hA1 :v1 , . . . , An :vn i
γ, T.Ai y vi
γ, T y false
γ, T y true
γ, T1 y v1
γ, if(T, e1 , e2 ) ⊲true T1 y v1
γ, T y v
γ, T2 y v2
γ, if(T, e1 , e2 ) ⊲false T2 y v2
γ, T1 y v1
γ, T2 y v2
γ, T1 ∪ T2 y 1 · v1 ⊎ 2 · v2
γ, T y v
v=∅
γ, empty T y true
γ, ∅ y ∅
γ, {T } y {ǫ.v}
γ, T y {ℓ1 .v1 , . . . , ℓn .vn }
γ, sum T y v1 +̂ . . . +̂ vn
γ, T y v
v 6= ∅
γ, empty T y false
γ, T y v
γ, x ∈ v, Θ y∗ v′
[
γ, {e | x ∈ T } ⊲ Θ y v′
γ, x ∈ v, Θ y∗ v ′
∗
γ, x ∈ ∅, Θ y ∅
γ, x ∈ v1 , Θ y∗ v1′
γ, x ∈ v2 , Θ y∗ v2′
γ, x ∈ v1 ⊎ v2 , Θ y∗ v1′ ⊎ v2′
ℓi ∈ dom(Θ)
γ[x 7→ vi ], Θ(ℓi ) y vi′
γ, x ∈ {ℓi .vi }, Θ y∗ ℓi · vi′
Figure 3. Trace replay.
Theorem 2.2. If γ is prefix-labeled and γ, e ⇓ v, T then v and
T are both prefix-labeled. Moreover, if γ and v are prefix-labeled
and γ, x ∈ v, e ⇓∗ v ′ , Θ then v ′ and Θ are prefix-labeled, and in
addition dom(Θ) = dom(v) ≤ dom(v ′ ).
Proof. By induction on derivations. The key cases are those for
union, which is straightforward, and for comprehensions. For the
latter case we need the second part, to show that whenever γ and
v are prefix-labeled, if γ, x ∈ v, e ⇓∗ v ′ , Θ then v ′ and Θ are
prefix-labeled, and in addition dom(Θ) = dom(v) and dom(v) is
a sub-prefix code of dom(v ′ ).
The prefix code property is needed later in the slicing algorithms, when we will need it to match elements of collections produced by comprehensions with corresponding elements of the trace
set Θ. From now on, we assume that all values, environments, and
traces are prefix-labeled, so any labeled set is assumed to have the
prefix code property.
We will use the following query as a running example.
[
Q = {if(x.B = 3, {hA:x.A, B:x.Ci}, {}) | x ∈ R}
Definition 2.1. A prefix code over Σ is a set of sequences L ⊆ Σ∗
such that for every x, y ∈ L, if x ≤ y then x = y. A sub-prefix
code of a prefix code L is a prefix code L′ such that for all x ∈ L
there exists y ∈ L′ such that y ≤ x. We write L′ ≤ L to indicate
that L′ is a sub-prefix code of L. We say that L and L′ are prefixdisjoint when no element of L is a prefix of an element of L′ and
vice versa.
This is a simple selection query; it identifies records in R that have
B-value of 3, and returns record hA:x.A, B:x.Ci containing x’s A
value and its C value renamed to B. The result of Q on the input R
in the introduction is Q(R) = {[r2 ].hA:2, B:8i, [r3 ].hA:4, B:9i},
and the trace is:
S
T = { | x ∈ R} ⊲ { [r1 ].if(x.B = 3, , ) ⊲false {},
[r2 ].if(x.B = 3, , ) ⊲true { },
[r3 ].if(x.B = 3, , ) ⊲true { } }
Let v be a collection v = {ℓ1 .v1 , . . . , ℓn .vn }. We define the
domain of v to be dom(v) = {ℓ1 , . . . , ℓn }. . We say that a value
or value environment is prefix-labeled if for every collection v
occurring in it, the labels ℓ1 , . . . , ℓn are distinct and dom(v) is a
prefix code. Similarly, we say that a trace is prefix-labeled if every
labeled trace set Θ = {ℓ1 .T1 , . . . , ℓn .Tn } is prefix-labeled.
4
where indicates omitted (easily inferrable) subexpressions.
Proposition 2.6 (Consistency). If γ, e ⇓ v, T then γ, T y v.
Replay We introduce a judgment γ, T y v for replaying a
trace on a (possibly different) environment γ. The rules for replaying NRC traces are presented in Figure 3. Many of the rules are
straightforward or analogous to the corresponding evaluation rules.
Here we only discuss the replay rules for conditional and comprehension traces.
For the conditional rules, the basic idea is as follows. If replaying the trace T of the test yields the same boolean value b as
recorded in the trace, we replay the trace of the taken branch. If the
test yields a different value, then replay fails.
S
To replay a comprehension trace {e | x ∈ T } ⊲ Θ, rule
RC OMP first replays trace T to update the set of elements v
over which we will iterate. We define a separate judgment γ, x ∈
v, Θ y∗ v ′ to iterate over set v and replay traces on the corresponding elements. For elements ℓi ∈ dom(Θ), we replay the
corresponding trace Θ(ℓi ). Replay fails if v contains any labels not
present in Θ.
Replaying a trace can fail if either the branch taken in a conditional test differs from that recorded in the trace, or the intermediate
set obtained from rerunning a comprehension trace includes values
whose labels are not present in the trace. This means, in particular,
that changes to the input can be replayed if they only change base
values or delete set elements, but changes leading to additions of
new labels to sets involved in comprehensions typically cannot be
replayed.
Returning to the running example, suppose we change field B
of row r1 of R from 2 to 5. This change has no effect on the
control flow choice taken in Q, and replaying the trace T succeeds.
Likewise, changing the A or C field of any column of R has no
effect, since these values do not affect the control flow choices in
T . However, changing the B field of a row to 3 (or changing it from
3 to something else) means that replay will fail.
Finally, trace replay is faithful to ordinary evaluation in the following sense: if T is generated by running e in γ and we successfully replay T on γ ′ then we obtain the same value (and same trace)
as if we had rerun e from scratch in γ ′ , and vice versa:
Proposition 2.7 (Fidelity). If γ, e ⇓ v, T , then for any γ ′ , v ′ we
have γ ′ , e ⇓ v ′ , T if and only if γ ′ , T y v ′ .
Observe that consistency is a special case of fidelity (with γ ′ =
γ, v ′ = v. Moreover, the “if” direction of fidelity holds even though
replay can fail, because we require that γ ′ , e ⇓ v ′ , T holds for the
same trace T . If γ ′ , e ⇓ v ′ , T ′ is derivable but only for a different
trace T ′ , then replay fails.
3. Trace Slicing
The goal of the trace slicing algorithm we consider is to remove
information from a trace and input that is not needed to recompute a part of the output. To accommodate these requirements, we
introduce traces and values with holes and more generally, we consider patterns that represent relations on values capturing possible
changes. In this section, we limit attention to pairs, and consider
slicing for a class of simple patterns. We consider records and more
expressive enriched patterns in the next section.
We extend traces with holes ✷
T
::=
··· | ✷
and define a subtrace relation ⊑ that is essentially a syntactic
precongruence on traces and trace sets such that ✷ ⊑ T holds
and Θ ⊆ Θ′ implies Θ ⊑ Θ′ . (The definition of ⊑ is shown in
full in the companion technical report.) Intuitively, holes denote
parts of traces we do not care about, and we can think of a trace
T with holes as standing for a set of possible complete traces
{T ′ | T ⊑ T ′ } representing different ways of filling in the holes.
The syntax of simple patterns p, set patterns sp, and pattern
environments ρ is:
2.2 Key properties
Before moving on to consider trace slicing, we identify some properties that formalize the intuition that traces are consistent with and
faithfully record execution.
Evaluation and traced evaluation are deterministic:
p
sp
ρ
Proposition 2.3 (Determinacy). If γ, e ⇓ v, T and γ, e ⇓ v ′ , T ′
then v = v ′ and T = T ′ . If γ, T y v and γ, T y v ′ then v = v ′ .
::=
::=
::=
✷ | ✸ | c | (p1 , p2 ) | sp
∅ | {ℓ1 .p1 , . . . , ℓn .pn }
[x1 7→ p1 , . . . , xn 7→ pn ]
Essentially, a pattern is a value with holes in some positions. The
meaning of each pattern is defined through a relation hp that says
when two values are equivalent with respect to a pattern. This relation is defined in Figure 5. A hole ✷ indicates that the part of
the value is unimportant, that is, for slicing purposes we don’t care
about that part of the result. Its associated relation h✷ relates any
two values. An identity pattern ✸ is similar to a hole: it says that
the value is important but its exact value is not specified, and its
associated relation h✸ is the identity relation on values. Complete
set patterns {ℓ1 .p1 , . . . , ℓn .pn } specify the labels and patterns for
all of the elements of a set; that is, such a pattern relates only sets
that have exactly the labeled elements specified and whose corresponding values match according to the corresponding patterns.
We define the union of two set patterns as {ℓi .pi } ⊎ {ℓ′i .p′i } =
{ℓi .pi , ℓ′i .p′i } provided their domains ℓ~i and ℓ~′i are prefix-disjoint.
We define a (partial) least upper bound operation on patterns p ⊔ p′
such that for any v, v ′ we have v hp⊔p′ v ′ if and only if v hp v ′
and v hp′ v ′ ; the full definition is shown in Figure 4. We define
the partial ordering p ⊑ p′ as p ⊔ p′ = p′ . We say that a value v
matches pattern p if p ⊑ v. Observe that this implies v hp v. We
extend the ⊔ and ⊑ operations to pattern environments ρ pointwise,
that is, (ρ ⊔ ρ′ )(x) = ρ(x) ⊔ ρ(x′ ).
We define several additional operations on patterns that are
needed for the slicing algorithm. Consider the following singleton
When the trace is irrelevant, we write γ, e ⇓ v to indicate that
γ, e ⇓ v, T for some T .
Traced evaluation is type-safe and produces well-typed traces,
and trace replay is also type-safe.
Theorem 2.4. If γ : Γ and Γ ⊢ e : τ and γ, e ⇓ v, T then v : τ
and Γ ⊢ T : τ . If γ : Γ and Γ ⊢ T : τ and γ, T y v then v : τ .
Traces can be represented using pointers to share common
subexpressions; using this DAG representation, traces can be stored
in space polynomial in the input. (This sharing happens automatically in our implementation in Haskell.)
Proposition 2.5. For a fixed e, if γ, e ⇓ v, T then the sizes of v
and of the DAG representation of T are at most polynomial in |γ|.
Proof. Most cases are straightforward. The only non-trivial case is
for comprehensions, where we need a stronger induction hypothesis: if γ, x ∈ v, e ⇓∗ v ′ , Θ then the sizes of v ′ and of the DAG
representation of Θ are at most polynomial in |γ|.
Furthermore, traced evaluation produces a trace that replays
to the same value as the original expression run on the original
environment. We call this property consistency.
5
✷⊔p=p⊔✷
✸⊔p=p⊔✸
c⊔c
(p1 , p2 ) ⊔ (p′1 , p′2 )
{ℓi .pi } ⊔ {ℓi .p′i }
=
=
=
=
=
p
p[✸/✷]
c
(p1 ⊔ p′1 , p2 ⊔ p′2 )
{ℓi .pi ⊔ p′i }
✸[✸/✷] = ✷[✸/✷]
c[✸/✷]
(p1 , p2 )[✸/✷]
{ℓi .pi }[✸/✷]
=
=
=
=
✸
c
(p1 [✸/✷], p2 [✸/✷])
{ℓi .pi [✸/✷]}
patterns describing the subsets obtained from the first and second
subtraces in the union pattern, respectively.
The rule SC OMP uses a similar idea to let-binding. We slice the
trace set Θ using an auxiliary judgment p, x.Θ ց∗ ρ, Θ0 , p0 , obtaining a sliced input environment, sliced trace set Θ0 , and pattern
p0 describing the set of values to which x was bound. We then use
p0 to slice backwards through the subtrace T that constructed the
set. The auxiliary slicing judgment for trace sets has three rules:
a trivial rule SE MPTY∗ when the set is empty, rule SS NG∗ that
uses label projection p[ℓ] to handle a singleton trace set, and a rule
SU NION∗ handling larger trace sets by decomposing them into subsets. This is essentially a structural recursion over the trace set, and
is deterministic even though the rules can be used to decompose the
trace sets in many different ways, because ⊎ is associative. Rule
SU NION∗ also requires that we restrict the set pattern to match the
domains of the corresponding trace patterns.
The slicing rules SS UM and SE MPTY P follow the same idea as
for primitive operations at base type: we require that the whole set
value be preserved exactly. As discussed by Perera et al. [27] with
primitive operations, this is a potential source of overapproximation, since (for example) for an emptiness test, all we really need is
to preserve the number of elements in the set, not their values. The
last rule shows how to slice pair patterns when the pattern is ✸: we
slice both of the subtraces by ✸ and combine the results.
It is straightforward to show that the slicing algorithm is welldefined for consistent traces. That is, if γ, T y v and p ⊑ v then
there exists ρ, S such that p, T ց ρ, S holds, where ρ ⊑ γ and
S ⊑ T . We defer the correctness theorem for slicing using simple
patterns to the end of the next section, since it is a special case of
correctness for slicing using enriched patterns.
Continuing our running example, consider the pattern p =
{[r2 ].hA:✷, B:8i, [r3 ].✷}. The slice of T with respect to this pattern is of the form:
S
T ′ = { | x ∈ R} ⊲ { [r1 ].✷,
[r2 ].if(x.B = 3, , ) ⊲true { },
[r3 ].✷ }
where:
Figure 4. Least upper bound for simple patterns
v hp v ′
v h✷ v
′
v h✸ v
v1 hp2 v2′
v1 hp1 v1′
(v1 , v2 ) h(p1 ,p2 ) (v1′ , v2′ )
c hc c
vi hpi vi′ (i ∈ {1, . . . , n})
′
{ℓ1 .v1 , . . . , ℓn .vn } h{ℓ1 .p1 ,...,ℓn .pn } {ℓ1 .v1′ , . . . , ℓn .vn
}
γ hρ γ ′ ⇐⇒ ∀x ∈ dom(ρ). γ(x) hρ(x) γ ′ (x)
Figure 5. Simple pattern equivalence.
extraction operation p.ǫ and label projection operation p[ℓ]:
({ǫ.p}).ǫ
sp[ℓ]
=
=
p
{ℓ′ .v | ℓ.ℓ′ .v ∈ sp}
✷.ǫ = ✷ ✸.ǫ = ✸
✷[ℓ] = ✷ ✸[ℓ] = ✸
These operations are only used for the above cases; they have no
effect on constant or pair patterns. For sets, p[ℓ] extracts the subset
of p whose labels start with ℓ, truncating the initial prefix ℓ, while if
p is ✷ or ✸ then again p[ℓ] returns the same kind of hole. Moreover,
dom(p[ℓ]) is a prefix code if dom(p) is.
Suppose L is a prefix code. We define restriction of a set pattern
p to L as follows:
sp|L = {ℓ.ℓ′ .p ∈ sp | ℓ ∈ L}
✷|L = ✷
and the slice of R is R′ = {[r1 ].✷, [r2 ].hA:✷, B:3, C:8i, [r3 ].✷}.
(That is, R′ is the value of ρ(R), where ρ is the pattern environment produced by slicing T with respect to p.) Observe that the
value of A is not needed but the value of B must remain 3 in order to preserve the control flow behavior of the trace on r2 . The
holes indicate that changes to r1 and r3 in the input cannot affect r2 . However, a trace matching T ′ cannot be replayed if any
of r1 , r2 , r3 are deleted from the input or if the A field is removed
from an input record, because the replay rules require all of the labels mentioned in collections or records in T ′ to be present. We
now turn our attention to enriched patterns, which mitigate these
drawbacks.
✸|L = ✸
It is easy to see that dom(p|L ) ⊆ dom(p) so dom(p|L ) is a prefix
code if dom(p) is, so this operation is well-defined on collections:
Lemma 3.1. If sp is prefix-labeled and L is a prefix code then
sp[ℓ] and sp|L are prefix-labeled.
We show other properties of patterns in Appendix B.
Backward Slicing The rules for backward slicing are given in
Figure 6. The judgment p, T ց ρ, T ′ slices trace T with respect
to a pattern p to yield the slice T ′ and sliced input environment
ρ. The sliced input environment records what parts of the input
are needed to produce p; this is needed for slicing operations such
as let-binding or comprehensions. The main new ideas are in the
rules for collection operations, particularly comprehensions. The
slicing rules SC ONST, SP RIM, SVAR, SL ET, SPAIR, SP ROJi , and
SI F follow essentially the same idea as in our previous work [1, 27].
We focus discussion on the new cases, but we review the key ideas
for these operations here in order to make the presentation selfcontained.
The rules for collections use labels and set pattern operations to
effectively undo the evaluation of the set pattern. The rule SE MPTY
is essentially the same as the constant rule. The rule SS NG uses
the singleton extraction operation p.ǫ to obtain a pattern describing the single element of a singleton set value matching p. The
rule SU NION uses the two projections p[1] and p[2] to obtain the
4. Enriched Patterns
So far we have considered only complete set patterns of the form
{ℓ1 .p1 , . . . , ℓn .pn }. These patterns relate pairs of values that have
exactly n elements labeled ℓ1 , . . . , ℓn , each of which matches
p1 , . . . , pn respectively.
This is awkward, as we already can observe in our running example above: to obtain a slice describing how one record was computed, we need to use a set pattern that lists all of the indexes in the
output. Moreover, the slice with respect to such a pattern may also
include labeled subtraces explaining why the other elements exist
in the output (and no others). This information seems intuitively
irrelevant to the value at ℓ1 , and can be a major overhead if the collection is large. We also have considered only binary pairs; it would
be more convenient to support record patterns directly.
6
p, T ց ρ, S
✷, T ց [], ✷
SH OLE
p, c ց [], c
SC ONST
✸, T1 ց ρ1 , S1
p2 , T2 ց ρ2 [x 7→ p1 ], S2
p1 , T1 ց ρ1 , S1
SL ET
p2 , let x = T1 in T2 ց ρ1 ⊔ ρ2 , let x = S1 in S2
✸, Tn ց ρn , Sn
p[1], T1 ց ρ1 , S1
p[2], T2 ց ρ2 , S2
p, T1 ∪ T2 ց ρ1 ⊔ ρ2 , S1 ∪ S2
✸, T ց ρ, S
SE MPTY P
p, sum T ց ρ, sum S
SU NION
p,
[
p, x ց [x 7→ p], x
p, ∅ ց [], ∅
SVAR
(p, ✷), T ց ρ, S
SP ROJ 1
p, T.#1 ց ρ, S.#1
p.ǫ, T ց ρ, S
SS NG
p, {T } ց ρ, {S}
SE MPTY
p, x.Θ ց∗ ρ′ , Θ′ , p′
p′ , T ց ρ, S
[
SC OMP
{e | x ∈ T } ⊲ Θ ց ρ ⊔ ρ′ , {e | x ∈ S} ⊲ Θ′
✸, T ց ρ, S
SS UM
p, empty T ց ρ, empty S
∅, x.∅ ց∗ [], ∅, ∅
SP RIM
p1 , T1 ց ρ1 , S1
p2 , T2 ց ρ2 , S2
SPAIR
(p1 , p2 ), (T1 , T2 ) ց ρ1 ⊔ ρ2 , (S1 , S2 )
p, T ′ ց ρ′ , S ′
b, T ց ρ, S
SI F
p, if(T, e1 , e2 ) ⊲b T ′ ց ρ′ ⊔ ρ, if(S, e1 , e2 ) ⊲b S ′
(✷, p), T ց ρ, S
SP ROJ 2
p, T.#2 ց ρ, S.#2
p, x.Θ ց∗ ρ, Θ0 , p0
···
p, f(T1 , . . . , Tn ) ց ρ1 ⊔ · · · ⊔ ρn , f(S1 , . . . , Sn )
✸, T1 ց ρ1 , S1
✸, T2 ց ρ2 , S2
SD IAMOND
✸, (T1 , T2 ) ց ρ1 ⊔ ρ2 , (S1 , S2 )
p[ℓ], T ց ρ[x 7→ p0 ], S
SS NG∗
p, x.{ℓ.T } ց∗ ρ, {ℓ.S}, {ℓ.p0 }
SE MPTY∗
p|dom(Θ1 ) , x.Θ1 ց∗ ρ1 , Θ′1 , p1
p|dom(Θ2 ) , x.Θ2 ց∗ ρ2 , Θ′2 , p2
∗
p, x.Θ1 ⊎ Θ2 ց ρ1 ⊔ ρ2 , Θ′1 ⊎ Θ′2 , p1 ⊎ p2
SU NION∗
Figure 6. Backward trace slicing.
In this section we sketch how to enrich the language of patterns
to allow for partial set and partial record patterns, as follows:
p
::=
✸⊎✷ =✷⊎✸ =✷⊎✷
✸⊎✸
sp ⊎ ∅ = ∅ ⊎ sp
{ℓi .pi } ⊎ ✷ = ✷ ⊎ {ℓi .pi }
{ℓi .pi } ⊎ ✸ = ✸ ⊎ {ℓi .pi }
{ℓi .pi } ⊎ ({ℓ′j .p′j } ∪˙ ✷)
˙ ✸)
{ℓi .pi } ⊎ ({ℓ′j .p′j } ∪
({ℓi .pi } ∪˙ ✷) ⊎ ({ℓ′j .p′j } ∪˙ ✷)
˙ ✸) ⊎ ({ℓ′j .p′j } ∪˙ ✷)
({ℓi .pi } ∪
˙ ✷) ⊎ ({ℓ′j .p′j } ∪
˙ ✸)
({ℓi .pi } ∪
˙ ✸)
({ℓi .pi } ∪˙ ✸) ⊎ ({ℓ′j .p′j } ∪
✷ | ✸ | c | sp | rp
rp
::=
hi | hAi : pi i | hAi : pi ; ✷i | hAi : pi ; ✸i
sp
::=
∅ | {ℓi .pi } | {ℓi .pi } ∪˙ ✷ | {ℓi .pi } ∪˙ ✸
Record patterns are of the form hAi : pi i, listing the fields and the
patterns they must match, possibly followed by ✷ or ✸, which the
remainder of the record must match. The pattern {ℓi .pi } ∪˙ ✷ stands
for a set with labeled elements matching patterns p1 , . . . , pn , plus
some additional elements whose values we don’t care about. For
example, we can use the pattern {ℓ1 .p1 } ∪˙ ✷ to express interest
in why element ℓ1 matches p1 , when we don’t care about the rest
of the set. The second partial pattern, {ℓi .pi } ∪˙ ✸, has similar
behavior, but it says that the rest of the sets being considered must
be equal. For example, {ℓ.✷} ∪˙ ✸ says that the two sets are equal
except possibly at label ℓ. This pattern is needed mainly in order to
ensure that we can define a least upper bound on enriched patterns,
since we cannot express ({ℓ1 .v1 , . . . , ℓn .pn } ∪˙ ✷) ⊔ ✸ otherwise.
We define dom({ℓi .pi } ∪˙ ✷) = dom({ℓi .pi } ∪˙ ✸) =
{ℓ1 , . . . , ℓn }. Disjoint union of enriched patterns sp ⊎ sp′ is defined only if the domains are prefix-disjoint, so that the labels of the
result still form a prefix code; this operation is defined in Figure 7.
We now extend the definitions of p[✸/✷] and ⊔ to account for
extended patterns. We extend the [✸/✷] substitution operation as
follows:
hAi : pi i[✸/✷]
hAi : pi ; ✷i[✸/✷]
hAi : pi ; ✸i[✸/✷]
({ℓi .pi } ∪˙ ✷)[✸/✷]
˙ ✸)[✸/✷]
({ℓi .pi } ∪
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
✷
✸
sp
˙ ✷
{ℓi .pi } ∪
˙ ✸
{ℓi .pi } ∪
({ℓi .pi } ⊎ {ℓ′j .p′j }) ∪˙ ✷
({ℓi .pi } ⊎ {ℓ′j .p′j }) ∪˙ ✸
({ℓi .pi } ⊎ {ℓ′j .p′j }) ∪˙ ✷
({ℓi .pi } ⊎ {ℓ′j .p′j }) ∪˙ ✷
({ℓi .pi } ⊎ {ℓ′j .p′j }) ∪˙ ✷
({ℓi .pi } ⊎ {ℓ′j .p′j }) ∪˙ ✸
Figure 7. Enriched pattern union
We extend the singleton extraction p.ǫ, label projection p[ℓ], and
restriction p|L operations on set patterns as follows:
˙ ✸).ǫ = ({ǫ.p} ∪
˙ ✷).ǫ
({ǫ.p} ∪
˙ ✷)
ℓ · ({ℓi .pi } ∪
ℓ · ({ℓi .pi } ∪˙ ✸)
˙ ✷)[ℓ]
({ℓi .pi } ∪
({ℓi .pi } ∪˙ ✸)[ℓ]
˙ ✷)|L
({ℓi .pi } ∪
({ℓi .pi } ∪˙ ✸)|L
=
=
=
=
=
=
=
p
(ℓ · {ℓi .pi }) ⊎ ✷
(ℓ · {ℓi .pi }) ⊎ ✸
({ℓi .pi }[ℓ]) ⊎ ✷
({ℓi .pi }[ℓ]) ⊎ ✸
({ℓi .pi }|L ) ⊎ ✷
({ℓi .pi }|L ) ⊎ ✸
Note that in many cases, we use the disjoint union operation ⊎
on the right-hand side; this ensures, for example, that we never
produce results of the form ∅ ∪˙ ✷ or ∅ ∪˙ ✸; these are normalized to
✷ and ✸ respectively, and this normalization reduces the number
of corner cases in the slicing algorithm.
We define a record pattern projection operation p.A as follows:
hAi : pi [✸/✷]i
hAi : pi [✸/✷];✸i
hAi : pi [✸/✷]; ✸i
{ℓi .pi [✸/✷]} ∪˙ ✸
{ℓi .pi [✸/✷]} ∪˙ ✸
hA1 :p1 , . . . , An :pn i.Ai
hA1 :p1 , . . . , An :pn ; i.Ai
hA1 :p1 , . . . , An :pn ; ✷i.B
hA1 :p1 , . . . , An :pn ; ✸i.B
We handle the additional cases of the ⊔ operation in Figure 8,
and extend the hp relation as shown in Figure 9. Note that taking
the least upper bound of a partial pattern with a complete pattern
yields a complete pattern, while taking the least upper bound of
partial patterns involving ✸ again relies on the [✸/✷] substitution
operation.
=
=
=
=
pi
✷.A = ✷
pi
✸.A = ✸
✷ (B ∈
/ {A1 , . . . , An })
✸ (B ∈
/ {A1 , . . . , An })
We extend the slicing judgment to accommodate these new
patterns in Figure 10. The rules SR EC and SP ROJA are similar to
those for pairs, except that we use the field projection operation
in the case for a record trace, and we use partial record patterns
7
{ℓi .pi , ℓ′j .qj } ⊔ ({ℓi .p′i } ∪˙ ✷)
=
{ℓi .pi ⊔ p′i , ℓ′j .qj }
˙ ✸)
{ℓi .pi , ℓ′j .qj } ⊔ ({ℓi .p′i } ∪
=
{ℓi .pi ⊔ p′i , ℓ′j .qj [✸/✷]}
˙
({ℓi .pi , ℓ′j .qj } ∪˙ ✷) ⊔ ({ℓi .p′i , ℓ′′
k .rk } ∪ ✷)
=
˙
{ℓi .pi ⊔ p′i , ℓ′j .qj , ℓ′′
k .rk } ∪ ✷
˙
({ℓi .pi , ℓ′j .qj } ∪˙ ✷) ⊔ ({ℓi .p′i , ℓ′′
k .rk } ∪ ✸)
=
˙
{ℓi .pi ⊔ p′i , ℓ′j .qj [✸/✷], ℓ′′
k .rk } ∪ ✸
˙
({ℓi .pi , ℓ′j .qj } ∪˙ ✸) ⊔ ({ℓi .p′i , ℓ′′
k .rk } ∪ ✸)
=
˙
{ℓi .pi ⊔ p′i , ℓ′j .qj [✸/✷], ℓ′′
k .rk [✸/✷]} ∪ ✸
hAi : pi , Bj : qj i ⊔ (hAi : p′i ; ✷i)
=
hAi : pi ⊔ p′i , Bj : qj i
p′i ; ✸i)
=
hAi : pi ⊔ p′i , Bj : qj [✸/✷]i
(hAi : pi , Bj : qj ; ✷i) ⊔ (hAi : p′i , Ck : rk ; ✷i)
=
hAi : pi ⊔ p′i , Bj : qj , Ck : rk ; ✷i
(hAi : pi , Bj : qj ; ✷i) ⊔ (hAi : p′i , Ck : rk ; ✸i)
=
hAi : pi ⊔ p′i , Bj : qj [✸/✷], Ck : rk ; ✸i
=
hAi : pi ⊔ p′i , Bj : qj [✸/✷], Ck : rk [✸/✷]; ✸i
hAi : pi , Bj : qj i ⊔ (hAi :
(hAi : pi , Bj : qj ; ✸i) ⊔ (hAi :
p′i , Ck
: rk ; ✸i)
Figure 8. Least upper bound for enriched patterns (excluding some symmetric cases).
v1 hp1 v1′
hAi : vi i hhA
i :pi i
v1 hp1 v1′
hAi : vi′ i
′
vn hpn vn
···
hAi : vi , Bj :wj i hhA
i :pi ;✷i
v1 hp1 v1′
hAi : vi , Bj :wj i hhA
i :pi ;✸i
{ℓi .vi } ⊎ v h{ℓ
v1 hp1 v1′
{ℓi .vi } ⊎ v
hAi : vi′ , Ck :wk′ i
Lemma 4.2 (Projection and ⊑).
′
vn hpn vn
···
v1 hp1 v1′
1. If p1 ⊑ v1 and p2 ⊑ v2 and v1 hp1 v1′ and v2 hp2 v2′ then
v1 ⊎ v2 hp1 ⊎p2 v1′ ⊎ v2′ , provided all of these disjoint unions
are defined.
2. If p ⊑ v1 ⊎ v2 and L1 ≤ dom(v1 ) and L2 ≤ dom(v2 ) and
L1 , L2 are prefix-disjoint, then p|L1 ⊑ v1 and p|L2 ⊑ v2 .
′
vn hpn vn
···
1.
2.
3.
4.
hAi : vi′ , Bj :wj i
···
′
vn hpn vn
˙
i .pi }∪✷
{ℓi .vi′ } ⊎ v′
Lemma 4.3 (Projection and hp ).
1. If p ⊑ {ǫ.v} and v hp.ǫ v ′ then {ǫ.v} hp {ǫ.v ′ }.
2. If p ⊑ 1 · v1 ⊎ 2 · v2 and v1 hp[1] v1′ and v2 hp[2] v2′ then
1 · v1 ⊎ 2 · v2 hp 1 · v1′ ⊎ 2 · v2′ .
3. If p ⊑ ℓ · v and v hp[ℓ] v ′ then ℓ · v hp ℓ · v ′ .
4. If p ⊑ hAi : vi i and v1 hp.A1 v1′ , . . . , vn hp.An vn′ then
hAi : vi i hp hAi : vi′ i.
′
···
vn hpn vn
{ℓi .vi′ } ⊎ v
h{ℓ .p }∪✸
i i ˙
Figure 9. Enriched pattern equivalence
p, T ց ρ, S
Proofs are collected in Appendix B.
We now state the key correctness property for slicing. Intuitively, it says that if we slice T with respect to output pattern p,
obtaining a slice ρ and S, then p will be reproduced on recomputation under any change to the input and trace that is consistent with
the slice — formally, that means that the changed trace T ′ must
match the sliced trace S, and the changed input γ ′ must match γ
modulo ρ.
hA:p; ✷i, T ց ρ, S
SR EC
p, T.A ց ρ, S.A
p.A1 , T1 ց ρ1 , S1
···
p.An , Tn ց ρn , Sn
p, hA1 :T1 , . . . , An :Tn i ց ρ1 ⊔ · · · ⊔ ρn , hA1 :S1 , . . . , An :Sn i
SP ROJ A
p, x.Θ ց∗ ρ, Θ′ , p0
✷, x.Θ ց∗ [], ∅, ✷
SH OLE∗
✸, x.∅ ց∗ [], ∅, ∅
If p ⊑ {ǫ.v} then p.ǫ ⊑ v.
If p ⊑ 1 · v1 ⊎ 2 · v2 then p[1] ⊑ v1 and p[2] ⊑ v2 .
If p ⊑ ℓ · v then p[ℓ] ⊑ v.
If p ⊑ hAi : vi i then p.Ai ⊑ vi .
Theorem 4.4 (Correctness of Slicing).
SD IAMOND∗
1. Suppose γ, T y v and p ⊑ v and p, T ց ρ, S. Then for all
γ ′ hρ γ and T ′ ⊒ S such that γ ′ , T ′ y v ′ we have v ′ hp v.
2. Suppose γ, x ∈ v0 , Θ y∗ v and p ⊑ v and p, x.Θ0 ց
ρ, Θ′0 , p0 , where Θ0 ⊆ Θ. Then for all γ ′ hρ γ and v0′ hp0 v0
and Θ′ ⊒ Θ′0 such that γ ′ , x ∈ v0 , Θ′ y∗ v ′ we have v ′ hp v.
Figure 10. Backward trace slicing over enriched patterns.
hA : p; ✷i in the case for a field projection trace. The added
rules SH OLE∗ and SD IAMOND∗ handle the possibility that a partial
pattern reduces to ✷ or ✸ through projection; we did not need to
handle this case earlier because a simple set pattern is either a hole
(which could be handled by the rule ✷, T ց [], ✷) or a complete
set pattern showing all of the labels of the result.
Both simple and extended patterns satisfy a number of lemmas
that are required to prove the correctness of trace slicing.
Returning to our running example, we can use the enriched
pattern p′ = {[r2 ].hB:8; ✷i} ∪˙ ✷ to indicate interest in the B field
of r2 , without naming the other fields of the row or the other row
indexes. Slicing with respect to this pattern yields the following
slice:
[
T ′′ = { | x ∈ R} ⊲ {[r2 ].if(x.B = 3, , ) ⊲true { }}
˙ This indicates
and the slice of R is R′′ = {[r2 ].hB:3, C:8; ✷i}∪✷.
that (as before) the values of r1 and r3 and of the A field of r2
Lemma 4.1 (Properties of union and restriction).
8
are irrelevant to r2 in the result; unlike T ′ , however, we can also
potentially replay if r1 and r3 have been deleted from R. Likewise,
we can replay if the A field has been removed from a record, or if
some other field such as D is added. This illustrates that enriched
patterns allow for smaller slices than simple patterns, with greater
flexibility concerning possible updates to the input.
A natural question is whether slicing computes the (or a) smallest possible answer. Because our definition of correct slices is based
on recomputation, minimal slices are not computable, by a straightforward reduction from the undecidability of minimizing dependency provenance [1, 12].
p, T ց
ց ρ, e
✷, T ց
ց [], ✷
p, c ց
ց [], c
✸, T1 ց
ց ρ1 , e1
···
✸, Tn ց
ց ρn , en
p, f(T1 , . . . , Tn ) ց
ց ρ1 ⊔ · · · ⊔ ρn , f(e1 , . . . , en )
p, x ց
ց [x 7→ p], x
p2 , T2 ց
ց ρ2 [x 7→ p1 ], e2
p1 , T1 ց
ց ρ1 , e1
p2 , let x = T1 in T2 ց
ց ρ1 ⊔ ρ2 , let x = e1 in e2
5. Query and Differential Slicing
p1 , T1 ց
ց ρ1 , e1
We now adapt slicing techniques to provide explanations in terms
of query expressions, and show how to use differences between
slices to provide precise explanations for parts of the output.
···
pn , Tn ց
ց ρn , en
hA1 :p1 , . . . , An :pn i, hA1 :T1 , . . . , An :Tn i ց
ց ρ1 ⊔ · · · ⊔ ρn , hA1 :e1 , . . . , An :en i
p.A, T ց
ց ρ, S
p, T.A ց
ց ρ, S.A
5.1 Query slicing
p1 , T1 ց
ց ρ1 , e′1
true, T ց
ց ρ, e′
p1 , if(T, e1 , e2 ) ⊲true T1 ց
ց ρ1 ⊔ ρ, if(e′ , e′1 , ✷)
Our previous work [27] gave an algorithm for extracting a program
slice from a trace. We now adapt this idea to queries. A trace slice
shows the parts of the trace that need to be replayed in order to compute the desired part of the output; similarly, a query slice shows the
part of the query expression that is needed in order to ensure that the
desired part of the output is recomputed. As with traces, we allow
holes ✷ in programs to allow deleting subexpressions, and define
⊑ as a syntactic precongruence such that ✷ ⊑ e. We also define a
least upper bound operation e ⊔ e′ on partial query expressions in
the obvious way, so that ✷ ⊔ e = e.
We define a judgment p, T ց
ց ρ, e that traverses T and “unevaluates” p, yielding a partial input environment ρ and partial program e. The rules are illustrated in Figure 11. Many of the rules are
similar to those for trace slicing; the main differences arise in the
cases for conditionals and comprehensions, where we collapse the
sliced expressions back into expressions, possibly inserting holes or
merging sliced expressions obtained by running the same code in
different ways (as in a comprehension that contains a conditional).
Again, it is straightforward to show that if γ, e ⇓ v, T and
p ⊑ v then there exist ρ, e′ such that p, T ց
ց ρ, e′ , where ρ ⊑ γ
and e′ ⊑ e. The essential correctness property for query slices is
similar to that for trace slices: again, we require that rerunning any
sufficiently similar query on a sufficiently similar input produces a
result that matches p. The proof of this result is in Appendix D.
p2 , T2 ց
ց ρ2 , e′2
false, T ց
ց ρ, e′
p2 , if(T, e1 , e2 ) ⊲false T2 ց
ց ρ2 ⊔ ρ, if(e′ , ✷, e′2 )
∅, ∅ ց
ց [], ∅
p[1], T1 ց
ց ρ1 , e1
p[2], T2 ց
ց ρ2 , e2
p, T1 ∪ T2 ց
ց ρ1 ⊔ ρ2 , e1 ∪ e2
p[ǫ], T ց
ց ρ, e
p, {T } ց
ց ρ, {e}
✸, T ց
ց ρ, e
p, empty T ց
ց ρ, empty e
✸, T ց
ց ρ, e
p, sum T ց
ց ρ, sum e
p, x.Θ ց
ց∗ ρ′ , e′ , p0
p0 , T ց
ց ρ, e′0
[
[
′
p, {e | x ∈ T } ⊲ Θ ց
ց ρ ⊔ ρ , {e′ | x ∈ e′0 }
✸, T1 ց
ց ρ1 , e1
···
✸, Tn ց
ց ρn , en
✸, hA1 : T1 , . . . , An : Tn i ց
ց ρ1 ⊔ · · · ⊔ ρn , hA1 : e1 , . . . , An : en i
✸, ∅ ց
ց [], ∅
p, x.Θ ց
ց ρ, e, p′
✷, x.Θ ց
ց∗ [], ✷, ✷
Theorem 5.1 (Correctness of Query Slicing).
∅, x.∅ ց
ց∗ [], ✷, ∅
✸, x.∅ ց
ց∗ [], ✷, ∅
p[ℓ], T ց
ց ρ[x 7→ p0 ], e
p, x.{ℓ.T } ց
ց∗ ρ, e, {ℓ.p0 }
1. Suppose γ, T y v and p ⊑ v and p, T ց
ց ρ, e. Then for all
γ ′ hρ γ and e′ ⊒ e such that γ ′ , e′ ⇓ v ′ we have v ′ hp v.
2. Suppose γ, x ∈ v0 , Θ y∗ v and p ⊑ v and p, Θ ց
ց ρ, e0 , p0 .
Then for all γ ′ hρ γ and v0′ hp0 v0 and e′0 ⊒ e0 such that
γ ′ , x ∈ v0′ , e′0 ⇓∗ v ′ we have v ′ hp v.
p|dom(Θ1 ) , x.Θ1 ց
ց∗ ρ1 , e1 , p1
p|dom(Θ2 ) , x.Θ2 ց
ց∗ ρ2 , e2 , p2
p, x.Θ1 ⊎ Θ2 ց
ց∗ ρ1 ⊔ ρ2 , e1 ⊔ e2 , p1 ⊎ p2
Combining with the consistency property (Proposition 2.6), we
have:
Figure 11. Unevaluation (selected rules).
Corollary 5.2. Suppose γ, e ⇓ v, T and p ⊑ v and p, T ց
ց ρ, e′ .
Then for all γ ′ hρ γ and e′′ ⊒ e′ such that γ ′ , e′′ ⇓ v ′ , we have
v hp v ′ .
5.2 Differential slicing
We consider a pattern difference to be a pair of patterns (p, p′ )
where p ⊑ p′ . Intuitively, a pattern difference selects the part of a
value present in the outer component p′ and not in the inner component p. For example, the pattern difference (hB:✷; ✷i, hB:8; ✷i)
selects the value 8 located in the B component of a record. We can
also write this difference as hB: 8 ; ✷i, using ? to highlight the
boundary between the inner and outer pattern. Trace and query pattern differences are defined analogously.
Continuing our running example, the query slice for the pattern
p′ considered above is
[
Q′ = {if(x.B = 3, {hA:✷, B:x.Ci}, {}) | x ∈ R}
since only the computation of the A field in the output is irrelevant
to p′ .
9
It is straightforward to show by induction that slicing is monotonic in both arguments:
A related point: one may wonder whether it makes sense to select a particular copy of hB:3i in the output, since in a conventional multiset, multiple copies of the same value are indistinguishable. We believe it is important to be able to distinguish different
copies of a value, which may have different explanations. Treating
n copies of a value as a single value with multiplicity n would obscure this distinction and force us to compute the slices of all of the
copies even if only a single explanation is required. This is why we
have chosen to work with indexed sets, rather than pure multisets.
Lemma 5.3 (Monotonicity). If p ⊑ p′ and T ⊑ T ′ and p′ , T ′ ց
ρ′ , S ′ then there exist ρ, S such that p, T ց ρ, S and ρ ⊑ ρ′ and
S ⊑ S ′ . In addition, p, S ′ ց ρ, S.
This implies that given a pattern difference and a trace, we can
compute a trace difference using the following rule:
p 2 , T ց ρ 2 , S2
p 1 , S2 ց ρ 1 , S1
(p1 , p2 ), T ց (ρ1 , ρ2 ), (S1 , S2 )
Joins So far all examples have involved a single table R. Consider a simple join query
It follows from monotonicity that ρ1 ⊑ ρ2 and S1 ⊑ S2 , thus,
pattern differences yield trace differences. Furthermore, the second
part of monotonicity implies that we can compute the smaller slice
S1 from the larger slice S2 , rather than re-traverse T . It is also
possible to define a simultaneous differential slicing judgment, as
an optimization to ensure we only traverse the trace once.
Query slicing is also monotone, so differential query slices can
be obtained in exactly the same way. Revisiting our running example one last time, consider the differential pattern {[r2 ].hB: 8 ; ✷i}∪˙
✷. The differential query slice for the pattern p′ considered above
is
[
Q′′ = {if(x.B = 3, {hA:✷, B: x.C i}, {}) | x ∈ R}
Q3 = {hA:x.A, B:y.Ci | x ∈ R, y ∈ S, x.B = y.B}
and consider the following table S, and the result Q3 (R, S).
id
[s1 ]
S=
[s2 ]
[s3 ]
B
2
3
4
C
4
4
5
id
[r1 , s1 ]
Q3 (R, S) =
[r2 , s2 ]
[r3 , s2 ]
A
1
2
4
B
4
4
5
The full trace of this query execution is as follows:
S
T3 = S
{ | x ∈ R} ⊲ {
[r1 ]. { | y ∈ S} ⊲ {
6. Examples and Discussion
[r2 ].
S
{ | y ∈ S} ⊲ {
In this section we present some more complex examples to illustrate key points.
[r3 ].
S
{ | y ∈ S} ⊲ {
Renaming Recall the swapping query from the introduction,
written in NRC as
[
Q1 = {{if(x.A > x.B, hA:x.B, B:x.Ai, hxi)} | x ∈ R}
[s1 ].if(x.B
[s2 ].if(x.B
[s3 ].if(x.B
[s1 ].if(x.B
[s2 ].if(x.B
[s3 ].if(x.B
[s1 ].if(x.B
[s2 ].if(x.B
[s3 ].if(x.B
=
=
=
=
=
=
=
=
=
y.B,
y.B,
y.B,
y.B,
y.B,
y.B,
y.B,
y.B,
y.B,
,
,
,
,
,
,
,
,
,
) ⊲true { },
) ⊲false {},
) ⊲false {}},
) ⊲false {},
) ⊲true { },
) ⊲false {}},
) ⊲false {},
) ⊲true { },
) ⊲false {}}}
Slicing with respect to {[r1 , s1 ].hA:1; ✷i, [r2 , s2 ].hB:4; ✷i} ∪˙ ✷
yields trace slice
This query illustrates a key difference between our approach and
the how-provenance model of Green et al. [20]. As discussed in
[13], renaming operations are ignored by how-provenance, so the
how-provenance annotations of the results of Q1 are the same as for
a query that simply returns R. In other words, the choice to swap
the fields when A > B is not reflected in the how-provenance,
which shows that it is impossible to extract where-provenance (or
traces) from how-provenance. Extracting where-provenance from
traces appears straightforward, extending our previous work [1].
This example also illustrates how traces and slices can be used
for partial recomputation. The slice for output pattern {[1, r1 ].hB:2i}∪˙
✷, for example, will show that this record was produced because
the A component of hA:1, B:2, C:7i at index [r1 ] in the input was
less than or equal to the B component. Thus, we can replay after
any change that preserves this ordering information.
Union Consider query
[
Q2 = {{hB:x.Bi} | x ∈ R} ∪ {hB:3i}
that projects the B fields of elements of R and adds another copy
of hB:3i to the result. This yields
Q2 (R) = {[1, r1 ].hB:2i, [1, r2 ].hB:3i, [1, r3 ].hB:3i, [2].hB:3i}
This illustrates that the indexes may not all have the same length,
but still form a prefix code. If we slice with respect to {[1, r2 ].hB:3i}∪˙
✷ then the query slice is:
[
Q′2 = {{hB:x.Bi} | x ∈ R} ∪ ✷
and R′ = {[r2 ].hB:3; ✷i} ∪˙ ✷ whereas if we slice with respect to
{[2].hB:3i} ∪˙ ✷ then the query slice is Q′′2 = ✷ ∪ {hB:3i} and
R′′ = ✷, indicating that this part of the result has no dependence
on the input.
10
S
T3′ = S
{ | x ∈ R} ⊲ {
[r1 ]. S{ | y ∈ S} ⊲ {
[r2 ]. { | y ∈ S} ⊲ {
[s1 ].if(x.B = y.B, , ) ⊲true { },
[s2 ].if(x.B = y.B, , ) ⊲true { }}
and input slice R3′ = {[r1 ].hA:1, B:2; ✷i, [r2 ].hB:3; ✷i} and
S3′ = {[s1 ].hB:2; ✷i, [s2 ].hB:3; C:4i}.
Workflows NRC expressions can be used to represent workflows, if primitive operations are added representing the workflow
steps [2, 21]. To illustrate query slicing and differential slicing for a
workflow-style query, consider the following more complex query:
Q4 = {f (x, y) | x ∈ T, y ∈ T, z ∈ U, p(x, y), q(x, y, z)}
where f computes some function of x and y and p and q are selection criteria. Here, we assume T and U are collections of data files
and f, p, q are additional primitive operations on them. This query
exercises most of the distinctive features of our approach; we can
of course translate it to the NRC core calculus used in the rest of the
paper. If dom(T ) = {t1 , . . . , t10 } and dom(U ) = {u1 , . . . , u10 }
then we might obtain result {[t3 , t4 , u5 ].v1 , [t6 , t8 , u10 ].v2 }. If we
focus on the value v1 using the pattern {[t3 , t4 , u5 ].v1 } ∪˙ ✷, then
the program slice we obtain is Q4 itself, while the data slice might
be T ′ = {[t3 ].✸, [t4 ].✸} ∪˙ ✷, U ′ = {[u5 ].✸} ∪˙ ✷, indicating that
if the values at t3 , t4 , u5 are held fixed then the end result will still
be v1 . The trace slice is similar, and shows that v1 was computed
by applying f with x bound to the value at t3 in T , y bound to t4 ,
and z bound to u5 , and that p(x, y) and q(x, y, z) succeeded for
these values.
If we consider a differential slice using pattern difference
{[t3 , t4 , u5 ]. v1 } ∪˙ ✷ then we obtain the following program difference:
[
Qδ4 = { f (x, y) | x ∈ T, y ∈ T, z ∈ U, p(x, y), q(x, y, z)}
This shows that most of the query is needed to ensure that the
result at [t3 , t4 , u5 ] is produced, but the subterm f (x, y) is only
needed to compute the value v1 . This can be viewed as a querybased explanation for this part of the result.
tracing, making it feasible for in-memory execution of workflows
represented in NRC.
8. Related and future work
Program slicing has been studied extensively [17, 30, 31], as has the
use of execution traces, for example in dynamic slicing. Our work
contrasts with much of this work in that we regard the trace and
underlying data as being of interest, not just the program. Some
of our previous work [12] identified analogies between program
slicing and provenance, but to our knowledge, there is no other
prior work on slicing in databases.
Lineage and why-provenance were motivated semantically in
terms of identifying witnesses, or parts of the input needed to ensure that a given part of the output is produced by a query. Early
work on lineage in relational algebra [15] associates each output
record with a witness. Buneman et al. studied a more general notion called why-provenance that maps an output part to a collection of witnesses [7, 8]. This idea was generalized further to the
how-provenance or semiring model [18, 20], based on using algebraic expressions as annotations; this approach has been extended
to handle some forms of negation and aggregation [4, 19]. Semiring
homomorphisms commute with query evaluation; thus, homomorphic changes to the input can be performed directly on the output
without re-running the query. However, this approach only applies
to changes describable as semiring homomorphisms, such as deletion.
Where-provenance was also introduced by Buneman et al. [7,
8]. Although the idea of tracking where input data was copied
from is natural, it is nontrivial to characterize semantically, because
where-provenance does not always respect semantic equivalence.
In later work, Buneman et al. [6] studied where-provenance for
the pure NRC and characterized its expressiveness for queries and
updates. It would be interesting to see whether their notion of
expressive completeness for where-provenance could be extended
to richer provenance models, such as traces, possibly leading to an
implementation strategy via translation to plain NRC.
Provenance has been studied extensively for scientific workflow
systems [5, 29], but there has been little formal work on the semantics of workflow provenance. The closest work to ours is that of
Hidders et al. [21], who model workflows by extending the NRC
with nondeterministic, external function calls. They sketch an operational semantics that records runs that contain essentially all of
the information in a derivation tree, represented as a set of triples.
They also suggest ways of extracting subruns from runs, but their
treatment is partial and lacks strong formal guarantees analogous
to our results.
There have been some attempts to reconcile the database and
workflow views of provenance; Hidders et al. [21] argued for the
use of Nested Relational Calculus (NRC) as a unifying formalism for both workflow and database operations, and subsequently
Kwasnikowska and Van den Bussche [22] showed how to map this
model to the Open Provenance Model. Acar et al. [2] later formalized a graph model of provenance for NRC. The most advanced
work in this direction appears to be that of Amsterdamer et al. [3],
who combined workflow and database styles of provenance in the
context of the PigLatin system (a MapReduce variant based on
nested relational queries). Lipstick allows analyzing the impact of
restricted hypothetical changes (such as deletion) on parts of the
output, but to our knowledge no previous work provides a formal
guarantee about the impact of changes other than deletion.
In our previous work [12], we introduced dependency provenance, which conservatively over-approximates the changes that
can take place in the output if the input is changed. We developed definitions and techniques for dependency provenance in full
NRC including nonmonotone operations (empty, sum) and primi-
7. Implementation
To validate our design and experiment with larger examples, we
extended our Haskell implementation Slicer of program slicing for
functional programs [27] with the traces and slicing techniques
presented in this paper. We call the resulting system NRCSlicer;
it supports a free combination of NRC and general-purpose functional programming features. NRCSlicer interprets expressions inmemory without optimization. As reported previously for Slicer,
we have experimented with several alternative tracing and slicing
strategies, which use Haskell’s lazy evaluation strategy in different
ways. The alternatives we consider here are:
• eager: the trace is fully computed during evaluation.
• lazy: the value is computed eagerly, but the trace is computed
lazily using Haskell’s default lazy evaluation strategy.
To evaluate the effectiveness of enriched patterns, we measured the
time needed for the eager and lazy techniques to trace and slice
the workflow example Q4 in the previous section. We considered a
instantiation of the workflow where the data values are simply integers and with input tables T, U = {1, . . . , 50}, and defined the
operations f (x, y) as x ∗ y, p(x, y) as x < y, and q(x, y, z) as
x2 + y 2 = z 2 . This is not a realistic workflow, and we expect that
the time to evaluate the basic operations of a realistic workflow following this pattern would be much larger. However, the overheads
of tracing and slicing do not depend on the execution time of primitive operations, so we can still draw some conclusions from this
simplistic example.
The comprehension iterates over 503 = 125,000 triples, producing 20 results. We considered simple and enriched patterns selecting a single element of the result. We measured evaluation time,
and the overhead of tracing, trace slicing, and query slicing. The
experiments were conducted on a MacBook Pro with 2GB RAM
and a 2.8GHz Intel Core Duo, using GHC version 7.4.
eager-simple
eager-enriched
lazy-simple
lazy-enriched
eval
0.5
0.5
0.5
0.5
trace
1.5
1.5
0.7
0.7
slice
2.5
<0.1
1.3
<0.1
qslice
1.6
<0.1
1.7
<0.1
The times are in seconds. The “eval” column shows the time needed
to compute the result without tracing. The “trace”, “slice”, and “qslice” columns show the added time needed to trace and compute
slices. The full traces in each of these runs have over 2.1 million
nodes; the simple pattern slices are almost as large, while the enriched pattern slices are only 95 nodes. For this example, slicing is
over an order of magnitude faster using enriched patterns. The lazy
tracing approach required less total time both for tracing and slicing (particularly for simple patterns). Thus, Haskell’s built-in lazy
evaluation strategy offers advantages by avoiding explicitly constructing the full trace in memory when it is not needed; however,
there is still room for improvement. Again, however, for an actual
workflow involving images or large data files, the evaluation time
would be much larger, dwarfing the time for tracing or slicing.
Our implementation is a proof-of-concept that evaluates queries
in-memory via interpretation, rather than compilation; further work
would be needed to adapt our approach to support fine-grained
provenance for conventional database systems. Nevertheless, our
experimental results do suggest that the lazy tracing strategy and
use of enriched patterns can effectively decrease the overhead of
11
tive functions. Dependency provenance cannot predict exactly how
the output will be affected by a general modification to the source,
but it can guarantee that some parts of the output will not change if
certain parts of the input are fixed. Our notion of equivalence modulo a pattern is a generalization of the equal-except-at relation used
in that work. Motivated by dependency provenance, an earlier technical report [10] presented a model of traced evaluation for NRC
and proved elementary properties such as fidelity. However, it did
not investigate slicing techniques, and used nondeterministic label
generation instead of our deterministic scheme; our deterministic
approach greatly simplifies several aspects of the system, particularly for slicing.
There are several intriguing directions for future work, including developing more efficient techniques for traced evaluation and
slicing that build upon existing database query optimization capabilities. It appears possible to translate multiset queries so as to
make the labels explicit, since a fixed given query increases the
label depth by at most a constant. Thus, it may be possible to evaluate queries with label information but without tracing first, then
gradually build the trace by slicing backwards through the query,
re-evaluating subexpressions as necessary. Other interesting directions include the use of slicing techniques for security, to hide confidential input information while disclosing enough about the trace to
permit recomputation, and the possibility of extracting other forms
of provenance from traces, as explored in the context of functional
programs in prior work [1].
ciety University Research Fellowship, by the EU FP7 DIACHRON
project, and EPSRC grant EP/K020218/1. Acar is partially supported by an EU ERC grant (2012-StG 308246—DeepSea) and an
NSF grant (CCF-1320563).
References
[1] U. A. Acar, A. Ahmed, J. Cheney, and R. Perera. A core calculus for
provenance. Journal of Computer Security, 21:919–969, 2013. Full
version of a POST 2012 paper.
[2] U. A. Acar, P. Buneman, J. Cheney, N. Kwasnikowska,
J. Van den Bussche, and S. Vansummeren. A graph model of
data and workflow provenance. In TAPP, 2010.
[3] Y. Amsterdamer, S. B. Davidson, D. Deutch, T. Milo, J. Stoyanovich,
and V. Tannen. Putting lipstick on pig: Enabling database-style workflow provenance. PVLDB, 5(4):346–357, 2011.
[4] Y. Amsterdamer, D. Deutch, and V. Tannen. Provenance for aggregate
queries. In PODS, pages 153–164. ACM, 2011.
[5] R. Bose and J. Frew. Lineage retrieval for scientific data processing: a
survey. ACM Comput. Surv., 37(1):1–28, 2005.
[6] P. Buneman, J. Cheney, and S. Vansummeren. On the expressiveness
of implicit provenance in query and update languages. ACM Transactions on Database Systems, 33(4):28, November 2008.
[7] P. Buneman, S. Khanna, and W. Tan. Why and where: A characterization of data provenance. In ICDT, number 1973 in LNCS, pages
316–330. Springer, 2001.
[8] P. Buneman, S. Khanna, and W. Tan. On propagation of deletions and
annotations through views. In PODS, pages 150–158, 2002.
9. Conclusion
[9] P. Buneman, S. A. Naqvi, V. Tannen, and L. Wong. Principles of programming with complex objects and collection types. Theor. Comp.
Sci., 149(1):3–48, 1995.
The importance of provenance for transparency and reproducibility
is widely recognized, yet there has been little explicit discussion
of correctness properties formalizing intuitions about how provenance is to provide reproducibility. In self-explaining computation,
traces are considered to be explanations of a computation in the
sense that the trace can be used to recompute (parts of) the output
under hypothetical changes to the input. This paper develops the
foundations of self-explaining computation for database queries,
by defining a tracing semantics for NRC, proposing a formal definition of correctness for tracing (fidelity) and slicing, and defining
a correct (though potentially overapproximate) algorithm for trace
slicing. Trace slicing can be used to obtain smaller “golden trail”
traces that explain only a part of the input or output, and explore the
impact of changes in hypothetical scenarios similar to the original
run. At a technical level, the main contributions are the careful use
of prefix codes to label multiset elements, and the development of
enriched patterns that allow more precise slices. Our design is validated by a proof-of-concept implementation that shows that laziness and enriched patterns can significantly improve performance
for small (in-memory) examples.
In the near term, we plan to combine our work on selfexplaining functional programs [27] and database queries (this
paper) to obtain slicing and provenance models for programming
languages with query primitives, such as F# [14] or Links [23].
Ultimately, our aim is to extend self-explaining computation to
programs that combine several execution models, including workflows, databases, conventional programming languages, Web interaction, or cloud computing.
[10] J. Cheney, U. A. Acar, and A. Ahmed. Provenance traces. CoRR,
arXiv.org/abs/0812.0564, 2008.
[11] J. Cheney, U. A. Acar, and R. Perera. Toward a theory of selfexplaining computation. In In search of elegance in the theory and
practice of computation: a Festschrift in honour of Peter Buneman,
number 8000 in LNCS, pages 193–216. Springer, 2013.
[12] J. Cheney, A. Ahmed, and U. A. Acar. Provenance as dependency
analysis. Mathematical Structures in Computer Science, 21(6):1301–
1337, 2011.
[13] J. Cheney, L. Chiticariu, and W. C. Tan. Provenance in databases:
Why, how, and where. Foundations and Trends in Databases,
1(4):379–474, 2009.
[14] J. Cheney, S. Lindley, and P. Wadler. A practical theory of languageintegrated query. In ICFP, pages 403–416, New York, NY, USA, 2013.
ACM.
[15] Y. Cui, J. Widom, and J. L. Wiener. Tracing the lineage of view data in
a warehousing environment. ACM Trans. Database Syst., 25(2):179–
227, 2000.
[16] J. Dean and S. Ghemawat. MapReduce: simplified data processing on
large clusters. Commun. ACM, 51(1):107–113, 2008.
[17] J. Field and F. Tip. Dynamic dependence in term rewriting systems
and its application to program slicing. Information and Software
Technology, 40(11–12):609–636, 1998.
[18] J. N. Foster, T. J. Green, and V. Tannen. Annotated XML: queries and
provenance. In PODS, pages 271–280, 2008.
[19] F. Geerts and A. Poggi. On database query languages for K-relations.
J. Applied Logic, 8(2):173–185, 2010.
Acknowledgments We are grateful to Peter Buneman, Jan Van
den Bussche, and Roly Perera for comments on this work and to
the anonymous reviewers for detailed suggestions. Effort sponsored
by the Air Force Office of Scientific Research, Air Force Material
Command, USAF, under grant number FA8655-13-1-3006. The
U.S. Government and University of Edinburgh are authorized to reproduce and distribute reprints for their purposes notwithstanding
any copyright notation thereon. Cheney is supported by a Royal So-
[20] T. J. Green, G. Karvounarakis, and V. Tannen. Provenance semirings.
In PODS, pages 31–40, 2007.
[21] J. Hidders, N. Kwasnikowska, J. Sroka, J. Tyszkiewicz, and J. Van den
Bussche. A formal model of dataflow repositories. In DILS, 2007.
[22] N. Kwasnikowska and J. Van den Bussche. Mapping the NRC
dataflow model to the open provenance model. In IPAW, pages 3–16,
2008.
12
[23] S. Lindley and J. Cheney. Row-based effect types for database integration. In TLDI, pages 91–102. ACM Press, 2012.
[24] P. Missier, B. Ludäscher, S. Dey, M. Wang, T. McPhillips, S. Bowers,
M. Agun, and I. Altintas. Golden trail: Retrieving the data history that
matters from a comprehensive provenance repository. International
Journal of Digital Curation, 7(1):139–150, 2011.
[25] L. Moreau. The foundations for provenance on the web. Foundations
and Trends in Web Science, 2(2–3), 2010.
[26] C. Olston, B. Reed, U. Srivastava, R. Kumar, and A. Tomkins. Pig
latin: a not-so-foreign language for data processing. In SIGMOD,
pages 1099–1110. ACM, 2008.
[27] R. Perera, U. A. Acar, J. Cheney, and P. B. Levy. Functional programs
that explain their work. In ICFP, pages 365–376. ACM, 2012.
[28] A. Sabelfeld and A. Myers. Language-based information-flow security. IEEE Journal on Selected Areas in Communications, 21(1):5–19,
2003.
[29] Y. Simmhan, B. Plale, and D. Gannon. A survey of data provenance
in e-science. SIGMOD Record, 34(3):31–36, 2005.
[30] F. Tip. A survey of program slicing techniques. J. Prog. Lang., 3(3),
1995.
[31] M. Weiser. Program slicing. In ICSE, pages 439–449. IEEE Press,
1981.
13
Γ⊢T :τ
b ∈ {true, false}
Γ ⊢ b : bool
n∈N
Γ ⊢ n : int
f : (b1 , . . . , bn ) → b ∈ Σ
Γ ⊢ T1 : b1
···
Γ ⊢ f(T1 , . . . , Tn ) : b
Γ ⊢ T1 : τ1
Γ, x : τ1 ⊢ T2 : τ2
Γ ⊢ let x = T1 in T2 : τ2
b ∈ {true, false}
Γ ⊢ T : {τ }
Γ ⊢ T1 : τ1
···
Γ ⊢ Tn : τn
Γ ⊢ hA1 : T1 , . . . , An : Tn i : hA1 : τ1 , . . . , An : τn i
Γ ⊢ T ∪ T ′ : {τ }
Γ, x : τ ⊢ e : {τ ′ }
Γ ⊢ T : {τ }
Γ ⊢ sum T : int
Γ ⊢ empty T : bool
Γ ⊢ T : {τ }
Γ, x ∈ {τ } ⊢ Θ : {τ ′ }
[
Γ ⊢ {e | x ∈ T } ⊲ Θ : {τ ′ }
Γ, x : τ ⊢ T : {τ ′ }
Γ, x ∈ {τ } ⊢ {ℓ.T } : {τ ′ }
′
Γ, x ∈ {τ } ⊢ ∅ : {τ }
Γ⊢T :τ
Γ ⊢ {T } : {τ }
Γ ⊢ ∅ : {τ }
Γ ⊢ T : {int}
Γ, x ∈ {τ } ⊢ Θ : {τ ′ }
x:τ ∈Γ
Γ⊢x:τ
Γ ⊢ T : hA1 : τ1 , . . . , An : τn i
Γ ⊢ T.Ai : τi
Γ ⊢ T′ : τ
Γ ⊢ T : bool
Γ ⊢ e1 : τ
Γ ⊢ e2 : τ
Γ ⊢ if(T, e1 , e2 ) ⊲b T ′ : τ
Γ ⊢ T ′ : {τ }
Γ ⊢ Tn : bn
Γ, x ∈ {τ } ⊢ Θ1 : {τ ′ }
Γ, x ∈ {τ } ⊢ Θ2 : {τ ′ }
Γ, x ∈ {τ } ⊢ Θ1 ∪ Θ2 : {τ ′ }
Figure 12. Well-typed traces.
T′ ⊑ T
T ′′ ⊑ T ′
✷⊑T
hA1 :
T ⊑T
T1′ ⊑ T1
′
T1 , . . . , An :
T′ ⊑ T
{T ′ } ⊑ {T }
T
′′
T′ ⊑ T
T1′ ⊑ T1
T2′ ⊑ T2
′
′
T1 ∪ T2 ⊑ T1 ∪ T2
···
f(T1′ , . . . , Tn′ )
⊑T
···
Tn′ ⊑ Tn
⊑ hA1 : T1 , . . . , An : Tn i
Tn′ i
T1′ ⊑ T1
T′ ⊑ T
T .A ⊑ T.A
′
T′ ⊑ T
sum T ′ ⊑ sum T
T1′ ⊑ T1
Tn′ ⊑ Tn
⊑ f(T1 , . . . , Tn )
let x =
T0′ ⊑ T0
′
if(T0 , e1 , e2 ) ⊲b T ′
T′ ⊑ T
empty T ′ ⊑ empty T
T1′
in
T2′
T2′ ⊑ T2
⊑ let x = T1 in T2
T′ ⊑ T
⊑ if(T0 , e1 , e2 ) ⊲b T
∅⊑∅
T′ ⊑ T
Θ′ ⊑ Θ
[
[
′
′
{e | x ∈ T } ⊲ Θ ⊑ {e | x ∈ T } ⊲ Θ
Θ′ ⊑ Θ ⇐⇒ ∀ℓ ∈ dom(Θ′ ).Θ′ (ℓ) ⊑ Θ(ℓ)
Figure 13. Subtrace relation.
A. Auxiliary definitions
Figure 12 summarizes the typing rules for traces. Figure 13 defines the subtrace relation.
B. Proofs of pattern properties
We prove the required properties for enriched patterns. The corresponding properties for the sublanguage of simple patterns follow
immediately since simple patterns are closed under the relevant operations.
Lemma B.1. If p, p′ ⊑ v then p ⊔ p′ exists and is the least upper bound of p and p′ .
Proof. If both p and p′ match some pattern q, then it is straightforward to show by induction on q that p ⊔ p′ is defined and p ⊔ p′ ⊑ q.
Specifically, if p or p′ is ✷ or ✸ then we are done; if p and p′ are both constants then we are done; otherwise, in each case, the toplevel
structure of p and p′ must match q, so that we can apply one of the rules for ⊔ on smaller terms that match part of q. When q = v, the desired
result follows. The second part (that ⊔ is a least upper bound) also follows directly since clearly, p, p′ ⊑ p ⊔ p′ and if p, p′ ⊑ q then a similar
argument shows that p ⊔ p′ ⊑ q.
Lemma B.2. For any v, v ′ , p, v hp[✸/✷] v ′ holds if and only if v hp v ′ holds and v = v ′ .
Proof. Straightforward induction on the derivation of v hp[✸/✷] v ′ .
Lemma B.3. For any v, v ′ , p, p′ , we have v hp⊔p′ v ′ if and only if v hp v ′ and v hp′ v ′ . Moreover, p ⊑ p′ if and only if for all v, v ′ , we
have v hp′ v ′ implies v hp v ′ .
Proof. For the first part, we proceed by induction on the total size of p, p′ . The cases where one of p, p′ is ✷ or ✸ are straightforward; for ✸
we also need Lemma B.2. The cases involving constants, pairs, or complete set and record patterns are also straightforward.
There are several similar cases involving partial set or record patterns. We illustrate two representative cases:
1. If p = {ℓi .pi , ℓ′i .qi } and p′ = {ℓi .p′i } ∪˙ ✷ then p ⊔ p′ = {ℓi .pi ⊔ p′i , ℓ′i .qi }. First, suppose v hp⊔p′ v ′ . This means v = {ℓi .vi , ℓ′i .wi }
and v ′ = {ℓi .vi′ , ℓ′i .wi′ }, where vi hpi ⊔p′i vi′ and wi hqi wi′ . Therefore, by induction, vi hpi vi′ and vi hp′i vi′ , so we can conclude that
{ℓi .vi , ℓ′i .wi } hp {ℓi .vi′ , ℓ′i .wi′ } and {ℓi .vi , ℓ′i .wi } hp′ {ℓi .vi′ , ℓ′i .wi′ }, as required.
14
Conversely, if we assume v hp v ′ and v hp′ v ′ , then we must have v = {ℓi .vi , ℓ′i .wi } and v ′ = {ℓi .vi′ , ℓ′i .wi′ }, where vi hpi vi′ and
vi hp′i vi′ and wi hqi wi′ . Thus, by induction we have vi hpi ⊔p′i vi′ so we can conclude {ℓi .vi , ℓ′i .wi } hp⊔p′ {ℓi .vi′ , ℓ′i .wi′ }.
2. If p = {ℓi .pi , ℓ′i .qi } ∪˙ ✷ and p′ = {ℓi .p′i , ℓ′′i .ri } ∪˙ ✸ then p ⊔ p′ = {ℓi .pi ⊔ p′i , ℓ′i .qi [✸/✷], ℓ′′i .ri } ∪˙ ✸. First, suppose v hp⊔p′ v ′ .
This means v = {ℓi .vi , ℓ′i .wi , ℓ′′i .ui } ⊎ w0 and v ′ = {ℓi .vi′ , ℓ′i .wi′ , ℓ′′i .u′i } ⊎ w0 , where vi hpi ⊔p′i vi′ and wi hqi [✸/✷] wi′ and
ui hri u′i . Therefore, by induction, vi hpi vi′ and vi hp′i vi′ , and we also have wi hqi wi′ and wi = wi′ , so we can conclude that
{ℓi .vi , ℓ′i .wi , ℓ′′i .ui } ⊎ w0 hp {ℓi .vi′ , ℓ′i .wi′ , ℓ′′i .u′i } ⊎ w0 and {ℓi .vi′ , ℓ′i .wi , ℓ′′i .ui } ⊎ w0 hp′ {ℓi .vi′ , ℓ′i .wi′ , ℓ′′i .u′i } ⊎ w0 , as required.
Conversely, if we assume v hp v ′ and v hp′ v ′ , then we must have v = {ℓi .vi , ℓ′i .wi , ℓ′′i .ui } ⊎ v0 and v ′ = {ℓi .vi′ , ℓ′i .wi′ , ℓ′′i .u′i } ⊎ v0′
where vi hpi vi′ and vi hp′i vi′ and wi hqi wi′ and ui hri u′i . In addition, we must have that wi = wi′ and v0 = v0′ since v and v ′
must be equal at all labels not in ℓi , ℓ′′i . Thus, by induction we have vi hpi ⊔p′i vi′ and (using Lemma B.2) we can also easily show that
wi hqi [✸/✷] wi , so we can conclude that {ℓi .vi , ℓ′i .wi , ℓ′′i .u′i } ⊎ v0 hp⊔p′ {ℓi .vi′ , ℓ′i .wi , ℓ′′i .ui } ⊎ v0 .
The second part follows immediately from the definition of p ⊑ p′ as p ⊔ p′ = p′ .
Lemma B.4 (Properties of union and restriction).
1. If p1 ⊑ v1 and p2 ⊑ v2 and v1 hp1 v1′ and v2 hp2 v2′ then v1 ⊎ v2 hp1 ⊎p2 v1′ ⊎ v2′ , provided all of these disjoint unions are defined.
2. If p ⊑ v1 ⊎ v2 and L1 ≤ dom(v1 ) and L2 ≤ dom(v2 ) and L1 , L2 are prefix-disjoint, then p|L1 ⊑ v1 and p|L2 ⊑ v2 .
Proof. For part 1, assume p1 ⊑ v1 , p2 ⊑ v2 , v1 hp1 v1′ and v2 hp2 v2′ , and assume that the domains of p1 and p2 , v1 and v2 , and v1′ and v2′
are prefix-disjoint respectively, so that the unions exist. There are several cases. If p1 or p2 is ✷ then the conclusion is immediate. If both are
✸ then v1 = v1′ and v2 = v2′ so v1 ⊎ v2 = v1′ ⊎ v2′ .
Most of the remaining cases are straightforward; we illustrate with the case p1 = {ℓi .pi } ∪˙ ✷ and p2 = {ℓ′i .qi } ∪˙ ✸. In this case,
v1
=
{ℓi .vi1 } ⊎ v01
v2
=
{ℓ′i .vi2 } ⊎ v02
v1′
v2′
=
{ℓi .wi1 } ⊎ w01
=
{ℓ′i .wi2 } ⊎ w02
and we also know that vi1 hpi wi1 and vi2 hqi wi2 for each i. Therefore,
{ℓi .vi1 } ⊎ v01 ⊎ {ℓ′i .vi2 } ⊎ v02
=
h{ℓ
′
˙
i .pi ,ℓi .qi }∪✷
=
{ℓi .vi1 , ℓ′i .vi2 } ⊎ v01 ⊎ v02
{ℓi .wi1 , ℓ′i .wi2 } ⊎ w01 ⊎ w02
{ℓi .wi1 } ⊎ w01 ⊎ {ℓ′i .wi1 } ⊎ w02
For part 2, assume p ⊑ v1 ⊎ v2 and Li ≤ dom(vi ) for i ∈ {1, 2}. We proceed by case analysis on p. The cases p = ✷ and p = ✸ are
immediate since ✷|L = ✷ and ✸|L = ✸. If p = {ℓi .pi } is a complete set pattern, then p|L1 selects just those elements of p that have a prefix
in Li , and since every element of vq has its label’s prefix in L1 we must have that v1 = {ℓ′i .vi′ } where pi ⊑ vi′ for each i and p|L1 = {ℓ′i .p′i },
which is what we need to show. The cases for p = {ℓi .pi } ∪˙ ✷ and p = {ℓi .pi } ∪˙ ✸ are similar.
A symmetric argument suffices to show p|L2 ⊑ v2 .
Lemma B.5 (Projection and ⊑).
1.
2.
3.
4.
If p ⊑ {ǫ.v} then p.ǫ ⊑ v.
If p ⊑ 1 · v1 ⊎ 2 · v2 then p[1] ⊑ v1 and p[2] ⊑ v2 .
If p ⊑ ℓ · v then p[ℓ] ⊑ v.
If p ⊑ hAi : vi i then p.Ai ⊑ vi .
Proof. For part (1), suppose p ⊑ {ǫ.v}. We proceed by case analysis on p. The cases for ✷ and ✸ are trivial. If p = {ǫ.p′ } then p.ǫ = p′ so
the conclusion follows. The cases for p = {ǫ.p′ } ∪˙ ✷ or p = {ǫ.p′ } ∪˙ ✸ are similar.
For part (2), suppose p ⊑ 1 · v1 ⊎ 2 · v2 . Suppose v1 = {ℓi .vi } and v2 = {ℓ′i .vi′ }. If p is a complete set pattern, then it must be of the
form {1.ℓi .pi , 2.ℓ′i .qi } , where pi ⊑ vi and qi ⊑ vi′ . The desired conclusion follows since p[1] = {ℓi .pi }, and a symmetric argument shows
that p[2] ⊑ v2 . The cases for partial set patterns are similar, since the ✷ or ✸ is preserved by the projection operation.
For part (3), suppose p ⊑ ℓ · v. The proof is analogous to the previous case.
For part (4), suppose p ⊑ hAi : vi i. If p = ✸ or ✷, the result is immediate; otherwise, p is a record pattern. If it is a total record pattern
hAi : pi i then clearly p.Ai = pi ⊑ vi . Otherwise, it is a partial pattern, in which case either p.Ai is a pattern pi mentioned in p, in which
case we are done, or p.Ai = ✸ or p.Ai = ✷, and the conclusion follows immediately.
Lemma B.6 (Projection and hp ).
1.
2.
3.
4.
If p ⊑ {ǫ.v} and v hp.ǫ v ′ then {ǫ.v} hp {ǫ.v ′ }.
If p ⊑ 1 · v1 ⊎ 2 · v2 and v1 hp[1] v1′ and v2 hp[2] v2′ then 1 · v1 ⊎ 2 · v2 hp 1 · v1′ ⊎ 2 · v2′ .
If p ⊑ ℓ · v and v hp[ℓ] v ′ then ℓ · v hp ℓ · v ′ .
If p ⊑ hAi : vi i and v1 hp.A1 v1′ , . . . , vn hp.An vn′ then hAi : vi i hp hAi : vi′ i.
15
Proof. For part (1), suppose p ⊑ {ǫ.v} and v hp.ǫ v ′ . If p = ✷ or ✸ then the result is immediate (this is the case for all three parts of the
lemma). If p = {ǫ.p′ }, p = {ǫ.p′ } ∪˙ ✷, or p = {ǫ.p′ } ∪˙ ✸ then p.ǫ = p′ so v hp′ v ′ . We can conclude {ǫ.v} hp {ǫ.v ′ }.
For part (2), suppose p ⊑ 1 · v1 ⊎ 2 · v2 and v1 hp[1] v1′ and v2 hp[2] v2′ . As usual, the cases p = ✷ and p = ✸ are trivial. Suppose
v1 = {ℓi .vi′ } and v2 = {ℓ′i .vi′′ }. If p is a complete set pattern it must be of the form {1.ℓi .pi , 2.ℓ′i .qi }, and v1 h{ℓi .pi } v1′ and v2 h{ℓ′ .qi } v2′ .
i
It is straightforward to show that 1·v1 h{1·ℓi .pi } 1·v1′ and 2·v2 h{2·ℓ′ .q } 2·v2′ , so by previous results we have 1·v1 ⊎2·v2 hp 1·v1′ ⊎2·v2′ .
i i
The cases for partial patterns follow the same reasoning, making use of the fact that projection preserves the partial pattern.
For part (3), suppose p ⊑ ℓ · v. The argument is similar to the previous case.
For part (4) suppose p ⊑ hAi : vi i and vi hp.Ai vi′ for each i. If p is ✷, ✸, or a complete record pattern then the conclusion is immediate.
Otherwise, if p = hBi : qi i ∪˙ ✷, the conclusion is immediate since each component of the records hAi : vi i and hAi : vi′ i either match the
appropriate qi or need not match because of the hole. Finally, if p = hBi : qi i ∪˙ ✸, the conclusion follows since each pair of corresponding
components of the records hAi : vi i and hAi : vi′ i either match the appropriate qi or are equal because p.Ai = ✸ if Ai is not among the
Bj .
C. Proof of correctness of trace slicing
Lemma C.1. If γ, x ∈ v1 , Θ y∗ v2 and v1′ ⊆ v1 then there exists v2′ ⊆ v2 such that γ, x ∈ v1′ , Θ y∗ v2′ .
Proof. The proof is straightforward by induction on derivations. The cases for v1 = ∅ and v1 = {ℓ.v} are immediate; if v1 = w1 ⊎ w2 then
we proceed by induction using the subsets w1′ = w1 ∩ v1′ and w2′ = w2 ∩ v1′ .
We prove correctness of the full trace slicing algorithm, with enriched patterns, since the correctness for simple patterns follows as a
special case.
Theorem C.2 (Correctness of Slicing).
1. Suppose γ, T y v and p ⊑ v and p, T ց ρ, S. Then for all γ ′ hρ γ and T ′ ⊒ S such that γ ′ , T ′ y v ′ we have v ′ hp v.
2. Suppose γ, x ∈ v0 , Θ y∗ v and p ⊑ v and p, x.Θ0 ց∗ ρ, Θ′0 , p0 , where Θ0 ⊆ Θ. Then for all γ ′ hρ γ and v0′ hp0 v0 and Θ′ ⊒ Θ′0
such that γ ′ , x ∈ v0′ , Θ′ y∗ v ′ we have v ′ hp v.
Proof. For part (1), the proof is by induction on the structure of slicing derivations, using inversion to extract information from other
derivations. The cases for variables, constants, primitive operations, and let-binding are exactly as in previous work [1]. The cases for
conditionals are similar to the those for variant types and case constructs in previous work.
We show the cases for records and set operations, which are new to this paper (records are handled similarly to pairs in our previous work,
so the cases for pairs are omitted).
• Field projection. If the last step in the slicing derivation is
hAi :p; ✷i, T ց ρ, S
p, T.Ai ց ρ, S.Ai
then the evaluation derivation must be of the form
γ, T y hA1 :v1 , . . . , An :vn i
γ, T.Ai y vi
Let γ ′ hρ γ and T ′ ⊒ S.Ai be given, where γ ′ , T ′ y v ′ . Then T ′ must have the form T ′′ .Ai for some T ′′ ⊒ S so the replay derivation
is of the form
γ ′ , T ′′ y hA1 :v1′ , . . . , An :vn′ i
γ ′ , T ′′ .Ai y vi′
The induction hypothesis applies since it is easy to show that hAi : p; ✷i ⊑ hA1 :v1 , . . . , An :vn i. Therefore
hA1 :v1′ , . . . , An :vn′ i hhAi :p;✷i hA1 :v1 , . . . , An :vn i .
From this it is obvious that vi′ hp vi .
• Record. If the last step in the slicing derivation is
p.A1 , T1 ց ρ1 , S1
···
p.An , Tn ց ρn , Sn
p, hA1 :T1 , . . . , An :Tn i ց ρ1 ⊔ · · · ⊔ ρn , hA1 :S1 , . . . , An :Sn i
then the evaluation derivation must be of the form
γ, T1 y v1
···
γ, Tn y vn
γ, hA1 :T1 , . . . , An :Tn i y hA1 :v1 , . . . , An :vn i
Let γ ′ hρ1 ⊔···⊔ρn γ and T ′ ⊒ hA1 :S1 , . . . , An :Sn i be given, where γ ′ , T ′ y v ′ . Then T ′ must have the form hA1 :T1′ , . . . , An :Tn′ i,
where Ti′ ⊒ Si for each i, so the replay derivation is of the form
γ ′ , T1′ y v1′
···
γ ′ , Tn′ y vn′
′
′
γ , hA1 :T1 , . . . , An :Tn i y hA1 :v1′ , . . . , An :vn′ i
′
16
By Lemma B.5 we know p.Ai ⊑ vi for each i, and γ ′ hρi γ for each i, so by induction vi′ hp.Ai vi for each i. Using Lemma B.6 we
can conclude that hA1 :v1′ , . . . , An :vn′ i hp hA1 :v1 , . . . , An :vn i.
• Empty set. This case is trivial, similar to the usual case for constants.
• Singleton. This case follows immediately from the relevant properties of p.ǫ, using similar reasoning to the record projection case.
• Union. If the last step in the slicing derivation is
p[1], T1 ց ρ1 , S1
p[2], T2 ց ρ2 , S2
p, T1 ∪ T2 ց ρ1 ⊔ ρ2 , S1 ∪ S2
then the evaluation derivation must be of the form
γ, T1 y v1
γ, T2 y v2
γ, T1 ∪ T2 y 1 · v1 ⊎ 2 · v2
Let γ ′ hρ1 ⊔ρ2 γ and T ′ ⊒ T1 ∪ T2 be given, where γ ′ , T ′ y v ′ . Then T ′ must have the form T1′ ∪ T2′ where Ti′ ⊒ Si so the replay
derivation is of the form
γ ′ , T1′ y v1′
γ ′ , T2′ y v2′
γ ′ , T1′ ∪ T2′ y 1 · v1′ ⊎ 2 · v2′
By Lemma B.5 we know p[1] ⊑ v1 and p[2] ⊑ v2 , and γ ′ hρi γ, so by induction we have v1′ hp[1] v1 and v2′ hp[2] v2 , and therefore by
Lemma B.6 we can conclude 1 · v1′ ⊎ 2 · v2′ hp 1 · v1 ⊎ 2 · v2 .
• Sum and emptiness. In both cases, since the slice ensures that the whole argument to the sum or emptiness test is preserved, the argument
is straightforward. For example, for emptiness suppose the derivation is of the form:
✸, T ց ρ, S
p, empty T ց ρ, empty S
Then there are two cases. If replay derivation is of the form
γ, T y ∅
γ, empty T y true
Suppose γ ′ hρ γ and T ′ ⊒ empty S with γ ′ , T ′ y v ′ . Then T ′ is of the form empty T ′′ with T ′′ ⊒ S, so the replay derivation must be
of the form:
γ ′ , T ′′ y v ′′
′
γ , empty T ′′ y v ′
By induction, v ′′ h✸ ∅, which implies v ′′ = ∅ so v ′ = true hp true (since p ⊑ true). The cases where the argument to empty T
evaluates to a nonempty set, and for sum T , are similar.
• Comprehension. If the derivation is of the form
p, x.Θ ց∗ ρ′ , Θ0 , p0
p0 , T ց ρ, S
[
[
p, {e | x ∈ T } ⊲ Θ ց ρ ⊔ ρ′ , {e | x ∈ S} ⊲ Θ0
then the replay derivation must be of the form
γ, T y v0
γ, x ∈ v0 Θ y∗ v
[
γ, {e | x ∈ T } ⊲ Θ y v
S
S
Suppose γ ′ hρ⊔ρ′ γ and T ′ ⊒ {e | x ∈ S} ⊲ Θ0 with γ ′ , T ′ y v ′ . Then T ′ must be of the form {e | x ∈ T ′′ } ⊲ Θ′ for some T ′′ ⊒ S
′
and Θ ⊒ Θ0 so the replay derivation must be of the form
γ ′ , T ′′ y v0′
γ ′ , x ∈ v0′ , Θ′ y∗ v ′
[
γ ′ , {e | x ∈ T ′′ } ⊲ Θ′ y v ′
Since p0 ⊑ v0 and γ ′ hρ γ and γ ′ hρ′ γ we have v0′ hp0 v0 by induction. Thus, by the second induction hypothesis, since p ⊑ v and
γ ′ hρ′ γ and v0′ hp0 v0 we have v ′ hp v.
For part (2), the proof is again by induction on the structure of derivations.
• If the slicing derivation is of the form
∅, x.∅ ց∗ [], ∅, ∅
then the conclusion is immediate, since rerunning an empty trace set always yields the empty set.
• If the slicing derivation is of the form
✸, x.∅ ց∗ [], ∅, ∅
then the conclusion is immediate as before, since rerunning an empty trace set always yields the empty set.
17
• If the slicing derivation is of the form
✷, x.Θ ց∗ [], ∅, ✷
then the conclusion is immediate, since any two values match according to ✷.
• If the slicing derivation is of the form
p[ℓ], T ց ρ[x 7→ p0 ], S
p, x.{ℓ.T } ց∗ ρ, {ℓ.S}, {ℓ.p0 }
then observe that Θ ⊇ {ℓ.T } by assumption, so Θ(ℓ) = T . So, the replay derivation must have the form
ℓ ∈ dom(Θ)
γ[x 7→ v0 ], T y v
γ, x ∈ {ℓ.v0 }, Θ y∗ ℓ · v
Now suppose that γ ′ hρ γ and v0′ h{ℓ.p0 } {ℓ.v0 } and Θ′ ⊒ {ℓ.S} are given where γ ′ , x ∈ v0′ , Θ′ y∗ v ′ . It follows that v0′ = {ℓ.v0′′ }
and v0′′ hp0 v0 , so γ ′ [x 7→ v0′′ ] hρ[x7→p0 ] γ[x 7→ v0 ]. Moreover, by inversion the derivation must have the form:
ℓ ∈ dom(Θ′ )
γ ′ [x 7→ v0′′ ], Θ′ (ℓ) y v ′
′
γ , x ∈ {ℓ.v0′′ }, Θ′ y∗ ℓ · v ′
Since by Lemma B.5 p[ℓ] ⊑ v and Θ′ (ℓ) ⊒ S (which holds because Θ′ ⊒ {ℓ.S}), we have by induction that v ′ hp[ℓ] v, and using
Lemma B.6 we can conclude that ℓ · v ′ hp ℓ · v, as desired.
• Suppose the slicing derivation is of the form:
p|dom(Θ1 ) , x.Θ1 ց∗ ρ1 , Θ′1 , p1
p|dom(Θ2 ) , x.Θ2 ց∗ ρ2 , Θ′2 , p2
p, x.Θ1 ⊎ Θ2 ց∗ ρ1 ⊔ ρ2 , Θ′1 ⊎ Θ′2 , p1 ⊎ p2
and suppose that γ, x ∈ v0 , Θ y∗ v where Θ ⊇ Θ1 ⊎ Θ2 . Suppose that γ ′ hρ1 ⊔ρ2 γ and v0′ hp1 ⊎p2 v0 and Θ′ ⊒ Θ′1 ⊎ Θ′2 are given,
where γ ′ , x ∈ v0′ , Θ′ y∗ v ′ . We need to show that v ′ hp v.
Since v0′ hp1 ⊎p2 v0 , it is straightforward to show that there must exist v1 , v2 , v1′ , v2′ such that v1 ⊎ v2 = v0 , v1′ ⊎ v2′ = v0′ , v1 hp1 v1
and v2′ hp2 v2 . Furthermore, by Lemma C.1 we know that γ, x ∈ vi , Θ y∗ wi and γ ′ , x ∈ vi′ , Θ′ y∗ wi′ for some w1 , w2 , w1′ , w2′ .
Therefore, we can conclude that:
γ ′ , x ∈ v1′ , Θ′ y∗ w1′
γ ′ , x ∈ v2′ , Θ′ y∗ w2′
′
′
′
′
γ , x ∈ v1 ⊎ v2 , Θ y∗ w1′ ⊎ w2′
γ, x ∈ v1 , Θ y∗ w1
γ, x ∈ v2 , Θ y∗ w2
γ, x ∈ v1 ⊎ v2 , Θ y∗ w1 ⊎ w2
Furthermore, since v1 ⊎ v2 = v0 , by determinacy we know that w1 ⊎ w2 = v and similarly w1′ ⊎ w2 = v ′ . By induction since pi ⊑ vi
and Θi ⊆ Θ, we know that wi′ hp|dom(Θ ) wi , so we know that
i
v ′ = w1′ ⊎ w2′ hp|dom(Θ
i)
⊎p|dom(Θ )
2
w 1 ⊎ w2 = v .
To conclude, since p ⊑ w1 ⊎ w2 and dom(Θ1 ) ≤ w1 and dom(Θ2 ) ≤ w2 , it follows that p = p|dom(Θ1 ) ⊎ p|dom(Θ2 ) , so we can
conclude w1 ⊎ w2 hp w1′ ⊎ w2′ as desired.
This exhausts all cases and completes the proof.
D. Proof of correctness of query slicing
Lemma D.1. If γ, x ∈ v1 , e ⇓∗ v2 and v1′ ⊆ v1 then there exists v2′ ⊆ v2 such that γ, x ∈ v1′ , e ⇓∗ v2′ .
Proof. The proof is straightforward by induction on derivations. The cases for v1 = ∅ and v1 = {ℓ.v} are immediate; if v1 = w1 ⊎ w2 then
we proceed by induction using the subsets w1′ = w1 ∩ v1′ and w2′ = w2 ∩ v1′ .
We prove Theorem 5.1 by strengthening the induction hypothesis as follows:
Theorem D.2 (Correctness of Query Slicing).
1. Suppose γ, T y v and p ⊑ v and p, T ց
ց ρ, e. Then for all γ ′ hρ γ and e′ ⊒ e such that γ ′ , e′ ⇓ v ′ we have v ′ hp v.
2. Suppose γ, x ∈ v0 , Θ y∗ v and p ⊑ v and p, Θ0 ց
ց∗ ρ, e0 , p0 , where Θ0 ⊆ Θ. Then for all γ ′ hρ γ and v0′ hp0 v0 and e′0 ⊒ e0 such
′
′
′
∗ ′
′
that γ , x ∈ v0 , e0 ⇓ v we have v hp v.
Proof. The proof is by induction on the structure of query slicing derivations. Many of the cases are essentially the same as for trace slicing.
The cases for conditionals are straightforward, since in either case the sliced trace and environment retain enough information to force the
same branch to be taken on recomputation. We show the details of the cases involving conditionals and comprehensions.
For part (1), we consider a conditional and comprehension rule:
• If the slicing derivation is of the form:
p 1 , T1 ց
ց ρ1 , e′1
true, T ց
ց ρ, e′
p1 , if(T, e1 , e2 ) ⊲true T1 ց
ց ρ1 ⊔ ρ, if(e′ , e′1 , ✷)
18
then the replay derivation must be of the form:
γ, T y true
γ, T1 y v1
γ, if(T, e1 , e2 ) ⊲true T1 y v1
Suppose γ ′ hρ1 ⊔ρ γ and e′′ ⊒ if(e′ , e′1 , ✷) are given where γ ′ , e′ ⇓ v ′ . Then e′′ = if(e′′0 , e′′1 , e′′2 ), where e′′0 ⊒ e′ and e′′1 ⊒ e′1 , so
there are two cases for the evaluation derivation. If it has the form
γ ′ , e′0 ⇓ false
γ ′ , e′2 ⇓ v2′
γ ′ , if(e′0 , e′1 , e′2 ) ⇓ v2′
then since γ ′ hρ γ and γ ′ hρ1 γ, by induction we would have that false htrue true, which is absurd. So this case cannot arise.
Otherwise, the derivation must have the form:
γ ′ , e′0 ⇓ true
γ ′ , e′1 ⇓ v1′
′
′
′
′
γ , if(e0 , e1 , e2 ) ⇓ v1′
Since γ ′ hρ γ and γ ′ hρ1 γ, by induction we have that v1′ hp1 v1 as desired.
• Comprehension. If the derivation is of the form
p0 , T ց
ց ρ, e′0
p, x.Θ ց
ց∗ ρ′ , e′1 , p0
[
[
p, {e | x ∈ T } ⊲ Θ ց
ց ρ ⊔ ρ′ , {e′1 | x ∈ e′0 }
then the replay derivation must be of the form
γ, T y v0
γ, x ∈ v0 , Θ y∗ v
[
γ, {e | x ∈ T } ⊲ Θ y v
S
S
Suppose γ ′ hρ⊔ρ′ γ and e′′ ⊒ {e′1 | x ∈ e′0 } with γ ′ , e′′ ⇓ v ′ . Then e′′ must be of the form {e′′1 | x ∈ e′′0 } for some e′′1 ⊒ e′1 and
′′
′
e0 ⊒ e0 , so the evaluation derivation must be of the form
γ ′ , e′′0 ⇓ v0′
γ ′ , x ∈ v0′ , e′′1 ⇓∗ v ′
[
γ ′ , {e′′1 | x ∈ e′′0 } ⇓ v ′
Since p0 ⊑ v0 and γ ′ hρ γ and γ ′ hρ′ γ we have v0′ hp0 v0 by induction. Thus, by the second induction hypothesis, since p ⊑ v and
γ ′ hρ′ γ and v0′ hp0 v0 we have v hp v ′ .
For part (2), we consider the singleton and union rules:
• If the slicing derivation is of the form
p[ℓ], T ց
ց ρ[x 7→ p0 ], e′
p, x.{ℓ.T } ց
ց∗ ρ, e′ , {ℓ.p0 }
then recall that by assumption, Θ(ℓ) ⊇ {ℓ.T }, so Θ(ℓ) = T , so the replay derivation must have the form
ℓ ∈ dom(Θ)
γ[x 7→ v0 ], T y v
γ, x ∈ {ℓ.v0 }, Θ y∗ ℓ · v
Now suppose that γ ′ hρ γ and v0′ h{ℓ.p0 } {ℓ.v0 } and e′′ ⊒ e′ are given where γ ′ , x ∈ v0′ , e′′ ⇓∗ v ′ . It follows that v0′ = {ℓ.v0′′ } and
v0′′ hp0 v0 , so γ ′ [x 7→ v0′′ ] hρ[x7→p0 ] γ[x 7→ v0 ]. Moreover, by inversion the derivation must have the form:
γ ′ [x 7→ v0′′ ], e′′ ⇓ v ′
γ , x ∈ {ℓ.v0′′ }, e′′ ⇓∗ ℓ · v ′
′
Since by Lemma B.5 p[ℓ] ⊑ v we have by induction that v ′ hp[ℓ] v, and using Lemma B.6 we can conclude that ℓ · v ′ hp ℓ · v, as desired.
• Suppose the slicing derivation is of the form:
p|dom(Θ1 ) , x.Θ1 ց
ց∗ ρ1 , e′1 , p1
p|dom(Θ2 ) , x.Θ2 ց
ց∗ ρ2 , e′2 , p2
∗
′
p, x.Θ1 ⊎ Θ2 ց
ց ρ 1 ⊔ ρ 2 , e0 , p 1 ⊎ p 2
and suppose that γ, x ∈ v0 , Θ y∗ v where Θ ⊇ Θ1 ⊎ Θ2 . Suppose that γ ′ hρ1 ⊔ρ2 γ and v0′ hp1 ⊎p2 v0 and e′0 ⊒ e0 are given, where
γ ′ , x ∈ v0′ , e′0 ⇓∗ v ′ . We need to show that v ′ hp v.
Since v0′ hp1 ⊎p2 v0 , it is straightforward to show that there must exist v1 , v2 , v1′ , v2′ such that v1 ⊎ v2 = v0 , v1′ ⊎ v2′ = v0′ , v1 hp1 v1 and
v2′ hp2 v2 . Furthermore, by Lemmas C.1 and D.1 we know that γ, x ∈ vi , Θ y∗ wi and γ ′ , x ∈ vi′ , e′0 ⇓∗ wi′ for some w1 , w2 , w1′ , w2′ .
Therefore, we can conclude that:
γ ′ , x ∈ v1′ , e′0 ⇓∗ w1′
γ ′ , x ∈ v2′ , e′0 ⇓∗ w2′
γ, x ∈ v1 , Θ y∗ w1
γ, x ∈ v2 , Θ y∗ w2
∗
′
′
′
′
γ, x ∈ v1 ⊎ v2 , Θ y w1 ⊎ w2
γ , x ∈ v1 ⊎ v2 , e0 ⇓∗ w1′ ⊎ w2′
Furthermore, since v1 ⊎ v2 = v0 , by determinacy we know that w1 ⊎ w2 = v and similarly w1′ ⊎ w2 = v ′ . By induction since pi ⊑ vi
and Θi ⊆ Θ, we know that wi′ hp|dom(Θ ) wi , so we know that
i
v ′ = w1′ ⊎ w2′ hp|dom(Θ
i)
19
⊎p|dom(Θ )
2
w 1 ⊎ w2 = v .
To conclude, since p ⊑ w1 ⊎ w2 and dom(Θ1 ) ≤ w1 and dom(Θ2 ) ≤ w2 , it follows that p = p|dom(Θ1 ) ⊎ p|dom(Θ2 ) , so we can
conclude w1 ⊎ w2 hp w1′ ⊎ w2′ as desired.
This exhausts all cases and completes the proof.
20
| 6cs.PL
|
Barely CAT(-1) groups are acylindrically hyperbolic
Anthony Genevois and Arnaud Stocker
December 14, 2017
arXiv:1712.04736v1 [math.GR] 13 Dec 2017
Abstract
In this paper, we show that, if a group G acts geometrically on a geodesically complete
CAT(0) space X which contains at least one point with a CAT(-1) neighborhood, then G
must be either virtually cyclic or acylindrically hyperbolic. As a consequence, the fundamental group of a compact Riemannian manifold whose sectional curvature is nonpositive
everywhere and negative in at least one point is either virtually cyclic or acylindrically hyperbolic. This statement provides a precise interpretation of an idea expressed by Gromov
in his paper Asymptotic invariants of infinite groups.
Contents
1 Introduction
1
2 Preliminaries
3
3 Barely CAT(-1) groups
4
4 Cubical case
11
5 A non relatively hyperbolic barely CAT(-1) group
12
References
13
1
Introduction
A key concept in geometric group theory is the notion of “curvature”. Usually, exhibiting
some negative curvature in the geometry of a given group provides interesting information on the algebraic properties of our group. The first instance of such a geometry
in group theory was small cancellation (see [LS77], or [MW02] for a more geometric
approach), next generalised by Gromov with the seminal concept of hyperbolic groups
[Gro87]. There, Gromov also suggests the definition of relatively hyperbolic groups, allowing some non hyperbolic subspaces but concentrating them into a controlled collection of
subgroups. We refer to the survey [Hru10] and references therein for more information
on relatively hyperbolic groups and their historical development. More recently, Osin
introduced acylindrically hyperbolic groups [Osi16] in order to unify several classes of
groups considered as “negatively-curved”.
Although acylindrically hyperbolic groups generalise relatively hyperbolic groups,
many arguments used in the context of relatively hyperbolic groups turn out to hold in
the context of acylindrically hyperbolic groups as well [DGO17]. One of the most impressive algebraic consequence is that acylindrically hyperbolic groups are SQ-universal. The
first motivating examples of such groups were mapping class groups [BF02] and outer
1
automorphism groups of free groups [BF10]. However, since then a large amount of articles have been dedicated to the recognition of acylindrically hyperbolic groups among
familiar classes of groups, including graph products [MO15], 3-manifold groups [MO15],
Cremona groups [Lon15], small cancellation groups [GS14, AH16, Gen16b], groups of
deficiency at least two [Osi15], Artin groups [CW16, CM16], diagram groups [Gen16c]
and graph braid groups [Gen17a]. So far, acylindrical hyperbolicity is the most general
convincing notion of groups admitting some “negatively-curved behaviour”.
In this paper, we are interested in some “nonpositively-curved” groups, namely
CAT(0) groups, ie., groups acting geometrically on CAT(0) spaces. These spaces generalise Riemannian manifolds whose curvature is everywhere nonpositive, also known as
Hadamard manifolds. Loosely speaking, what we show is that a CAT(0) group which is
“barely hyperbolic” turns out to be acylindrically hyperbolic. More precisely, the main
result of the article is:
Theorem 1.1. Let G be a group acting geometrically on some geodesically complete
proper CAT(0) space X. Suppose that X contains at least one point with a CAT(-1)
neighborhood. Then G must be either virtually cyclic or acylindrically hyperbolic.
As a consequence of this criterion, one immediately gets the following statement:
Corollary 1.2. The fundamental group of a compact Riemannian manifold whose sectional curvature is nonpositive everywhere and negative in at least one point is either
virtually cyclic or acylindrically hyperbolic.
In fact, this statement has motivated the present work, following a remark made by
Gromov in his article Asymptotic invariants of infinite groups [Gro93].
“In fact, the fundamental groups Γ of manifolds V with K ≤ 0, such
that K < 0 at some point v ∈ V generalize in a certain way the hyperbolic
groups”. ([Gro93], page 147.)
Our strategy to prove Theorem 1.1 is to show the existence of a contracting (or
equivalently, rank-one) geodesic ray in our CAT(0) space, which implies the existence of
a contracting (or equivalently, rank-one) isometry in our group, and which finally implies
that our group must be either virtually cyclic or acylindrically hyperbolic. Consequently,
Theorem 1.1 is related to the
Rank Rigidity Conjecture. Let X be a proper geodesically complete CAT(0) space
and G an infinite group acting geometrically on X. If X is irreducible, then X is a
higher rank symmetric space, or a Euclidean building of dimension at least two, or G
contains a rank-one isometry.
In one direction, our theorem is a consequence of the conjecture. And in the oppositive direction, our theorem may be applied to verify the conjecture in particular cases.
However, we do not know examples where our criterion applies and where the Rank
Rigidity Conjecture remains unknown.
The article is organised as follows. In Section 2, we recall briefly several known characterisations of acylindrical hyperbolicity among CAT(0) groups, including the criterion
which we will use to prove Theorem 1.1 in Section 3. In Section 4, we state and prove a
combinatorial version of Theorem 1.1 in the context of CAT(0) cube complexes, under
more general assumptions. Finally, in Section 5, we give an example of a barely CAT(-1)
group which is not relatively hyperbolic, showing that the conclusion of Theorem 1.1
cannot be strengthened to get some relative hyperbolicity.
2
2
Preliminaries
CAT(κ) spaces. In this paragraph, we briefly recall the definition of CAT(κ) spaces
for κ ≤ 0 and the notion of angle in these spaces which will be used in the paper; we refer
to [BH99] for more information and proofs of the statements of this section. For κ ≤ 0
let (Mκ2 , dκ ) be the simply connected Riemannian 2-manifold of constant curvature equal
to κ. That is
(
Mκ2
=
√1 H2
−κ
E2
if κ < 0
if κ = 0.
A triangle in a geodesic metric space X consists of three points p, q, r ∈ X (the vertices)
together with a choice of geodesics segments [p, q], [q, r] and [r, p]. Such a triangle will
be denoted ∆(p, q, r) (this notation is not accurate because the geodesic segments are, in
general, not unique). A comparison triangle in Mκ2 for ∆(p, q, r) is a triangle ∆(p̄, q̄, r̄) in
Mκ2 such that dκ (p̄, q̄) = d(p, q), dκ (q̄, r̄) = d(q, r) and dκ (r̄, p̄) = d(r, p). Such a triangle
always exists and is unique up to isometry. A point x̄ ∈ [p̄, q̄] is called a comparison
point for x ∈ [p, q] if d(p, x) = dκ (p̄, x̄). Similarly, we can define comparison points on
[q̄, r̄] and [p̄, r̄].
Definition 2.1. Let (X, d) be a geodesic metric space and κ ≤ 0.
• A triangle ∆(p, q, r) in X satisfies the CAT(κ) inequality if for all x, y ∈ ∆(p, q, r)
and all comparison points x̄, ȳ ∈ ∆(p̄, q̄, r̄),
d(x, y) ≤ dκ (x̄, ȳ).
• X is a CAT(κ) space if all triangles satisfies the CAT(κ) inequality.
Basic examples of CAT(κ) spaces, which motivated the above definition, are the
simply connected Riemannian manifolds of sectionnal curvature ≤ κ.
In CAT(κ) spaces, κ ≤ 0, one can define a notion of angle:
Definition 2.2. Let (X, d) be a CAT(κ) space, κ ≤ 0 and p, x, y three distinct points
in X. Let cx , cy be geodesic paths from p to x and y respectively. Then the limit
lim ∠(0)
p (cx (t), cy (t)),
t→0
(0)
exists, where ∠p (cx (t), cy (t)) is the angle at p̄ in a comparison triangle for ∆(p, cx (t), cy (t))
in M02 . This limit is called the Alexandrov angle ∠p (x, y) at p between x and y.
The Alexandrov angle ∠p (x, y) is always smaller than the angle ∠κp (x, y) at p̄ in a
comparison triangle ∆(p̄, x̄, ȳ) ⊂ Mκ2 and satisfies the triangular inequality.
Acylindrically hyperbolic CAT(0) groups. We do not define acylindrically hyperbolic groups here, and refer to [Osi16] and references therein for more information
on this class of groups. We only mention the following characterisation of acylindrical
hyperbolicity in the context of CAT(0) groups (see [Gen17b, Theorem 6.40] and the
references mentioned there).
Theorem 2.3. Let G be a non virtually cyclic group acting geometrically on a CAT(0)
space. Then G is acylindrically hyperbolic if and only if X contains a contracting
geodesic ray.
3
Recall that, given a metric space (S, d), a subspace Y ⊂ S is contracting if there exists
some B ≥ 0 such that the nearest-point projection onto Y of any ball which is disjoint
from Y has diameter at most B. The strategy to prove our main theorem will be to
construct a contracting geodesic line. To be precise, we will construct a slim geodesic
line, as define in [CS15].
Definition 2.4. Let X be a CAT(0) space. A geodesic line γ ⊂ X is slim if there
exists some δ ≥ 0 such that, for every x ∈ X and every y ∈ γ, the distance between the
projection of x onto γ and the geodesic [x, y] is at most δ.
According to [CS15, Theorem 2.4], being contracting and being slim for a geodesic line
in a CAT(0) space are equivalent.
3
Barely CAT(-1) groups
In this section, we prove the main result of our article. Namely:
Theorem 3.1. Let G be a group acting geometrically on some geodesically complete
proper CAT(0) space X. Suppose that X contains at least one point with a CAT(-1)
neighborhood. Then G must be either virtually cyclic or acylindrically hyperbolic.
For convenience, from now on we will refer to points admitting CAT(-1) neighborhoods
as CAT(-1) points. Our theorem will be a consequence of Propositions 3.2 and 3.6
below. The first proposition shows that X must contain some geodesic line passing
through infinitely many CAT(-1) points in such a way that the distance between any
two consecutive such points is uniformly bounded. Next, our second proposition states
that such a line must be slim, or equivalently, contracting. The conclusion finally follows
from Theorem 2.3.
Proposition 3.2. Let X be a geodesically complete, cocompact and proper CAT(0)
space which contains at least one point with a CAT(-1) neighborhood. Then there exists
some geodesic ray r : [0, +∞) → X and an increasing sequence (tn ) of nonnegative reals
tending to +∞ such that r(tn ) has a CAT(-1) neighborhood for every n ≥ 0 and such
that the auxiliary sequence (tn+1 − tn ) is bounded.
The proposition will follow from the following two lemmas:
Lemma 3.3. Let X be a CAT(0) space and R, δ, > 0 some fixed constants. Set
k = δ + 1. For every geodesic ray r, the segment [r(0), x] intersects B(r(t), ) for any
t ∈ [0, δ] and x ∈ B(r(t + kR), R).
Proof. We note o = r(0), z = r(t), y = r(t + kR). Consider x ∈ B(y, R) and a
comparison triangle ∆(ō, ȳ, x̄) in E2 for the geodesic triangle ∆(o, y, x). Let ω̄ be the
intersection between [ō, x̄] and the parallel line to (ȳ, x̄) passing through z̄ and let ω be
the corresponding point on [0, x]. Then, Thales theorem gives
d(ō, z̄)
d(ω̄, z̄)
=
,
d(ō, ȳ)
d(x̄, ȳ)
and with the CAT(0) inequality we get
d(ω, z) ≤ d(ω̄, z̄) =
δR
d(ō, z̄)
d(x̄, ȳ) ≤
< .
d(ō, ȳ)
t + kR
4
Lemma 3.4. Let X be a cocompact proper CAT(0) space which contains at least one
point with a CAT(-1) neighborhood and R, , c > 0 some fixed constants. There exists
some k = k(R, , c) such that, if o, z, y, x ∈ X are points satisfying:
1. o, z, y are aligned in this order;
2. d(x, y) ≤ R and d(z, y) ≥ k;
3. B(z, c) is a CAT(-1) neighborhood of z;
then the segment [o, x] intersects the ball B(z, ).
Proof. By contradiction, assume that, for all n ∈ N, there exist on , xn , yn , zn ∈ X such
that:
• on , zn , yn are aligned in this order;
• d(xn , yn ) ≤ R and d(zn , yn ) ≥ n;
• B(zn , c) is a CAT(-1) neighborhood of zn ;
• [on , xn ] ∩ B(zn , ) = ∅.
This last assumption, together with the previous lemma, implies that
n ≤ d(zn , yn ) ≤
d(on , zn )
+1 ·R
ε
so d(on , zn ) −−−−−→ +∞. Since X is cocompact, up to translation, we may assume
n→+∞
that the sequence (zn ) stays in a compact K ⊂ X; and since d(on , zn ) and d(zn , yn )
both tend to +∞, we deduce from Arzelà-Ascoli theorem that, up to subsequence, the
segments [on , yn ] tend to a geodesic line l. Similarly, because [on , xn ] is included into
the R-neighborhood of [on , yn ] as a consequence of the convexity of the distance, up to
a subsequence the segments [on , xn ] converge to a geodesic line l0 parallel to l. These
lines are distinct since d(zn , [on , xn ]) ≥ for all n. The contradiction comes from the
Flat Strip Theorem [BH99, Theorem II.2.13] and the following fact, stating that a limit
of CAT(-1) points must be CAT(-1) as well:
Fact 3.5. Let (zn ) be a sequence of points converging to some z ∈ X such that B(zn , c)
is CAT(-1) for all n. Then B(z, 2c ) is a CAT(-1) neighborhood of z.
Indeed, for n large enough, B(z, 2c ) ⊂ B(zn , c). The conclusion follows from the convexity of balls.
Now we are ready to prove Proposition 3.2. For o, z ∈ X and R > 0 we define the
shadow Oo (z, R) to be the set of y ∈ X such that the segment [o, y] meets the ball
centered at z of radius R.
Proof of Proposition 3.2. Let Y = X/G be a compact quotient of X where G < Isom(X)
and z ∈ X a point such that B(z, 2) is a CAT(-1) neighborhood for some > 0. Choose
R ≥ 2 + diam(Y ) and K ≥ max(R + 2, k(R, , 2)) where k(R, , 2) is the constant
from Lemma 3.4. Finally, we fix a base point o ∈ X.
We now construct by induction a sequence of points (zn ) in the G-orbit of z such that
for all n ≥ 1:
1. Oo (zn , ) ⊂ · · · ⊂ Oo (z1 , );
2. there exist zn1 , . . . , znn ∈ [o, zn ] such that znk ∈ B(zk , ), ≤ d(znk , znk+1 ) ≤ K + R +
and znn = zn .
5
Figure 1
Step n = 1. Simply take z = z1 .
From step n to step n + 1. Assume that z1 , . . . , zn are constructed. Since X is
geodesically complete, we can extend the segment [o, zn ] to a geodesic ray rn and consider
the point yn on rn such that d(yn , zn ) = K. By the choice of K and Lemma 3.4, we
have
Oo (yn , R) ⊂ Oo (zn , ).
Moreover, since the orbit of z is diam(Y )-dense, by choice of R, there is some zn+1 ∈ G·z
such that B(zn+1 , ) ⊂ B(yn , R). It follows that
Oo (zn+1 , ) ⊂ Oo (yn , R) ⊂ Oo (zn , ),
and so, [o, zn+1 ] meets every ball B(zk , ), k = 1, . . . , n. Finally, if one fixes some
k
zn+1
∈ B(zk , ) ∩ [o, zn+1 ], then for k ≤ n − 1 we have
k+1
k+1
k
k
d(zn+1
, zn+1
) ≤ d(zn+1
, zk ) + d(zk , yk+1 ) + d(yk+1 , zn+1
) ≤ + K + R,
and
k+1
k+1
k
k
d(zn+1
, zn+1
) ≥ d(zk , yk+1 ) − d(zn+1
, zk ) − d(yk+1 , zn+1
) ≥ K − R − ≥ .
By Arzelà-Ascoli theorem, up to subsequence, the sequence of segments [z1 , zn ] converges
to some geodesic ray r uniformly on compact sets. By construction, r meets each
closed ball B̄(zn , ) at some point r(tn ) such that ≤ tn+1 − tn ≤ R + K + and
B(r(tn ), ) ⊂ B(zn , 2) is a CAT(-1) neighborhood.
Now, we focus on the second step of the proof of Theorem 3.1. Namely:
Proposition 3.6. Let X be a CAT(0) space and γ : R → X a geodesic line passing
through infinitely many CAT(-1) points in such a way that the distance between any two
consecutive such points is uniformly bounded from above and from zero. Then γ is a
slim geodesic.
Inspired from [KL95], our proof is based on the Gauss-Bonnet formula. We begin by
recalling a few definitions.
Definition 3.7. Let X be a polygonal complex. A corner (v, C) is the data of a vertex
v ∈ X and a polygon C ⊂ X containing v. Given a vertex v ∈ X (resp. a polygon
R ⊂ X), we denote by corner(v) (resp. corner(R)) the set of corners based at v (resp.
the set of corners supported by R).
6
Definition 3.8. An angled polygonal complex (X, ∠) is the data of a polygonal complex
X and a map
∠ : {corners of X} → R.
The curvature of a vertex v ∈ X is defined as
κ(v) = 2π − π · χ (link(v)) −
X
∠(v),
c∈corner(v)
and the curvature of a polygon R ⊂ X as
X
κ(R) =
∠(c) − π · |∂R| + 2π.
c∈corner(R)
It is worth noticing that, if R is a triangle, then its curvature coincides with minus its
deficiency, denoted by def(R), ie., the difference between π and the sum of its angles.
As proved in [MW02, Theorem 4.6], this formalism allows us to recover a combinatorial
version of the well-known Gauss-Bonnet formula.
Combinatorial Gauss-Bonnet formula. Let (X, ∠) be an angled polygonal complex.
Then
X
X
κ(R) = 2π · χ(X).
κ(v) +
v vertex
R polygon
Now we are ready to prove Proposition 3.6.
Proof of Proposition 3.6. Let x ∈
/ γ and y ∈ γ be two points, and let z ∈ γ denote
the projection of x onto γ. Fix some constants , L, R > 0 such that a ball of radius
centered at some CAT(-1) point is CAT(-1) and such that the distance between any
two consecutive CAT(-1) points along γ is at most L and at least R. Without loss
of generality, we may suppose that R > 2. Suppose that there exist N consecutive
CAT(-1) points x1 , . . . , xN ∈ γ between z and y such that B(xi , ) ∩ [x, y] = ∅ for every
1 ≤ i ≤ N . For every 1 ≤ i ≤ N , fix a point yi ∈ [z, x] ∪ [x, y] whose projection onto
γ is xi ; notice that, because the projection of [x, z] is reduced to the singleton {z},
necessarily yi ∈ [x, y]. Finally, for every 1 ≤ i ≤ N , let ai ∈ [xi−1 , xi ], bi ∈ [xi , yi ] and
ci ∈ [xi , xi+1 ] be the unique points of the corresponding segments at distance from xi
(for convenience, we set x0 = z and xN +1 = y). The configuration is summarised by
Figure 2.
Denote by ∆ the union of the geodesics
• [x, y], [y, z] and [x, z];
• [yi , bi ] for every 1 ≤ i ≤ N ;
• [bi , ai ] and [bi , ci ] for every 1 ≤ i ≤ N ;
• [yi , ci ], [yi , ai+1 ] and [yi , bi+1 ] for every 0 ≤ i ≤ N − 1 (setting y0 = x and x0 = z).
Roughly speaking, ∆ is a triangulation of the geodesic triangle T (x, y, z). Now, consider
a comparison triangle T (x̄, ȳ, z̄) of T (x, y, z). For every 1 ≤ i ≤ N , let xi , yi , ai , ci denote
the preimages of xi , yi , ai , ci respectively under the map T (x̄, ȳ, z̄) → T (x, y, z). This
map has a natural extension
T (x̄, ȳ, z̄) ∪
N
[
[y i , xi ] → T (x, y, z) ∪
i=1
N
[
[yi , xi ],
i=1
7
Figure 2
sending each segment [y i , xi ] to the geodesic [yi , xi ] by an affine map. We denote by bi
the preimage of bi under this map, for every 1 ≤ i ≤ N . By triangulating T (x̄, ȳ, z̄) as
T (x, y, z), one gets a planar triangle complex ∆, and a map f : ∆ → ∆ extending the
comparison map T (x̄, ȳ, z̄) → T (x, y, z).
From now on, the triangles (ci , ai , bi ) for 1 ≤ i ≤ N , and their images under f , will be
referred to as CAT(-1) triangles. Notice that a CAT(-1) triangle in ∆ is contained into
a ball of radius centered at a CAT(-1) point, so it lies in a CAT(-1) subspace. For
every triangle δ of ∆, we assign angles to its corners by the following rules:
• for non CAT(-1) triangles, we assign the angles of a comparison triangle in E2 ;
• for the CAT(-1) triangle (āi , b̄i , c̄i ) we assign to the vertex āi (respectively c̄i ) the
corresponding angle in a comparison triangle in H2 for (ai , bi , xi ) (respectively
(ci , bi , xi )) and to the vertex b̄i the corresponding angle in a comparison triangle
in H2 for (ai , bi , ci ).
We emphasize that the angle ∠c̄i āi , b̄i , for instance, denotes the angle as defined
above, and not the corresponding angle in the Euclidean triangulation ∆.
The plan is to apply the Gauss-Bonnet formula in order to bound the number N . So
we need to investigate the curvatures of the vertices and triangles of ∆.
Claim 3.9. The curvature of any vertex of ∆, different from x̄, ȳ, z̄ or the b̄i for 1 ≤
i ≤ N , is nonpositive.
Fix some 1 ≤ i ≤ N . We want to compute the curvature of c̄i . Notice that
∠ci (ai , bi ) + ∠ci (bi , y i ) + ∠ci (y i , ai+1 )
is at least
∠ci (ai , bi ) + ∠ci (bi , yi ) + ∠ci (yi , ai+1 ) ≥ ∠ci (ai , ai+1 ) = π.
Consequently, κ(ci ) ≤ 0. The same argument implies that κ(ȳi ) and κ(āi ) are nonpositive. This concludes the proof of our claim.
Let (a0i , b0i , x0i ) and (c00i , b00i , x00i ) be comparison triangles in H2 of (ai , bi , xi ) and (ci , bi , xi )
respectively. These are isosceles triangles with angles at least π2 at x0i and x00i respectively.
¯
In particular, by definition of the angles in ∆,
∠āi (b̄i , c̄i ) = ∠a0i (b0i , x0i ) = ∠b0i (a0i , x0i ),
8
∠c̄i (b̄i , āi ) = ∠c00i (b00i , x00i ) = ∠b00i (c00i , x00i ),
where ∠a0i (b0i , x0i ), ∠b0i (a0i , x0i ), ∠c00i (b00i , x00i ), ∠b00i (c00i , x00i ) are the angles measured in H2 .
Claim 3.10. For 1 ≤ i ≤ N , we have κ(b̄i ) ≤ ∠b0i (a0i , x0i ) + ∠b00i (x00i , c00i ) − ∠b̄i (āi , c̄i ).
Fix some 1 ≤ i ≤ N . The same argument as before shows that
u := ∠bi (y i , y i−1 ) + ∠bi (y i−1 , ai ) + ∠b0i (a0i , x0i ) + ∠bi (y i , ci ) + ∠b00i (c00i , x00i )
is at least 2π. Since
κ(b̄i ) = 2π − (u − ∠b0i (a0i , x0i ) − ∠b00i (x00i , c00i ) + ∠b̄i (āi , c̄i )),
this proves the claim.
Next, notice that the curvature of a triangle of ∆ which is not CAT(-1) is zero, since its
angles come from a Euclidean triangle. Therefore,
X
κ(δ) =
δ triangle
X
κ(δ).
δ CAT(-1)
¯ Recall that the deficiency def(δi ) of δi is the
Let δi be the triangle (āi , b̄i , c̄i ) in ∆.
difference between π and the sum of its angles, that is
def(δi ) = π − (∠āi (b̄i , c̄i ) + ∠c̄i (āi , b̄i ) + ∠b̄i (āi , c̄i ))
= π − (∠a0i (b0i , x0i ) + ∠c00i (b00i , x00i ) + ∠b̄i (āi , c̄i ))
= π − (∠b0i (a0i , x0i ) + ∠b00i (c00i , x00i ) + ∠b̄i (āi , c̄i )).
Claim 3.11.
s
N
X
(κ(b̄i ) − def(δi )) ≤ −N π − 4 arccos
i=1
cosh()
cosh() + 1
!!
.
We will use the following hyperbolic trigonometry formula:
Lemma 3.12. [Bus10, Theorem 2.2.1(ii)] Let ∆ be a geodesic triangle in H2 with sides
of length a, b, c and respective opposite angles α, β, γ. Then
cos(γ) = sin(α) sin(β) cosh(c) − cos(α) cos(β).
As a consequence of Claim 3.10,
κ(b̄i ) − def(δi ) ≤ ∠b0i (a0i , x0i ) + ∠b00i (x00i , c00i ) − ∠b̄i (āi , c̄i ) + ∠b0i (a0i , x0i ) + ∠b00i (c00i , x00i )
+ ∠b̄i (āi , c̄i ) − π
≤ 2(∠b0i (a0i , x0i ) + ∠b00i (x00i , c00i )) − π.
By applying Lemma 3.12 to the triangle (a0i , b0i , x0i ), we deduce that
0 ≥ cos ∠x0i (a0i , b0i )
≥ sin ∠a0i (b0i , x0i ) sin ∠b0i (a0i , x0i ) cosh(d(a0i , b0i )) − cos ∠a0i (b0i , x0i ) cos ∠b0i (a0i , x0i )
≥
1 − cos2 ∠b0i (a0i , x0i ) cosh(d(ai , bi )) − cos2 ∠b0i (a0i , x0i )
So we get, since ∠b0i (a0i , x0i ) ≤ π2 , that
s
cos ∠b0i (a0i , x0i )
≥
cosh(d(ai , bi ))
.
1 + cosh(d(ai , bi ))
9
Notice that d(ai , bi ) is greater than since the triangle (a0i , b0i , x0i ) is isoscele with angle at
cosh(x)
is increasing. Therefore,
least π2 at x0i and d(x0i , a0i ) = . Also notice that x 7→ cosh(x)+1
since cosinus is a decreasing function on [0, π], we obtain that
s
∠b0i (a0i , x0i ) ≤ arccos
!
cosh()
.
1 + cosh()
Similarly, one has
s
∠b00i (c00i , x00i )
≤ arccos
!
cosh()
.
1 + cosh()
Thus,
s
κ(b̄i ) − def(δi ) ≤ 4 arccos
cosh()
1 + cosh()
!
− π.
This concludes the proof of the claim.
Finally, by applying the Gauss-Bonnet formula to ∆, one deduces that
s
2π =
X
κ(v) +
v vertex
X
κ(δ) ≤ 3π − N π − 4 arccos
δ triangle
hence
s
N π − 4 arccos
r
Notice that π − 4 arccos
strictly greater than
Hence,
1
2)
cosh()
cosh()+1
since x 7→
cosh()
cosh() + 1
cosh()
cosh() + 1
≤ π.
is strictly positive (or equivalently,
is increasing and takes value
π
N ≤ η() :=
r
π − 4 arccos
,
!!
cosh(x)
cosh(x)+1
!!
cosh()
cosh()+1
cosh()
cosh()+1
1
2
is
at zero.
.
So far, we have proved that, given our points x ∈
/ γ and y ∈ γ, there exist at most η()
CAT(-1) points along γ separating y and the projection z of x onto γ such that the balls
of radii centered at these points are all disjoint from the geodesic [x, y]. Therefore,
either [z, y] contains at most η() consecutive CAT(-1) points, so that
d(z, [x, y]) ≤ d(z, y) ≤ (η() + 1) · L,
or there exist more than η() consecutive CAT(-1) points in [z, y], so that, if c ∈ [z, y]
denotes the (η() + 1)-th CAT(-1) point (starting from z to y), then
d(z, [x, y]) ≤ d(z, c) + ≤ η() · L +
since [x, y] must intersect the ball B(c, ). Consequently, the geodesic γ is K-slim where
K = (η() + 1) · L + .
Proof of Theorem 3.1. According to Proposition 3.2, there exists a geodesic ray r and
an increasing sequence (tn ) of nonnegative reals tending to +∞ such that the point r(tn )
has a CAT(-1) neighborhood for every n ≥ 0 and such that the sequence (tn+1 − tn )
is bounded. For every n ≥ 0, let gn ∈ G be an isometry translating r(tn ) into some
fixed compact fundamental domain. Because X is proper, we deduce from Arzelà-Ascoli
theorem that the sequence of rays (gn · rn ) subconverges to a geodesic line γ. Moreover,
since the limit of a sequence of CAT(-1) points must be a CAT(-1) point according to
10
Fact 3.5, we know that γ passes through infinitely many CAT(-1) points in such a way
that the distance between any two consecutive such points is uniformly bounded from
above and from zero. According to Proposition 3.6, this geodesic must be slim, and
finally contracting according to [CS15, Theorem 2.4]. The desired conclusion follows
from Theorem 2.3.
Let us conclude this section with an open question. Define barely CAT(-1) groups as
groups acting geometrically on geodesically complete proper CAT(0) spaces containing
points with CAT(-1) neighborhoods. It would be interesting to determine if barely
CAT(-1) groups satisfy stronger hyperbolic properties than just being acylindrically
hyperbolic. Notice that, in Section 5, an example of a barely CAT(-1) group which is
not relatively hyperbolic is given. A naive question is:
Question 3.13. Must an acylindrically hyperbolic CAT(0) group be barely CAT(-1)?
For instance, when is a right-angled Artin group barely CAT(-1)?
4
Cubical case
In this section, one restricts our attention to CAT(0) cube complexes. It is worth
noticing that, endowed with its usual CAT(0) metric, a CAT(0) cube complex cannot
contain a point admitting a CAT(-1) neighborhood, as each cube is isometric to a (flat)
Euclidean cube. A fortiori, Theorem 3.1 does not apply. A naive attempt to get more
hyperbolicity would be to identify each cube with a fixed hyperbolic cube, but the cube
complex one obtains in that way may be no longer CAT(0). Nevertheless, in [Gro87,
Section 4.2.C], Gromov notices that, by endowing each cube of a given CAT(0) cube
complex with the metric of a hyperbolic cube, the geodesic metric space thus obtained
turns out to be CAT(-1) provided the vertices of our cube complex have no cycles of
length four in their links. This observation motivates our next definition.
Definition 4.1. In a CAT(0) cube complex, a CAT(-1) vertex is a vertex without cycles
of length four in its link.
Now, thanks to Caprace and Sageev’s work on the Rank Rigidity Conjecture [CS11],
we are able to find a cubical analogue of Theorem 3.1. Before stating our result, let us
recall a few definitions.
• If a group G acts on some CAT(0) cube complex X, the action is essential if, for
every halfspace D ⊂ X and every vertex x ∈ X, the orbit G · x does not lie in any
neighborhood of D.
• If a group G acts on some metric space X, the action is non-uniformly weakly
acylindrical if there exists some constant L ≥ 0 such that, for any pair of points
x, y ∈ X at distance at least L apart, the intersection stab(x) ∩ stab(y) is finite.
Notice that, for instance, any proper action is non-uniformly weakly acylindrical. Also,
most of the time, assuming that an action on a CAT(0) cube complex is essential is not
really restrictive; see [CS11, Proposition 3.5].
Proposition 4.2. Let G be a group acting essentially and non-uniformly weakly acylindrically on some finite-dimensional CAT(0) cube complex X. Suppose that one the
following two conditions holds:
• G does not fix a point in X ∪ ∂∞ X;
• or X is locally finite and G acts cocompactly on X.
11
If X contains a CAT(-1) vertex, then G must be either virtually cyclic or acylindrically
hyperbolic.
Proof. If X contains a CAT(-1) vertex, then it cannot split as a Cartesian product of two
cube complexes. Therefore, [CS11, Theorem 6.3] implies that G contains a contracting
isometry, say g ∈ G. As the action G y X is non-uniformly weakly acylindrical, it
follows from [Gen16c, Theorem 3.13] and [Gen16a, Theorem 17] that such an isometry
is necessarily WPD, ie., for every d ≥ 0 and every x ∈ X, there exists some constant
N ≥ 0 such that the set
n
o
h ∈ G | d(x, hx), d g N x, hg N x ≤ d
is finite. It follows from [BBF15, Theorems H and I] that G must be either virtually
cyclic or acylindrically hyperbolic.
As a particular case of the previous proposition, one gets the following statement:
Corollary 4.3. A group acting geometrically and essentially on a CAT(0) cube complex which contains a CAT(-1) vertex must be either virtually cyclic or acylindrically
hyperbolic.
Example 4.4. Let Σg be a compact orientable surface of genus g ≥ 2. We think of Σg
as obtained from a hyperbolic regular right-angled 4g-gon by identifying parallel sides
in the usual way. Each side of this polygon defines a loop in Σg , which we refer to as a
canonical loop. Similarly, we think of the torus Σ1 as obtained from a Euclidean square
by identifying parallel sides in the usual way; and we refer to the image in Σ1 of any side
of our square as a canonical loop. Up to rescaling our metrics, we may suppose without
loss of generality that any canonical loop in any of our surfaces has a fixed length. Now,
let S be a space obtained by gluing compact orientable surfaces along canonical loops
by isometries; let Σ denote the underlying collection of surfaces. The universal cover Se
of S is naturally a polygonal complex. Moreover, if one adds a central vertex to each
polygon which is not a square and if one links this new vertex to half of the vertices of
the corresponding polygon, one obtain a square complex, which turns out to be CAT(0)
provided that the graph underlying our graph of spaces is triangle-free. By noticing
that each of our central vertices are CAT(-1), it follows from Proposition 4.2 that the
fundamental group π1 (S) must be acylindrically hyperbolic, if the collection Σ contains
at least one surface of genus at least two.
5
A non relatively hyperbolic barely CAT(-1) group
Let Σ denote the graph of spaces given by Figure 3. More explicitely, let S be a compact
orientable surface of genus two and T1 , T2 , T3 , T4 , T12 , T23 , T34 seven tori. We denote by
a1 , a2 , a3 , a4 the canonical loops of S (ordered by following a fixed cyclic order on the
boundary of the polygon corresponding to S); by aii , bii the canonical loops of the torus
Ti ; and by bi,i+1
, bi,i+1
i
i+1 the canonical loops of the torus Ti,i+1 . The space Σ is obtained
from these eight surfaces by gluing each Ti to S by identifying ai and aii , and by gluing
i
i
Ti,i+1 to Ti and Ti+1 by identifying bii,i+1 and bi,i+1
i+1 respectively to bi and bi+1 .
It follows from [BH99, Proposition II.11.6] that Σ is locally CAT(0). Moreover, each
subsurface is locally isometrically embedded, so that any point on the surface of genus
two which does not belong to any torus admits a CAT(-1) neighborhood. Next, notice
e of Σ is geodesically complete, as a union of Euclidean and
that the universal cover Σ
e Consequently, the fundamental group G of
hyperbolic planes (which are convex in Σ).
Σ is barely CAT(-1).
12
Figure 3: A non relatively hyperbolic barely CAT(-1) group.
Fact 5.1. The group G is not relatively hyperbolic.
First of all, notice that G admits the following presentation:
*
a1 , a2 , a3 , a4
b1 , b2 , b3 , b4
[a1 , a2 ][a3 , a4 ] = [b1 , b2 ] = [b2 , b3 ] = [b3 , b4 ] = 1
[a1 , b1 ] = [a2 , b2 ] = [a3 , b3 ] = [a4 , b4 ] = 1
+
Suppose that our group G is hyperbolic relatively to some finite collection of subgroups.
As a consequence of [Osi06, Theorems 4.16 and 4.19], every non cyclic free abelian
subgroup of G must be contained into some peripheral subgroup, so, for every 1 ≤ i ≤ 4
(resp. 1 ≤ j ≤ 3), there exists some peripheral subgroup Hi (resp. Kj ) containing both
ai and bi (resp. bj and bj+1 ). Notice that
bi ∈ Hi ∩ Ki for every 1 ≤ i ≤ 4, and bi ∈ Ki ∩ Ki+1 for every 1 ≤ i ≤ 3.
Since the collection of peripheral subgroups must be malnormal according to [Osi06,
Theorems 1.4 and 1.5], it follows that H1 = H2 = H3 = H4 = K1 = K2 = K3 .
Let H denote this common peripheral subgroup. Since H contains all the generators,
necessarily G = H. This proves our fact.
References
[AH16]
G. Arzhantseva and M. Hagen. Acylindrical hyperbolicity of cubical smallcancellation groups. arXiv:1603.05725, 2016.
[BBF15] Mladen Bestvina, Ken Bromberg, and Koji Fujiwara. Constructing group
actions on quasi-trees and applications to mapping class groups. Publications
mathématiques de l’IHÉS, 122(1):1–64, 2015.
[BF02]
M. Bestvina and K. Fujiwara. Bounded cohomology of subgroups of mapping
class groups. Geometry and Topology, 6:69–89, 2002.
13
[BF10]
Mladen Bestvina and Mark Feighn. A hyperbolic Out(Fn )-complex. Groups,
Geometry, and Dynamics, 4(1):31–58, 2010.
[BH99]
Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature, volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
[Bus10]
Peter Buser. Geometry and spectra of compact Riemann surfaces. Springer
Science & Business Media, 2010.
[CM16]
I. Chatterji and A. Martin. A note on the acylindrical hyperbolicity of groups
acting on CAT(0) cube complexes. arXiv:1610.06864, 2016.
[CS11]
Pierre-Emmanuel Caprace and Michah Sageev. Rank rigidity for CAT(0) cube
complexes. Geom. Funct. Anal., 21(4):851–891, 2011.
[CS15]
Ruth Charney and Harold Sultan. Contracting boundaries of CAT(0) spaces.
J. Topol., 8(1):93–117, 2015.
[CW16]
M. Calvez and B. Wiest. Acylindrical hyperbolicity and Artin-Tits groups of
spherical type. arXiv:1606.07778, 2016.
[DGO17] François Dahmani, Vincent Guirardel, and Denis Osin. Hyperbolically embedded subgroups and rotating families in groups acting on hyperbolic spaces,
volume 245. American Mathematical Society, 2017.
[Gen16a] A. Genevois. Acylindrical action on the hyperplanes of a CAT(0) cube complex. arXiv:1610.08759, 2016.
[Gen16b] A. Genevois. Coning-off CAT(0) cube complexes. arXiv:1603.06513, 2016.
[Gen16c] A. Genevois. Contracting isometries of CAT(0) cube complexes and acylindricaly hyperbolicity of diagram groups. arXiv:1610.07791, 2016.
[Gen17a] A. Genevois. Algebraic characterisations of negatively-curved special groups
and applications to graph braid groups. arXiv:1709.01258, 2017.
[Gen17b] A. Genevois. Hyperbolicities in cat(0) cube complexes. arXiv:1709.08843,
2017.
[Gro87]
M. Gromov. Hyperbolic groups. Essays in group theory, 8(75-263):2, 1987.
[Gro93]
M. Gromov. Asymptotic invariants of infinite groups. In Geometric group
theory, Vol. 2 (Sussex, 1991), volume 182 of London Math. Soc. Lecture Note
Ser., pages 1–295. Cambridge Univ. Press, Cambridge, 1993.
[GS14]
D. Gruber and A. Sisto. Infinitely presented graphical small cancellation
groups are acylindrically hyperbolic. arXiv:1408.4488, 2014.
[Hru10]
G Christopher Hruska. Relative hyperbolicity and relative quasiconvexity for
countable groups. Algebraic & Geometric Topology, 10(3):1807–1856, 2010.
[KL95]
M. Kapovich and B. Leeb. On asymptotic cones and quasi-isometry classes
of fundamental groups of 3-manifolds. Geometric and Functional Analysis,
5(3):582–603, 1995.
[Lon15]
A. Lonjou.
Non simplicité du groupe de cremona sur tout corps.
arXiv:1503.03731, 2015.
14
[LS77]
C. Lyndon and P. E. Schupp. Combinatorial group theory. Ergebnisse der
Mathematik und ihrer Grenzgebiete 89. Springer, Berlin, 1977.
[MO15]
Ashot Minasyan and Denis Osin. Acylindrical hyperbolicity of groups acting
on trees. Mathematische Annalen, 362(3):1055–1105, 2015.
[MW02] J. McCammond and D. Wise. Fans and ladders in small cancellation theory.
Proc. London Math. Soc., 84(3):599–644, 2002.
[Osi06]
D. Osin. Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems. Mem. Amer. Math. Soc., 179(843):vi–100,
2006.
[Osi15]
D. Osin. On acylindrical hyperbolicity of groups with positive first `2 -Betti
number. Bulletin of the London Mathematical Society, 47(5):725, 2015.
[Osi16]
D. Osin. Acylindrically hyperbolic groups. Trans. Amer. Math. Soc., 368:851–
888, 2016.
15
| 4math.GR
|
arXiv:1702.03292v4 [math.AC] 19 Oct 2017
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
Abstract. In this paper we recall the object sectional matrix
which encodes the Hilbert functions of successive hyperplane sections of a homogeneous ideal. We translate and/or reprove recent
results in this language. Moreover, some new results are shown
about their maximal growth, in particular a new generalization
of Gotzmann’s Persistence Theorem, the presence of a GCD for a
truncation of the ideal, and applications to saturated ideals.
1. Introduction
Let K be a field of characteristic 0 and let P be the polynomial
ring K[x1 , . . . , xn ] with n indeterminates and the standard grading.
Given a finitely generated Z-graded P -module M = ⊕d∈N Md , the Md ’s
are finite-dimensional K-vector spaces. The Hilbert function of M,
HM : Z → N with HM (d) := dimK (Md ), is a very frequent and
powerful tool of investigation in Commutative Algebra.
From Macaulay [12], it is well known that the computation of the
Hilbert function of M may be reduced to the computation of the Hilbert
function of some K-algebras of type P/J, where J is a monomial ideal.
In particular, if I is a homogeneous ideal in P , then HP/I = HP/LT(I) .
Another very common practice consists of studying generic hyperplane sections which, in algebraic terms, means reducing modulo by a
generic linear form. The combination of Hilbert functions and hyperplane sections lead to the result by Green [10] (1988).
The sectional matrix of a homogenous ideal I was introduced by Bigatti and Robbiano in [6] (1997) unifying the concepts of the Hilbert
function of a homogeneous ideal I (along the rows) and of its hyperplane sections (along the columns). Sectional matrices did not receive
much attention, and in this paper we want to revive them. We extend
some results in this language, confirming the merit of this tool and
Date: October 20, 2017.
2010 Mathematics Subject Classification. Primary 13D40, 05E40, Secondary
13P99 .
Key words and phrases. Sectional Matrix, Hilbert Function, Hyperplane Section,
Generic Initial Ideal, Reduction Number, Extremal Behaviour.
1
2
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
suggesting that further investigation might cast a new light on many
aspects of Commutative Algebra.
In Section 2 we set our notation and recall the definition of sectional
matrix. In Sections 3 we recall its main properties converting the results
from [6] into terms of the quotient P/I (instead of the ideal I). In
particular, Theorem 3.11 shows Macaulay’s and Green’s inequalities. In
Section 4, we recall Gotzmann’s Persistence Theorem and the sectional
matrix analogue from [6]. Moreover, we generalize it into a sectional
version (Theorem 4.6).
In Section 5, we describe how to deduce information on the dimension and the degree of a homogeneous ideal in terms of the entries of its
sectional matrix. In Section 6, we show how the extremal behaviour implies the presence of a GCD for the truncation of homogeneous ideals.
In Section 7, we apply these results to the class of saturated ideals.
Finally, in Section 8, we present several examples comparing the information given by the sectional matrix, the generic initial ideal, and the
resolution of a homogeneous ideals.
The examples in this paper have been computed with CoCoA ([1],
[2], SectionalMatrix, PrintSectionalMatrix).
2. Definitions and notation
Let K be a field of characteristic 0 and P = K[x1 , . . . , xn ] be the
polynomial ring with n indeterminates with the standard grading. Let
I ⊆ P be a homogeneous ideal in P and A = P/I. Then A is a graded
P -module ⊕d∈N Ad , where Ad = Pd /Id .
The definition of the Hilbert function was extended in [6] to the
bivariate function encoding the Hilbert functions of successive generic
hyperplane sections: the sectional matrix of a homogeneous ideal
I in P . In this paper, we define, in the obvious way, the sectional
matrix for the quotient algebra P/I, and then we show how to adapt
the results given in [6] to the use of P/I.
Definition 2.1. Given a homogeneous ideal I in P = K[x1 , . . . , xn ],
we define the sectional matrix of I and of P/I to be the functions
{1, . . . , n} × N −→ N
MI (i, d) = dimK ((I + (L1 , . . . , Ln−i ))/(L1 , . . . , Ln−i ))d ,
MP/I (i, d) = dimK (Pd /(I + (L1 , . . . , Ln−i ))d ),
where L1 , . . . , Ln−i are generic linear forms. Notice that MP/I (n, d) =
HP/I (d) and MP/I (i, d) = d+i−1
− MI (i, d).
i−1
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
3
Remark 2.2. A generic linear form is a polynomial L = a1 x1 + · · · +
an xn in K(a1 , . . . , an )[x1 , ..., xn ]. In this paper we restrict our attention
to a field K of characteristic 0, so the equalities of the Definition 2.1
hold for any L′ = α1 x1 + · · · + αn xn with (α1 , . . . , αn ) in a non-empty
Zariski-open set in PnK . Therefore in this case it is common practice
to talk about “generic linear forms in K[x1 , . . . , xn ]” instead of dealing
with the explicit extension of K.
This small example will be used as a running example throughout
the paper.
Example 2.3. Let P = Q[x, y, z] and I = (x4 − y 2 z 2 , xy 2 − yz 2 − z 3 )
an ideal of P . Then the sectional matrix of P/I is
...
HP/(I+hL1 ,L2 i) (d) = MP/I (1, d) : 1 1 1 0 0 0 0 0 . . .
HP/(I+hL1 i) (d) = MP/I (2, d) : 1 2 3 3 2 1 0 0 . . .
HP/I (d) = MP/I (3, d) : 1 3 6 9 11 12 12 12 . . .
0
1
2
3
4
5
6
7
where the continuations of the lines are obvious in this example. The
general theory about truncation and continuation of the lines will be
described in Theorem 4.1 and Remark 4.3.
3. Background results on sectional matrices
In this section we recall the main properties of sectional matrices
from [6] translating them in terms of the quotient P/I. In particular,
we describe the persistence theorem and the connection with rgin.
Let σ be a term-ordering on P = K[x1 , . . . , xn ]. The leading term
ideal or initial ideal of an ideal I ⊆ P is the ideal, denoted LTσ (I),
generated by {LTσ (f ) | f ∈ I\{0} }. For any homogenous ideal I it is
well known that HP/I = HP/LTσ (I) . This nice property does not extend
to MP/I , but only one inequality holds, as Conca proved in [7]: we
write his result in sectional matrix notation.
Theorem 3.1 (Conca, 2003). Let I be a homogeneous ideal in P =
K[x1 , . . . , xn ] and σ a term-ordering. Then, for all i = 1, . . . , n and
d ∈ N, MP/I (i, d) ≤ MP/LTσ (I) (i, d)
Example 3.2. Recall I from Example 2.3. For σ =DegRevLex compare MP/I with MP/LTσ (I) : the σ-Gröbner basis of I is {xy 2 −yz 2 −z 3 ,
x4 −y 2 z 2 , x3 yz 2 −y 4 z 2 +x3 z 3 , y 5 z 2 −y 4 z 3 +x3 z 4 −x2 yz 4 −x2 z 5 }, thus we
4
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
have LTσ (I) = (xy 2, x4 , x3 yz 2 , y 5z 2 ).
...
0 1 2 3
4
5
6
7
MP/LTσ (I) (1, d) : 1 1 1 0 0 0 0 0 . . .
MP/LTσ (I) (2, d) : 1 2 3 3 2 1 1 0 . . .
MP/LTσ (I) (3, d) : 1 3 6 9 11 12 12 12 . . .
Then we observe that MP/I (2, 6) = 0 < 1 = MP/LTσ (I) (2, 6). Notice
that the third lines are equal for all term-orderings because the Hilbert
functions are the same: HP/I = HP/LTσ (I) .
In this paper we compare some of our results with the ones from
[4]. In order to make the comparison clearer to the reader we need to
introduce the notion of s-reduction number and to describe how it is
stated in terms of the sectional matrix. The definition of s-reduction
number has several equivalent formulations and we recall here the one
given in [4].
Definition 3.3. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ].
The s-reduction number, rs (P/I), is max{d | HP/(I+(L1 ,...,Ls )) (d)6=0},
where L1 , . . . , Ls are generic linear forms in P . In our language
rs (P/I) = max{d | MP/I (n−s, d)6=0}.
The reduction number r(P/I) is rdim(P/I) (P/I).
Notice that, for their definitions, the reduction number and the sectional matrix, use “complementary” indices s and n − s.
Example 3.4. In Example 2.3 and Example 3.2 we see that I and
LTσ (I) have the same 2-reduction number, r2 (P/I) = r2 (P/LTσ (I)) =
2, and different 1-reduction number: r1 (P/I) = 5 and r1 (P/LTσ (I)) =
6. From the equalities dim(P/I) = dim(P/LTσ (I)) = 1 it follows that
r(P/I)=5 and r(P/LTσ (I))=6.
Remark 3.5. Using Theorem 3.1 Conca in [7] proved the inequality
for the reduction numbers: r(P/I) ≤ r(P/LTσ (I)).
Going back to the problem of finding a monomial ideal with the same
sectional matrix as P/I, we recall the definitions of strongly stable ideal
and of gin. A monomial ideal J is said to be strongly stable if for
every power-product t ∈ J and every i, j such that i < j and xj |t, the
power-product xi ·t/xj is in J.
In [8] Galligo proved that, given a homogeneous ideal I in the polynomial ring K[x1 , . . . , xn ], with K a field of characteristic 0 and σ a
term-ordering such that x1 >σ x2 >σ · · · >σ xn , then there exists a nonempty Zariski-open set U ⊆ GL(n) and a strongly stable ideal J such
that for each g ∈ U, LTσ (g(I)) = J. This ideal is called the generic
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
5
initial ideal of I with respect to σ and it is denoted by ginσ (I).
In particular, when σ =DegRevLex, it is denoted by rgin(I).
Example 3.6. Consider the ideal I = (x4 − y 2 z 2 , xy 2 − yz 2 − z 3 ) from
Example 2.3. Then rgin(I) = (x3 , x2 y 2 , xy 4, y 6 ). See [3] for details
about the computation of gin in CoCoA.
Remark 3.7. Let I be a homogeneous ideal. If I has a minimal generator of degree d (then so does g(I)), then also rgin(I) has a minimal generator of degree d. The converse is not true in general: consider for example the ideal I = (z 5 , xyz 3 ) in Q[x, y, z], then rgin(I) =
(x5 , x4 y, x3 y 3) has a minimal generator of degree 6, and I doesn’t. In
particular, this shows that the highest degree of a minimal generating
set of rgin(I) may be strictly greater than that of I.
The following result represents the sectional matrix analogue of Macaulay’s
Theorem for Hilbert functions (HP/I = HP/LT(I) ): it reduces the study
of the sectional matrix of a homogeneous ideal to the combinatorial
behaviour of a monomial ideal.
Lemma 3.8. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ]. Then
MP/I (i, d) = MP/rgin(I) (i, d) = dimK (Pd /(rgin(I) + (xi+1 , . . . , xn ))d ).
Proof. See Lemma 5.5 in [6]. ⊓
⊔
Remark 3.9. Lemma 3.8 shows that when we have a strongly stable
ideal J in P (and in particular rgin(I) is strongly stable) the sectional
matrix of P/J is particularly easy to compute because sectioning J
by n−i generic linear forms is the same as sectioning J by the smallest n−i indeterminates, xi+1 , . . . , xn .
Using this combinatorial view, Bigatti and Robbiano in [6] proved
a combination of Macaulay’s and Green’s inequalities ([12], [10]) and
then an analogue of Gotzmann’s Persistence Theorem ([9]) for sectional
matrices.
We recall the definition of binomial expansion following the notation
of [6]. If I is a homogeneous ideal then the (n − 1)-binomial expansion
of HI (d) corresponds to a “description” of a lex-segment ideal L in degree d, and similarly the d-binomial expansion of HP/I (d) corresponds
to P/L. See for example [11, Proposition 5.5.13].
Definition 3.10. For h and i, two positive
integers,
we define
the
h(i)
h(i−1)
h(j)
i-binomial expansion of h as h = i + i−1 + · · · + j with
h(i) > h(i − 1) > · · · > h(j) ≥ j ≥ 1. Such expression exists and is
unique.
6
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
Moreover we define a family of functions related
to the expansion
in
h(i)+s
h(i−1)+s
h(j)+s
s
the following way: (hi )t := i+t + i−1+t + · · · + j+t .
1
−
For short, we will write (hi )+
instead
+ instead of (hi )1 , and (hi )
−1
of (hi )0 .
Here is Theorem 5.6 of [6] and again we convert the statement in
terms of the quotient P/I using the properties of the functions derived
from the binomial expansion.
Theorem 3.11 (Sectional matrices, 1997). Let I be a homogeneous
ideal in the polynomial ring P = K[x1 , . . . , xn ] and M := MP/I . Then
i
P
(a) M(i, d + 1) ≤
M(j, d) for all i = 1, . . . , n and d ∈ N.
j=1
(b) (Macaulay) M(i, d+1) ≤ (M(i, d)d )+
+ for all i = 1, . . . , n and d ∈ N.
(c) M(i − 1, d) − M(i − 2, d) ≤ ((M(i, d) − M(i − 1, d))d)− for all i =
3, . . . , n and d ∈ N.
(d) (Green) M(i − 1, d) ≤ (M(i, d)d )− for all i = 2, . . . , n and d ∈ N.
Proof. This is Theorem 5.6 of [6], using the conversions
(HI (d)n−1 )+ = (HP/I (d)d )+
+
and
−
(HI (d)n−1 )−
− = (HP/I (d)d )
(see for example [11] Proposition 5.5.16 and Proposition 5.5.18). ⊔
⊓
For the extremal case of Bigatti, Geramita and Migliore in [5], Macaulay’s
inequality defined the maximal growth of the Hilbert function. We
analogously define maximal growth of the sectional matrix following
the extremal case in Theorem 3.11.a.
Definition 3.12. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ].
• The Hilbert function HP/I has maximal growth in degree d
if “Macaulay’s equality” holds: HP/I (d + 1) = (HP/I (d)d )+
+.
• The sectional matrix MP/I has i-maximal growth in degree d if “Bigatti-Robbiano’s equality” holds: MP/I (i, d + 1) =
i
P
MP/I (j, d).
j=1
Remark 3.13. For a homogeneous ideal I in P = K[x1 , . . . , xn ] if
MP/I has n-maximal growth in degree d then I, and rgin(I), have no
minimal generators of degree d+1.
More precisely, for any i ∈ {1, . . . , n}, Corollary 2.7 of [6] implies that
MP/I has i-maximal growth in degree d if and only if rgin(I) has no
minimal generators of degree d+1 in x1 , . . . , xi . (This is a generalization
of Lemma 2.17 in [4].)
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
7
4. Sectional Persistence Theorem
Gotzmann’s Persistence Theorem [9] says that, if the generators of
an ideal I have degree ≤ δ and the Hilbert function of P/I has maximal
growth in degree δ, then it has maximal growth for all higher degrees.
This is also true for sectional matrices. Here we recall the Persistence
Theorem 5.8 of [6], and in Theorem 4.6 we will generalize it for imaximal growth, for i ≤ n.
Theorem 4.1 (Persistence Theorem, 1997).
Let I be a homogeneous ideal in P = K[x1 , . . . , xn ]. If MP/I has
n-maximal growth in degree δ and I has no generators of degree >δ
then it has i-maximal growth for all i = 1, . . . , n and for all degrees
>δ.
Moreover, MP/I has n-maximal growth for all degrees ≥ reg(I).
Example 4.2. Consider the polynomial ring P = Q[x, y, z, t] and the
ideal I = (x4 − x2 yz, x5 + xy 3 z) of P . Then
0
MP/I
1
= 1
1
1
1
2
3
4
5
6
7
1 1 1 0 0 0 0
2 3 4 4 3 2 1
3 6 10 14 17 19 20
4 10 20 34 51 70 90
8
0
1
21
111
...
...
...
...
...
Therefore MP/I has 4-maximal growth starting from degree 7, whereas
a direct computation shows that HP/I has maximal growth starting
from degree 49.
Remark 4.3. In particular, the regularity is used by CoCoA for truncating the size of the sectional matrix, displaying the rows up to degree
reg(I)+1, so that the last column shows the persisting equalities. In
Example 2.3, we have rgin(I) = (x3 , x2 y 2 , xy 4, y 6 ), thus reg(I) = 6,
and in degree 7 we read the persisting equalities.
Remark 4.4. We emphasize that the regularity of a homogeneous
ideal I, the highest degree of the generators of rgin(I), is usually a
much lower number than the highest degree of the generators of the
lex-segment ideal with the same Hilbert function of I, as shown in
Example 4.2. This fact makes the persistence in Theorem 4.1 more
“practical” than Gotzmann’s.
With the next lemma we show that if MP/I has i-maximal growth
in degree δ for some i < n, this persists in higher degrees, even if it
does not have n-maximal growth.
8
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
Lemma 4.5. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ] generated in degree ≤ δ + 1. If there exists i ≤ n such that rgin(I) has no
minimal generators of degree δ + 1 in P(i) = K[x1 , . . . , xi ], then rgin(I)
has no minimal generators of any degree > δ in P(i) .
Proof. Let σ be DegRevLex and g a generic change of coordinates.
Suppose that the σ-Gröbner basis of g(I) has a polynomial f2 of degree
δ + 2. Then f2 comes from a minimal syzygy of rgin(I) = LTσ (g(I))
and hence this syzygy is linear (see Lemma 5.7 in [6]). This means
that there exists a minimal generator t1 of rgin(I) of degree δ + 1, and,
by the hypothesis, all minimal generators of rgin(I) must be in the
ideal (xi+1 , . . . , xn ). Let f1 be the Gröbner basis polynomial such that
t1 = LTσ (f1 ). Because we are using the reverse lexicographic termordering, this fact implies that f1 ∈ (xi+1 , . . . , xn ). As a consequence
any s-polynomial constructed with f1 is a difference of polynomials in
(xi+1 , . . . , xn ), so f2 ∈ (xi+1 , . . . , xn ). Thus any Gröbner basis element
of degree δ +2 is in (xi+1 , . . . , xn ), and therefore rgin(I) has no minimal
generators of degree δ + 2 in K[x1 , . . . , xi ]. Iterating this reasoning, we
can conclude that rgin(I) has no minimal generators of degree > δ in
K[x1 , . . . , xi ]. ⊓
⊔
Using Lemma 4.5, we can now extend Theorem 4.1.
Theorem 4.6 (Sectional Persistence Theorem).
Let I be a homogeneous ideal in P = K[x1 , . . . , xn ] generated in degree ≤ δ+1. If there exists i ∈ {1, . . . , n} such that MP/I has i-maximal
growth in degree δ then it has j-maximal growth for all j ∈ {1, . . . , i}
and for all degrees ≥ δ.
P
Proof. From MP/I (i, δ + 1) = ij=1 MP/I (j, δ) we have that rgin(I)
has no generators of degree δ + 1 in K[x1 , . . . , xi ] (see Remark 3.13).
Then, applying Lemma 4.5, we know that rgin(I) has no generators
of degree > δ in K[x1 , . . . , xi ]. Hence we can apply Theorem 4.1
to J(i) = rgin(I) ∩ K[x1 , . . . , xi ] and for all d > δ and j = 1, . . . , i
Pj
we get MP/rgin(I) (i, d+1) = MP/J(i) (j, d+1) =
k=1 MP/J(i) (k, d) =
Pj
⊓
k=1 MP/rgin(I) (k, d). The conclusion now follows from Lemma 3.8. ⊔
The Sectional Persistence Theorem says that, whereas the persistence of the n-th row starts at the regularity of the ideal, the persistence in the first rows may be detected in degree lower than the highest
degree of the generators.
Example 4.7. Consider the polynomial ring P = Q[x, y, z, w] and the
ideal I = (x2 , xy, xz(z + w), x(z 2 + w 2 )) of P . Then rgin(I) =
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
9
(x2 , xy, xz 2 , xzw, xw 3 ). Notice that I is generated in degree ≤ 3, and
its regularity is 4, so from Theorem 4.1 it follows that MP/I has imaximal growth in degree 4 for i = 1, . . . , 4.
MP/I
0
1
2
1
= 1
1
1
1
2
3
4
0 0 0 0
1 1 1 1
4 4 5 6
8 11 15 21
3
4
5
...
...
...
...
...
We see that MP/I has 2-maximal growth in degree 2 and 3-maximal
growth in degree 3:
MP/I (2, 3) = 1 = MP/I (1, 2) + MP/I (2, 2)
MP/I (3, 4) = 5 = MP/I (1, 3) + MP/I (2, 3) + MP/I (3, 3).
Hence from Theorem 4.6 it follows that MP/I has 2-maximal growth
for all degrees ≥ 2 and it has 3-maximal growth for all degrees ≥ 3.
5. Hilbert Polynomial, Hilbert Series, dimension,
multiplicity
In this section, we show how to read some algebraic invariants for a
homogeneous ideal I from its sectional matrix MP/I truncated at some
degree δ.
Lemma 5.1. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ] generated in degree ≤ δ + 1. If MP/I has i-maximal growth in degree δ,
then for all d ∈ N>0 we have
P
MP/I (i, δ+d) = ij=1 i−j+d−1
· MP/I (j, δ)
i−j
Proof. We prove the statement by induction on d. From Theorem 4.6
it follows that MP/I has k-maximal growth in degree δ, for all k =
P
1, . . . , i, thus for d = 1 MP/I (k, δ+1) = kj=1 MP/I (j, δ).
Let d>1 and suppose the stated equality holds for MP/I (k, δ + d)
for all k = 1, . . . , i. Then by the i-maximal growth from Theorem 4.6,
i X
i
k
X
X
k − j+d−1
MP/I (k, δ+d) =
MP/I (i, δ+d+1) =
·MP/I (j, δ)
k
−
j
j=1
k=1
k=1
then we swap the sums varying j in {1, . . . , i} and t = k − j with
k ∈ {j, .
. . , i}:
Pi
Pi−j t+d−1
P
· MP/I (j, δ) = ij=1 i−j+d−1
· MP/I (j, δ)
j=1
t=0
t
i−j
and this concludes the proof. ⊓
⊔
Proposition 5.2. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ]
generated in degree ≤ δ + 1, and let L1 , . . . , Ln be generic linear forms
10
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
in P . If MP/I has i-maximal growth in degree δ, then the Hilbert
polynomial of P/(I + (L1 , . . . , Ln−i )) is
i
X
i−j+x−δ−1
· MP/I (j, δ)
pi (x) =
i−j
j=1
In particular, if pi 6= 0 let k = min{j ∈ {1, . . . , i} | MP/I (j, δ) 6= 0},
then
pi (x) =
MP/I (k, δ) i−k
x + ... terms of lower degree .
(i − k)!
Proof. The first part of the corollary
trivial from Lemma 5.1: for
Pis
i
x > δ we have pi (x) = MP/I (i, x) = j=1 i−j+(x−δ)−1
· MP/I (j, δ).
i−j
(x−δ+i−k−1)...(x−δ+1)(x−δ)
i−k+x−δ−1
We conclude by observing that
=
(i−k)!
i−k
1
i−k
and therefore equal to (i−k)! x + (terms of lower degree). ⊔
⊓
Proposition 5.3. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ]
generated in degree ≤ δ + 1, and let L1 , . . . , Ln be generic linear forms
in P . If MP/I has i-maximal growth in degree δ, then the Hilbert series
of Ri = P/(I + (L1 , .., Ln−i )) is
X
i
δ
X
MP/I (j, δ) δ+1
d
MP/I (i, d)t +
HSRi (t) =
t .
i−j+1
(1
−
t)
j=1
d=0
P
d
Proof. By definition HSRi (t) = ∞
d=0 MP/I (i, d)t . By Lemma 5.1, we
have that
i
∞
∞ X
X
X
i − j+d − δ − 1
d
· MP/I (j, δ) td
MP/I (i, d)t =
i
−
j
j=1
d=δ+1
d=δ+1
swapping the sums and letting k = d − δ − 1 it becomes
Pi P∞ i−j+k k
Pi MP /I (j,δ) δ+1
δ+1
= j=1
.
=
t · MP/I (j, δ) · t
k=0
j=1 (1−t)i−j+1 t
i−j
Therefore,
we can conclude by adding the first part of the series,
Pδ
d
⊔
d=0 MP/I (i, d)t . ⊓
The following theorem shows that we can easily read the dimension
and the multiplicity of P/I from its sectional matrix. In particular,
this information may be found in the δ-th column with δ < reg(I) (see
Example 5.6).
Theorem 5.4. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ] generated in degree ≤ δ+1 such that Iδ 6=Pδ and let i = min{j | MP/I (j, δ)6=0}.
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
11
If MP/I (i, δ) = MP/I (i, δ + 1), then
dim(P/I) = n−i+1 and
deg(P/I) = MP/I (i, δ).
Proof. The hypothesis implies that MP/I has i-maximal growth in degree δ, so, by Theorem 4.6, has j-maximal growth in degree d, for all
d > δ and j = 1, . . . , i; this means that i = minj {MP/I (j, d) 6= 0}
for all d > δ and MP/I (i, d) = MP/I (i, δ) for all d ≥ δ. Now, let
δ ′ ≥ δ such that MP/I has n-maximal growth in degree δ ′ , for example
δ ′ = max{δ, reg(I)}. Applying Proposition 5.3, and setting the highest
power, (1 − t)n−i+1 , as common denominator, it follows that
MP/I (i, δ ′ ) + f (t)(1 − t)
HSP/I (t) =
(1 − t)n−i+1
for some polynomial f (t) ∈ K[t] and the fraction above is reduced.
Therefore, the degree of its denominator, n−i+1, is dim(P/I), and the
evaluation of the numerator in 1, MP/I (i, δ), is deg(P/I). ⊔
⊓
Remark 5.5. The hypothesis of Theorem 5.4 is equivalent to the
existence of an integer i such that MP/I (i, δ) = MP/I (i, δ + 1) for
δ > rn−i+1 (P/I): this kind of formulation should look more familiar to
the readers of [5] and [4].
Example 5.6. Following Example 4.7 we consider i = 2, δ = 2 and
then MP/I (1, 2) = 0 and MP/I (2, 2) = MP/I (2, 3) = 1 6= 0. We
then conclude dim(P/I) = n−i+1 = 4 − 2 + 1 = 3 and deg(P/I) =
MP/I (i, δ) = MP/I (2, 2) = 1.
Note that we deduced this information from the sectional matrix in
degree 2, strictly smaller than reg(I) = 4 and also smaller than 3, the
maximal degree of the generators of I.
Remark 5.7. For any homogeneous ideal I in P = K[x1 , . . . , xn ],
MP/I has i-maximal growth for all degrees ≥ reg(I) and for all i ∈
{1, . . . , n}. Therefore all the results in this section hold replacing their
hypotheses with “δ ≥ reg(I)”.
Example 5.8. In Example 2.3, with reg(I) = 6, we have i = 3 so
dim(P/I) = 3 − 3 + 1 = 1 and deg(P/I) = MP/I (3, 6) = 12,
6. Maximal Growth and Greatest Common Divisor
Now we make a subtle change: we consider the same scenario in
degree δ for an arbitrary homogeneous ideal I and see that the same
conclusion hold on the truncation hI≤δ i.
12
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
Proposition 6.1. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ].
If there exists δ such that Iδ 6=Pδ and MP/I (i, δ) = MP/I (i, δ+1), for
i = min{j>1 | MP/I (j, δ) 6= 0}, then we have
dim(P/hI≤δ i) = dim(P/hI≤δ+1 i) = n−i+1,
deg(P/hI≤δ i) = deg(P/hI≤δ+1 i) = MP/I (i, δ).
Proof. By construction MP/I (j, δ) = MP/hI≤δ i (j, δ) and MP/I (j, δ+1) =
MP/hI≤δ+1 i (j, δ+1) for all j = 1, . . . , n. From Remark 3.13 it follows
that rgin(I) has no minimal generators in degree δ+1 in x1 , . . . , xi , and
therefore nor does rgin(hI≤δ i) ⊆ rgin(I). It then follows that
MP/I (j, δ+1) = MP/hI≤δ i (j, δ+1)
for all j = 1, . . . , i. This implies that
MP/hI≤δ i (i, δ+1) = MP/hI≤δ i (i, δ) 6= 0,
MP/hI≤δ+1 i (i, δ+1) = MP/hI≤δ+1 i (i, δ) 6= 0,
and i = min{j | MP/hI≤δ i (j, δ) 6= 0} = min{j | MP/hI≤δ+1 i (j, δ) 6= 0}.
Now we can apply Theorem 5.4 and get the conclusions. ⊔
⊓
Corollary 6.2. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ].
If there exists δ such that Iδ 6=Pδ and MP/I has n-maximal growth in
degree δ, let i = min{j>1 | MP/I (j, δ)6=0}, then hI≤δ i = hI≤δ+1 i,
dim(P/hI≤δ i) = n−i+1, and deg(P/hI≤δ i) = MP/I (i, δ).
Proof. By hypothesis MP/I has n-maximal growth in degree δ, hence
by Remark 3.13 it follows that I has no minimal generators in degree δ+1, and therefore hI≤δ i = hI≤δ+1 i. Moreover, from Theorem 4.1
it has i-maximalP
growth in degree δ. Hence we have the equality
i
MP/I (i, δ+1) =
j=1 MP/I (j, δ) = MP/I (i, δ) 6= 0 and the conclusion follows from Proposition 6.1. ⊓
⊔
Example 6.3. Consider the polynomial ring P = Q[x, y, z, t] and the
ideal I = (x3 , x2 y, xy 2, xyz 2 , xyzt3 ) of P . Then
0
1
1
MP/I = 1
1
1
1
2
3
4
0
MP/I≤3
1
= 1
1
1
1
2
3
...
1 0 0 0 0 0 ...
3 1 1 1 1 1 ...
6 7 7 8 9 10 . . .
10 17 24 32 40 50 . . .
2
4
1 1 0 0
2 3 1 1
3 6 7 8
4 10 17 25
3
4
5
6
7
...
0 1
2
3
4
5
1 1 1 0 0 0
...
. . . MP/I≤4 = 1 2 3 1 1 1
1 3 6 7 7 8
...
1 4 10 17 24 32
...
...
...
...
...
...
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
13
We see that MP/I has i = 2-maximal growth in degree δ = 3. Then
by Proposition 6.1, we have that dim(P/hI≤3 i) = dim(P/hI≤4 i) = 3,
and deg(P/hI≤3i) = deg(P/hI≤4 i) = 1, regardless what happens in
MP/I (j, δ) for j > i = 2.
The following proposition is the generalization of Proposition 1.6 of
[5].
Proposition 6.4. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ].
If MP/I has n-maximal growth in degree δ, then reg(hI≤δ i) ≤ δ.
Proof. Let I = hI≤δ i. By construction MP/I (j, d) = MP/I (j, d) for
all j = 1, . . . , n and 0 ≤ d ≤ δ. By Remark 3.13, I has no minimal
generators in degree δ+1, and hence MP/I (j, δ+1) = MP/I (j, δ+1) for
all j = 1, . . . , n. This implies that MP/I has n-maximal growth in
degree δ, and then, by the Persistence Theorem 4.1, it has n-maximal
growth in all degrees > δ. By Lemma 3.8, MP/I = MP/rgin(I) and
hence by Lemma 4.5, rgin(I) has no minimal generators of degree > δ,
and then reg(I) ≤ δ. ⊓
⊔
In the rest of this section, we generalize some results of [5] and [4]
about the existence a common factor when there is a certain kind of
maximal growth.
Corollary 6.5. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ].
If there exists δ such that Iδ 6= {0} and MP/I (2, δ) = MP/I (2, δ+1)
(i.e. has 2-maximal growth in degree δ) then hI≤δ i has a GCD of degree
MP/I (2, δ). Furthermore, hI≤δ+1 i shares the same GCD.
Proof. From Iδ 6={0} it follows that xδ1 ∈ rgin(I), and then MP/I (1, δ) =
0. If MP/I (2, δ) = 0 then Iδ = Pδ has GCD = 1 of degree 0. Otherwise, by Proposition 6.1 dim(P/hI≤δ i) = dim(P/hI≤δ +1 i) = n−1 and
deg(P/hI≤δ i) = deg(P/hI≤δ+1 i) = k = MP/I (2, δ). This means that
hI≤δ i defines a hypersurface of degree k, i.e. hI≤δ i = (F ) ∩ J with
dim(J) < n−1 and deg(F ) = k. Therefore hI≤δ i ⊆ (F ) as claimed.
Similarly for hI≤δ+1 i. ⊓
⊔
Following the statement of Corollary 6.5, and along the line of ideas
in [5], we give a new definition for the potential GCD, based on the
sectional matrix instead of the Hilbert function.
Definition 6.6. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ]
such that Iδ 6= {0}. The M-potential degree of the GCD of Iδ is
k = MP/I (2, δ).
The following corollary is the generalization of Proposition 2.7 of [5]
and of Corollary 5.2 of [4].
14
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
Corollary 6.7. Let I be a homogeneous ideal in P = K[x1 , . . . , xn ]
and let δ be such that Iδ 6= {0}. Let k be the M-potential degree of the
GCD of Iδ . If MP/I has i-maximal growth in degree δ for some i ≥ 2,
then hI≤δ i and hI≤δ+1 i share the same GCD, F , of degree k.
Proof. By Theorem 4.6, if MP/I has i-maximal growth in degree δ,
then it has 2-maximal growth in degree δ. Therefore we conclude by
Corollary 6.5. ⊓
⊔
Let us see this result in action on an example.
Example 6.8. Consider the polynomial ring P = Q[x, y, z] and the
ideal I = (x3 + y 3 , x2 + 3xy + 2y 2 − xz − yz, x4 + x3 y, xy 4 − 16xyz 3 ,
y 5 − 3xy 3 z − 4y 4 z + 12xyz 3 − 25y 3z 2 + 100yz 4 ) of P . Then
...
1 1 0 0 0 0 0 0 0 0 ...
=
1 2 2 1 1 0 0 0 0 0 ...
1 3 5 6 6 4 3 2 1 1 ...
0
MP/I
1
2
3
4
5
6
7
8
9
We see that MP/I has 2-maximal growth in degree 3 and indeed both
I3 and I4 have a GCD of degree k = MP/I (2, 3) = MP/I (2, 4) = 1.
Indeed, a direct computation shows that the GCD is x + y.
7. Saturated ideals
A homogeneous ideal I in P =K[x1 , . . . , xn ] is saturated if the irrelevant maximal ideal m=(x1 , . . . , xn ) is not an associated prime ideal, i.e.
(I : m) = I. For any homogeneous ideal I of P , the saturation of I, denoted I sat , is defined by I sat := {f ∈ P | f mℓ ⊆ I for some integer ℓ}.
In this section we apply the results obtained previously to the case
of saturated ideals.
Remark 7.1. It is well known that for any homogeneous ideal J there
exists ℓ ∈ N such that J sat = J : mℓ and therefore Jd = (J sat )d for all
d ≫ 0.
Remark 7.2. (Bayer-Stillman) Let I be a homogeneous ideal of P .
Then rgin(I sat ) = rgin(I)xn →0 . This shows that if I is saturated, then
rgin(I) has no minimal generators involving xn .
Lemma 7.3. Let I be a saturated ideal in P = K[x1 , . . . , xn ]. Then
MP/I has n-maximal growth in degree δ if and only if it has (n−1)maximal growth in degree δ.
Proof. By Theorem 3.11.(a) n-maximal growth implies (n−1)-maximal
growth.
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
15
Suppose now MP/I has (n−1)-maximal growth. By Remark 3.9 this
implies that rgin(I) has no minimal generators in x1 , . . . , xn−1 in degree
δ+1. Moreover, since rgin(I) is saturated, from Remark 7.2 there are
no minimal generators divisible by xn . With no minimal generators in
degree δ+1 MP/rgin(I) , and therefore MP/I , has also n-maximal growth.
⊔
⊓
In general the truncation of a saturated ideal is not saturated, as
the following example shows. To guarantee that also the truncation is
saturated we need some additional hypothesis, see Lemma 7.5.
Example 7.4. Consider the polynomial ring P = Q[x, y, z, t] and the
ideal I = (yz − xt, z 3 − yt2 , xz 2 − y 2 t, y 3 − x2 z, x3 , x2 y 2). This ideal
is saturated, however, a direct computation shows that the truncation
I≤3 is not saturated.
The following lemma is the generalization of Lemma 1.4 of [5].
Lemma 7.5. Let I be a saturated ideal in P = K[x1 , . . . , xn ]. If MP/I
has (n−1)-maximal growth in degree δ then the ideal hI≤δ i is saturated.
Proof. From Lemma 7.3 it follows that MP/I has n-maximal growth
in degree δ.
sat
Let I = hI≤δ i and Ie = I . Notice that we have that I d ⊆ Ied for
all d ∈ N. We want to prove that our hypotheses imply I d = Ied for all
d ∈ N.
By Remark 7.1 we have that I d = Ied for all d ≫ 0.
Let f ∈ Ied be an element with d ≤ δ. Then f mℓ ⊆ I for some
integer ℓ. Since I ⊆ I, we have f mℓ ⊆ I, and, by hypothesis, I is
saturated, therefore f ∈ I. Now, I and I coincide in degree ≤ δ, hence
f ∈ I. This shows I d = Ied for all d ≤ δ.
By Lemma 4.5, I has no minimal generators in degree δ+1, and
hence Id = I d = Ied also for d = δ+1.
By contradiction, let d > δ+1 be the biggest integer such that I d (
e
Id . This means that in degree d+1
MP/I (n, d+1) = MP/Ie(n, d+1)
and in degree d
MP/I (n, d) > MP/Ie(n, d)
and MP/I (j, d) ≥ MP/Ie(j, d), for j = 1, . . . , n−1.
16
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
By definition I is generated in degree δ < d, hence, by Theorem 4.1,
we have that
n
X
MP/I (n, d+1) =
MP/I (j, d).
j=1
Now, using the equalities and inequalities above, we get
MP/Ie(n, d+1) = MP/I (n, d+1) =
n
X
j=1
MP/I (j, d) >
n
X
MP/Ie(j, d).
j=1
This is impossible by the inequalities in Theorem 3.11.(a). ⊔
⊓
Extending Corollary 6.2 to the case of saturated ideals, we can generalize Theorem 3.6 of [4] and Theorem 3.6 of [5].
Corollary 7.6. Let I be a saturated ideal in P = K[x1 , . . . , xn ]. If there
exists δ such that Iδ 6=Pδ and MP/I has (n−1)-maximal or n-maximal
growth in degree δ, let i = min{j>1 | MP/I (j, δ)6=0}, then hI≤δ i is
a saturated ideal of dimension n−i, of degree MP/I (i, δ) and it is δregular. Moreover, dim(P/I) ≤ n−i.
Proof. By Lemma 7.3, since I is saturated, having (n−1)-maximal or nmaximal growth is equivalent. By Lemma 7.5, hI≤δ i is a saturated ideal.
The conclusions then follows from Corollary 6.2 and Proposition 6.4.
⊔
⊓
Now we can generalize Corollary 5.2 of [4] and Corollary 2.9 of [5].
Corollary 7.7. Let I be a saturated ideal in P = K[x1 , . . . , xn ]. If
MP/I has (n−1)-maximal growth in degree δ and potential degree of
the GCD = k ≥ 1. Then hI≤δ i = hI≤δ+1 i is saturated and it has a
GCD of degree k.
Proof. From Lemma 7.3 it follows that MP/I has n-maximal growth
in degree δ. Hence Corollary 6.7 and Lemma 7.5 apply. ⊔
⊓
Similarly to Corollary 7.6, we might be tempted to extend Proposition 6.1 to the case of saturated ideals, or, equivalently, Corollary 7.6
to the case of 2-maximal growth. The example below shows that this
is not possible.
Example 7.8. As in [4], under the assumption of Proposition 6.1,
if I is saturated we can not conclude that hI≤δ i is saturated. For
this example, we consider a first set of 98 points on the conic Q with
equation (z − 3t)(z + 3t) = 0 in P3 and a second set of 16 points
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
17
outside Q. In this way we obtain a saturated homogeneous ideal I in
P = Q[x, y, z, t] and MP/I is
0
1
1
1
1
1
1 1 1 1 0 0 0 0 0
0
0
0
0
0
2 3 4 5 2 2 0 0 0
0
0
0
0
0
3 6 10 15 17 13 13 11 9
7
5
3
1
0
4 10 20 35 52 65 78 89 98 105 110 113 114 114
2
3
4
5
6
7
8
9
10
11
12
13
14
...
...
...
...
...
From MP/I we can read that hI≤5 i has a GCD of degree 2 (Corollary 6.5), however MP/I does not have 3-maximal growth in degree 5
(so Corollary 7.7 does not apply), indeed a direct computation shows
that I≤5 is not saturated.
Example 7.9. Consider the polynomial ring P = Q[x, y, z, t, h] and
the strongly stable ideal
I = (x5 , x4 y, x3 y 2 , x2 y 3 , xy 4, x4 z, x3 yz, x2 y 2z, x4 t, xy 3 z 3 ).
The ideal I is saturated and
MP/I
0
1
1
1
=
1
1
1
1
2
3
4
5
...
1 1 1
0
0
0
0 ...
3 4 5
1
1
1
1 ...
6 10 15 13 14 14 15 . . .
10 20 35 47 61 75 90 . . .
15 35 70 117 178 253 343 . . .
2
3
4
5
6
7
8
We can now apply Corollary 7.7 and see that hI≤5 i and hI≤6 i are saturated and have GCD of degree 1 (we can check that the GCD is x).
For this example the result in [4, Corollary 5.2], for detecting a GCD,
do not apply.
8. Sectional matrices, GIN, and resolutions
In this section, we will present some examples in order to compare
the sectional matrix with other algebraic invariants, such as the Hilbert
function H, the generic initial ideal and the minimal resolution.
We start from two homogeneous ideals with same Hilbert function
but different rgin, sectional matrix and Betti numbers.
Example 8.1. Consider P = Q[x, y, z] and let
I = (x2 , xy, xz, y 3 , y 2z, yz 2 , z 3 )
and
J = (x2 , xy, y 2, xz 2 , yz 2 , z 3 )
be two ideals in P . Both ideals are strongly stable and hence, they
coincide with their own rgin. These two ideals clearly have distinct
rgin, but they have the same Hilbert function (the last row in the
18
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
sectional matrices). They have different sectional matrix and different
graded Betti numbers.
...
1 1 0 0 0 ...
=
1 2 1 0 0 ...
1 3 3 0 0 ...
0
MP/I
1
2
3
...
1 1 0 0 0 ...
=
1 2 0 0 0 ...
1 3 3 0 0 ...
0
4
MP/J
1
2
3
4
The resolutions of P/I and P/J are respectively
0 → P (−4)⊕P (−5)3 → P (−3)3 ⊕P (−4)7 → P (−2)3 ⊕P (−3)4 → P → P/I → 0.
0 → P (−5)3 → P (−3)2 ⊕P (−4)6 → P (−2)3 ⊕P (−3)3 → P → P/J → 0.
In the following example, we show two ideals with the same sectional
matrix and same Betti numbers, but different generic initial ideal.
Example 8.2. Consider the polynomial ring P = Q[x, y, z] and the
ideals
I = (x5 , x4 y, x3y 2 , x2 y 3 , xy 4 , x4 z, x3 yz, x2 y 2 z, x3 z 2 , x2 yz 2 )
J = (x5 , x4 y, x3 y 2, x2 y 3, xy 4 , x4 z, x3 yz, x2 y 2 z, x3 z 2 , xy 3 z).
Both the ideals are strongly stable and hence, they coincide with their
own rgin. These two ideals clearly have distinct rgin, but they have the
same Hilbert function, the same sectional matrix and the same Betti
numbers.
...
0 1 2
3
4
5
6
1 1 1 1 1 0 0 ...
MP/I = MP/J =
1 2 3 4 5 1 1 ...
1 3 6 10 15 11 12 . . .
The resolution of P/I and P/J is
0 → P (−7)5 → P (−6)14 → P (−5)10 → P.
In the following example we show two ideals with the same sectional
matrix, but different generic initial ideal and different Betti numbers.
Example 8.3. Consider the polynomial ring P = Q[x, y, z] and the
ideals
I = (x5 , x4 y, x3 y 2 , x2 y 3 , xy 4, x4 z, x2 y 2 z, x3 z 2 , x2 yz 2 )
J = (x5 , x4 y, x3y 2 , x2 y 3 , xy 4 , x4 z, x2 y 2z, x3 z 2 , xy 3z).
We have that
rgin(I) = (x5 , x4 y, x3 y 2, x2 y 3, xy 4 , x4 z, x3 yz, x2 y 2 z, x3 z 2 , x2 yz 3 ),
rgin(J) = (x5 , x4 y, x3y 2 , x2 y 3, xy 4 , x4 z, x3 yz, x2 y 2 z, xy 3 z, x3 z 3 ).
EXTREMAL BEHAVIOUR IN SECTIONAL MATRICES
...
1 1 1 1 1 0 0 0 ...
=
1 2 3 4 5 1 1 1 ...
1 3 6 10 15 12 12 13 . . .
0
MP/I = MP/J
19
1
2
3
4
5
6
7
The resolution of P/I is
0 → P (−7)2 ⊕ P (−8) → P (−6)11 → P (−5)9 → P → P/I → 0.
The resolution of P/J is
0 → P (−7)3 ⊕P (−8) → P (−6)11 ⊕P (−7) → P (−5)9 → P → P/J → 0.
In the following example we show two ideals with the same rgin,
therefore the same sectional matrix and Hilbert function, but different
Betti numbers.
Example 8.4. Consider the polynomial ring P = Q[x, y, z], and the
ideals of P
I = (x4 , y 4, z 4 , xy 2 z 3 , x3 yz 2 , x2 y 3 z)
and
J = rgin(I).
These two ideals clearly have the same rgin, therefore the same sectional
matrix. However, J has more minimal generators than I, so they have
different resolutions.
0 1 2
3
4
5
6 7 8 ...
1 1 1 1 0 0 0 0 0 ...
MP/I = MP/J =
1 2 3 4 2 0 0 0 0 ...
1 3 6 10 12 12 7 0 0 . . .
The resolution of P/I is
0 → P (−9)7 → P (−7)3 ⊕P (−8)9 → P (−4)3 ⊕P (−6)3 → P → P/I → 0.
The resolution of P/J is
0 → P (−8)5 ⊕ P (−9)7 → P (−5)2 ⊕ P (−6)2 ⊕ P (−7)10 ⊕ P (−8)14 →
→ P (−4)3 ⊕ P (−5)2 ⊕ P (−6)5 ⊕ P (−7)7 → P → P/J → 0.
In summary:
Example rgin Sec. Mat. Betti n.
8.1
6=
6=
6
=
8.2
6=
=
=
8.3
6=
=
6=
8.4
=
=
6=
20
ANNA BIGATTI, ELISA PALEZZATO, AND MICHELE TORIELLI
Acknowledgements. The authors thank Professor L. Robbiano for
the valuable discussions and for suggesting us Example 7.4.
During the preparation of this paper the third author was supported
by the MEXT grant for Tenure Tracking system.
References
[1] J. Abbott, A. M. Bigatti, and G. Lagorio. CoCoA-5:
a system for doing Computations in Commutative Algebra. Available at
http://cocoa.dima.unige.it, 2017.
[2] J. Abbott and A.M. Bigatti. CoCoALib:
a C++ library for
doing
Computations
in
Commutative
Algebra.
Available
at
http://cocoa.dima.unige.it/cocoalib, 2017.
[3] John Abbott and Anna Maria Bigatti. Gröbner bases for everyone with
CoCoA-5 and CoCoALib. arXiv:1611.07306, to appear in the Proc. of the
MSJ SI 2015, 2016.
[4] J. Ahn and J. C. Migliore. Some geometric results arising from the Borel fixed
property. J. Pure Appl. Algebra, 209(2):337–360, 2007.
[5] A. M. Bigatti, A. V. Geramita, and J. C. Migliore. Geometric consequences
of extremal behavior in a theorem of Macaulay. Trans. Amer. Math. Soc.,
346(1):203–235, 1994.
[6] A. M. Bigatti and L. Robbiano. Borel sets and sectional matrices. Ann. Comb.,
1(1):197–213, 1997.
[7] A. Conca. Reduction numbers and initial ideals. Proc. Amer. Math. Soc.,
131(4):1015–1020, 2003.
[8] A. Galligo. A propos du théoreme de préparation de Weierstrass. In Fonctions
de plusieurs variables complexes, pages 543–579. Springer, 1974.
[9] G. Gotzmann. Eine Bedingung für die Flachheit und das Hilbertpolynom eines
graduierten Ringes. Math. Z., 158:61–70, 1978.
[10] M. Green. Restriction of linear series to hyperplanes, and some results of
Macaulay and Gotzmann, algebraic curves and projective geometry. In Proceedings, Trento, number 1389 in Lecture Notes in Mathematics. Springer,
1988.
[11] Martin Kreuzer and Lorenzo Robbiano. Computational commutative algebra
2. Springer Science & Business Media, 2005.
[12] F. S. Macaulay. Some properties of enumeration in the theory of modular
systems. Proc. Lond. Math. Soc., 26:531–555, 1927.
Department of Mathematics, Genoa University, Via Dodecaneso 35,
16146 Genoa, Italy.
E-mail address: [email protected]
Department of Mathematics, Genoa University, Via Dodecaneso 35,
16146 Genoa, Italy.
E-mail address: [email protected]
Department of Mathematics, Hokkaido University, Kita 10, Nishi 8,
Kita-Ku, Sapporo 060-0810, Japan.
E-mail address: [email protected]
| 0math.AC
|
FIXED-POINT PROPERTY FOR AFFINE ACTIONS ON A
HILBERT SPACE
arXiv:1705.02644v1 [math.GR] 7 May 2017
SHIN NAYATANI
Abstract. Gromov [7] showed that for fixed, arbitrarily large C, any uniformly
C-Lipschitz affine action of a random group in his graph model on a Hilbert
space has a fixed point. We announce a theorem stating that more general
affine actions of the same random group on a Hilbert space have a fixed point.
We discuss some aspects of the proof.
Introduction
In [10], Izeki, Kondo and the present author proved that a random group in the
Gromov graph model had fixed-point property, meaning that any isometric action
had a fixed point, for a large class of CAT(0) spaces, by using the method which
concerns the n-step energy of maps. Naor and Silberman [17] proved a similar
result for a class of p-uniformly convex geodesic metric spaces. (Note that CAT(0)
spaces are 2-uniformly convex.) In these studies it seemed that the condition that
actions are isometric was essential and without the condition the argument should
break dwon. Gromov [7], however, had shown that any uniformly C-Lipschitz
affine action of the same random group on a Hilbert space has a fixed point, where
C may be arbitrarily large but should be specified in advance. The purpose of
this article is to announce a fixed-point theorem for more general affine actions
of the same random group, allowing the Lipschitz constants of the affine maps to
have mild growth with respect to a certain length function on the group [11]. It is
worth while to mention the following: if the Lipschitz constants of the affine maps
are uniformly bounded, then the action reduces to an isometric one on a Banach
space by replacing the Hilbert norm by an equivalent one. On the other hand, our
case treats really non-isometric actions which cannot reduce to isometric ones.
A key of the proof is to verify the existence of a discrete harmonic map from
the group into the Hilbert space which is equivariant with respect to the given
action. In the case of isometric actions, the method of energy minimization coupled
with scaling ultralimit argument was effective. In the general affine case, this
method fails because a map minimizing local energy does not necessarily satisfy
the condition of harmonicity. We therefore employ the method of discrete tensioncontracting flow due to Gromov [7]. Indeed, we refine Gromov’s method and
derive the existence of a harmonic map still by coupling it with scaling ultralimit
argument.
Graduate School of Mathematics, Nagoya University, Chikusa-ku, Nagoya 464-8602, Japan,
[email protected].
1
This article is organized as follows. In §1, we define affine action, discuss the
rigidity of isometric actions and state Shalom’s theorem on the rigidity and existence of uniformly Lipschitz affine actions. In §2, after discussing Nowak’s fixedpoint theorem for uniformly Lipschitz affine actions of a random group in the
Gromov density model, we state our main fixed-point theorem. In §3, we discuss
discrete harmonic maps and state an existence theorem for such maps. We also
discuss the failure of the method of energy minimization. In §4, we introduce
Gromov’s discrete tension-contracting flow and outline the proof of the existence
of harmonic maps. In §5, we outline the proof of the main theorem. In Appendix,
we prove the existence of maps minimizing local energy which are equivariant with
respect to a given affine action.
1. Affine actions on a Hilbert space
Let H be a Hilbert space, and denote the algebra of all bounded linear operators
of H by B(H). Let Γ be a finitely generated infinite group, and let ρ : Γ y H be
an affine action. Thus, for γ ∈ Γ, ρ(γ) : H → H has the form
ρ(γ)(v) = A(γ)(v) + b(γ),
v ∈ H,
where A(γ) ∈ B(H) and b(γ) ∈ H. Since γ 7→ ρ(γ) is a homomorphism, we have
A(γγ ′ ) = A(γ)A(γ ′ ),
b(γγ ′ ) = b(γ) + A(γ)b(γ ′ ),
γ, γ ′ ∈ Γ.
Definition 1.1. An affine action ρ : Γ y H is called uniformly C-Lipschitz if
ρ(γ) : H → H is a C-Lipschitz map, or equivalently kA(γ)k ≤ C, for all γ ∈ Γ.
Note that C ≥ 1 necessarily.
Most basic example of a uniformly Lipschitz affine action is an isometric action.
Recall that a σ-compact, locally compact topological group G is said to have
property F H if any continuous isometric action ρ : G y H has a fixed point, that
is, there exists v ∈ H such that ρ(g)(v) = v for all g ∈ G. It is a celebrated result
of Delorme [4] and Guichardet [8] that property F H is equivalent to Kazhdan’s
property (T). Kazhdan [13] defined this property for locally compact groups in
terms of unitary representations, and proved that if Γ is a lattice in a Lie group
G, then Γ has property (T) if and only if G has property (T). As examples, simple
real Lie groups of real rank at least two have property (T). For n ≥ 2, Sp(n, 1) is
a simple Lie group of real rank one that has property (T). Thus these Lie groups
and their lattices have property F H.
In his unpublished work, Shalom proved the following theorem which exhibits
that higher-rank groups have stronger rigidity than Sp(n, 1) (cf. [2, 19]).
Theorem 1.2 (Shalom). (i) Any uniformly Lipschitz affine action of a simple
real Lie group of real rank at least two (or its lattices) on H has a fixed point.
(ii) Sp(n, 1) admits a uniformly Lipschitz affine action on H without fixed point.
Mimura [16] observes that the action in the statement (ii) is indeed metrically
proper. Hence, any infinite discrete subgroup of Sp(n, 1) also admits a uniformly
Lipschitz affine action on H without fixed point. This exhibits many infinite
hyperbolic groups which admit such affine actions.
Shalom proposed the following (cf. [19])
2
Conjecture 1 (Shalom). Any non-elementary hyperbolic group admits a uniformly
Lipschitz affine action on H without fixed point.
2. Fixed-point property of random groups w.r.t. uniformly
Lipschitz affine actions
In this section, we review two fixed-point theorems regarding uniformly Lipschitz affine actions of certain ramdom groups on a Hilbert space. Recall that in
±1
the Gromov density model G(m, l, d) of random groups, generators s±1
1 , . . . , sm
dl
and a density 0 < d < 1 are fixed, and choose (2m − 1) words, each of them
chosen uniformly and independently from the set of all reduced words of lenght l
±1
±1
±1
in s±1
1 , . . . , sm . The group Γ generated by s1 , . . . , sm and having those reduced
words as relations is a constituent of the model G(m, l, d). Given a group property
P (e.g. Kazhdan’s property (T)), we say that a random group in the Gromov density model has property P if the probability of Γ having property P tends to one
as l → ∞.
√
Theorem 2.1 (Nowak [18]). Fix 1 ≤ C < 2. Let Γ be a random group in
the Gromov density model with density 1/3 < d < 1/2. Then any uniformly CLipschitz affine action ρ : Γ y H has a fixed point.
Note that the random group Γ of the theorem is non-elementary hyperbolic
(hence infinite) [6, 20] and has property (T) [23, 15].
The proof of Theorem 2.1 is based on a fixed-point theorem for an isometric
action of a deterministic group on a Banach space, which we shall review. Let Γ
be a finitely generated group equipped with a finite, symmetric generating set S
not containing the identity element. Modifying the construction as in [23], one
constructs the link graph L(S); its vertices are the elements of S, generators s and
t span an edge (written s ∼ t) if s−1 t is a generator, and the edges are suitably
weighted. (For the account of the choice of weight, see [18, p. 703], [9, Proof of
Lemma 3.1].)
Let B be a Banach space with norm k · k and denote by κp (S, B) the optimal
constant in the p-Poincaré inequality for maps f : S → B:
X
X
kf (s) − f kp m(s) ≤ κp
kf (s) − f (t)kp m(s, t),
s∼t
s∈S
where m(s, t) is P
the weight of the edge (s, t), m(s) =
P
m(s)f
(s)/
s∈S
s∈S m(s), the mean value of f .
P
t∼s
m(s, t), and f =
Theorem 2.2 (Nowak [18]). Let B be a reflexive Banach space and let Γ and S
be as above. If the link graph L(S) is connected and for some 1 < p < ∞ and its
adjoint index p∗ , satisfying 1/p + 1/p∗ = 1, the corresponding Poincaré constants
satisfy
∗
max{2−1/p κp (S, B), 2−1/p κp∗ (S, B∗ )} < 1,
then any affine isometric action ρ : Γ y B has a fixed point.
Let ρ : Γ y H be a uniformly C-Lipschitz affine action, where H is a Hilbert
space, and introduce a new norm on H by |||v||| = supγ∈Γ kA(γ)(v)k for v ∈ H.
Then B = (H, ||| · |||) is a Banach space isomorphic to H, thus reflexive, and
3
ρ : Γ y B is an affine isometric action. Since the norms of B and B∗ (∼
= H) satisfy
k · k ≤ ||| · ||| ≤ Ck · k and C −1 k · k ≤ ||| · |||∗ ≤ k · k, respectively, it follows that
κ2 (S, B), κ2 (S, B∗ ) ≤ C κ2 (S, H) = C κ2 (S, R).
Therefore, we obtain the following
Corollary 2.3. Let Γ and S be as above, and suppose that the link graph L(S)
is connected. √ Then any uniformly C-Lipschitz affine action ρ : Γ y H with
C κ2 (S, R) < 2 has a fixed point.
Now let Γ be a random group in the Gromov density model with density 1/3 <
d < 1/2. By the argument due to Żuk [23], Kotowski and Kotowski [15], there is a
random group Γ′ in a different model so that Γ contains a quotient of Γ′ as a finite
index subgroup and the link graph L(S ′ ) of Γ′ has κ2 (S ′ , R) arbitrarily close to
one. Therefore, we may apply Corollary 2.3 to Γ′ and the conclusion of Theorem
2.1 holds for Γ′ and hence for Γ.
Gromov [7] claimed a result similar to Theorem 2.1 for a random group in
the graph model which was also invented by him. To state Gromov’s result, we
first review the construction of this model. Let Fm denote the free group on m
generators, and let S be the collection of these m elements and their inverses.
Let G = (V, E) be a finite connected graph, where V and E are the sets of
vertices and undirected edges, respectively. We denote the set of directed edges
→
−
→
−
→
−
by E . A map α : E → S satisfying α((v, u)) = α((u, v))−1 for all (u, v) ∈ E is
called an S-labelling of G. Let A(m, G) denote the set of all S-labellings of G,
consisting of (2m)#E elements, and equip it with the uniform probability measure.
→
−
→
→
−
→
For α ∈ A(m, G) and a path −
p = (−
e 1, . . . , →
e l ) in G, where −
e i ∈ E , define
→
→
→
α(−
p ) = α(−
e 1 ) · · · · · α(−
e l ) ∈ Fm . Then set
→
→
R = {α(−
c)|−
c are cycles in G},
α
Γα = Fm /normal closure of Rα .
Let λ1 (G, R) denote the second eigenvalue of the discrete Laplacian of G, acting
on real-valued functions on V . The girth of G, denoted by girth(G), is the minimal
length of a cycle (i.e. a closed path) in G.
A sequence {Gj }j∈N of finite graphs is called a sequence of (bounded-degree)
expanders if it satisfies
(i) #Vj → ∞ as j → ∞,
(ii) ∃d, ∀j, ∀u ∈ Vj , 2 ≤ deg(u) ≤ d (sparce),
(iii) ∃λ > 0, ∀j, λ1 (Gj , R) ≥ λ (highly-connected).
Such a {Gj }j∈N is said to have diverging girth if it further satisfies
(iv) girth(Gj ) → ∞ as j → ∞.
Now suppose that a sequence of expanders {Gj }j∈N with diverging girth is given.
Then the collection of groups G(m, Gj ) = {Γα | α ∈ A(m, Gj )} is the graph model
of random groups. Given a group property P, we say that a random group in the
graph model has property P if the probability of Γα having property P tends to
one as j → ∞. Gromov [7] and Silberman [21] proved that a random group in the
graph model had fixed-point property for Hilbert spaces with respect to isometric
actions, that is, it had property (T). This result was generalized to fixed-point
4
property for CAT(0) spaces [10] (see also [7]) and for p-uniformly convex geodesic
metric spaces [17]. In both generalizations the degrees of singularity of the relevant
geodesic metric spaces should be suitably bounded.
We now state
Theorem 2.4 (Gromov [7]). Fix C > 0. Let Γ be a random group in the graph
model associated with a sequence of expanders with diverging girth. Then any
uniformly C-Lipschitz affine action ρ : Γ y H has a fixed point.
We can relax the condition that the Lipschitz constants of the relevant affine
maps should be uniformly bounded. To state our result precisely, we introduce the
following
Definition 2.5. Let Γ be a finitely generated group equipped with a finite, symmetric generating set S, and let l : Γ → Z≥0 denote the word-length function with
respect to S. For each conjugacy class c of Γ, we define
lconj (c) = inf l(γ)
γ∈c
and call lconj : {conjugacy classes of Γ} → Z≥0 the conjugacy-length function of Γ
[3].
We now state
Theorem 2.6. Fix C > 0 and 0 ≤ σ < 1/10. Let Γ be a random group in the graph
model associated with a sequence of expanders with diverging girth and diameter
growing at most linearly in girth. Then any affine action ρ : Γ y H satisfying
(2.1)
∀γ ∈ Γ, kA(γ)k ≤ C lconj ([γ])σ ,
where [γ] denotes the conjugacy class containing γ, has a fixed point.
Remark 1. In order for a random group in the graph model to be non-elementary
hyperbolic (hence infinite), the relevant sequence of expanders should satisfy some
further conditions (cf. [7, 5, 1]). One of these conditions implies the diameter
growth condition in Theorem 2.6, which is therefore essentially superficial.
3. Discrete harmonic maps
Let Γ be a finitely generated group and fix a finite, symmetric generating set S.
Let µ be the standard random walk of Γ associated with S, that is,
1/#S if ∃s ∈ S, x′ = xs,
′ def
µ(x → x ) =
0
otherwise.
The barycenter map of H, bar : {finite-support probability measures on H} → H,
is given by
!
m
m
X
X
(3.1)
bar
ti Dirac(vi ) =
ti vi .
i=1
i=1
Let ρ : Γ y H be an affine action.
5
Definition 3.1. A ρ-equivariant map f : Γ → H is called harmonic if it satisfies
bar(f∗ µ(x → q )) = f (x)
(3.2)
for all x ∈ Γ. Notice that
bar(f∗ µ(x → q )) =
1 X
f (xs).
#S s∈S
Remark 2. Since the action ρ is affine, we may conclude that a ρ-equivariant f is
harmonic if it satisfies (3.2) for some x ∈ Γ. To see this, suppose that (3.2) holds
for x, and write any x′ ∈ Γ as x′ = γx. Then
1 X
1 X
f (x′ s) =
ρ(γ)(f (xs))
#S s∈S
#S s∈S
!
1 X
= ρ(γ)
f (xs) = ρ(γ)f (x)
#S s∈S
bar(f∗ µ(x′ → q )) =
= f (x′ ),
and (3.2) holds for x′ , too.
The action ρ has a fixed point if and only if a ρ-equivariant constant map, which
are trivially harmonic, exists. In contrast, we have the following existence result
for nonconstant harmonic maps when ρ has no fixed point.
Theorem 3.2. Let Γ be a finitely generated group equipped with a finite, symmetric
generating set S, and let ρ : Γ y H be an affine action, where H is a Hilbert space,
satisfying (2.1) for some C > 0 and σ ≥ 0. Suppose that ρ(Γ) has no fixed point.
Then there exist a (possibly new) affine action ρ′ : Γ y H′ , where H′ is a (possibly
new) Hilbert space, satisfying (2.1) for the same C, σ as above and a nonconstant
harmonic ρ′ -equivariant map f : Γ → H′ .
Before discussing the actual proof, we observe that the standard approach via
energy minimization coupled with scaling ultralimit argument would fail.
Definition 3.3. For a ρ-equivariant map f : Γ → H and x ∈ Γ, define the local
energy E(f )(x) of f at x by
(3.3)
def
E(f )(x) =
1X
kf (x) − f (x′ )k2 µ(x → x′ ).
2 x′ ∈Γ
As will be verified in the appendix, under the assumption that ρ(Γ) has no fixed
point, one can always find a nonconstant ρ-equivariant map f : Γ → H minimizing
the local energy at x, though the Hilbert space H and the affine action ρ : Γ y H
should possibly be renewed. We now focus on the question whether the map f
satisfies (3.2) for x. For v ∈ H and t ∈ R, let ft : Γ → H be the ρ-equivariant map
6
such that ft (e) = f (e) + tv. Then we have, taking x = e for simplicity,
1 X
kft (e) − ft (s)k2
2#S s∈S
1 X
kf (e) + tv − ρ(s)(f (e) + tv)k2
=
2#S s∈S
1 X
=
k(f (e) − f (s)) + t(v − A(s)(v))k2
2#S s∈S
1 X
=
k(f (e) − f (s))k2 + 2t hf (e) − f (s), v − A(s)(v)i + O(t2 ) ,
2#S s∈S
E(ft )(e) =
and therefore
0=
d
1 X
hf (e) − f (s), v − A(s)(v)i .
E(ft )(e)|t=0 =
dt
#S s∈S
If the action ρ is isometric, which means that A(s) is orthogonal, then
1 X
hf (e) − f (s), vi −
#S s∈S
1 X
=
hf (e) − f (s), vi −
#S s∈S
2 X
=
hf (e) − f (s), vi .
#S s∈S
R.H.S. =
1 X
A(s−1 )(f (e) − f (s)), v
#S s∈S
1 X
f (s−1) − f (e), v
#S s∈S
Since this vanishes for all v ∈ H, we conclude (3.2). However, if ρ is not isometric,
the above computation fails and we would not be able to conclude (3.2), that is,
that f is harmonic.
Instead, we use Gromov’s discrete tension-contracting flow developed in [7, §3.6]
and produce a harmonic f . Postponing the details to [11], we shall outline the
argument for the proof of Theorem 3.2. In the remainder of this section, let Γ be
a finitely generated group equipped with a finite, symmetric generating set S, and
let ρ : Γ y H be an affine action, where H is a Hilbert space.
For a ρ-equivariant map f : Γ → H, define new maps Hf : Γ → H and ∆f : Γ →
H by
def
Hf (x) =
=
1
2
1
2
X
x′ ∈Γ
f (x′ ) µ(x → x′ ) + f (x)
!
X
1
f (xs) + f (x) ,
#S s∈S
7
!
def
∆f (x) = (1 − H)f (x)
1X
=
(f (x) − f (x′ )) µ(x → x′ )
2 ′
x ∈Γ
1 X
=
(f (x) − f (xs)).
2#S s∈S
The maps Hf and ∆f are ρ-equivariant and A-equivariant, respectively. We call
H (resp. ∆) the averaging operator (resp. Laplacian). Note that f is harmonic if
and only if ∆f = 0, or Hf = f .
Proposition 3.4 (cf. Gromov [7]). We have
k∆Hf (x)k ≤
max
x′ ∈x(S∪{e})
k∆f (x′ )k
for all x ∈ Γ, and if the equality sign holds for some x then ∆f (x) is a constant
vector independent of x ∈ Γ.
Motivated by this proposition, we introduce the following
Definition 3.5 (cf. Gromov [7]). Let f : Γ → H be a ρ-equivariant map, and
define f0 := f and fi+1 := Hfi inductively. We say that f is (harmonically) stable
if
0 < ∃λ < 1, ∃i0 ∈ N, ∀i ≥ i0 , ∀x ∈ Γ, k∆fi+1 (x)k ≤ λ
max
x′ ∈x(S∪{e})
k∆fi (x′ )k.
It should be mentioned that the above definition of harmonic stability is slightly
modified from Gromov’s original one and it is more suitable for our purpose.
Remark 3. Suppose fi0 is harmonic, that is, ∆fi0 = fi0 − fi0 +1 = 0 for some i0 .
Then fi = fi0 and thus ∆fi = 0 for all i ≥ i0 . Therefore, f is stable.
Proposition 3.6. Suppose that ρ satisfies (2.1) for some C > 0 and σ ≥ 0.
(i) If a ρ-equivariant map f : Γ → H is stable, then {fi }i∈N converges pointwise to
a map f∞ : Γ → H, and f∞ is harmonic.
(ii) If a ρ-equivariant map f : Γ → H is not stable, then there exist a Hilbert space
H′ and a nonconstant harmonic map f ′ : Γ → H′ , equivariant with respect to an
affine action ρ′ : Γ y H′ satisfying (2.1) for the same C, σ as above.
This proposition implies Theorem 3.2. The proofs of the two propositions above
will be given in [11].
4. Proof of Theorem 2.6
In this section, we prove Theorem 2.6. Let Γ be a finitely generated group
equipped with a finite, symmetric generating set S. Let ρ : Γ y H be an affine
action, where H is a Hilbert space, and suppose that ρ satisfies
(4.1)
∀γ ∈ Γ, kA(γ)k ≤ C l(γ)σ
for some C > 0 and σ ≥ 0. (Note that this condition is weaker than (2.1).) For a
ρ-equivariant map f : Γ → H and x ∈ Γ, define the local n-step energy of f at x
8
by
def
E (n) (f )(x) =
1X
kf (x) − f (x′ )k2 µn (x → x′ ),
2 ′
x ∈Γ
n
where µ is the n-th convolution of µ.
Lemma 4.1. Suppose that σ < 1/2. Let f : Γ → H be a harmonic ρ-equivariant
map. Then we have
E (n) (f )(x) &C,σ,x n1−2σ E(f )(x)
for all x ∈ Γ.
The proof of this lemma will be given in [11].
So far, the group Γ has been any finitely generated group. The following lemma,
essentially due to Gromov and Silberman [7, 21], concerns a random Γ.
Lemma 4.2. Let Γ be a random group in the graph model associated with a sequence of expanders with diverging girth and diameter growing at most linearly in
girth, and let ρ : Γ y H be an affine action as above. Then for any ρ-equivariant
map f : Γ → H and any x ∈ Γ, we have
E (n) (f )(x) .C,σ,x,λ n8σ E(f )(x).
(4.2)
Here, n is a positive integer depending on f and x, and we may assume that n is
arbitrarily large, and λ is the positive constant as in the definition of a sequence
of expanders.
Proof. The proof of Proposition 2.14 in [21], which treats the case that the action
is isometric, mostly works for the non-isometric case. For the readers’ convenience
we outline Silberman’s argument, and explain how the term n8σ comes in.
The lemma is a consequence of the following statement. With probability tending to one as j → ∞, the group Γ corresponding to α ∈ A(m, Gj ) has the following
girth(Gj )
property: for any affine action ρ : Γ y H satisfying (4.1), any n <
, any
2
√
ρ-equivariant map f : Γ → H and any x ∈ Γ, there exists n < l ≤ n such that
(4.3)
E (l) (f )(x) .C,σ,x,λ diam(Gj )4σ E(f )(x).
girth(Gj )
,
2
Indeed, choosing n ≃
diam(Gj ), and therefore
we have diam(Gj ) . n ≤ l2 by the assumption on
E (l) (f )(x) .C,σ,x,λ l8σ E(f )(x).
p
Note that l & girth(Gj ); thus l diverges as j → ∞.
For the time being, fix a member Gj of the expander sequence defining the graph
model, and denote it by G = (V, E). Let µG and νG denote the standard random
walk on G and the standard probability measure on V given by
(
→
−
1
deg(u)
if (u, v) ∈ E ,
deg(u)
µG (u, v) =
and νG (u) =
,
2#E
0
otherwise,
respectively. For a map ϕ : V → H and n ∈ N, the n-step energy of ϕ is defined
by
X
1X
EµnG (ϕ) =
νG (u)
kϕ(u) − ϕ(v)k2 µnG (u → v).
2 u∈V
v∈V
9
Recall [21, Lemma 2.11] that we have
(4.4)
EµnG (ϕ) ≤
2
Eµ (ϕ)
λ1 (G, R) G
for all maps ϕ : V → H and all n ∈ N.
→
−
Let α : E → S be an S-labelling of G, and Γ the corresponding group. Let
ρ : Γ y H be an affine action, and ρe: Fm y H its lift. The strategy in proving
(4.3) is to transplant (4.4) onto Γ. In fact, we may work on Fm instead of Γ, and
so we shall transplant (4.4) onto Fm . In order to do this, we ‘push-forward’, using
α, the random walks µG and µnG on G to those on Fm as follows.
If u ∈ V and x ∈ Fm are fixed, α induces a corresponding graph morphism βu→x
from G to X = Cay(Fm , S), the Cayley graph of Fm with respect to S, as follows:
→
→
→
For v ∈ V , choose a path −
p = (−
e 1, . . . , −
e l ) from u to v in G, and set
def
→
→
→
βu→x (v) = x α(−
p ) = x α(−
e 1 ) · · · · · α(−
e l ).
To be precise, βu→x is well-defined only on the set of vertices whose graph distance
from u is less than g/2, where g = girth(G). We now define, for n < g/2, a random
walk µnG,α on X by
X
def
νG (u) (βu→x )∗ µnG (u → ·).
µnG,α (x → ·) =
u∈V
Note that the average over V is taken in order to produce a random walk independent of the individual vertices of G.
We can now transplant (4.4) onto Fm , and it is here that something different
occurs when the action ρ is non-isometric. Suppose n < g/2 and let f : Fm → H
be a ρe-equivariant map.1 When ρ is isometric,
(4.5)
EµnG,α (f )(x) = EµnG (f ◦ βu0 →x )(x)
holds for a fixed u0 ∈ V . Indeed,
X
1X
kf (x) − f (x′ )k2 [(βu→x )∗ µnG (u → ·)](x′ )
EµnG,α (f )(x) =
νG (u)
2 u∈V
x′ ∈Fm
X
X
1
=
νG (u)
kf ◦ βu→x (u) − f ◦ βu→x (v)k2 µnG (u → v).
2 u∈V
v∈V
If ρ is isometric, then we can replace βu→x by βu0 →x in the last expression and get
the right-hand side of (4.5). Now consider the general case that ρ is not necessarily
→
→
isometric. Let −
p and −
r be a path from u0 to v and a shortest path from u to u0 ,
→
−
→
→
respectively, and let q denote the path from u to v traveling along −
r and −
p in
this order. Then
→
f ◦ βu→x (v) = f (xα(−
q ))
→
−
→
= ρe(xα( r )x−1 )f (xα(−
p ))
→
−
−1
= ρe(xα( r )x )f ◦ β
(v),
u0 →x
1
Equivalently, f : Fm → H is the lift of a ρ-equivariant map Γ → H. In particular, the map
f ◦ βu→x is well-defined on the whole vertex set V .
10
and therefore
→
−
e
kf ◦ βu→x (u) − f ◦ βu→x (v)k ≤ kA(xα(
r )x−1 )kkf ◦ βu0 →x (u) − f ◦ βu0 →x (v)k,
e is the linear part of ρe. Since
where A
→
−
→
e
e
e −
e −1 )k
kA(xα(
r )x−1 )k ≤ kA(x)kk
A(α(
r ))kkA(x
→
≤ C 3 l(x)σ l(α(−
r ))σ l(x−1 )σ
≤ C 3 D σ l(x)2σ ,
where D = diam(G), we obtain
and likewise,
EµnG,α (f )(x) ≤ C 6 D 2σ l(x)4σ EµnG (f ◦ βu0 →x )(x),
EµG (f ◦ βu0 →x )(x) ≤ C 6 D 2σ l(x)4σ EµG,α (f )(x).
Together with (4.4), these imply
(4.6)
EµnG,α (f )(x) ≤
2 C 12 D 4σ l(x)8σ
EµG,α (f )(x).
λ1 (G, R)
In order to conclude (4.3) (provisionally on Fm instead of Γ), we must show that
with high probability the random walks µG,α and µnG,α in (4.6) can be replaced
√
by µX and µlX , n < l ≤ n, respectively, where µX is the standard random
walk of X. This will be done by verifying that with high probability the random
variables α 7→ µG,α and α 7→ µnG,α concentrate on their expectations and that these
expectations are computed in terms of µX and its convolutions.
We begin with the second issue. For n < g/2, the expectation µnG,X of the
random variable α 7→ µnG,α can be computed and expressed as a convex combination
of µlX , 0 ≤ l ≤ n:
n
X
(n)
n
(4.7)
w l µlX ,
µG,X =
l=0
where the weights
(n)
wl
satisfy
(4.8)
√
′
X
w
(n)
l
n<l≤n
≥ C′
for a certain absolute constant C > 0.
For the first issue, let j get large and observe that the random variables µGj , q
and µnGj , q , where n < gj /2, concentrate on their expectations µGj ,X and µnGj ,X , respectively. Indeed, one can verify that the map α 7→ µnGj ,α is Lipschitz with respect
to the Hamming distance on A(m, Gj ) with the Lipschitz constant depending only
on the fixed parameters d, m. Using this fact, one deduces that with probability
tending to one as j → ∞,
1
µGj ,α (x → x′ ) ≤ 2 µGj ,X (x → x′ ) and µnGj ,α (x → x′ ) ≥ µnGj ,X (x → x′ )
2
′
hold for all x, x ∈ X.
Now for any ρe-equivariant map f : Fm → H, we obtain
EµGj ,α (f )(x) ≤ 2 EµGj ,X (f )(x) = 2 EµX (f )(x)
11
and
EµnG
j
(f )(x) ≥
,α
≥
1
1 X
(n)
EµnG ,X (f )(x) ≥
w l EµlX (f )(x)
j
2
2√
n<l≤n
C
2
′
√min EµlX (f )(x).
n<l≤n
Together with (4.6), these imply that there exists
f and x) such that
√
n < l ≤ n (which depends on
8 C 12 Dj4σ l(x)8σ
EµX (f )(x).
C ′λ
Now let f : Γ → H be a ρ-equivariant map and set fe = f ◦ π. Let x ∈ Γ and
choose x
e ∈ π −1 (x) ⊂ Fm so that l(e
x) = l(x). Since the ball of radius less than
gj /2 with center x
e in X is isometrically isomorphic to that of the same radius with
center x in Cay(Γ, S), the above inequality (for fe, x
e) implies
EµlX (f )(x) ≤
8 C 12 Dj4σ l(x)8σ
E(f )(x),
E (f )(x) ≤
C ′λ
(l)
that is, (4.3).
Theorem 2.6 now follows by combining Theorem 3.2, Lemma 4.1 and Lemma
4.2.
Appendix
Let Γ be a finitely generated group equipped with a finite, symmetric generating
set S, and let ρ : Γ y H be an affine action, where H is a Hilbert space. In §3, we
referred to the following fact: if ρ(Γ) has no fixed point, then energy minimization
coupled with scaling ultralimit argument produces a nonconstant map from Γ to
a (possibly new) Hilbert space H′ which is equivariant with respect to a (possibly
new) affine action ρ′ : Γ y H′ and minimizes the local energy at a point. While
this fact would not be useful for our purpose of proving Theorem 3.2 as we observed
that we would not be able to conclude the resulting map is harmonic, it might be
so in other circumstances. Therefore, we shall verify the above fact by proving the
following
Proposition 4.3. Let Γ be a finitely generated group equipped with a finite, symmetric generating set S, and let ρ : Γ y H be an affine action, where H is a Hilbert
space. Suppose that ρ(Γ) has no fixed point. Fix x ∈ Γ. Then there exist a (possibly
new) affine action ρ′ : Γ y H′ , where H′ is a (possibly new) Hilbert space, and a
nonconstant ρ′ -equivariant map f : Γ → H′ minimizing the local energy at x. If ρ
satisfies (4.1) for some C > 0 and σ ≥ 0, then ρ′ also satisfies (4.1) for the same
C, σ.
Before proceeding to the proof, we review the definitions of ultrafilter and the
ultralimit of a sequence of metric spaces.
A nonempty subset ω ⊂ 2N is called an ultrafilter on N if it satisfies the following
conditions:
(i) ∅ ∈
/ ω.
12
(ii) A ∈ ω, A ⊂ B ⇒ B ∈ ω.
(iii) A, B ∈ ω ⇒ A ∩ B ∈ ω.
(iv) For any subset A ⊂ N, A ∈ ω or N \ A ∈ ω.
An ultrafilter ω on N is called non-principal if it satisfies also
(v) For any finite subset F ⊂ N, F ∈
/ ω (hence, N \ F ∈ ω).
Let ω be a non-principal ultrafilter on N. Let (aj )∞
j=1 ⊂ R be a sequence of
real numbers. We call α ∈ R an ω-limit of (aj ) and write ω- limj aj = α if
{j ∈ N | |aj − α| < ε} ∈ ω holds for any ε > 0. Let (Yj , dj , oj ) be a sequence of metric spaces with base point. On the set of sequences (yj ), where
yj ∈ Yj and dj (oj , yj ) is bounded independent of j, consider the equivalence relation [(yj ) ∼ (zj ) ⇔ ω- limj dj (yj , zj ) = 0], and denote the equivalence class of
(yj ) by y∞ = ω- limj yj . Let Y∞ denote the set of equivalence classes, and endow
it with the metric d∞ (y∞ , z∞ ) = ω- limj dj (yj , zj ). One writes (Y∞ , d∞ , o∞ ) =
ω- limj (Yj , dj , oj ), called the ω-limit of (Yj , dj , oj ). It is known that the metric
space (Y∞ , d∞ ) is necessarily complete.
Proof of Proposition 4.3 We shall follow [22] and [14] which treat the case that
the action is isometric.
Fix a non-principal ultrafilter ω on N. We divide the proof into two cases, according to whether E0 := inf E(f )(x) is strictly positive or not, where the infimum
is taken over all ρ-equivariant maps f : Γ → H.
Case 1. The case that E0 > 0.
This is a simpler case, and we only outline the argument. Let {fj }∞
j=1 be a
sequence of ρ-equivariant maps Γ → H such that E(fj )(x) ց E0 . Set vj =
fj (x) and define (H∞ , k · k∞ , v∞ ) = ω- limj (H, k · k, vj ). Then an affine action
ρ∞ : Γ y H∞ is induced and satisfies (4.1). Define a map f∞ : Γ → H∞ by
f∞ (y) = ω- limj fj (y) for y ∈ Γ. Then f∞ is ρ∞ -equivariant, and
E(f∞ )(x) = ω- lim E(fj )(x) = E0 ;
j
in particular, f∞ is nonconstant. On the other hand, one can verify that E(g)(x) ≥
E0 for all ρ∞ -equivariant maps g : Γ → H∞ . Thus, f∞ minimizes the local energy
at x.
Case 2. The case that E0 = 0.
Define δ : H → R≥0 by δ(v) = maxs∈S kρ(s)(v) − vk. While δ > 0 since ρ(Γ)
has no fixed-point, we have inf v∈H δ(v) = 0; indeed,
1 X
E(f )(x) =
kf (xs) − f (x)k2
2#S s∈S
1 X
=
kρ(x){ρ(sx−1 )(f (x)) − ρ(x−1 )(f (x))}k2 ,
2#S s∈S
which is clearly comparable to δ(ρ(x−1 )(f (x)))2 .
In order to proceed, we need the following elementary fact: let Y be a complete
metric space and ϕ : Y → R a strictly positive continuous function. Then there
exists y ∈ Y such that dY (z, y) ≤ ϕ(y) ⇒ ϕ(z) ≥ 21 ϕ(y). Let j ∈ N and apply
13
this fact to the function jδ : H → R. Then we get vj ∈ H such
that kw − vj k≤
1
jδ(vj ) ⇒ δ(w) ≥ 2 δ(vj ). Now let (H∞ , k · k∞ , v∞ ) = ω- limj H, δ(v1 j ) k · k, vj .
We shall define an affine action ρ∞ : Γ y H∞ . Let w∞ ∈ H∞ and write w∞ =
ω- limj wj . By definition, there exists M > 0 such that kwj − vj k ≤ Mδ(vj ) for
all j ∈ N. Then
kρ(s)(wj ) − vj k ≤ kρ(s)(wj ) − ρ(s)(vj )k + kρ(s)(vj ) − (vj )k
≤ C kwj − vj k + δ(vj )
≤ (CM + 1) δ(vj ),
where C = kA(s)k. It follows that ω- limj ρ(s)(wj ) exists, and it is easy to verify
that this limit is independent of the choice of wj . Hence, by defining ρ∞ (s)(w∞ ) =
ω- limj ρ(s)(wj ), we obtain a well-defined map ρ∞ (s) : H∞ → H∞ , which is clearly
C-Lipschitz. It is also easy to see that the affineness, that is, the property of preserving internally dividing points, of ρ(s) is inherited by ρ∞ (s). Let γ ∈ Γ and write
γ = s1 . . . sl , where s1 , . . . , sl ∈ S. Let w∞ = ω- limj wj ∈ H∞ . Then the ultralimit of ρ(γ)(wj ) = ρ(s1 ) . . . ρ(sl )(wj ) exists and equals to ρ∞ (s1 ) . . . ρ∞ (sl )(w∞ ).
Thus, defining ρ∞ (γ)(w∞ ) = ω- limj ρ(γ)(wj ), we have ρ∞ (γ) = ρ∞ (s1 ) . . . ρ∞ (sl )
and obtain an affine action ρ∞ : Γ y H∞ . It is clear that if ρ satisfies (4.1), then
ρ∞ also satisfies (4.1) with the same constants.
We now verify that δ∞ ≥ 21 , where δ∞ is the function δ with respect to ρ∞ .
To do so, take any w∞ = ω- limj wj ∈ H∞ , so that kwj − vj k ≤ M δ(vj ) for
some M > 0, and set As := {j ∈ N | kρ(s)(wj ) − wj k ≥ 12 δ(vj )} for s ∈ S.
For j > M, kwj − vj k ≤ j δ(vj ), and therefore δ(wj ) ≥ 21 δ(vj ), that is, j ∈
∪s∈S As . Thus ∪s∈S As ∈ ω. But this means As ∈ ω for some s ∈ S. Therefore,
kρ∞ (s)(w∞ ) − w∞ k∞ ≥ 21 , and δ∞ ≥ 12 . We thus recover the situation of Case 1.
References
[1] G. Arzhantseva and T. Delzant, Examples of random groups, preprint.
[2] U. Bader, A. Furman, T. Gelander and N. Monod, Property (T) and rigidity for actions on
Banach spaces, Acta Math. 198 (2007), 57–105.
[3] M. Coornaert and G. Knieper, Growth of conjugacy classes in Gromov hyperbolic groups,
Geom. Funct. Anal. 12 (2002), 464–478.
[4] P. Delorme, 1-cohomologie des représentations unitaires des groupes de Lie semi-simples
et résolubles. Produits tensoriels continus et représentations, Bull. Soc. Math. France 105
(1977), 281–336.
[5] E. Ghys, Groupes Aléatoires [d’après Misha Gromov,...], Séminaire Bourbaki, 55ème année,
2002–2003, n◦ 916.
[6] M. Gromov, Asymptotic invariants of infinite groups, in Geometric group theory,
ed. G. Niblo, M. Roller, Cambridge University Press, Cambridge, 1993.
[7] M. Gromov, Random walk in random groups, Geom. Funct. Anal. 13 (2003), 73–146.
[8] A. Guichardet, Étude de la 1-cohomologie et de la topologie du dual pour les groupes de Lie
à radical abélien, Math. Ann. 228 (1977), 215–232.
[9] H. Izeki, T. Kondo, and S. Nayatani, Fixed-point property of random groups, Annals of
Global Analysis and Geom. 35 (2009), 363–379
[10] H. Izeki, T. Kondo, and S. Nayatani, N -step energy of maps and fixed-point property of
random groups, Groups, Geometry, and Dynamics 6 (2012), 701–736.
[11] H. Izeki, T. Kondo, and S. Nayatani, in preparation.
14
[12] H. Izeki and S. Nayatani, Combinatorial harmonic maps and discrete-group actions on
Hadamard spaces, Geom. Dedicata 114 (2005), 147–188.
[13] D. Kazhdan, Connection of the dual space of a group with the structure of its closed subgroups, Funct. Anal. Appl. 1 (1967), 63–65.
[14] T. Kondo, Fixed point theorems via a scaling limit argument (in Japanese), RIMS Kokyuroku
1720 (2010), 139–149.
[15] M. Kotowski and M. Kotowski, Random groups and property (T): Zuk’s theorem revisited,
J. Lond. Math. Soc. 88 (2013), 396–416.
[16] M. Mimura, private communication, April 14, 2015.
[17] A. Naor and L. Silberman, Poincaré inequalities, embeddings, and wild groups, Compos. Math. 147 (2011), 1546–1572.
[18] P. W. Nowak, Poincaré inequalities and rigidity for actions on Banach spaces,
J. Eur. Math. Soc. 17 (2015), no. 3, 689–709.
[19] P. W. Nowak, Group actions on Banach spaces. Handbook of group actions. Vol. II, 121–149,
Adv. Lect. Math. 32, Int. Press, Somerville, MA, 2015.
[20] Y. Ollivier, Sharp phase transition theorems for hyperbolicity of random groups,
Geom. Funct. Anal. 14 (2004), 595–679.
[21] L. Silberman, Addendum to “Random walk on random groups” by M. Gromov,
Geom. Funct. Anal. 13 (2003), 147–177.
[22] L. Silberman, note formerly available at the author’s homepage.
[23] Zuk, A.: Property (T) and Kazhdan constants for discrete groups. Geom. Funct. Anal. 13
(2003), 643–670.
15
| 4math.GR
|
CMSIS-NN: Efficient Neural Network Kernels for
Arm Cortex-M CPUs
arXiv:1801.06601v1 [cs.NE] 19 Jan 2018
Liangzhen Lai
Arm Inc.
[email protected]
Naveen Suda
Arm Inc.
[email protected]
Vikas Chandra
Arm Inc.
[email protected]
Abstract
Deep Neural Networks are becoming increasingly popular in always-on IoT edge
devices performing data analytics right at the source, reducing latency as well as
energy consumption for data communication. This paper presents CMSIS-NN,
efficient kernels developed to maximize the performance and minimize the memory
footprint of neural network (NN) applications on Arm Cortex-M processors targeted
for intelligent IoT edge devices. Neural network inference based on CMSIS-NN
kernels achieves 4.6X improvement in runtime/throughput and 4.9X improvement
in energy efficiency.
1
Introduction
Connected devices – otherwise known as the Internet of Things (IoT) – have been rapidly proliferating
over the past few years and are predicted to reach 1 trillion across various market segments by 2035 [1].
These IoT edge devices typically consist of sensors collecting data – such as audio, video, temperature,
humidity, GPS location and acceleration – which is then processed and communicated with other
nodes or the cloud. Currently, the data from the sensors are processed by analytics tools in the cloud
to enable a wide range of applications, such as industrial monitoring and control, home automation
and health care. However, as the number of the IoT nodes increases, this places a considerable
burden on the network bandwidth, as well as adding latency to the IoT applications. Furthermore,
dependency on the cloud makes it challenging to deploy IoT applications in regions with limited or
unreliable network connectivity. One solution to this problem is edge computing [2], performed right
at the source of data, i.e. the IoT edge node, thus reducing latency as well as saving energy for data
communication.
In terms of accuracy, deep neural networks have demonstrated near-human performance for many
complex machine learning applications such as image classification, speech recognition and natural
language processing. A typical neural network (NN) for image classification consists of multiple
layers of convolution based feature extractors, followed by fully-connected layers for classification,
as shown in Fig. 1. Due to the computational complexity and resource requirements, the execution
of NNs has predominantly been confined to cloud computing with high-performance server CPUs
or specialized hardware (e.g. GPU or accelerators), which adds latency to the IoT applications.
Classification using a small neural network right at the source of the data, i.e. the IoT edge, reduces
the overall latency and energy consumption of data communication between the IoT edge and the
cloud.
In this work, we explore the performance optimization of neural networks on resource-constrained
microcontroller based platforms, targeted for intelligent IoT edge nodes. To enable this, we have
developed optimized software kernels for deploying NNs on Arm Cortex-M CPUs. Using these
software kernels, we demonstrate a convolutional neural network (CNN) for CIFAR-10 dataset on
an off-the-shelf Arm Cortex-M7 platform classifying 10.1 images per second with an accuracy of
79.9%.
…
…
Convolution
Pooling
Fully-connected
layer
Figure 1: Structure of a typical deep neural network.
2
Overview
The overview of the neural network kernels is shown in Fig. 2. The kernel code consists of two parts:
NNFunctions and NNSupportFunctions. NNFunctions include the functions that implement popular
neural network layer types, such as convolution, depthwise separable convolution, fully-connected
(i.e. inner-product), pooling and activation. These functions can be used by the application code
to implement the neural network inference applications. The kernel APIs are also kept simple,
so that they can be easily retargeted for any machine learning framework. NNSupportFunctions
include utility functions, such as data conversion and activation function tables, which are used in
NNFunctions. These utility functions can also be used by the application code to construct more
complex NN modules, such as Long Short Term Memory (LSTM) or Gated Recurrent Unit (GRU).
For some kernels, such as fully-connected and convolution, different versions of the kernel functions
are implemented. A basic version is provided that works universally, ‘as-is’, for any layer parameters.
We have also implemented other versions which include further optimization techniques with either
transformed inputs or with some limitations on the layer parameters. Ideally, a simple script can be
used to parse the network topology and automatically determine the appropriate functions to be used.
Figure 2: Overview of the neural network kernel structure.
3
Fixed-Point Quantization
Traditionally, NN models are trained using 32-bit floating point data representation. However, such
high precision is generally not required during inference. Research has shown that NNs work well
even with low-precision fixed-point representation [4, 5, 6]. Fixed-point quantization helps to avoid
the costly floating-point computation and reduces the memory footprint for storing both weights and
activations, which is critical for resource-constrained platforms. Although precision requirements for
different networks or network layers can vary [7], it is hard for the CPU to operate on data types with
varying bit-width. In this work, we develop the kernels that support both 8-bit and 16-bit data.
The kernels adopt the same data type format as used in CMSIS [3], i.e. q7_t as int8, q15_t as
int16 and q31_t as int32. The quantization is performed assuming a fixed-point format with a
power-of-two scaling, i.e. the represented value will be A × 2n , where A is the integer value and n
is an integer number that indicates the location of the radix point. We pass the scaling factors for
the bias and outputs as parameters to the kernels and the scaling is implemented as bitwise shift
2
operations because of the power-of-two scaling. We use this type of quantization – instead of the
8-bit quantization used in TensorFlow [4] – to avoid the need for floating-point de-quantization in
between layers, as some Arm Cortex-M CPUs may not have a dedicated floating point unit (FPU),
thus limiting their floating-point computation capabilities. The other benefit of such quantization is
that we can use simpler table look-up based activation, which is discussed in Section 4.5.
4
Software Kernels
In this section, we describe the implementation and optimization of the proposed software kernels
for Arm Cortex-M CPUs. The Cortex-M [8] family of processors are 32-bit RISC processor cores
that are designed for energy efficiency, and typically used as microcontrollers for deeply embedded
applications. In this work, we focus on enabling neural networks on Cortex-M based systems that
support SIMD instructions, especially 16-bit Multiply-and-Accumulate (MAC) instructions (e.g.
SMLAD) which are very useful for NN computation.
4.1
Support Functions
Most NNFunctions use the 16-bit MAC instructions, hence data transformation is required to convert
the 8-bit data type (i.e. q7_t) into 16-bit data type (i.e. q15_t). CMSIS provides a utility function,
arm_q7_to_q15, to perform the data transformation. The illustration and pseudo code is shown in
Fig. 3. The data transformation is done in two steps: the first step expands the 8-bit data into 16-bit
data by using the sign extension instruction (__SXT B16); the second step rearranges the data so
that the output follows the same order as the input.
Figure 3: Illustration and pseudo code of the data transform from q7_t to q15_t in CMSIS
arm_q7_to_q15 function (assuming big-endian data format).
The performance of data transformation is critical, as it is used in the inner loop inside the computation
kernels. While the first step of sign extension is essential, the second step of rearranging the data
can be omitted if both operands follow the same ordering. To better exploit this, we created another
version of the data transformation routine without the data reordering, as shown in Fig. 4. The routine
is discussed in detail in Section 4.2.
Figure 4: Illustration and pseudo code for data transformation from q7_t to q15_t without reordering.
Output and input data are ordered differently.
3
4.2
Matrix Multiplication
Matrix multiplication is the most important computation kernel in neural networks [9]. The implementation in this work is based on the mat_mult kernels in CMSIS. Similar to CMSIS implementation,
the matrix multiplication kernel is implemented with 2 × 2 kernels, illustrated in Fig. 5. This enables
some data reuse and saves on the total number of load instructions. The accumulation is done with
the q31_t data type and both operands are of q15_t data type. We initialize the accumulator with
the corresponding bias value. The computation is performed using the dedicated MAC instruction
__SM LAD.
Figure 5: The inner-loop of matrix multiplication with 2 × 2 kernel. Each loop computes the dot
product results of 2 columns and 2 rows, i.e. 4 outputs.
If the input activations or network weights are q7_t type, data expansion may be required to convert
them to q15_t type. As discussed in Section 4.1, if both inputs and weights are q7_t type, we can
use data transformation without reordering (as shown in Fig. 4) to improve performance. However,
the data alignment can be tricky when the number of elements is not a multiple of 4.
The other scenario is with q7_t weights and q15_t activations. In this case, the weights can be
pre-processed with the second and third byte swapped for every 32-bit word, i.e. converting [1, 2, 3, 4]
into [1, 3, 2, 4]. With this pre-processing, the data transformation without reordering (as shown in
Fig. 4) will generate q15_t data in the original order, i.e. converting [1, 3, 2, 4] back to [1, 2, 3, 4]. This
pre-processing is only for the network weights and can be reused for different inputs. Alternatively,
the pre-processing can be performed offline when generating the network model.
We consider the fully-connected layer to have a batch size of one, so the main computation becomes
matrix-vector multiplication. Similar to matrix-matrix multiplication, matrix-vector multiplication
can also be implemented with a 1 × 2 kernel size to improve performance. Supporting a large kernel
can further improve performance, but may be limited by the total number of registers. Arm Cortex-M
cores have 16 architectural registers, including PC and LR. This limited number of registers can be a
challenge if we want to implement larger kernels.
Since the weights are kept constant and re-used during inference, we can reorder the matrix weights
so that row data are interleaved and can be read with only one pointer access. This weight reordering
is illustrated in Fig. 6. During the matrix-vector multiplication, the same q7_to_q15 function, without
reordering, is used to expand the q7_t data into q15_t as shown in Fig. 7. In this way, we can fit the
1 × 4 kernels using the available registers.
4
Figure 6: The weight reordering process to support a 1 × 4 kernel (i.e., 1 column and 4 rows). Weights
are interleaved every four rows and shuffled every four entries in the main part. Only the leftover
columns are interleaved, whereas leftover rows remain the same.
Figure 7: The computation of a 1 × 4 kernel. The data order of the weights and activation vectors is
matched after the q7_to_q15 operation without reordering. Each inner loop iteration processes two
1 × 4 MAC operations.
5
4.3
Convolution
A convolution layer extracts a new feature map by computing a dot product between filter weights
and a small receptive field in the input feature map. Typically, a CPU-based implementation of
convolution is decomposed into input reordering and expanding (i.e. im2col, image-to-column)
and matrix multiplication operations. im2col is a process of transforming the image-like input into
columns that represent the data required by each convolution filter. An example of im2col is shown
in Fig.8.
Figure 8: Example of im2col on a 2D image with a 3x3 kernel, padding size of 1 and stride size of 2.
One of the main challenges with im2col is the increased memory footprint, since the pixels in the
input image are repeated in the im2col output matrix. To alleviate the memory footprint issue while
retaining the performance benefits from im2col, we implemented a partial im2col for our convolution
kernels. The kernel will only expand a limited number of columns (e.g. 2), sufficient to get the
maximum performance boost from the matrix-multiplication kernels while keeping memory overhead
minimal.
The image data format can also affect the performance of convolution, especially im2col efficiency [10]. With a batch size of one, the convolution operation is a 2D convolution (i.e. the
convolution window can move in two directions) on 3D data, as shown in Fig. 9. The two most
common image data formats are Channel-Width-Height (CHW), i.e. channel last, and Height-WidthChannel (HWC), i.e. channel first. The dimension ordering is the same as that of the data stride. In
an HWC format, the data along the channel is stored with a stride of 1, data along the width is stored
with a stride of the channel count, and data along the height is stored with a stride of (channel count
× image width).
1st Stride
Channel
Kernel size
2nd Stride
Height
*
Width
Figure 9: Convolution on 3D data. The image has three dimensions: height, width and channel.
The data layout has no impact on the matrix-multiplication operations, as long as the dimension order
of both weights and images is the same. The im2col operations are performed along the width and
height dimensions only. The HWC-style layout enables efficient data movement, as data for each
6
pixel (i.e. at the same x,y location) is stored contiguously and can be copied efficiently with SIMD
instructions. To validate this, we implemented both CHW and HWC versions and compared the
runtime on a Cortex-M7. The results are highlighted in Fig. 10, where we fixed the HWC input to be
16x16x16 and swept the number of output channels. When the output channel value is zero, it means
that the software performs only im2col and no matrix-multiplication operation. Compared to CHW
layout, HWC has less im2col runtime with the same matrix-multiplication performance. Therefore,
we implement the convolution kernels assuming that the data layout is in HWC format.
CHW
HWC
30
Runtime (ms)
25
20
15
10
5
0
0
4
8
12
16
Output channel count
Figure 10: Experiment results with CHW and HWC data layout. Both data layout styles have the
same matrix-multiplication runtime. HWC has less im2col runtime.
Recently, MobileNets using depthwise separable convolution have been proposed as an efficient
alternative to the standard 3D convolution operation [11]. Depthwise separable convolution has been
used to achieve compact network architectures, which are particularly useful for resource-constrained
devices [12]. CMSIS-NN kernels also include support for a depthwise separable convolution layer.
4.4
Pooling
Pooling layers are typically inserted in between convolution layers to reduce the feature dimensions
and thus the number of parameters and computations in the network. Similar to convolution, pooling
is a window-based operation with a given kernel size, stride and padding. Unlike convolution, pooling
typically operates within the same channel, and is independent of data in the other channels. Pooling
is usually performed with a stride size greater than 1, so the output features will have smaller width
and height.
There are two common types of pooling layers: average pooling, which calculates the average value
of all pixels within the window, and max pooling, which calculates the maximum value. Pooling
can be implemented as a nested for-loop over each window, e.g. pooling layers in Caffe [13]. One
efficient alternative is to split the pooling operation into x-pooling (i.e. along the width) and then
y-pooling (i.e. along the height). This way, the max/average operations along the x-direction can
be reused along the y-direction, allowing the total number of operations to be reduced. We call this
approach split x-y pooling. One potential issue with split x-y pooling is the data arrangement, as
additional memory may be required to store the intermediate results after x-pooling.
Our pooling kernels are implemented with split x-y pooling as, based on our experiments, the split x-y
pooling is significantly faster than the window-based pooling. To eliminate the need for additional
memory, the kernels perform the pooling operations in situ. This makes the pooling layer a destructive
operation on the input. An example of an in situ max pooling implementation on a 1D array is
illustrated in Fig. 11. The same operation can be extended for high dimensional data where each
element can represent an array. For example, when performing x-pooling with HWC image data,
each element block in Fig. 11 can represent an array of size equal to the number of channels, i.e.
all the channel data for one pixel. If each block represents an entire image row, this will effectively
perform y-pooling. Compared to window-based pooling, the in situ split x-y pooling achieves 4.5X
speed-up with no additional memory overhead.
7
Figure 11: Example of max pooling with in situ operations.
4.5
Activation Functions
The role of an activation function is to add non-linearity in the network. The most commonly used
activations functions are ReLU, sigmoid and tanh.
4.5.1
ReLU
Rectified-Linear Unit (ReLU) layer is a common activation function for neural networks. The
operation is f (x) = max(0, x), which is 0 when x < 0 and linear with a slope of 1 when x > 0. A
simple implementation (e.g. in Caffe) loops over all elements and make them 0 if they are negative.
Since we mostly use q7_t type data, ReLU can be implemented using a similar concept as SWAR
(SIMD within a register). The key is to identify the sign bit of the q7_t number and make the
number 0 if it is negative. Our ReLU kernel implementation is shown in Fig. 12. The basic idea is
to use the MSB of the q7_t number as the sign bit and extend it into a mask by using the byte-level
subtraction instruction (__QSU B8). This SWAR approach provides about 4X speed-up compared to
the conventional approach of going over each element.
Figure 12: Illustration of the optimized ReLU kernel. The MSB of each q7_t data is used to construct
a data mask. If the number is negative, i.e. the sign bit is 1, the mask is 0xFF, which will make the
output 0.
4.5.2
Sigmoid and Tanh
Sigmoid and tanh are other common activation functions that are typically used in recurrent neural
networks. Computing these activation functions requires the use of dedicated math functions and can
be computationally expensive on Cortex-M CPUs, hence we implement them using a table-lookup
approach with fixed-point input and output. There are two possible ways to do this, the first of which
is to use a unified table for all ranges with fixed input granularity. In this implementation, the MSB of
the input is used to identify the corresponding entries in the look-up table, while the LSB can be used
for linear interpolation, if needed. This is similar to the sine and cosine table look-up in CMSIS.
The other option is to implement two separate tables to cover different regions of the functions. This
can improve accuracy, as both sigmoid and tanh functions are highly non-linear. Ideally, there should
be a table with finer granularity for an input region around 0 and another table with coarser granularity
8
for inputs with larger absolute values. In this case, the function first determines which table to use
based on the input data range, e.g. by looking at the MSB. After that, similar to the unified table
approach, the LSB can be used for calculating the table entries and interpolation.
Unlike periodic functions such as sine or cosine, it is important to determine the table range for
sigmoid and tanh functions. We determine that range of [−8, 8] works well enough for both functions,
as sigmoid(8) = 0.9997 and tanh(8) = 0.9999. We can also generate other versions of the look-up
table with different ranges, e.g. [−4, 4].
5
Experimental Results
We tested the CMSIS-NN kernels on a CNN trained on the CIFAR-10 dataset, consisting of 60,000
32x32 color images divided into 10 output classes. The network topology is based on the built-in
example provided in Caffe, with three convolution layers and one fully-connected layer. All the
layer weights and activation data are quantized to q7_t format. The layer parameters and the detailed
runtime results using the CMSIS-NN kernels are shown in the Table 1. The runtime is measured on a
NUCLEO-F746ZG Mbed board [14] with an Arm Cortex-M7 core running at 216 MHz.
Table 1: Layer parameters and performance for the CIFAR-10 CNN.
Layer 1
Layer 2
Layer 3
Layer 4
Layer 5
Layer 6
Layer 7
Total
Layer Type
Convolution
Max Pooling
Convolution
Max Pooling
Convolution
Max Pooling
Fully-connected
Filter Shape
5x5x3x32 (2.3 KB)
N.A.
5x5x32x32 (25 KB)
N.A.
5x5x32x64 (50 KB)
N.A.
4x4x64x10 (10 KB)
87 KB weights
Output Shape
32x32x32 (32 KB)
16x16x32 (8 KB)
16x16x32 (8 KB)
8x8x32 (2 KB)
8x8x64 (4 KB)
4x4x64 (1 KB)
10
55 KB activations
Ops
4.9 M
73.7 K
13.1 M
18.4 K
6.6 M
9.2 K
20 K
24.7 M
Runtime
31.4 ms
1.6 ms
42.8 ms
0.4 ms
22.6 ms
0.2 ms
0.1 ms
99.1 ms
The entire image classification takes about 99.1 ms per image (the equivalent of 10.1 images per
second). The compute throughput of the CPU is about 249 MOps per second for running this network.
The pre-quantized network achieves an accuracy of 80.3% on the CIFAR-10 test set. The 8-bit
quantized network running on Arm Cortex-M7 core achieves 79.9% accuracy. Maximum memory
footprint using the CMSIS-NN kernels is ∼133 KB, where convolutions are implemented with partial
im2col to save memory, followed by matrix-multiplication. Memory footprint without partial im2col
would be ∼332 KB and the neural network would not fit on the board.
To quantify the benefits of CMSIS-NN kernels over existing solutions, we also implemented a
baseline version using a 1D convolution function (arm_conv from CMSIS-DSP), Caffe-like pooling
and ReLU. For the CNN application, Table 2 summarizes the comparison results of the baseline
functions and the CMSIS-NN kernels. The CMSIS-NN kernels achieve 2.6X to 5.4X improvement
in runtime/throughput over the baseline functions. The energy efficiency improvement is also in line
with the throughput improvement.
Table 2: Throughput and energy efficiency improvments by layer types
Layer type
Baseline runtime
New kernel runtime
Convolution
Pooling
ReLU
Total
443.4 ms
11.83 ms
1.06 ms
456.4ms
96.4 ms
2.2 ms
0.4 ms
99.1 ms
9
Improvement
Throughput Energy Efficiency
4.6X
4.9X
5.4X
5.2X
2.6X
2.6X
4.6X
4.9X
6
Conclusion
We developed CMSIS-NN to maximize the performance and minimize the memory footprint of
neural networks on Arm Cortex-M CPUs. Neural network inference based on CMSIS-NN kernels
achieved 4.6X improvement in runtime/throughput and 4.9X improvement in energy efficiency for a
convolutional neural network targeting the CIFAR-10 dataset. The CMSIS-NN kernels are available
at https://github.com/ARM-software/CMSIS_5. The application code can directly use these
kernels to implement neural network algorithms on Arm Cortex-M CPUs. Alternatively, these kernels
can be used as primitives by machine learning frameworks to deploy trained models.
References
[1] Philip Sparks. The route to a trillion devices. https://community.arm.com/iot/b/blog/posts/whitepaper-the-route-to-a-trillion-devices.
[2] Weisong Shi, Jie Cao, Quan Zhang, Youhuizi Li, and Lanyu Xu. Edge computing: Vision and
challenges. IEEE Internet of Things Journal, 3(5):637–646, 2016.
[3] https://github.com/ARM-software/CMSIS_5.
[4] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro,
Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale
machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467,
2016.
[5] Liangzhen Lai, Naveen Suda, and Vikas Chandra. Deep convolutional neural network inference
with floating-point weights and fixed-point activations. arXiv preprint arXiv:1703.03073, 2017.
[6] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep
convolutional networks. In International Conference on Machine Learning, pages 2849–2858,
2016.
[7] Naveen Suda, Vikas Chandra, Ganesh Dasika, Abinash Mohanty, Yufei Ma, Sarma Vrudhula,
Jae-sun Seo, and Yu Cao. Throughput-optimized opencl-based fpga accelerator for largescale convolutional neural networks. In Proceedings of the 2016 ACM/SIGDA International
Symposium on Field-Programmable Gate Arrays, pages 16–25. ACM, 2016.
[8] http://www.arm.com/products/processors/cortex-m.
[9] https://petewarden.com/2015/04/20/why-gemm-is-at-the-heart-of-deep-learning/.
[10] Chao Li, Yi Yang, Min Feng, Srimat Chakradhar, and Huiyang Zhou. Optimizing memory
efficiency for deep convolutional neural networks on gpus. In High Performance Computing,
Networking, Storage and Analysis, SC16: International Conference for, pages 633–644. IEEE,
2016.
[11] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias
Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural
networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.
[12] Yundong Zhang, Naveen Suda, Liangzhen Lai, and Vikas Chandra. Hello edge: Keyword
spotting on microcontrollers. arXiv preprint arXiv:1711.07128, 2017.
[13] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick,
Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature
embedding. arXiv preprint arXiv:1408.5093, 2014.
[14] Nucleo-f746zg development board. http://www.st.com/en/evaluation-tools/nucleo-f746zg.html.
10
| 9cs.NE
|
1
Green Stability Assumption: Unsupervised Learning
for Statistics-Based Illumination Estimation
arXiv:1802.00776v1 [cs.CV] 2 Feb 2018
Nikola Banić and Sven Lončarić
Abstract—In the image processing pipeline of almost every
digital camera there is a part dedicated to computational color
constancy i.e. to removing the influence of illumination on the
colors of the image scene. Some of the best known illumination
estimation methods are the so called statistics-based methods.
They are less accurate than the learning-based illumination
estimation methods, but they are faster and simpler to implement
in embedded systems, which is one of the reasons for their
widespread usage. Although in the relevant literature it often
appears as if they require no training, this is not true because
they have parameter values that need to be fine-tuned in order to
be more accurate. In this paper it is first shown that the accuracy
of statistics-based methods reported in most papers was not
obtained by means of the necessary cross-validation, but by using
the whole benchmark datasets for both training and testing. After
that the corrected results are given for the best known benchmark
datasets. Finally, the so called green stability assumption is proposed that can be used to fine-tune the values of the parameters of
the statistics-based methods by using only non-calibrated images
without known ground-truth illumination. The obtained accuracy
is practically the same as when using calibrated training images,
but the whole process is much faster. The experimental results
are presented and discussed. The source code is available at
http://www.fer.unizg.hr/ipg/resources/color constancy/.
Index Terms—Chromaticity, color constancy, Gray-edge, Grayworld, green, illumination estimation, Shades-of-Gray, standard
deviation, unsupervised learning, white balancing.
I. I NTRODUCTION
R
EGARDLESS of the influence of the scene illumination,
the human visual system can recognize object colors
through its ability known as color constancy [1]. In the image
processing pipeline of almost every digital camera there is
also a part dedicated to computational color constancy [2].
It first estimates the scene illumination and then uses it to
chromatically adapt the image i.e. to correct the colors. For
a more formal problem statement an often used image f
formation model written under Lambertian assumption is [3]
Z
fc (x) =
I(λ, x)R(λ, x)ρc (λ)dλ
(1)
ω
where c ∈ {R, G, B} is a color channel, x is a given image
pixel, λ is the wavelength of the light, ω is the visible spectrum, I(λ, x) is the spectral distribution of the light source,
This work has been supported by the Croatian Science Foundation under
Project IP-06-2016-2092.
Copyright (c) 2017 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to [email protected].
The authors are with the Image Processing Group, Department of Electronic Systems and Information Processing, Faculty of Electrical Engineering and Computing, University of Zagreb, 10000 Zagreb, Croatia (email:
[email protected]; [email protected]).
R(λ, x) is the surface reflectance, and ρc (λ) is the camera
sensitivity of color channel c. Assuming uniform illumination
for the sake of simplicity makes it possible to remove x from
I(λ, x) and then the observed light source color is given as
Z
eR
e = eG =
I(λ)ρ(λ)dλ.
(2)
ω
eB
The direction of e provides enough information for successful chromatic adaptation [4]. Still, calculating e is an illposed problem because only image pixel values f are given,
while both I(λ) and ρ(λ) are unknown. The solution to this
problem is to make additional assumptions. Different assumptions have given rise to numerous illumiantion estimation
methods that can be divided into two main groups. First
of these groups contains low-level statistics-based methods
such as White-patch [5], [6] and its improvements [7], [8],
[9], Gray-world [10], Shades-of-Gray [11], Grey-Edge (1st
and 2nd order) [12], Weighted Gray-Edge [13], using bright
pixels [14], using bright and dark colors [15]. The second
group includes learning-based methods such as gamut mapping (pixel, edge, and intersection based) [16], using neural networks [17], using high-level visual information [18],
natural image statistics [19], Bayesian learning [20], spatiospectral learning (maximum likelihood estimate, and with gen.
prior) [21], simplifying the illumination solution space [22],
[23], [24], using color/edge moments [25], using regression
trees with simple features from color distribution statistics [26], performing various kinds of spatial localizations [27],
[28], using convolutional neural networks [29], [30], [31].
Statistics-based illumination estimation methods are less
accurate than the learning-based ones, but they are faster and
simpler to implement in embedded systems, which is one of
the reasons for their widespread usage [32]. Although in the
relevant literature it often appears as if they require no training,
this is not true because they have parameter values that need
to be fine-tuned in order to give higher accuracy. In this paper
it is first shown that in most papers on illumination estimation
the accuracy of statistics-based methods was not obtained by
means of the necessary cross-validation, but by using the
whole benchmark datasets for both training and testing, which
leads to an unfair comparison between the methods. After that
the corrected results are given for the best known benchmark
datasets by performing the same cross-validation framework as
for other learning-based methods. Finally, the so called green
stability assumption is proposed that can be used to fine-tune
the values of the parameters of the statistics-based methods by
using only non-calibrated images without known ground-truth
illumination. The obtained accuracy is practically the same as
2
when using calibrated training images, but the whole process
is much faster and it can be directly applied in practice.
The paper is structured as follows: Section II briefly describes the best known statistics-based methods, Section III
shows that their accuracy data should be revisited, Section IV
proposes the green stability assumption, Section V presents
the results, and finally, Section VI concludes the paper.
II. B EST KNOWN STATISTICS - BASED METHODS
Some of the best known statistics-based illumination estimation methods are centered around the Gray-world assumption
and its extensions. Under this assumption the average scene
reflectance is achromatic [10] and e is therefore calculated as
R
f(x)dx
R
= ke
(3)
dx
where k ∈ [0, 1] is the reflectance amount with 0 meaning
no reflectance and 1 meaning total reflectance. By adding
the Minkowski norm p to Eq. (3), the Gray-world method
is generalized into the Shades-of-Gray method [11]:
1
R
p
(f(x)) dx p
R
= ke.
(4)
dx
Having p = 1 results in Gray-world, while p → ∞ results
in White-patch [5], [6]. In [12] Eq. (4) was extended to the
general Gray-world by introducing local smoothing:
R σ
1
p
(f (x)) dx p
R
= ke
(5)
dx
where fσ = f ∗ Gσ and Gσ is a Gaussian filter with standard
deviation σ. Another significant extension is the Grey-edge
assumption, under which the scene reflectance differences
calculated with derivative order n are achromatic [12] so that
Z
p1
p
∂ n (fσ (x))
dx
= ke.
(6)
∂nx
The described Shades-of-Gray, general Gray-world, and
Gray-edge methods have parameters and the methods’ accuracy depends on how the values of these parameters are tuned.
Nevertheless, in the literature it often appears as if they require
no training [3], [15], which is then said to be an advantage.
It may be argued that the parameter values are in most cases
the same, but this is easily disproved. In [15] for methods
mentioned in this section the best fixed parameter values were
given for ten different datasets. These values are similar for
some datasets, but overall they span two orders of magnitude.
With such high differences across different datasets in mind,
it is obvious that the parameter values have to be learned.
III. VALIDATION REVISITED
A. Angular error
Before recalculating the accuracy of the methods from the
previous section, some introduction to the used measures
is needed. From various proposed illumination estimation
accuracy measures [33], [34], [35], the angular error is most
commonly used. It represents the angle between the illumination estimation vector and the ground-truth illumination.
All angular errors obtained for a given method on a chosen
dataset are usually summarized by different statistics. Because
of the non-symmetry of the angular error distribution, the most
important of these statistics is the median angular error [36].
Angular errors below 3◦ are considered acceptable [37], [38].
The ground-truth illuminations of benchmark dataset images
are obtained by reading off calibration objects put in the image
scene, e.g. a gray ball or a color checker. When a method is
tested, these objects are masked out to prevent possible bias.
B. The need for cross-validation
When it comes to accuracy obtained on benchmark datasets,
the ones available in [3], at [39], and in [27] are the most
widely copied and referenced. If, for example, the results
obtained for Shades-of-Gray on the GreyBall dataset [40] are
checked [39], the reported mean and median angular errors of
6.1◦ and 5.3◦ , respectively, are obtained by setting p to 12 on
all 15 folds. However, by performing cross-validation i.e. by
looking for the best p on 14 training folds, applying this to
the test fold, and repeating it all 15 times clearly shows that p
differs for various training sets. Overall, the mean and median
errors for combined results of all test folds are 7.8◦ and 7.2◦ ,
respectively, which differs from the reported results. Similar
differences can be shown for other methods from Section II as
well. For the sake of fair comparison with other illumination
estimation methods, these accuracies are properly recalculated
and they are provided in Section V together with other results.
IV. T HE PROPOSED ASSUMPTION
A. Practical application
The methods mentioned in Section II are some of the most
widely used illumination estimation methods [32] and this
means that their parameters should preferably be appropriately
fine-tuned before putting them in production. The best way
to do this is to use a benchmark dataset, but because of
dependence of Eq. (2) on ρ(λ), a benchmark dataset would
be required for each used camera sensor. Since putting the
calibration objects into image scenes and later extracting the
ground-truth illumination is time consuming, it would be
better, if possible, to perform some kind of unsupervised
learning on non-calibrated images without known ground-truth
illumination. This would save time and be of practical value.
B. Motivation
When for a dataset the ground-truth illuminations are
unknown, an alternative is to make assumptions about the
nature of illumination estimations produced by statistics-based
methods when their parameters are fine-tuned and then to
meet the conditions of the assumptions. When considering the
nature of illumination estimations, a good starting point is the
observation that some statistics-based illumination estimations
appear ”to correlate roughly with the actual illuminant” [25].
Fig. 1 shows this for the images of the GreyBall dataset [40].
The points in Fig. 1 can be considered to occupy a space
around a line in the rb-chromaticity [22] which is connected
3
Fig. 1: The rb-chromaticities of the ground-truth
illuminations and Shades-of-Gray illumination estimations
for GreyBall dataset images [40] (best viewed in color).
to the fact that the green chromaticity of the ground-truth illuminations is relatively stable and similar for all illuminations.
For the GreyBall dataset the standard deviations of the red,
green, and blue chromaticity components of the ground-truth
illuminations are 0.0723, 0.0106, and 0.0750, respectively, and
similar results are obtained for all other datasets. For Shadesof-Gray illumination estimations shown in Fig. 1 the red,
green, and blue chromaticity components of the ground-truth
illuminations are 0.0842, 0.0253, and 0.0770, respectively,
which means that there is also a trend of green chromaticity
stability, although the standard deviation is greater than in the
case of ground-truth illumination. This means that if a set of
illumination estimations is to resemble the set of ground-truth
illuminations, the estimations’ green chromaticity standard
deviation should also be smaller and closer to the one of the
ground-truth. As a matter of fact, if for example the Shadesof-Gray illumination estimations for p = 2 and p = 15 shown
in Fig. 2 are compared, the standard deviations of their green
chromaticities are 0.0253 and 0.0158, respectively, while their
median angular errors are 6.2◦ and 5.3◦ , respectively. Similar
behaviour where lower green chromaticity standard deviation
is to some degree followed by lower median angular error can
be seen on all datasets and for all methods from Section II.
each method M ∈ M, where M contains all methods from
Section II, the Cartesian product of discrete sets of evenly
spread values for individual parameters of M was calculated
(i)
to get n tuples pM , i ∈ {1, 2, . . . , n}. Gray-world and Whitepatch have no parameters, but they were implicitly included as
(i)
special cases of Shades-of-Gray. Second, each pM was used
to set the parameter values of M and then M was applied to
all images of the GreyBall dataset to obtain an illumination
estimation for each of them. Third, for these illumination
estimations the standard deviation of their green chromaticities
σi and their median
angular error mi were calculated. Fourth,
for every of n2 possible pairs of indices i, j ∈ {1, 2, . . . , n}
such that i < j a new difference pair {∆σk , ∆mk } was
calculated such that ∆σk = σi − σj and ∆mk = mi − mj .
Finally, all such difference pairs created for all M ∈ M were
put together into set of pairs P. If members of pairs in P are
interpreted as coordinates, then their plot is shown in Fig. 3.
Fig. 3: Relation between difference in standard deviations of
illumination estimations’ green chromaticity and the
difference in illumination estimations’ median angular error.
C. Green stability assumption
The value of Person’s linear correlation coefficient for the
points in Fig. 3 is 0.7408, which indicates a strong positive
linear relationship [41]. In other words, the difference between
the standard deviations of green chromaticities of illumination
estimations produced by the same method when using different
parameter values is strongly correlated to the difference between median angular errors of these illumination estimations.
The same correlations for NUS datasets [15] are in Table I.
TABLE I: Correlation between difference in green
chromaticity standard deviation and difference in median
angular errors for NUS datasets [15].
Dataset
Correlation
Fig. 2: The rb-chromaticities of different Shades-of-Gray
illumination estimations for GreyBall dataset images [40]
(best viewed in color).
For a deeper insight into this behaviour, another experiment was conducted on the GreyBall dataset [40]. First, for
C1
0.9255
C2
0.6381
Fuji
0.8977
N52
0.9443
Oly
0.8897
Pan
0.9644
Sam
0.8902
Sony
0.9095
Based on this empirical results and observations, it is
possible to introduce the green stability assumption: the parameter values for which a method’s illumination estimations’
green chromaticity standard deviation is lower simultaneously
lead to lower illumination estimation errors. Like many other
assumptions, this assumption does not always hold, but it can
still be useful in cases when the ground-truth illuminations for
4
a set of images taken with a given sensor are not available.
These images should also be taken under similar illuminations
as the mentioned datasets that were used for empirical results.
For a specific case when the parameter values of a chosen
method are fine-tuned and only non-calibrated images are
available, the green stability assumption can be expressed more
formally. If n is the number of images in the training set, pi
is the i-th vector of parameter values, ei,j is the method’s
illumination estimation obtained for the j-th image when pi
is used for parameter values, ei,j
G is the green component of
ei,j , and eiG is the mean green component of illumination
estimations for all images obtained with parameters pi , then
under the green stability assumption the index i∗ of such pi∗
that should result in minimal angular errors is obtained as
v
2
uP
i,j
u n
i
e
−
e
t
G
G
j=1
.
(7)
i∗ = arg min
n
−
1
i
Since Eq. (7) performs minimization of standard deviation,
it can also be written without the square and the denominator.
TABLE II: Combined accuracy on eight NUS dataset (lower
Avg. is better). The used format is the same as in [28].
Algorithm
Mean Med. Tri.
Originally reported results
Shades-of-Gray [11]
3.67 2.94 3.03
General Gray-World [4]
3.20 2.56 2.68
1st-order Gray-Edge [12]
3.35 2.58 2.76
2nd-order Gray-Edge [12]
3.36 2.70 2.80
Revisited results
Shades-of-Gray [11]
3.48 2.63 2.81
General Gray-World [4]
3.37 2.49 2.61
1st-order Gray-Edge [12]
3.12 2.19 2.39
2nd-order Gray-Edge [12]
3.15 2.23 2.42
Green stability assumption results
Shades-of-Gray [11]
3.44 2.65 2.81
General Gray-World [4]
3.40 2.63 2.76
1st-order Gray-Edge [12]
3.29 2.36 2.55
2nd-order Gray-Edge [12]
3.29 2.44 2.59
Best Worst Avg.
25% 25%
0.98
0.85
0.79
0.89
7.75
6.68
7.18
7.14
3.01
2.63
2.67
2.76
0.81
0.73
0.71
0.74
7.62
7.58
7.11
7.13
2.76‘
2.61
2.42
2.46
0.83
0.77
0.79
0.83
7.41
7.42
7.36
7.30
2.75
2.69
2.58
2.63
TABLE III: Accuracy on the original GreyBall dataset
(lower median is better).
method
mean (◦ )
median (◦ )
Originally reported results
Shades-of-Gray [11]
6.14
5.33
General Gray-World [4]
6.14
5.33
1st-order Gray-Edge [12]
5.88
4.65
2nd-order Gray-Edge [12]
6.10
4.85
Revisited results
Shades-of-Gray [11]
7.80
7.15
General Gray-World [4]
7.61
6.85
1st-order Gray-Edge [12]
6.14
5.32
2nd-order Gray-Edge [12]
6.89
5.84
Green stability assumption results
Shades-of-Gray [11]
6.80
5.30
General Gray-World [4]
6.80
5.30
1st-order Gray-Edge [12]
5.97
4.64
2nd-order Gray-Edge [12]
6.69
5.17
TABLE IV: Accuracy on the linear GreyBall dataset (lower
median is better).
method
mean (◦ )
median (◦ )
Originally reported results
Shades-of-Gray [11]
11.55
9.70
General Gray-World [4]
11.55
9.70
1st-order Gray-Edge [12]
10.58
8.84
2nd-order Gray-Edge [12]
10.68
9.02
Revisited results
Shades-of-Gray [11]
13.32
11.57
General Gray-World [4]
13.69
12.11
1st-order Gray-Edge [12]
11.06
9.54
2nd-order Gray-Edge [12]
10.73
9.21
Green stability assumption results
Shades-of-Gray [11]
12.68
10.50
General Gray-World [4]
12.68
10.50
1st-order Gray-Edge [12]
13.41
11.04
2nd-order Gray-Edge [12]
12.83
10.70
trimean (◦ )
10.23
10.23
9.18
9.40
12.10
12.55
9.81
9.49
11.25
11.25
11.87
11.44
newly calculated accuracy results for methods mentioned in
Section II and to test the effectiveness of the proposed green
stability assumption: the GreyBall dataset [40], its approximated linear version, and eight linear NUS dataset [15]. The
ColorChecker dataset [20], [42] was not used because of
its confusing history of wrong usage despite warnings from
leading experts [43]. Except the original GreyBall dataset,
all other contain linear images, which is preferred because
illumination estimation is in cameras usually performed on
linear images [2] similar to the model described by Eq. (1).
The tested methods include all the ones from M. During
cross-validation on all datasets the same folds were used as in
other publications. The source code for recreating the numerical results given in the following subsection is publicly available at http://www.fer.unizg.hr/ipg/resources/color constancy/.
B. Accuracy
Tables II, III, and IV show the previously reported accuracies, the newly recalculated accuracies, and the accuracies
obtained by using the green stability assumption. The results
clearly confirm the potential and the practical applicability
of the green stability assumption. This also demonstrates the
success of unsupervised learning for illumination estimation.
trimean (◦ )
5.51
5.51
5.11
5.28
7.21
6.92
5.49
6.06
5.77
5.77
5.10
5.72
V. E XPERIMENTAL RESULTS
A. Experimental setup
The following benchmark datasets have been used to
demonstrate the difference between previously reported and
VI. C ONCLUSIONS AND FUTURE RESEARCH
In most relevant papers the accuracy results for some of
the most widely used statistics-based methods were calculated without cross-validation. Here it was shown that crossvalidation is needed and the accuracy results were revisited.
When statistics-based methods are fine-tuned, the best way to
do this is by using images with known ground-truth illumination. Based on several observations and empirical evidence,
the green stability assumption has been proposed that can be
successfully used to fine-tune the parameters of statistics-based
methods when only non-calibrated images without groundtruth illumination are available. This makes the whole finetuning process much simpler, faster, and more practical. It is
also an unsupervised learning approach to color constancy.
In future, other similar bases for further assumptions for
unsupervised learning for color constancy will be researched.
5
R EFERENCES
[1] M. Ebner, Color Constancy, ser. The Wiley-IS&T Series in Imaging
Science and Technology. Wiley, 2007.
[2] S. J. Kim, H. T. Lin, Z. Lu, S. Süsstrunk, S. Lin, and M. S. Brown, “A
new in-camera imaging model for color computer vision and its application,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 34, no. 12, pp. 2289–2302, 2012.
[3] A. Gijsenij, T. Gevers, and J. Van De Weijer, “Computational color constancy: Survey and experiments,” Image Processing, IEEE Transactions
on, vol. 20, no. 9, pp. 2475–2489, 2011.
[4] K. Barnard, V. Cardei, and B. Funt, “A comparison of computational
color constancy algorithms. i: Methodology and experiments with synthesized data,” Image Processing, IEEE Transactions on, vol. 11, no. 9,
pp. 972–984, 2002.
[5] E. H. Land, The retinex theory of color vision. Scientific America.,
1977.
[6] B. Funt and L. Shi, “The rehabilitation of MaxRGB,” in Color and
Imaging Conference, vol. 2010, no. 1. Society for Imaging Science
and Technology, 2010, pp. 256–259.
[7] N. Banić and S. Lončarić, “Using the Random Sprays Retinex Algorithm
for Global Illumination Estimation,” in Proceedings of The Second
Croatian Computer Vision Workshopn (CCVW 2013). University of
Zagreb Faculty of Electrical Engineering and Computing, 2013, pp. 3–
7.
[8] ——, “Color Rabbit: Guiding the Distance of Local Maximums in
Illumination Estimation,” in Digital Signal Processing (DSP), 2014 19th
International Conference on. IEEE, 2014, pp. 345–350.
[9] ——, “Improving the White patch method by subsampling,” in Image
Processing (ICIP), 2014 21st IEEE International Conference on. IEEE,
2014, pp. 605–609.
[10] G. Buchsbaum, “A spatial processor model for object colour perception,”
Journal of The Franklin Institute, vol. 310, no. 1, pp. 1–26, 1980.
[11] G. D. Finlayson and E. Trezzi, “Shades of gray and colour constancy,” in
Color and Imaging Conference, vol. 2004, no. 1. Society for Imaging
Science and Technology, 2004, pp. 37–41.
[12] J. Van De Weijer, T. Gevers, and A. Gijsenij, “Edge-based color
constancy,” Image Processing, IEEE Transactions on, vol. 16, no. 9,
pp. 2207–2214, 2007.
[13] A. Gijsenij, T. Gevers, and J. Van De Weijer, “Improving color constancy by photometric edge weighting,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 34, no. 5, pp. 918–929, 2012.
[14] H. R. V. Joze, M. S. Drew, G. D. Finlayson, and P. A. T. Rey,
“The Role of Bright Pixels in Illumination Estimation,” in Color and
Imaging Conference, vol. 2012, no. 1. Society for Imaging Science
and Technology, 2012, pp. 41–46.
[15] D. Cheng, D. K. Prasad, and M. S. Brown, “Illuminant estimation for
color constancy: why spatial-domain methods work and the role of the
color distribution,” JOSA A, vol. 31, no. 5, pp. 1049–1058, 2014.
[16] G. D. Finlayson, S. D. Hordley, and I. Tastl, “Gamut constrained illuminant estimation,” International Journal of Computer Vision, vol. 67,
no. 1, pp. 93–109, 2006.
[17] V. C. Cardei, B. Funt, and K. Barnard, “Estimating the scene illumination
chromaticity by using a neural network,” JOSA A, vol. 19, no. 12, pp.
2374–2386, 2002.
[18] J. Van De Weijer, C. Schmid, and J. Verbeek, “Using high-level visual
information for color constancy,” in Computer Vision, 2007. ICCV 2007.
IEEE 11th International Conference on. IEEE, 2007, pp. 1–8.
[19] A. Gijsenij and T. Gevers, “Color Constancy using Natural Image
Statistics.” in CVPR, 2007, pp. 1–8.
[20] P. V. Gehler, C. Rother, A. Blake, T. Minka, and T. Sharp, “Bayesian
color constancy revisited,” in Computer Vision and Pattern Recognition,
2008. CVPR 2008. IEEE Conference on. IEEE, 2008, pp. 1–8.
[21] A. Chakrabarti, K. Hirakawa, and T. Zickler, “Color constancy with
spatio-spectral statistics,” Pattern Analysis and Machine Intelligence,
IEEE Transactions on, vol. 34, no. 8, pp. 1509–1519, 2012.
[22] N. Banić and S. Lončarić, “Color Cat: Remembering Colors for Illumination Estimation,” Signal Processing Letters, IEEE, vol. 22, no. 6, pp.
651–655, 2015.
[23] ——, “Using the red chromaticity for illumination estimation,” in Image
and Signal Processing and Analysis (ISPA), 2015 9th International
Symposium on. IEEE, 2015, pp. 131–136.
[24] N. Banić and S. Lončarić, “Color Dog: Guiding the Global Illumination
Estimation to Better Accuracy,” in VISAPP, 2015, pp. 129–135.
[25] G. D. Finlayson, “Corrected-moment illuminant estimation,” in Proceedings of the IEEE International Conference on Computer Vision, 2013,
pp. 1904–1911.
[26] D. Cheng, B. Price, S. Cohen, and M. S. Brown, “Effective learningbased illuminant estimation using simple features,” in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2015,
pp. 1000–1008.
[27] J. T. Barron, “Convolutional Color Constancy,” in Proceedings of the
IEEE International Conference on Computer Vision, 2015, pp. 379–387.
[28] J. T. Barron and Y.-T. Tsai, “Fast Fourier Color Constancy,” in Computer
Vision and Pattern Recognition, 2017. CVPR 2017. IEEE Computer
Society Conference on, vol. 1. IEEE, 2017.
[29] S. Bianco, C. Cusano, and R. Schettini, “Color Constancy Using CNNs,”
in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition Workshops, 2015, pp. 81–89.
[30] W. Shi, C. C. Loy, and X. Tang, “Deep Specialized Network for
Illuminant Estimation,” in European Conference on Computer Vision.
Springer, 2016, pp. 371–387.
[31] Y. Hu, B. Wang, and S. Lin, “Fully Convolutional Color Constancy
with Confidence-weighted Pooling,” in Computer Vision and Pattern
Recognition, 2017. CVPR 2017. IEEE Conference on. IEEE, 2017, pp.
4085–4094.
[32] Z. Deng, A. Gijsenij, and J. Zhang, “Source camera identification using
Auto-White Balance approximation,” in Computer Vision (ICCV), 2011
IEEE International Conference on. IEEE, 2011, pp. 57–64.
[33] A. Gijsenij, T. Gevers, and M. P. Lucassen, “Perceptual analysis of
distance measures for color constancy algorithms,” JOSA A, vol. 26,
no. 10, pp. 2243–2256, 2009.
[34] G. D. Finlayson and R. Zakizadeh, “Reproduction angular error: An
improved performance metric for illuminant estimation,” perception, vol.
310, no. 1, pp. 1–26, 2014.
[35] N. Banić and S. Lončarić, “A Perceptual Measure of Illumination
Estimation Error,” in VISAPP, 2015, pp. 136–143.
[36] S. D. Hordley and G. D. Finlayson, “Re-evaluating colour constancy
algorithms,” in Pattern Recognition, 2004. ICPR 2004. Proceedings of
the 17th International Conference on, vol. 1. IEEE, 2004, pp. 76–79.
[37] G. D. Finlayson, S. D. Hordley, and P. Morovic, “Colour constancy
using the chromagenic constraint,” in Computer Vision and Pattern
Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on,
vol. 1. IEEE, 2005, pp. 1079–1086.
[38] C. Fredembach and G. Finlayson, “Bright chromagenic algorithm for
illuminant estimation,” Journal of Imaging Science and Technology,
vol. 52, no. 4, pp. 40 906–1, 2008.
[39] T. G. A. Gijsenij and J. van de Weijer. (2017) Color Constancy
— Research Website on Illuminant Estimation. [Online]. Available:
http://colorconstancy.com/
[40] F. Ciurea and B. Funt, “A large image database for color constancy
research,” in Color and Imaging Conference, vol. 2003, no. 1. Society
for Imaging Science and Technology, 2003, pp. 160–164.
[41] D. J. Rumsey, U Can: statistics for dummies. John Wiley & Sons,
2015.
[42] B. F. L. Shi. (2015, May) Re-processed Version of the Gehler
Color Constancy Dataset of 568 Images. [Online]. Available:
http://www.cs.sfu.ca/colour/data/
[43] S. E. Lynch, M. S. Drew, and k. G. D. Finlayson, “Colour Constancy
from Both Sides of the Shadow Edge,” in Color and Photometry in
Computer Vision Workshop at the International Conference on Computer
Vision. IEEE, 2013.
| 1cs.CV
|
ENDOMORPHISM FIELDS OF ABELIAN VARIETIES
arXiv:1606.02803v2 [math.NT] 4 Jun 2017
ROBERT GURALNICK AND KIRAN S. KEDLAYA
Abstract. We give a sharp divisibility bound, in terms of g, for the degree of the field
extension required to realize the endomorphisms of an abelian variety of dimension g over
an arbitrary number field; this refines a result of Silverberg. This follows from a stronger
result giving the same bound for the order of the component group of the Sato-Tate group
of the abelian variety, which had been proved for abelian surfaces by Fité–Kedlaya–Rotger–
Sutherland. The proof uses Minkowski’s reduction method, but with some care required in
the extremal cases when p equals 2 or a Fermat prime.
1. Introduction
For A an abelian variety over a field K, the endomorphism field of A is the minimal
algebraic extension L of K such that EndpAL q “ EndpAL q. The purpose of this paper is to
establish a bound on the degree rL : Ks in terms of the dimension of A; more precisely, we
compute the LCM of all possible degrees as A, K vary while dimK A remains fixed.
Before stating our result, we state a prior result of Silverberg [8] which already contains
many of the main ideas. For g a positive integer and p a prime, define
^
8 Z
ÿ
2g
rpg, pq :“
.
i
pp
´
1qp
i“0
Theorem 1.1 (Silverberg). For A an abelian variety of dimension g over a number
K,
śfield
rpg,pq
the endomorphism field of A is a finite Galois extension of K of degree dividing 2ˆ p p
.
The proof [8, Theorem 4.1] is elegantly simple: one verifies that for each prime ℓ ą 2, the
Galois group of the endomorphism field extension is isomorphic (via its action on ℓ-torsion
points) to a subquotient of the group Spp2g, Fℓq. The bound is then obtained by taking the
greatest common divisor of the orders of these finite groups. This echoes the method used
by Minkowski to bound the order of a finite group of integer matrices (as exposed in [3]); we
will return to this analogy a bit later.
From this proof, it is not clear whether one should expect the bound of Theorem 1.1 to be
sharp. However, a moment’s thought shows that it is not sharp for g “ 1: the bound is 24 ˆ 3
but the optimal bound is obviously 2, with the worst case being an elliptic curve whose CM
field is not contained in K. More seriously, for g “ 2, the bound of Theorem 1.1 is 28 ˆ 32 ˆ 5,
but (as will be explained shortly) the work of Fité–Kedlaya–Rotger–Sutherland [2] implies
that the optimal bound is 24 ˆ 3. This raises the question of identifying the discrepancy
between Theorem 1.1 and the optimal bound, and this is achieved by our main result.
Date: May 30, 2017.
Thanks to Francesc Fité, Gaël Rémond, Jean-Pierre Serre, Alice Silverberg, and Andrew Sutherland for
feedback. Guralnick was partially supported by NSF grants DMS-1302886 and DMS-1600056. Kedlaya was
supported by NSF grant DMS-1501214 and UC San Diego (Warschawski Professorship).
1
Theorem 1.2. For A an abelian variety of dimension g ą 0 over a number field K, the
degree over K of the endomorphism field of A divides
$
’
if p “ 2;
&rpg, pq ´ g ´ 1
ź 1
r pg,pq
1
p
,
r pg, pq :“ maxt0, rpg, pq ´ 1u if p is a Fermat prime;
’
%rpg, pq
p
otherwise.
Moreover, for fixed g, p (but varying over all K), this value of r 1 pg, pq is best possible.
To give one more example, for g “ 3, the bound from Theorem 1.1 is 211 ˆ 34 ˆ 5 ˆ 7
whereas the optimal bound is 26 ˆ 33 ˆ 7. It is easy to see that the factor of 7 is necessary,
e.g., by considering twists of the Klein quartic (see [5, §4]).
As in [2], our approach uses the relationship between the endomorphism field L of A
and the Sato-Tate group of A; the latter is a compact Lie group whose component group
surjects canonically onto GalpL{Kq, and the bound we ultimately prove is for the order of the
component group (see Theorem 5.4). The Sato-Tate group is constructed as a compact form
of a certain linear algebraic group over Q, the algebraic Sato-Tate group, which allows us to
bound the order of the component group using a variant of Minkowski’s method. One key
point is that the extremal cases occur for CM abelian varieties, for which the connected part
of the algebraic Sato-Tate group is a torus which splits over a CM field; what distinguishes a
Fermat prime p in this context is that Qpµp q contains no proper subfield which is CM. (For
p “ 2, the same statement about Qpµ4 q plays an analogous role.)
To prove that Theorem 1.2 is sharp, we use the relationship between twisting of abelian
varieties and Sato-Tate groups; this reduces the problem to exhibiting abelian varieties admitting actions by large finite groups, which we achieve using CM abelian varieties and the
same wreath product construction as in Minkowski’s theorem. Note that the fields of definition of the resulting abelian varieties are controlled by class groups of abelian number fields,
so we are unable to establish lower bounds over any fixed number field.
To conclude this introduction, we comment on the subtler problem of giving bounds by
size, rather than divisibility. As described in [3, §6], results of Weisfeiler and Feit can be
combined with the classification of finite simple groups to show that for n ą 10, the largest
finite subgroups of GLpn, Qq have order 2n n! (and are unique up to conjugacy). For abelian
varieties of sufficiently large dimension g, one would expect that the largest possible endomorphism field extension, and the largest possible component group of the Sato-Tate group, are
obtained by twisting a power of an elliptic curve with j-invariant 0 using an automorphism
group of order 6g g! (note that these examples already occur over Q). For the endomorphism
field, this expectation has been confirmed by work of Rémond [6]; it is highly likely that
a similar analysis applies to the component group (because the cases where the two differ
tend not to have enough CM to trouble the bounds), but this would require some additional
argument.
2. Group schemes
We start with some notation. Our notation choices are not entirely typical; they are made
to help us distinguish between groups and group schemes.
Definition 2.1. For G a group scheme (over some base), we write GX for the base extension
of G to the base scheme X, and GpXq for the group of X-valued points of G. We say that
2
G is pointful over X if GpXq occupies a Zariski-dense subset of G. By convention, all group
schemes we consider will be smooth and linear; the standard linear groups will be considered
as schemes over Z, and we will write GLpn, Xq instead of GLpnqpXq and so on.
For G a group scheme over a connected base, let G˝ denote the identity connected component, and write π0 pGq :“ G{G˝ for the group of connected components, viewed as a finite
group scheme over the same base. If G is pointful, then so are both G˝ and G{G˝ .
For L{K a finite extension of fields and G a group scheme over Spec L, write ResLK G for
the Weil restriction of scalars of G to Spec K.
Example 2.2. For n a positive integer, the n-torsion subscheme of the multiplicative group
over Q is the group scheme Spec Qrxs{pxn ´ 1q which is obviously defined over Q. However,
it is only pointful for n “ 1, 2.
Definition 2.3. For G a group scheme, let OutpGq be the group scheme of outer automorphisms of G, i.e., the cokernel of the map G Ñ AutpGq induced by conjugation. Note that for
G a group scheme over a field k which is not algebraically closed, an element of AutpGqpkq
may map trivially to OutpGqpkq even though it does not come from the image of Gpkq; that
is, the scheme-theoretic notion of an outer automorphism disagrees with the group-theoretic
notion because the latter is not stable under base extension.
3. Minkowski’s method
We next formulate our version of Minkowski’s reduction method. We implicitly follow [3],
but see also [7] for another detailed treatment (both considering only finite groups).
Throughout §3, let G be a group scheme over a number field K.
Definition 3.1. By convention G is a scheme of finite type over K, so it can be extended
to a group scheme over oK r1{Ns for some positive integer N. In particular, it makes sense
to form the base extension of G to Fq :“ oK {q for all but finitely many prime ideals q of oK .
Let H be a finite pointful subquotient group scheme of G. By the previous paragraph,
HpKq is isomorphic to a subquotient of GpFq q for all but finitely many q; in particular, for
each prime p, any p-Sylow subgroup of HpKq is isomorphic to a subquotient of a p-Sylow
subgroup of GpFq q for all but finitely many q.
To translate this into a numerical bound, define the nonnegative integers rpG, pq by the
formula
ź
prpG,pq “ suptgcd #GpFq qu
p
S
qPS
where S runs over all cofinite
of prime ideals of oK . Then the preceding discussion implies
ś sets
rpG,pq
that H has order dividing p p
; in particular, this bound applies to the component group
of any pointful subgroup scheme of G.
We collect some remarks related to this construction.
Remark 3.2. Suppose that there exists a finite pointful subquotient group scheme H of G
such that the p-part of #H equals the upper bound prpG,pq . Let P be a p-Sylow subgroup of
H; then for infinitely many q, P has the same cardinality as a p-Sylow subgroup of GpFq q,
so by Sylow’s theorem the two must be isomorphic. That is, the isomorphism type of P is
uniquely determined by G.
3
Remark 3.3. Let H be a pointful subgroup scheme of G. Then #π0 pHq divides
for
ź
pδpG,pq “ inf tgcd #H ˝ pFq qu.
S
p
ś
p
prpG,pq´δpG,pq
qPS
Remark 3.4. In light of Wedderburn’s theorem, for any number field K and any twisted
form G of GLpnqK one has rpG, pq “ rpGLpnqK , pq.
Remark 3.5. For each prime p, the group Cp ≀ Stn{pp´1qu embeds into GLpn, Qq, as then do
its p-Sylow subgroups. The original theorem of Minkowski asserts that the conjugates of the
latter are the largest possible p-subgroups of GLpn, Qq. However, if we compare this to the
values
Z
^ Z
^ Z
^
n
n
n
rpGLpnqQ , pq “
`
` 2
` ¨¨¨
pp ą 2q
p´1
ppp ´ 1q
p pp ´ 1q
Yn] Yn]
`
` ¨¨¨ ,
rpGLpnqQ , 2q “ n ` 2
2
4
we see that theX Minkowski
bound is only sharp for p ą 2; for p “ 2, the Minkowski bound
\
n
is too large by 2 , and one must supplement using some extra analysis involving quadratic
forms over finite fields [3, §5]. For this reason, we do not know whether the example Cp ≀
Stn{pp´1qu is optimal also for subquotients when p “ 2.
Remark 3.6. For K a number field, the Chebotarev density theorem implies that rpGLpnqK , pq
depends only on K X Qpµp8 q. For
mpK, pq :“ mintm ě 1 : K X Qpµpm q “ K X Qpµp8 qu
for p ą 2 we have
(3.1)
tpK, pq :“ rQpµpmpK,pq q : K X QpµpmpK,pq qs,
^ Z
^ Z
^
n
n
n
`
` 2
` ¨¨¨
rpGLpnqK , pq “ mpK, pq
tpK, pq
ptpK, pq
p tpK, pq
Z
and this bound is again optimal (see [3, §5.3] for a detailed discussion).
For p “ 2, the situation depends crucially on whether Qpζ4 q Ď K. If so, then tpK, 2q “ 1,
Yn] Yn]
`
` ¨¨¨ ,
(3.2)
rpGLpnqK , 2q “ mpK, 2qn `
2
4
Sn . If not, the situation is more complicated;
and this bound is optimal, achieved by C2mpK,2q ≀ ?
we limit ourselves?to observing that for K “ Qp ´2q we have rpGLpnq
K , 2q “ rpGLpnqQ , 2q,
X \
while for K “ Qp 2q we have rpGLpnqK , 2q “ rpGLpnqQ , 2q ` n2 .
Remark 3.7. Although we do not need this for our main result, we note in passing the
following corollary of Remark 3.6 (which only affects the prime p “ 2): the
divisibility
ś optimal
rpGLpgqQpiq ,2q
rpGLp2gq,pq
bound for the order of a finite subgroup of Spp2g, Qq is 2
. This
pą2 p
comes down to the fact that any irreducible finite subgroup of Spp2g, Qq is centralized by
some totally imaginary number field [4, Lemma 2.3].
Remark 3.8. It is natural to use Minkowski’s method as a starting point for bounding
the order of finite subgroups of any reductive group over any field. This has been discussed
extensively by Serre [7].
4
4. Comparison of Minkowski bounds
For our purposes, it will be important to compare the Minkowski bounds for various
group/subgroup pairs. The key points will be to identify discrepancies for p “ 2, and to
isolate cases for p ą 2 where the bound for the subgroup matches that of the full group.
Remark 4.1. A trivial but useful observation along these lines is that for n ě 1, p ą 2,
d ą 1,
rpGLpdnqQ , pq ą rpGLpnqQ , pq
whenever rpGLpdnqQ , pq ą 0.
A slightly less trivial observation is that for n ě 1,
rpGLpnqQpiq , 2q ą rpGLpnqQ , 2q.
Lemma 4.2. Let K be a number field of degree d over Q. For each integer n ě 1 and each
odd prime p,
rpGLpdnqQ , pq ě rpGLpnqK , pq ` rpSd , pq;
moreover, if n ą 1 and mpK, pq ą 1, or if d ě ppp ´ 1q, then the inequality is strict.
Proof. There is nothing to check when d “ 1, so we may assume d ě 2. Put m “ mpK, pq,
t “ tpK, pq; then the desired inequality is
Z
^ Z
^
Z ^ Z ^
Yn] Z n ^
dn
dn
d
d
`
` ¨¨¨ ě m
`
` ¨¨¨`
` 2 ` ¨¨¨ .
p´1
ppp ´ 1q
t
pt
p
p
d
equals 1t times the integer pm´1 which is
From the equality dt “ pm´1 pp ´ 1q, we see that p´1
no less than m (and strictly greater than m if m ą 1). Consequently, by writing the difference
between the two sides as
Z
^
Z ^ Z ^
Y n ] Z dn ^ Z n ^
dn
d
d
´m
`
´
` ¨¨¨`
` 2 ` ¨¨¨ ,
p´1
t
ppp ´ 1q
pt
p
p
we see that this difference does not decrease if we increase n by 1 (and strictly increases if
m ą 1). If n “ 1, then the desired inequality becomes rpGLpdqQ , pq ě rpSd , pq, which holds
because Sd embeds into GLpd, Qq; this equality is strict whenever d ě ppp ´ 1q. This proves
the claim.
Corollary 4.3. For K a number field of degree d, for p ą 2 we have
(4.1)
rpGLpdnqQ , pq ě rpAutpK{Qq ˙ ResK
Q GLpnqK , pq.
Moreover, if K Ę Qpµp q, K is not the degree-p subextension of Qpµp2 q, and rpGLpdnqQ , pq ‰
0, then the equality is strict.
Proof. The inequality (4.1) holds because AutpK{Qq ˙ ResK
Q GLpnqK embeds into GLpdnqQ .
We thus only need to obtain a contradiction assuming that K Ę Qpµp q, K is not the degree-p
subextension of Qpµp2 q, rpGLpdnqQ , pq ą 0, and equality holds in (4.1).
Note that (4.1) also follows from Lemma 4.2, so equality must also hold in the latter. By
Remark 3.6, if K 1 :“ K X Qpµp8 q has degree d1 ‰ d, then rpGLpnqK , pq “ rpGLpnqK 1 , pq; we
then get the strict inequality by applying Lemma 4.2 to the field K 1 and invoking Remark 4.1
(using the condition that rpGLpdnqQ , pq ą 0). We must therefore have K “ K 1 and hence
K Ď Qpµp8 q; since we assumed K Ę Qpµp q, this implies that mpK, pq ą 1. To have equality
in Lemma 4.2, we must then have n “ 1 and d ă ppp ´ 1q.
5
Since mpK, pq ą 1, K must contain the degree-p subextension of Qpµp2 q, necessarily
strictly by hypothesis; hence d{p is an integer strictly greater than 1. By the bound on d, we
cannot then have Qpµp q Ď K, so rpGLp1qK , pq “ 0. Meanwhile, K is an abelian extension
of Q, so rpAutpK{Qq, pq “ 1. However, since d ě 2p, rpGLpdqQ , pq ě 2, yielding the desired
contradiction.
Corollary 4.4. For any integers n, d ą 1, for each odd prime p for which rpGLpndqQ , pq ą 0,
we have
rpGLpdnqQ , pq ą rpGLpnqQ , pq ` rpSd , pq.
Proof. XWe\ first reduce to the case where d is even. Namely, we may do this by replacing d
with 2 d2 except if d is odd, rpGLpdnqQ , pq ą 0, and rpGLppd ´ 1qnqQ , pq “ 0. This implies
pd ´ 1qn ă p ´ 1 ď dn, which implies on one hand that n ă p ´ 1 and rpGLpnqQ , pq “ 0, and
on the other hand that d ´ 1 ă p ´ 1 and so rpSd , pq “ 0; this yields the claimed inequality.
For any number field K of even degree d, by Lemma 4.2 we have
rpGLpdnqQ , pq ě rpGLpnqK , pq ` rpSd , pq ě rpGLpnqQ , pq ` rpSd , pq,
so it suffices to confirm that both equalities cannot hold simultaneously. Since we are free
to choose K, we take it to contain the quadratic subfield of Qpµp q; then by Remark 3.6, we
have rpGLpnqK , pq ą rpGLpnqQ , pq unless both quantities equal zero. It thus suffices to rule
out the equality rpGLpdnqQ , pq “ rpSd , pq; since rpGLpdqQ , pq ě rpSd , pq, this follows from
Remark 4.1.
For p “ 2, we have the following analogue of Corollary 4.3.
Remark 4.5. For K{F an extension of number fields of degree d, we obviously have
rpGLpdnqF , 2q ě rpAutpK{F q ˙ ResK
F GLpnqK , 2q.
In case F “ Qpiq, one can show using Remark 3.2 that equality holds only when d “ 1;
however, we will not need this.
For p “ 2, we have the following analogue of Corollary 4.4.
Lemma 4.6. For any integers n, d ą 1 such that g “ dn{2 is an integer,
rpGLpgqQpiq , 2q ě rpGLpnqQ , 2q ` rpSd , 2q
with equality only for pn, dq “ p2, 2q.
Proof. We are claiming that
Z ^ Z ^
Z ^ Z ^
Yn] Yn]
dn
dn
d
d
dn `
`
` ¨¨¨ ě n` 2
`
` ¨¨¨`
`
` ¨¨¨
4
8
2
4
2
4
X \
with equality only for pn, dq “ p2, 2q. For d “ 2, this inequality becomes n ě n2 ` 1; for
n “ 2, it becomes 2d ě 4. It thus remains to check the cases where n, d ě 3.
We may write the difference between the two sides as
ˆZ ^ Z ^˙ ˆZ ^ Z ^˙
Yn] Yn]
dn
d
dn
d
pd ´ 1qn `
´
`
´
` ¨¨¨´ 2
´
´ ¨¨¨ ;
4
2
8
4
2
4
in particular, for n ě 2 fixed, this difference increases if we increase d by 2. Using the previous
paragraph, we deduce the claim when d is even. For d odd, we may argue that
rpGLpgqQpiq , 2q ě rpGLpg ´ n{2qQpiq , 2q ě rpGLpnqQ , 2q ` rpSd´1, 2q “ rpGLpnqQ , 2q ` rpSd , 2q
6
with strict inequality if n ą 2.
5. Abelian varieties and Sato-Tate groups
We now specialize to the cases of interest for abelian varieties.
Definition 5.1. For A an abelian variety over a number field K, let ASTpAq denote the
algebraic Sato-Tate group of A in the sense of [1, Definition 9.5]. The key properties that we
need are the following.
‚ The group scheme ASTpAq is a pointful subgroup scheme of Spp2gqQ whose connected
part is reductive. The connected part is closely related to the Mumford-Tate group
of A.
‚ There exists a torus T Ă ASTpAq˝C which acts on C2g with weights 1, ´1 each with
multiplicity g. In particular, the fixed space of ASTpAq is the zero subspace.
‚ The component group π0 pAq surjects onto GalpL{Kq for L the endomorphism field of
A; this map is a bijection whenever the Mumford-Tate group is completely explained
by endomorphisms (which holds in all cases when g ď 3 but can fail for g ě 4, as
originally shown by Mumford). Moreover, for K 1 a finite extension of K, ASTpAK 1 q
is the inverse image of GalpLK 1 {K 1 q Ď GalpL{Kq in ASTpAq.
‚ Any decomposition of Q2g into indecomposable ASTpAq-modules corresponds to a
product-up-to-isogeny decomposition of A.
‚ The group ASTpAq is a torus if and only if A is isogenous to a product of abelian
varieties with CM defined over K.
Example 5.2. Put
?
?
?
M1 :“ Qpiq, M2 :“ Qp ´2q, M3 :“ Qp ´3q, M4 :“ Qp ´6q
and let K be the compositum of these four fields. Let A be the product of four elliptic curves
E1 , . . . , E4 with CM by M1 , . . . , M4 , respectively. Then ASTpAq is a torus of dimension 3.
Remark 5.3. The Sato-Tate group of A, as studied for abelian surfaces in [2], is a maximal
compact subgroup of ASTpA, Cq; it therefore has the same component group as ASTpAq.
On one hand, this means that the argument using algebraic Sato-Tate groups in the proof of
Theorem 5.4 directly applies also to the component groups of Sato-Tate groups; on the other
hand, the conclusion of Theorem 5.4 in the case g “ 2 is a consequence of [2, Theorem 4.3],
which will save a bit of case analysis.
We prove the upper bound assertion of Theorem 1.2 by establishing the following result.
Theorem 5.4. For A an abelian variety of dimension g over a number field K, the component group of the algebraic
ś 1 Sato-Tate group of A (or equivalently, the Sato-Tate group of A)
has order dividing p pr pg,pq .
Proof. Put G “ ASTpAq. It suffices to check the claimed divisibility for the p-part of #π0 pGq;
this is immediate from Remark 3.5 (applied with n “ 2g) unless p “ 2 or p is a Fermat prime.
In light of Remark 5.3, we may assume further that g ě 3.
Note that for any fixed p, r 1 pg, pq is superadditive in g; we may thus reduce to the case
where A is indecomposable, which as noted above implies that G acts indecomposably on
V “ Q2g . Using Corollary 4.4 and Lemma 4.6, we may also deduce the claim in case G˝ does
7
not act isotypically on V (the exceptional case of Lemma 4.6 cannot occur for g ě 3). We
may thus assume hereafter that G˝ acts isotypically on V .
Let W be an irreducible G˝ -representation occurring in V , and put
D :“ EndG˝ pW q,
M :“ EndG˝ pV q,
F :“ ZpDq “ ZpMq.
By Schur’s lemma, D is a division algebra, M is a matrix ring over D, and T :“ imagepG˝ Ñ
M ˆ q is a torus which splits over F . If we define H :“ kerpG Ñ OutpG˝ qq, we obtain an
induced injective morphism H{G˝ ãÑ M ˆ {T . Meanwhile, since V is G˝ -isotypical, G{H acts
faithfully on the set of isomorphism classes of G˝ -constituents of W bQ Q; this implies that
the map G Ñ AutpF {Qq induces an injective morphism G{H ãÑ AutpF {Qq.
Define the following positive integers:
a :“ rF : Qs;
b :“ rankF D “ the G˝ -multiplicity of W bF F ;
c :“ rankD M “ the G˝ -multiplicity of W in V ;
2g
d :“
“ the Q-dimension of a G˝ -constituent of V bQ Q.
abc
In this notation, G{H injects into Sa while H{G˝ injects into a subgroup of a twisted form
of GLpbcqF . In light of Remark 3.4, it follows that the p-adic valuation of #π0 pGq is at most
(5.1)
rpAutpF {Qq, pq ` rpGLpbcqF , pq ď rpGLpabcqQ , pq.
If d ą 1, then Remark 4.1 immediately yields the desired bound (even for p “ 2). We may
thus assume hereafter that d “ 1, which implies that G˝ is abelian and hence a torus. In this
case, F must be totally imaginary; it is in fact the CM field associated to the unique simple
isogeny factor of A.
If p ą 2 is a Fermat prime, then the only way to violate the desired bound would be
to have equality in (5.1), which can only occur if F Ď Qpµp q in light of Corollary 4.3 (the
degree-p subextension of Qpµp2 q cannot be a compositum of CM fields because its degree
is odd). Since F is totally imaginary, this would force F “ Qpµp q. However, under these
conditions, we may invoke Remark 3.3: the reduction of G˝ itself always has order divisible
by p, yielding exactly the correct bound.
If p “ 2 and F X Qpµ28 q “ Q, then rpGLpnqF , 2q “ rpGLpnqQ , 2q (see Remark 3.6) and
so we may invoke Lemma 4.6 to deduce the desired result (again assuming that g ě 3). If
instead Qpµ4 q Ď F , then Remark 4.5 gives a valuation bound which is only off by 2, and
again this discrepancy is accounted for by Remark 3.3 (the reduction of ?
G˝ itself always
has
?
1
order divisible?by 4). Otherwise, F :“ F X Qpµ8 q must equal one of Qp 2q or Qp ´2q. In
case F 1 “ Qp ´2q, Remark 3.6 implies that rpGLpnqF 1 , 2q “ rpGLpnqQ , 2q; we thus obtain
an upper bound of
rpAutpF 1 {Qq, 2q ` rpGLpabc{2qF 1 , 2q “ 1 ` rpGLpgqQ , 2q
and again Lemma 4.6
? settles the question (for g ě 3).
In case F 1 “ Qp 2q, we must argue a bit more carefully. Since F 1 is Galois and totally
real, we must have F 1 ‰ F and rpAutpF {Qq, 2q “ 1 ` rpAutpF {F 1 q, 2q. By Remark 3.6, we
have
Yn]
rpGLpnqF , 2q “ rpGLpnqF 1 , 2q “ rpGLpnqQ , 2q `
.
2
8
Note finally that the reduction of G˝ always has order divisible by 2. From (5.1), we now
obtain an upper bound of
Z ^
bc
1
rpAutpF {F q, 2q ` rpGLpbcqQ , 2q `
2
which by Lemma 4.6 (and the fact that a ě 4) is itself bounded above by
Yg ]
Yg]
ď rpGLpgqQ , 2q `
.
rpGLpgqQ , 2q `
a
4
This gives the desired bound once more.
6. Lower bounds
To conclude, we establish the lower bound assertion of Theorem 1.2 by twisting powers of
CM abelian varieties.
Definition 6.1. We briefly recall [2, Definition 2.20]. Let A be an abelian variety over a
number field K, let L{K be a finite Galois extension, and let f : GalpL{Kq Ñ AutpAL q be
a 1-cocycle. Then there exists an abelian variety Af over K equipped with an isomorphism
AfL – AL such that the action of τ P GK on Af pKq – AfL pKq corresponds to the action of
f pτ qτ on ApKq – AL pKq. The isomorphism AfL – AL induces an isomorphism EndpAfL q –
EndpAL q in which corresponding elements α P EndpAfL q, β P EndpAL q satisfy the relation
(6.1)
τ pαq “ f pτ qτ pβqf pτ q´1 .
We use the twisting setup in the following setting.
Definition 6.2. Fix a prime p and a positive integer m. Let A0 be an abelian variety of
dimension g0 over some number field K, such that A0 has complex multiplication by the
ring of integers oM of a subfield M of Qpµpm q; put d “ rQpµpm q : Ms. (Note that we cannot
hope to fix the field K, because its degree over Q is related to the class number of M.) Let
G0 Ď GLpd, Mq be a subgroup of order pm stable under GalpM{Qq and identify G0 with a
subgroup of AutpAd0,KM q. Put A1 “ Adn
0 for some positive integer n; then G1 “ G0 ≀ Sn may
be identified with a subgroup of AutpA1,KM q stable under GalpKM{Kq. Let G be the image
of G1 under the map GLpdn, Mq Ñ PGLpdn, Mq.
Choose an Sn -extension L0 {K linearly disjoint from KM, so that L0 M{KM is again an
Sn -extension. Note that for a “generic” Cpm -extension L1 {L0 M, the Galois closure L2 of L1
over M will have Galois group G0 ≀ Sn . Using class field theory, we may further ensure that
L2 is also Galois over K and that there exists a 1-cocycle f : GalpL2 {Kq Ñ AutpA1,L2 q
whose restriction to GalpL2 {KMq is the preceding identification of the latter with G0 ≀ Sn Ď
AutpA1,KM q.
Put A “ Af1 and let L be the endomorphism field of A. Then KM Ď L and (6.1) implies
the existence of a surjective morphism from GalpL{KMq to G, but not in general to G1 .
Theorem 6.3. For each prime p, there exists an abelian variety A of dimension g over some
1
number field K such that the p-part of rL : Ks is at least prg,p .
Proof. Suppose first that p ´ 1 is not a power of 2 (i.e., p is odd and not a Fermat prime).
Then there exists a subfield M of Qpµp q whose index ℓ is an odd prime divisor of p ´ 1.
9
Y ]
g
then yields the desired result; note that in
Applying Definition 6.2 with m “ 1, n “ p´1
this case, there is no harm to take K to contain M, which simplifies the analysis somewhat.
Suppose next that p is a Fermat prime; we may assume that rg,p ě 1. The previous
construction breaks down because p ´ 1 admits
Y no
] odd prime divisor, so we can only apply
g
Definition 6.2 with m “ 1, M “ Qpµp q, n “ p´1 . Note that we now lose one factor of p to
the quotient map G1 Ñ G, so again we get the desired result.
Suppose finally that p “ 2. Apply Definition 6.2 with m “ 2, M “ Qpiq, n “ g; note that
in this case we may even take K “ Q. This time, we lose two factors of 2 to the quotient map
G1 Ñ G, but gain one back from the extension M{Q. This proves the claim once more.
References
[1] G. Banaszak and K.S. Kedlaya, An algebraic Sato-Tate group and Sato-Tate conjecture, Indiana Univ.
Math. J. 64 (2015), 245–274.
[2] F. Fité, K. S. Kedlaya, V. Rotger, A. V. Sutherland, Sato-Tate distributions and Galois endomorphism
modules in genus 2, Compositio Math. 148 (2012), 1390–1442.
[3] R.M. Guralnick and M. Lorenz, Orders of finite groups of matrices, in Groups, Rings and Algebras:
Proceedings of a Conference in Honor of Donald S. Passman, Contemporary Math. 420, 2006, 141–162.
[4] M. Kirschmer, Finite symplectic matrix groups, Experiment. Math. 20 (2011), 217–228.
[5] B. Poonen, E.F. Schaefer, and M. Stoll, Twists of Xp7q and primitive solutions to x2 ` y 3 “ z 7 , Duke
Math. J. 137 (2007), 103–158.
[6] G. Rémond, Degré de définitions des endomorphismes d’une variété abélienne, preprint available at
https://www-fourier.ujf-grenoble.fr/~remond/.
[7] J.-P. Serre, Bounds for the orders of the finite subgroups of Gpkq, in Group Representation Theory, EPFL
Press, Lausanne, 2007, 405–450.
[8] A. Silverberg, Fields of definition for homomorphisms of abelian varieties, J. Pure Appl. Algebra 77
(1992), 253–262.
10
| 4math.GR
|
Towards Provably Invisible Network Flow
Fingerprints
Ramin Soltani∗ , Dennis Goeckel∗ , Don Towsley† , and Amir Houmansadr†
∗
Electrical and Computer Engineering Department, University of Massachusetts, Amherst,
arXiv:1711.10079v1 [cs.NI] 28 Nov 2017
{soltani, goeckel}@ecs.umass.edu
†
College of Information and Computer Sciences, University of Massachusetts, Amherst,
{towsley, amir}@cs.umass.edu
Abstract
Network traffic analysis reveals important information even when messages are encrypted. We consider
active traffic analysis via flow fingerprinting by invisibly embedding information into packet timings of flows. In
particular, assume Alice wishes to embed fingerprints into flows of a set of network input links, whose packet
timings are modeled by Poisson processes, without being detected by a watchful adversary Willie. Bob, who
receives the set of fingerprinted flows after they pass through the network modeled as a collection of independent
and parallel M/M/1 queues, wishes to extract Alice’s embedded fingerprints to infer the connection between
input and output links of the network. We consider two scenarios: 1) Alice embeds fingerprints in all of the
flows; 2) Alice embeds fingerprints in each flow independently with probability p. Assuming that the flow rates
are equal, we calculate the maximum number of flows in which Alice can invisibly embed fingerprints while
having those fingerprints successfully decoded by Bob. Then, we extend the construction and analysis to the
case where flow rates are distinct, and discuss the extension of the network model.
Keywords: Computer Networks, Network Security, Network Analysis, Network Security Analysis, Network De-Anonymization, Anonymous Networks, Covert Communication, Fingerprinting, Flow Fingerprinting, Network Flow Fingerprinting, Watermarking, Flow Watermarking, Network Flow Watermarking, Timing Channel, Covert Channel, Queuing Networks, M/M/1 queues, Timing Capacity of Queues,
Shared Processor Queues, Queues in Tandem.
This work has been supported by the National Science Foundation under grants CNS-1564067 and CNS-1525642.
This work has been presented at the 51st Annual Asilomar Conference on Signals, Systems, and Computers, November 2017.
Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
1
Fig. 1. Alice may embed fingerprints in flows. Bob receives the potentially fingerprinted flow after it passes through a network of M/M/1
queues, which are independent and parallel. Each queue is shared between a fingerprinted flow (shown by a blue solid line) and interfering
flows (shown by green dotted lines).
I. I NTRODUCTION
Security in computer networks has emerged as an important area of research. Although encryption
hides information sent from a transmitter, network traffic analysis can extract important information
from the size, count, and timings of the packets. For instance, when attackers relay their flows through
compromised nodes called stepping stones, traffic analysis can trace back the attackers [1], [2]. Also,
traffic analysis can find the correlations in traffic patterns to link incoming/outgoing flows and break
anonymity [3].
Flow watermarking and flow fingerprinting are two active traffic analysis methods that work by
perturbing packet timings of flows according to specific patterns to embed information in them. In flow
watermarking, the information embedded in a flow is one bit, i.e., either the flow is marked or not.
However, in many applications, more than one bit of information is required to be embedded in the
packet timings of the flows. Flow fingerprinting provides the solution for such applications by embedding
several bits of information in the flows such as the information about the party that has embedded the
fingerprint, the source of the flow, or the location at which the flow has been fingerprinted [4].
Active traffic analysis has become an important area of research due to the increasing use of encryption. Wang et al. [5] proposed to embed flow watermarks in inter-packet delays to detect stepping
stones, and Wang et al. [6] used an interval-based flow watermark to compromise anonymized VoIP
conversations. Houmansadr et al. proposed the first non-blind watermark, RAINBOW [7], offering
significantly higher invisibility compared to prior designs, and SWIRL [8] was designed to resist
aggregated-flows attacks. Houmansadr et al. [9] was the first to introduce flow fingerprinting, and
TagIt [10] introduced the first blind flow fingerprint.
As previous active traffic analysis designs are based on ad hoc heuristics (such as moving packets
into secret time intervals), they do not offer any theoretical guarantees on the invisibility-performance
2
trade-off. In this work, we take a systematic approach to design a flow fingerprinting system with
provable information-theoretic guarantees on invisibility and performance (e.g., number of fingerprints).
Consider a network containing m independent, parallel, work conserving, and First In First Out (FIFO)
queues with independent exponential service times where the ith queue conveys the ith flow (fi ) from
the ith input link to the ith output link, and conveys interfering flows independent of fi (See Fig. 1).
The network is anonymous to Alice and Bob such that they do not know the connections between input
and output links. Alice has access to the input links and is able to buffer the packets and release them
when she desires to embed fingerprints in packet timings of the flows. Adversary Willie is between
Alice and the network and observes f1 , f2 , . . . , fm after they are accessed by Alice and wishes to detect
if Alice is embedding fingerprints in the flows or not. Bob observes the packet timings of the flows on
the output links and wishes to extract Alice’s fingerprints.
We consider the following problem: given the time interval [0, T ], can Alice embed flow identifier
fingerprints invisible to Willie in the packet timings such that Bob can extract them successfully to deanonymize the network and, if yes, what is the maximum number m of flows that can be fingerprinted?
For the case where the packet timings of the flows are governed by Poisson process, we calculate the
asymptotic expression for the number of flows that can be fingerprinted as a function of T , for two
scenarios: 1) Alice embeds fingerprints in all of the flows; 2) Alice embeds fingerprints in each flow
independently with some small probability p.
The remainder of the paper is organized as follows. In Section II, we present the system model,
definitions and the metrics employed for the two scenarios of interest. Then, we provide constructions
and analyses for the two fingerprinting scenarios in Sections III and Sections IV. In Section V we
discuss the results, the extension of the scenarios to distinct flow rates, and the extension of the network
model to more general networks. Finally, we conclude in Section VI.
II. S YSTEM M ODEL , D EFINITIONS , A ND M ETRICS
A. System Model
(I)
(I)
(I)
Alice has access to a set of input links L1 , L2 , . . . , Lm of a network, and is able to buffer packets
(I)
and release them when she desires. The packet flow conveyed over Li
is denoted by fi (1 ≤ i ≤ m),
and F = {f1 , f2 , . . . , fm } is the set of flows accessed by Alice. Bob receives the flows f1 , f2 , . . . , fm
(O)
(O)
(O)
from the output links L1 , L2 , . . . , Lm of the network, respectively. The network is anonymous such
that Alice and Bob do not know the connections between input and output links; they wish to infer this
in the interval [0, T ], and thus de-anonymize the network.
3
Alice embeds a unique flow identifier fingerprint in each flow by altering its packet timings according
to a secret codebook of fingerprints shared with Bob, and Bob extracts the fingerprints from the observed
flows. Warden Willie, who is between Alice and the network, observes the input links and wishes to
detect if Alice embeds fingerprints in them or not (see Fig. 1). Willie knows the fingerprinting scheme
Alice will employ if she chooses to embed fingerprints, but he does not have access to the codebook
of fingerprints.
Alice, Bob, and Willie know that the packet timings of the flows f1 , f2 , . . . , fm are modeled by
Poisson processes with rates λ1 , . . . , λm , respectively. The network consists of m independent single
server queues with exponential service times, i.e., M/M/1 queues, which are work conserving and
discipline First In First Out (FIFO). Each queue has multiple inputs and outputs such that the ith queue
(I)
(qi ) conveys fi from the input link Li
(O)
to the output link Li , and also conveys interfering Poisson
flows independent of fi . We denote the sum of the rates of the interfering flows on qi by λ0i . The service
rate of qi is µi , and the queues are stable, i.e., λi + λ0i ≤ µi .
We consider two scenarios. In Scenario 1 (analyzed in Section III), the flow rates are equal (λi = λ),
and Alice embeds fingerprints in all of the flows of F. In Scenario 2 (analyzed in Section IV), the flow
rates are equal, but Alice embeds fingerprints in each flow independently with probability p. For each
scenario, we calculate the number of flows in which Alice can invisibly and reliably embed fingerprints,
as described precisely next.
B. Definitions
Willie’s hypotheses are H0 (Alice did not embed fingerprints) and H1 (Alice embedded fingerprints).
We denote by PFA the probability of rejecting H0 when it is true (type I error or false alarm), and PMD
the probability of rejecting H1 when it is true (type II error or mis-detection). We assume that Willie
uses classical hypothesis testing with equal prior probabilities and seeks to minimize his probability of
(w)
error, Pe
=
PFA +PMD
;
2
the generalization to arbitrarily prior probabilities is available in [11].
Definition 1. (Invisibility) Alice’s fingerprinting is invisible (covert) if and only if she can lower bound
(w)
Willies’ probability of error (Pe ) by
1
2
− for any > 0, asymptotically. This definition is similar to
that of covertness developed in [11], and used in covert communication [12]–[15]
Definition 2. (Reliability) Fingerprinting for each flow is reliable if and only if Pf ≤ ζ for any ζ > 0,
where Pf is the probability of the failure event which occurs when
•
Alice cannot successfully embed a fingerprint since she does not have a packet to release when
she needs to;
4
•
Alice runs out of fingerprints because the number of fingerprints in her codebook is less than the
number of flows in which she wishes to embed fingerprints; or
•
Bob cannot extract a fingerprint successfully.
Definition 3. (Lambert W-function) The Lambert W-function is the inverse function of f (W ) = W eW .
Definition 4. (The Kullback–Leibler divergence) If f and g are probability measures over a set S, then
the Kullback–Leibler divergence between f and g is:
Z
f (x)
dx
D(f ||g) =
f (x) log
g(x)
S
(1)
In this paper, we use standard Big-O notation [16, ch. 3].
III. S CENARIO 1: A LL FLOWS ARE FINGERPRINTED
In this section we consider Scenario 1: Alice embeds fingerprints in all of the input flows of a network
with equal rates in the time interval [0, T ], and Bob extracts the fingerprints from the output flows to infer
the connection between input and output flows of the network. Because Alice is able to buffer packets
and release them when she desires, she changes the packet timings of the flows to embed fingerprints in
them according to a secret fingerprinting codebook shared with Bob. Each of the fingerprints is a flow
identifier and consists of a sequence of inter-packet delays to be employed to embed the corresponding
fingerprint. To successfully change the packet timings of a flow according to the chosen codeword,
Alice must have a packet in her buffer to transmit at the appropriate times. To account for this, Alice
uses a two phase scheme for each flow fi , similar to the one adopted in [13, Section IV]. First, Alice
slows down fi to buffer packets; then, during the fingerprinting phase, she releases the packets from her
buffer with the inter-packet delays prescribed by the codeword corresponding to the fingerprint, while
buffering the arriving packets of fi .
We calculate the asymptotic expression for the number of flows that can be fingerprinted as a function
of T using this strategy.
Theorem 1. Consider the setting in Section II-A. In a set F containing m flows with rates λi = λ,
Alice and Bob can invisibly and reliably track all of the flows in the time interval [0, T ], as long as
m = O(T /W (T )), where W (·) is the Lambert-W function.
Proof. Construction: Per above, Alice divides the time interval of length T into two phases: a buffering
phase of length T1 and a fingerprinting phase of length T2 such that T = T1 + T2 . During the buffering
phase, Alice slows down the packets of each flow fi , from rate λ to rate λ − ∆, in order to build up
5
Fig. 2. Two phase scheme: Alice divides the duration of time T into two phases with lengths T1 and T2 = T − T1 . In the first phase,
Alice slows down each flow from the rate λ to the rate λ − ∆, and buffers the excess packets. In the next phase, she transmits packets
at rate λ according to the inter-packet delays in the codeword corresponding to the fingerprint to be embedded.
Fig. 3. Codebook generation: Alice and Bob share a secret codebook which specifies the sequence of inter-packet delays corresponding
to each fingerprint. To generate each codeword, a number N is generated according to the Poisson distribution with parameter λT2 , and
then N points are placed uniformly and randomly in the time interval [0, T2 ].
packets in her buffer, ensuring that with high probability, she will not run out of packets during the
fingerprinting phase of length T2 = T − T1 (see Fig. 2).
Alice and Bob share a codebook to which Willie does not have access.The codebook construction is
similar to that of [13], [14]. To build the codebook, a set of m codewords {C(Wl )}l=m
l=1 are independently
generated according to realizations of a Poisson process with parameter λ. In particular, to generate a
codeword C(Wl ), first a random variable N is generated according to a Poisson distribution with mean
λT2 . Then, N inter-packet delays are generated by placing N points uniformly and independently on
an interval of length T2 [17] (see Fig. 3). Therefore, each codeword of the codebook is a series of
inter-packet delays and corresponds to a unique flow identifier fingerprint. To embed a fingerprint in a
flow fi , Alice applies the inter-packet delays of the chosen codeword to the packets of the flow fi .
Analysis: (Invisibility) The analysis of invisibility follows from that of covertness in [13, Theorem
p
2]. In the first phase, Alice slows down the flows from rate λ to rate λ − 2λ/mT1 , where > 0,
(w)
while lower bounding Willie’s error probability (Pe ) by
1
2
− . During the second phase, the packet
timings for each flow is an instantiation of a Poisson process with rate λ and hence the traffic pattern
is indistinguishable from the pattern that Willie expects to observe. Hence, the scheme is invisible.
(Reliability) By [17, Definition 2], Bob can successfully extract the fingerprint from fi as long as T2
6
is large and:
log m
< C(qi ),
T2
(2)
where C(qi ) is the capacity of qi for conveying information through packet delays. By [18, Proposition
1], C(qi ) = λ log ((µi − λ0i )/λ), where λ0i is the sum of rates of the interfering flows passing through
qi . Define
C = λ log min{µi −
i
λ0i }/λ
.
(3)
Since (2) holds for all 1 ≤ i ≤ m, for large T2 :
log m
< min{C(qi )} = C.
i
T2
(4)
Note that Alice does not run out of fingerprints since the number of fingerprints in her codebook
equals to the number of flows. Finally, similar to the reliability analysis in [13], we can show that if
T mα/2
T1 =
,
1 + mα/2
T2 = T − T1 =
T
,
1 + mα/2
(5)
(6)
where
α = (2erf −1 (1 − ζ))2 ,
(7)
then Pf ≤ ζ. Thus Alice’s fingerprinting is reliable.
(Number of flows) By (4) and (6), we require
log m < CT2 =
CT
.
1 + mα2
(8)
Next, we show that if
1
m = min
2
2
α
TC
TC
−1 ,
,
W (T C)
W (T C)
(9)
then (8) is satisfied. Consider the following fact:
Fact 1. For x, y > 0, if x log x = y, then x = y/W (y), where W (·) is the lambert-W function (see
Definition 3).
Proof. By Definition 3, W (y)eW (y) = y. Therefore, W (y) = log Wy(y) . Consequently,
x log x =
y
y
y
log
=
W (y) = y
W (y)
W (y)
W (y)
7
If m ≥ 1 + mα/2 , since m <
TC
,
W (T C)
then Fact 1 yields:
T C > m log m ≥ (1 + mα/2 ) log(m).
2
TC
2
If m < 1 + mα/ , since m < α W (T C) − 1 , then:
1 + mα/2
2
2
2
T C > (1 + mα/ ) log (1 + mα/ ) = (1 + mα/ ) log m + log
≥ (1 + mα/2 ) log m.
m
Consequently, Alice and Bob can invisibly and reliably track m = O(T /W (T )) flows. Note that by (5),
T1 → ∞ as T → ∞, as required by the proof for invisibility of the second phase. Also, by (9) and (6)
we can show that T2 → ∞ as T → ∞, as required by the proof for reliability.
IV. S CENARIO 2: E ACH FLOW IS FINGERPRINTED INDEPENDENTLY WITH PROBABILITY p
In this section we consider Scenario 2. In a set containing m network input flows with equal rates,
Alice embeds fingerprints into each flow independently with probability p in the time interval [0, T ],
and Bob extracts the fingerprints from the output flows to infer the connection between input and output
flows of the network. Similar to Scenario 1, we show that employing a two phase scheme, Alice can
embed a unique flow identifier fingerprint in the chosen flows by altering their packet timings according
to a secret fingerprint codebook shared between Alice and Bob but unknown to Willie. We calculate
the asymptotic expression for the number of flows that can be fingerprinted as a function of T using
this strategy
Theorem 2. Consider the setting in Section II-A. In a set F containing m flows with rates λi = λ, if
Alice embeds fingerprints in each flow independently with probability p, Alice and Bob can invisibly
√
CT − CT α
flows in the time interval [0, T ], where C and α are given in (3)
and reliably track O e
and (7), respectively, as long as
√
e2CT − CT α
,
m=
22
(10)
p = 2 e−CT .
(11)
Proof. Construction: The construction is similar to that of Scenario 1. Alice’s codebook contains M
fingerprints where
M = (1/2)e
CT √
1+α/ln 1+e CT α
(
),
(12)
To decide whether to embed a fingerprint in a flow or not, Alice generates independent Bernoulli random
variables X1 , . . . , Xm with P(Xi = 1) = p, and she embeds a fingerprint in fi if and only if Xi = 1.
8
Analysis: (Invisibility) We analyze the invisibility of the first and second phases separately. In the
first phase, the joint pdfs of Willie’s observations under H0 (Alice did not embed fingerprints), and H1
(Alice embedded fingerprints) are:
P0 =
m
Y
Pλ (ni ),
i=1
P1 =
m
Y
(pPλ−∆ (ni ) + (1 − p)Pλ (ni )) ,
i=1
(w)
When Willie applies the optimal hypothesis test to minimize Pe [15, Eq.1]:
r
1
1
(w)
D(P1 ||P0 ),
Pe ≥ −
2
8
(13)
where D(P1 ||P0 ) is the relative entropy between P1 and P0 . Denote by Ep [·] the expected value with
respect to the probability measure (pPλ−∆ (ni ) + (1 − p)Pλ (ni )). Then:
m
X
D(P1 ||P0 ) =
D(pPλ−∆ (ni ) + (1 − p)Pλ (ni )||Pλ (n))
i=1
n
λ−∆
∆T1
= mEp ln pe
+ (1 − p) ,
λ
n
(a)
λ−∆
∆T1
≤ mEp pe
−p ,
λ
(b)
= mp2 (e∆
2 T /λ
1
− 1).
(14)
h
i
λ−∆ k
2
where (a) is true since ln(1+x) ≤ x for all x ∈ R, and (b) is true since Ep
= pe−∆T1 e∆ T1 /λ +
λ
q
λ
2
2
(1 − p)e−∆T1 . Let ∆ =
ln(1 + 2mp
2 ). Then, (14) yields D(P1 ||P0 ) ≤ /2. By (13), Willies’
T1
(w)
probability of error (Pe ) is lower bounded by
1
2
− , and thus the first phase is invisible. The analysis
of the invisibility for the second phase is the same as that of Scenario 1. Thus, the fingerprinting scheme
is invisible.
(Reliability) Similar to the reliability analysis of Theorem 1, we can show that the probability that
Alice runs out of packets for each flow is upper bounded by ζ as long as
T1 =
Tα
ln(1 +
2
)
2mp2
T2 = T − T1 =
By (10), (11),
2
2mp2
√
=e
CT α
+α
,
(15)
T
1 + α/ ln(1 +
2
)
2mp2
,
, and thus T1 , T2 → ∞ as T → ∞.
Next, we show that Bob can successfully extract Alice’s fingerprints. By (12) and (16),
2
1 + α/ ln(1 + 2mp
2)
log M
√
= C.
<C
T2
1 + α/ln 1 + e CT α
9
(16)
where the last step is true since
2
2mp2
√
=e
CT α
. Hence, (4) is satisfied, and thus Bob successfully extracts
each fingerprint.
Furthermore, since M −mp is an increasing function of T , by the weak law of large numbers (WLLN)
we can show that Alice does not run out of fingerprints. Hence, Alice’s fingerprinting is reliable.
√
(Number of flows) By (10), (11), mp = (1/2)eCT −
CT α
, and by the WLLN,
∀γ > 0 : lim P(Nf − mp > γ) = 0.
(17)
T →∞
Therefore, Alice and Bob can invisibly and reliably track Nf = O(eCT −
√
CT α
) flows.
V. D ISCUSSION
A. Source of the gain in Scenario 2
The result for Scenario 2 indicates a much larger fingerprint dictionary can be generated and employed
covertly than in Scenario 1. Note that (11) implies that in Scenario 2, a small portion of the flows are
fingerprinted. Intuitively, because Willie has to investigate a large number of flows to look for alterations
in the timings of a relatively (very) small random subset of those flows, in particular in the first phase,
this makes covertness much easier to achieve and leads to the significant gains observed.
B. Extension to distinct rates
When Scenarios 1 and 2 are extended to distinct flow rates, Alice can build a codebook in which
the rate of the codewords is λmin = min (λ1 , . . . , λm ). To embed a fingerprint in a flow fi , Alice first
scales the corresponding codeword (τ1 , . . . , τN ) by a factor λi /λmin and applies the inter-packet delays
τN
i τ1
, . . . , λλimin
) to the first N + 1 packets of the flow. If Alice receives more than N + 1 packets in
( λλmin
the fingerprinting phase, she releases the excess packets according to random independent inter-packet
delays generated from the pdf of an exponential random variable with mean λ−1
i . Bob rescales the flow
fi by a factor of λmin /λi and uses the codebook to extract the corresponding fingerprint.
q
p
2
We can show that if ∆i = 2λi /mT1 and ∆i = (λi /T1 ) ln(1 + 2mp
2 ) in the first phases of
Scenarios 1 and 2, respectively, then Alice’s buffering is invisible. Note that the fingerprinted flow fi
in the second phase is a realization of a Poisson process with rate λi , and thus it is indistinguishable
from the pattern that Willie expects to observe. Hence, the scheme is invisible.
Note that the time to transmit a fingerprint in fi is T2 λmin /λi . Therefore Bob can successfully extract
Alice’s codeword from fi as long T2 is large and
log M
< λi C 0 (qi ) = log
T2 λmin /λi
10
µi − λ0i
.
λi
(18)
Since (18) is true for all 1 ≤ i ≤ m,
log M
≤ min
i
T2
λmin 0
C (qi ) = C 0 .
λi
Finally, we can obtain an expression for the number of flows by replacing C with C 0 in the results of
Theorems 1 and 2.
C. Extension of the network
By [19], we can extend our model to m parallel routes where each route consists of multiple M/M/1
queues in tandem. On each route ri , queues are shared between a main flow fi and interfering flows and
the interfering flows are independent. Furthermore, by [20, Corollary 3.3], we can relax the condition
of independent interference for queues on each route and extend our model to a feedforward multiclass
product form network [21] containing m parallel routes where each route ri conveys flow fi and consists
of multiple M/M/1 queues in tandem shared between a main flow fi and interfering flows.
VI. C ONCLUSION
In this paper, we presented the construction and analysis for embedding fingerprints in packet timings
of flows. In a setting where a set of flows visit Alice, adversary Willie, a network of m independent
parallel M/M/1 queues with background traffic, and Bob respectively, we established a construction
where Alice alters the packet timings in the time interval [0, T ], according to a secret codebook shared
with Bob, to embed flow identifier fingerprints in them without being detected by Willie. We considered
two scenarios: 1) Alice embeds fingerprints in all of the flows; 2) Alice embeds fingerprint in each flow
independently with probability p, and calculated the asymptotic expression for the number of flows that
can be fingerprinted as a function of T .
R EFERENCES
[1] S. Staniford-Chen and L. T. Heberlein, “Holding intruders accountable on the internet,” in Security and Privacy, 1995. Proceedings.,
1995 IEEE Symposium on, pp. 39–49, IEEE, 1995.
[2] Y. Zhang and V. Paxson, “Detecting stepping stones,” in USENIX Security Symposium, vol. 171, p. 184, 2000.
[3] P. Syverson, G. Tsudik, M. Reed, and C. Landwehr, “Towards an analysis of onion routing security,” in Designing Privacy Enhancing
Technologies, pp. 96–114, Springer, 2001.
[4] A. Houmansadr, Design, analysis, and implementation of effective network flow watermarking schemes. PhD thesis, University of
Illinois at Urbana-Champaign, 2012.
[5] X. Wang and D. S. Reeves, “Robust correlation of encrypted attack traffic through stepping stones by manipulation of interpacket
delays,” in Proceedings of the 10th ACM conference on Computer and communications security, pp. 20–29, ACM, 2003.
[6] X. Wang, S. Chen, and S. Jajodia, “Tracking anonymous peer-to-peer voip calls on the internet,” in Proceedings of the 12th ACM
conference on Computer and communications security, pp. 81–91, ACM, 2005.
[7] A. Houmansadr, N. Kiyavash, and N. Borisov, “Rainbow: A robust and invisible non-blind watermark for network flows,” in NDSS,
2009.
[8] A. Houmansadr and N. Borisov, “Swirl: A scalable watermark to detect correlated network flows,” in NDSS, 2011.
[9] A. Houmansadr and N. Borisov, “The need for flow fingerprints to link correlated network flows,” in International Symposium on
Privacy Enhancing Technologies Symposium, pp. 205–224, Springer, 2013.
11
[10] F. Rezaei and A. Houmansadr, “Tagit: Tagging network flows using blind fingerprints,” Proceedings on Privacy Enhancing
Technologies, vol. 2017, no. 4, pp. 290–307, 2017.
[11] B. Bash, D. Goeckel, and D. Towsley, “Limits of reliable communication with low probability of detection on AWGN channels,”
Selected Areas in Communications, IEEE Journal on, vol. 31, pp. 1921–1930, September 2013.
[12] R. Soltani, B. Bash, D. Goeckel, S. Guha, and D. Towsley, “Covert single-hop communication in a wireless network with distributed
artificial noise generation,” in Communication, Control, and Computing (Allerton), 2014 52nd Annual Allerton Conference on,
pp. 1078–1085, IEEE, 2014.
[13] R. Soltani, D. Goeckel, D. Towsley, and A. Houmansadr, “Covert communications on poisson packet channels,” in 2015 53rd Annual
Allerton Conference on Communication, Control, and Computing (Allerton), pp. 1046–1052, IEEE, 2015.
[14] R. Soltani, D. Goeckel, D. Towsley, and A. Houmansadr, “Covert communications on renewal packet channels,” in 2016 54th Annual
Allerton Conference on Communication, Control, and Computing (Allerton), IEEE, 2016.
[15] R. Soltani, D. Goeckel, D. Towsley, B. Bash, and S. Guha, “Covert wireless communication with artificial noise generation,” arXiv
preprint arXiv:1709.07096, 2017.
[16] T. H. Cormen, Introduction to algorithms. MIT press, 2009.
[17] V. Anantharam and S. Verdu, “Bits through queues,” Information Theory, IEEE Transactions on, vol. 42, no. 1, pp. 4–18, 1996.
[18] X. Liu and R. Srikant, “The timing capacity of single-server queues with multiple flows,” DIMACS Series in Discrete Mathematics
and Theoretical Computer Science, 2004.
[19] P. Mimcilovic, “Mismatch decoding of a compound timing channel,” in Forty-Fourth Annual Allerton Conference on Communication,
Control, and Computing, 2006.
[20] F. P. Kelly, Reversibility and stochastic networks. Cambridge University Press, 2011.
[21] F. Baskett, K. M. Chandy, R. R. Muntz, and F. G. Palacios, “Open, closed, and mixed networks of queues with different classes of
customers,” Journal of the ACM (JACM), vol. 22, no. 2, pp. 248–260, 1975.
12
| 7cs.IT
|
Particle Swarm Optimized Fuzzy Controller for Indirect Vector
Control of Multilevel Inverter Fed Induction Motor
Sanjaya Kumar Sahu1,T. V. Dixit 2 and D.D. Neema3
1
Electrical Engineering, Bhilai Institute of Technology
Durg, Chhattisgarh491001, India
2
Electrical Engineering, Sarguja University
Ambikapur, Chhattisgarh-497001, India
3
Electrical and Electronics Engineering, Chhattisgarh Institute of Technology
Rajanandgaon, Chhattisgarh-491445, India
Abstract
The Particle Swarm Optimized (PSO) fuzzy controller has been
proposed for indirect vector control of induction motor. In this
proposed scheme a Neutral Point Clamped (NPC) multilevel
inverter is used and hysteresis current control technique has been
adopted for switching the IGBTs. A Mamdani type fuzzy
controller is used in place of conventional PI controller. To
ensure better performance of fuzzy controller all parameters such
as membership functions, normalizing and de-normalizing
parameters are optimized using PSO. The performance of
proposed controller is investigated under various load and speed
conditions. The simulation results show its stability and
robustness for high performance derives applications.
Keywords: Multilevel Inverter, Hysteresis Current Control,
Particle Swarm Optimization (PSO), Fuzzy Logic Controller
(FLC).
1. Introduction
Three phase induction motors are widely used in the
industrial purpose because they show better performance
during heavy loads as well as cost effective. However the
drawbacks associated with induction motor are its nonlinear behaviour, controllability and its complexity in
developing mathematical model [1]. By vector control or
field oriented control (FOC) theory, induction motor can
be controlled like a separately excited dc motor. As a
result field and torque of the induction machine can be
controlled
independently
by
manipulating
the
corresponding field oriented quantities. There are two
methods of vector control - direct and indirect vector
control [1]–[3]. In this paper the indirect control method is
adopted, where the slip angle, the d-axis and q-axis stator
currents in synchronous reference frame are computed
from the torque and rotor flux and used for vector control.
A multi-level inverter is a power electronic circuit
built to synthesize stepped approximation of a sinusoidal
wave output voltage or current from a number of DC
voltages [4]. The multilevel inverters gained the attention
in industrial drive application due to following features
[5]:
a)
Improves the waveform quality as the level of
inverter increases.
b) Reduces the size and rating of filter components.
c) High efficiency due to low switching frequency.
d) Lower dv/dt across switches and generate lower
distorted output voltages.
e) Draw input current with very low distortion.
f) Generate smaller common-mode voltage which
reduces the stress in the motor bearings.
The several multilevel inverter topologies are: The Neutral
point clamped (NPC) inverter, Flying Capacitor Inverter
(FCI), and Cascade H-Bridge (CHB) inverter [6]. The
NPC inverters are very popular for high voltage and high
power applications. Theoretically, NPC topology with any
number of levels can be realized. But some of the
problems like complexity of switching algorithm, voltage
unbalance
across
capacitors,
voltage
clamping
requirements, and circuit layouts have limits on the level
in practical multilevel inverters [7]. In an N-level NPC,
each phase leg consists of
power switches,
clamping diodes. The DC bus requires
bulk capacitors. The line voltage has
Levels. At any given time there are (N-1) switches in each
leg which are in ON state. Voltage rating of each of the
device is assumed to be(
⁄
).
In recent years, Fuzzy logic has emerged as an
important artificial intelligence tool to characterize and
control a system, whose model is not known or ill defined.
This paper involves the development of novel
methodology to optimize the performance of Mamdani
type fuzzy logic controller based on a pre defined
objective function. The predefined objective function is
optimized by optimizing the normalization parameter, denormalization parameter and the membership functions of
the Fuzzy logic controller.
Recently, there has been a huge interest in the Particle
Swarm Optimization (PSO) due to its great potential as an
evolutionary algorithm, which is based on the social
behaviour of flocks of birds and schools of fish [8].Since it
is population based and self adaptive, it has gained an
increasing popularity as an efficient alternative to the
genetic algorithm (GA) in solving optimization problem.
Similar to other population-based optimization method
such as the GA, the PSO algorithm starts with random
initialization of a population of individuals in the search
space. Each particle in the search space is adjusted by its
own flying experience and the other particles flying
experience to find the global best solution at each
generation [8]. This paper proposes an implementation of
the PSO method for an off-line tuning of the normalization
parameter, de-normalization parameter and membership
functions of the fuzzy controller.
2.2 Three-Level NPC Inverter
The three-level NPC inverter with two DC link capacitors
C1 and C2 in series and a neutral point O is shown in Fig.1.
2. Modeling
Fig. 1 Three Level NPC Inverter
2.1 Induction Motor
Table 1: Switching Levels in a Three-Level NPC Inverter
The three-phase squirrel cage induction motor
mathematical equations in synchronous rotating reference
frames are as follows [1]–[3]:
Vdse Rs idse pdse e qse
(1)
Vqse Rs iqse pqse e dse
(2)
0 Rr idre pdre (e r )qre
(3)
0 Rr iqre pqre (we wr )dre
(4)
dse Ls idse Lmidre
(5)
e
m qr
Li L i
(6)
dre Lr idre Lmidse
(7)
qre Lr iqre Lmiqse
(8)
Where,
e
qs
e
s qs
And electromagnetic torque
3P
Te
Lm (iqse idre idse iqre )
22
d
wr r
dt
dwr
Te jm
Bm wr Tl
dt
S i1
Si 2
Si 3
Si 4
i th Pole voltage Vio
ON
ON
OFF
OFF
Vdc / 2
OFF
ON
ON
OFF
0
OFF
OFF
ON
ON
Vdc / 2
Each phase of the three-level NPC inverter has two pair of
switching devices Si1 , Si 2 and Si 3 , Si 4 in series, where
i a, b, c phases. The center of each pair is clamped to
the neutral of the DC link capacitors through the clamping
diodes D1, D2, D3, D4, D5 and D6. Table I enumerates the
switching states for the semiconductor devices for the ithphase of this inverter. In Table-1, the switching symbols +,
0 and - respectively denote that the ith-phase terminal is
connected to the positive bus, the neutral point and the
negative bus.
2.3 Three-Level Hysteresis Current Controller
(9)
(10)
(11)
Fig. 2: Three-Level Hysteresis Switching Scheme
An analytical solution of different multilevel PWM
techniques for three-level NPC has been presented [9],
[10]. Among these techniques, the hysteresis band is used
very often because of its simplicity of implementation, fast
response current and robust structure [11], [12]. Hysteresis
band controller is used to track the line current references.
The current errors between the reference and measured
currents are used to develop three valid switching states in
each inverter leg by the hysteresis band controller.
Actual Current Reference Current
Dead Zone
2.4 Indirect Vector Control
Hysteresis
Band
ia*+h
ia*+δ
ia*
ia*-δ
ia*- Vdc/2
Vdc/2
h
-Vdc/2
+1 states) or in the upper band (through 0 and -1 states).
Here U 1 , means the switch state is Vdc / 2 ; U 0 means
the switch state is 0; and U 1 , means the switch state is
Vdc / 2 . Similarly the b-phase and c-phase switching
function for the three-phase voltage source inverter can be
obtained.
-Vdc/2
Fig.
The indirect vector control is a technique that controls
the dynamic speed of Induction motor. Unlike direct
vector control, in indirect vector control, the unit vectors
are generated in an indirect manner. Fig.4 is the phasor
diagram that explains the fundamental principle of indirect
vector control. The d s q s axes are fixed on the stator and
3: Three-Level Hysteresis Current Control
d r q r axes are fixed on the rotor which rotates at a speed
To develop a switching scheme for the three-level
inverter, the zero voltage level should be applied only at
appropriate instants. The switching logic must ensure that
V
there is no successive transition between dc and Vdc
2
2
states, as this will increase the frequency of switching. A
dead zone ‘δ’ is necessary in the hysteresis band ‘h’, to
avoid switching towards two-level scheme, because of
finite sampling rate of error. Without the dead zone, when
the error becomes zero and is not detected, the opposite
polarity of forcing function follows, resulting in a twolevel scheme. However, the introduction of dead zone
increases the tracking error and has to be chosen to a
minimum value, depending on the best sampling speed
that can be achieved [13].
r . Synchronously rotating axes d e q e are rotating ahead
of d r q r axes by the positive slip angle sl corresponding
to slip frequency sl . Thus
If U represents the input state to be applied, e
represents error (ia* ia ) and ce represents the change in
error the switching logic is governed by equation (12)
e e dt (r sl )dt
(13)
For decoupling control qr 0 or pqr 0 and r dr
.Substituting the above condition in equations (3), (4), (7)
and (8).
is
Vs
Vqse
Vdss
e
s
qs
V
sl
r
Vdse
iqse
iqss
dre r
e
ds
i
(q-d, e)
Synchronous
reference frame
r
idss
(q-d, r)
Rotor reference
frame
If e 0 then
(q-d, s)
Stator Reference
frame
U 1 for e h
U 0 for e
Fig. 4: Phasor diagram of Indirect Vector Control principle
U 0 for e h and ce 0
sl
U 1 for e h and ce 0
Rr Lm iqse
Lr r
(14)
Te
3 P Lm e
r iqs
2 2 Lr
(15)
U -1 for e -h
U 0 for e -
iqse
2 2 Lr Te
3 P Lm r
(16)
U 0 for - e - h and ce 0
idse
L
1
[r r pr ]
Lm
Rr
(17)
Else if e 0 then
U -1 for - e - h and ce 0
(12)
The above logic represented in Fig.2 and Fig.3, tracks
reference current either in the lower band (through 0 and
The equations (14-17) are used to produce an adequate
field orientation. These equations could be propagated to
the set point variables [14].
*
*
sl
Rr Lm iqse
(18)
Lr
*
r
2 2 Lr Te*
i
3 P Lm r*
e*
qs
*
(20)
If it is accepted that the rotor flux set point is constant then
its derivative is zero and the above equation is simplified
as
idse
*
r*
(21)
Lm
Using the above equations the block diagram of indirect
vector control of induction motor drive is as shown in
Fig.5.
+
*
-
r
Speed
controller
(G1)
Estimation of
i de s*
idse* , iqse* , sl*
T e*
(G2 )
i
sl*
e*
qs
Transformation
(q, d , e) (a, b, c)
(G3 )
*
i abc
Three-Level u Three- V
abc
abc
Level
Hystersis
Inverter
Comtroller
e
+
+
r
iabc
r
Fig. 5: Block diagram of Indirect Vector Control of IM
It contains three principal blocks, G1 used as speed
*
controller, G2 used for estimation of idse , iqse , sl* andG3 used
*
for current co-ordinate transformation (q, d, e) to (a, b, c).
3. Fuzzy Logic Speed Controller
The speed controller block G1 is proposed to be a
Mamdani type fuzzy controller having five blocks namely
normalizer, fuzzifier, inference mechanism, de-fuzzifier,
and de-normalizer as shown in Fig.6. [15].
e
Normalizer
ce
Fuzzifier
(22)
Input2 ce(t ) e(t ) e(t 1)
(23)
Inference
Mechanism
du
De-Fuzzifier
Where, is the actual speed and is the reference
speed. Two normalization parameters (k1, k2) for inputs (e,
ce) and one de-normalization parameter (k3) for output
(du) are defined. In normalization process the input
variables are scaled in the range of (-1, +1) and in denormalization process the output values of fuzzy controller
are converted to a value depending on the terminal control
element. The determination of normalization and denormalization parameters of fuzzy controller is important
for system stability.
*
3.2Fuzzifier and De-fuzzifier
780V
r*
Reference
speed
Input1 e(t ) * (t ) (t )
(19)
1 * Lr
[r
pr* ]
Lm
Rr
idse
shown in Fig. The error is the difference between the
reference speed and the actual rotor speed.
The fuzzifier processes the crisp input values (e, ce) and
convert them into fuzzy values. Also the fuzzy values
obtained in fuzzy inference mechanism are converted to
crisp output (du) value by a de-fuzzifier. Here, a triangular
fuzzy membership function is defined for each input and
output values by seven clusters. For seven clusters in the
membership functions, seven linguistic variables are
defined as: Negative Big (NB), Negative Medium (NM),
Negative Small (NS), Zero (Z), Positive Small (PS),
Positive Medium (PM), Positive Big (PB). Fig.11 shows
the membership functions used to fuzzify two input values
(e, ce) and de-fuzzify output (du) of the realized fuzzy
controller. The peak or bottom points of the membership
functions to be tuned are a1 and a2 for error (e), b1 and b2
for change in error (ce) and c1 and c2 for output (du).
Therefore the design of fuzzy controller requires the
optimization of nine parameters (k1, k2, k3, a1, a2, b1, b2, c1,
c2).
(e)
NB
NM
NS
a2
a1
NM
NS
1
b2
b1
NB
NM
NS
c2
c1
PS
PM
PB
0
a1
a2
1
Z
PS
PM
PB
0
b1
b2
1
PS
PM
PB
c1
c2
1
Z
De-Fuzzifier
Knowledge
Based rule
1
NB
(ce)
New Fuzzy
Parameters
By PSO
Fig. 6: Block Diagram of Fuzzy Controller
(du)
Z
3.1 Normalizer and De- normalizer
In closed loop control system the use of error (e) and the
change in error (ce) as controller input is a universal
approach. Therefore the fuzzy controller has two inputs,
error and change in error (e, ce) and one output (du) as
1
0
Fig. 7: Membership functions of Inputs and output
In this work the centre of gravity or centroid method is
used for de-fuzzification. As a result the control increment
is obtained by the equation [16].
4.1 Particle Swarm Optimization
m
du
di A i
(24)
i 1
m
A i
i 1
Here di is the distance between ith fuzzy set and the
centre, A i is area value of ith fuzzy set.
3.3 Knowledge Base and Inference Mechanism
The rule definition is subjective and based on expert’s
knowledge and experiences. It establishes the relationship
between outputs with inputs [17]. For the system with two
inputs and seven membership functions in each leads to
forty nine combination of these inputs, in which there are
forty nine rules. The rules are like:
R1. If e = NB and ce = NB Then du is NB
or
R2. If e = NB and ce = NM Then du is NB
or......
R49. If e = PB and ce = PB Then du is PB
Change in Error
(ce)
Table 2: Fuzzy Linguistic Rule Table
NB
NM
NS
Z
PS
PM
PB
NM
NB
NB
NB
NM
NS
Z
PS
Error (e)
NS
Z
PS
NB NB NM
NB NM NS
NM NS
Z
NS
Z
PS
Z
PS
PM
PS
PM PB
PM PB
PB
Particle Swarm Optimization is a population based
stochastic optimization technique, inspired by social
behaviour of bird flocking or fish schooling. In PSO
system the individuals called particles, fly around in a
multidimensional search space and change their position
with time. During its flight, each particle adjusts its
position according to its own experience and according to
the experience of neighbouring particle. The position or
value corresponding to its own experiences called Pbest and
corresponding to the experience of neighbouring particle is
called Gbest. The search for the optimal position advances
as the particles’ velocities and positions are updated. The
fitness of each particle’s position and iteration is
calculated using a pre defined objective (fitness) function
and the velocity of each particle is updated using the Pbest
and Gbest, which were previously defined. The velocity of
ith particle can be modified by the following equation.
vi (n 1) (n) vi (n) C1 r1(n) Pbest xi (n)
i
The rules are represented by a matrix called matrix
inference shown in Table 2. A feature of the rule base used
is the symmetry across the diagonal. This feature occurs in
systems where the physical behaviors of the system exhibit
symmetry, which is consistent in case of speed control of
Induction, motor.
The developed fuzzy logic uses the inference method
for each rule given by the relation
(25)
i du min i (e), i (ce) ; i 1, 2,......49
NB
NB
NB
NB
NB
NM
NS
Z
4. Fuzzy Logic Speed Controller Based on
Particle Swarm Optimization
(27)
C2 r2 (n) Gbest xi (n)
i
Where, vi (n) is the velocity of ith particle at iteration n,
vi (n 1)
r1 (n)
and
is the velocity of ith particle at iteration (n + 1),
r2 (n)
are random numbers with uniform
distribution in the interval [0, 1], (n) is the momentum
or inertial weight constant given by [18]
max min
n
nmax
(n) max
(28)
Here nmax is the maximum number of iteration, max
PM
NS
Z
PS
PM
PB
PB
PB
PB
Z
PS
PM
PB
PB
PB
PB
Therefore, the resulting membership function is given by
(26)
du max 1(du), 2 (du),.......49 (du)
and min are the maximum and minimum weights
respectively. Appropriate values of max and min are 0.9
and 0.4 respectively [19]. The values C1 and C2 are two
positive constants represent the social and cognitive
accelerations for the Pbest and Gbest positions, respectively.
Varying these parameters has the effect of varying the
strength of the pull towards the two bests. Values of
C1 C2 0 mean that both the cognitive and social
accelerations are absent, and particles keep moving at their
current speed until they hit a boundary of the search space
(assuming no inertia) [20].With C1 0 and C2 0 , each
particle searches for the best position in its neighbourhood,
and replaces the current best position if the new position is
better [20]. However, with C2 0 and C1 0 , the entire
swarm is attracted to a single point, Gbest. Furthermore,
having C1 C2 causes each particle to be attracted to its
own personal best position to a very high extent, resulting
in excessive wandering. On the other hand, C2 C1 results
in particles being more strongly attracted to the global best
position, thus causing particles to rush prematurely
towards optima [20]. It is demonstrated that the particle
swarm is only stable and guaranteed to converge to a
stable equilibrium point if the following conditions are
satisfied [21].
0 (C1 C2 ) 4
(29)
(C1 C2 )
1 ( n) 1
2
(30)
However, whether or not this point is actually the global
minimum cannot be guaranteed, and its acceptability as a
solution should be verified. The position of ith particle at
iteration n is xi (n) . The modified position at iteration
(n 1) is given by
xi (n 1) xi (n) vi (n 1)
(31)
Fig. 8: Flowchart of the PSO algorithm
4.2 Optimization of Fuzzy Controller
The Particle Swarm Optimization is applied to automate
and optimize the fuzzy controller design process. The
normalization parameters (k1, k2, k3) and the parameters of
the membership functions (a1, a2, b1, b2, c1, c2) are
optimized by optimizing a properly defined objective or
fitness function [22], [23]. In the context of optimization
our goal is to have a speed response with a short rise time,
small overshoot and near zero steady state error. In this
respect a multiple objective function is defined as
t
t
F e dt e tdt
0
(32)
0
Where, the first term is the measure of fast dynamic
response and the second term is the measure of steady
state error.
Thus the purpose of PSO algorithm is to minimize the
objective function. The PSO based approach to find the
minimum value of objective function is as shown in Fig.8.
The input parameters of the proposed PSO algorithm are:
nmax 100, n pop 30, nvar s 9, C1 0.5, C2 1.25,
max 0.9, min 0.4, 0 (a1, a2 , b1, b2 , c1, c2 ) 1,
a2 a1, b2 b1, c2 c1, 0 k1 6.67e 3,
0 k2 1, 0 k3 6, Stall generation =20 and Function
tolerance = 1e 6.
5. Simulation Results and Discussion
Complete simulation model for vector controlled Induction
motor (IM) drive of the proposed scheme is developed
using MATLAB/ SIMULINK. The motor parameters are:
Rated Power Prated 50HP , Rated Voltage V 480volt ,
Rated Frequency F 50Hz , Pair of poles P=2, Stator
Resistance Rs 0.087 , Rotor Resistance Rr 0.228 ,
Stator Inductance Ls 0.8mH , Rotor Inductance Lr 0.8mH ,
Mutual Inductance Lm 34.7mH , Moment of Inertia
J 1.662Kg.m2 .
Fig.9 shows the scores of the fitness function
corresponding to different generation in PSO. The PSO is
terminated at 41 generations as the termination criteria
reached. The termination criteria of the algorithm is either
the maximum generations reached or the weighted average
change in the fitness function value over Stall generations
is less than function tolerance. The values of nine
parameters used in fuzzy controller and their optimized
values by particle swarm optimization are shown in
Fig.10. The particle swarm optimized input and output
membership functions are in Fig.11.
7
Pole Voltage Vp (V)
x 10
Best: 1.65831e+007 Mean: 1.65847e+007
Mean Score
Best Score
Line Voltage Vab (V)
1.75
1.7
1.65
0
10
20
30
Generation
500
0
-500
40
Fig. 9: Fitness Score versus Generation
3
3.05
3.1
3.15
3.2
3.25
3.3
3.35
3.4
3.45
3.5
3.3
3.35
3.4
3.45
3.5
3.3
3.35
3.4
3.45
3.5
Time(sec)
1000
500
0
-500
-1000
Stator Currents(A)
Score
1.8
3
3.05
3.1
3.15
3.2
3.25
Time(sec)
50
0
-50
3
3.05
3.1
3.15
3.2
3.25
Time(sec)
Fig. 12: Inverter Voltages and Currents
6
50
Speed
(rad/sec)
3.4746
4
-50
0
1
2
1
2
k1
k2
c2
Speed Error
(rad/sec)
0.5
0.6951
c1
DOM
DOM
NB
1
0.5
0
-1
DOM
1NB
0.5
0
-1
NS Z PS
-0.5
0 Error (e)
PM
0.5
-0.5
NM
Change in0Error (ce)
PM
0.5
-0.5
0
Output (du)
PB
1
PM
NS Z PS
PB
1
NS Z PS
NM
6
3
Time(sec)
4
5
6
k3
Fig.12 shows the pole voltage, line voltage and the line
currents (stator line currents) of the three-level inverter
under steady state condition. The line currents are
sinusoidal with almost negligible ripple. Fig.13 shows the
speed tracking performance of the motor following a
trapezoidal speed reference. The speed tracking
experiment is on no load condition. The motor speed
almost tracks the reference speed in both the direction. The
ripple content in the torque during the transition is
comparatively reduced with PSO fuzzy controller as
compared to simple fuzzy controller as shown in Fig.14
NM
5
Fig. 13: Trapezoidal Speed Tracking
Fig. 10: Conventional and PS Optimized fuzzy
Controller parameters
NB
1
0.5
0
-1
4
0
-1
0
0.5
Fig. 11: PS Optimized Input Output Membership
Functions of fuzzy controller
PB
1
Torque Developed
(Nm)
b2
0.2
0.0505
0.5
0.7936
0.2
0.0538
b1
3
Time(sec)
1
200
Torque Developed
(Nm)
a2
1
0.5383
a1
0.0067
6.367e-4
0
0.5
0.7008
2
2
1
Motor Speed
Reference Speed
0
3
0.2
0.0357
Values of Parameters
5
5
Fuzzy Parameters
Particle Swarm Optimized Fuzzy Parameters
200
Torque developed during trapizoidal speed tracking with Fuzzy
100
0
-100
-200
0
1
2
3
4
5
Time(sec)
Torque developed during trapizoidal speed tracking with PSO Fuzzy
6
100
0
-100
-200
0
1
2
3
Time(sec)
4
5
6
Fig. 14: Torque Developed in Trapezoidal Speed Tracking
Fig. 15 and Fig.16 show the performance of motor
for the constant reference speed of
with
constant load torque
in both fuzzy and PSO fuzzy
speed controller. The ripple in speed and torque, when the
motor achieves the reference speed is nearly zero in case
of PSO fuzzy as compared to fuzzy controller. Fig.17 and
Fig.18 show the performance of motor when the load
torque is suddenly changed from
to
and
then from
to
at constant reference speed
1
1.5
2
3
3.5
-1
0
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
0
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
0
5
Torque
(Nm)
2
2.5
Time(sec)
3
3.5
4
4.5
5
0
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
Fig. 19: Performance under variable speed and
constant torque with fuzzy
1
1.5
2
2.5
Time(sec)
3
3.5
Motor Speed
Reference Speed
4
4.5
5
120
Speed
(Rad./Sec)
0.5
80
50
Motor Speed
Reference Speed
4
4.5
5
0
0.5
1
1.5
2
400
2.5
Time(sec)
3
3.5
4
4.5
5
0
Torque Developed
Load Torque
200
0
0
1.5
Torque Developed
Load Torque
0
0
-1
0
1
-200
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
Torque
(Nm)
Speed
(Rad./Sec)
Speed Error
(Rad./Sec)
1
0
0
0.5
200
Fig. 15: Performance under constant speed and
constant torque with fuzzy
120
80
Torque Developed
Load Torque
200
0
0
Motor Speed
Reference Speed
120
0
400
Torque
(Nm)
2.5
Time(sec)
all these cases the performances are better with PSO
fuzzy as compared to fuzzy controller.
Speed
(Rad./Sec)
0.5
1
Motor Speed
Reference Speed
4
4.5
5
Torque
(Nm)
0
0
Speed Error
(Rad./Sec)
Speed
(Rad./Sec)
120
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
Torque Developed
Load Torque
200
0
-200
0.5
1
1.5
2
1
2.5
Time(sec)
3
3.5
0
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
200
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
1.5
2
2.5
Time(sec)
3
3.5
-1
0
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4.5
5
Torque Developed
Load Torque
200
0
0
4
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
3.5
4
Fig. 18: Performance under constant speed and
variable torque with PSO fuzzy
Fig.19 and Fig.20 show the performance of the motor
when the reference speed is suddenly changed from
to
and then from
to
with a constant load torque of
Fig.21 and Fig.22 show the performance of the motor
with variable speed and variable torque. The speed is
increased from
to
and then
decreased from
to
with a
variable load torque. The load torque is increased from
to
and then it is decreased to
In
5
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
Torque Developed
Load Torque
4
4.5
5
Fig. 21: Performance under variable speed and
variable torque with fuzzy
Motor Speed
Reference Speed
100
50
0
0
5
4.5
0
Motor Speed
Reference Speed
4
4.5
5
0
400
3
200
0
Speed(Rad./Sec)
1
2.5
Time(sec)
-200
Torque(Nm)
Speed
(Rad./Sec)
Speed Error
(Rad./Sec)
0.5
2
50
120
1
1.5
Motor Speed
Reference Speed
0
0
Fig. 17: Performance under constant speed and
variable torque with fuzzy
0
0
1
100
Torque Developed
Load Torque
Torque(Nm)
-1
0
0
0
Torque
(Nm)
0.5
Fig. 20: Performance under variable speed and
constant torque with PSO fuzzy
Motor Speed
Reference Speed
4
4.5
5
Speed(Rad./Sec)
0
0
400
Torque
(Nm)
0
120
Speed Error
(Rad./Sec)
Speed
(Rad./Sec)
Fig. 16: Performance under constant speed and
constant torque with PSO fuzzy
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
0.5
1
1.5
2
2.5
Time(sec)
3
3.5
4
4.5
5
200
0
-200
0
Torque Developed
Load Torque
4
4.5
5
Fig. 22: Performance under variable speed and
variable torque with PSO fuzzy
6. Conclusions
The optimal fuzzy controller has been designed off-line
using techniques of Particle Swarm Optimization for
indirect vector control of multilevel inverter fed Induction
motor. It achieves good pursuit of reference speed, starting
without overshoot and rapid rejection of disturbances with
a low drop-out speed. By comparison with fuzzy logic
controller, it testifies that this method is not only robust,
but also can improve dynamic performance of the system.
References
[1] B. K. Bose, Modern Power Electronics and AC Drives, Inc.
Upper Saddle River, NJ-07458: Prentice-Hill PTR
Companies, 2002.
[2] P. Vas, Vector control of AC machines, New York:
Clarendon, 1990.
[3] R. Krishnan, Electric motor drives modeling, analysis and
control. New Delhi: PHI Pvt. Ltd, 2003.
[4] A. Nabae, I. Takahasi, and H. Akagi, “A neutral point
clamped PWM inverter”, IEEE Transaction on Industry
Applications, Vol. 1A, No. 17,Sept./Oct 1981, pp. 518–523.
[5] J. Rodrguez, J.-S. Lai, and F. Z. Peng, “Multilevel inverters:
A survey of topologies, controls, and applications”, IEEE
Transaction on Industrial Electronics, Vol. 49, No. 4, Aug.
2002.
[6] S. Lai and F. Z. Peng, “Multilevel converters- a new breed
of power converters”, IEEE Transaction on Industry
Applications, Vol. 32, No. 3, May 1996, pp. 509–517.
[7] X. Yuan and I. Barbi, “A new diode clamping multilevel
inverter,” IEEE, conference, 1999, pp. 495–501.
[8] J. Kennedy and R. C. Eberhart, “Particle swarm
optimization,” Proceedings of the IEEE Int. Conf. on Neural
Networks, Perth, Australia,1995, pp. 1942–1948.
[9] M. Kazmierkowski and L. Malesani, “Current control
techniques for three-phase voltage-source pwm converters:
A survey,” IEEE Transaction Ind. Electron, Vol. 45, No. 5,
1998, pp. 691–703.
[10] A. Shukla, A. Ghosh, and A. Joshi, “Hysteresis modulation
of multilevel inverters,” IEEE Transaction on Power
Electronics, Vol. 26, No. 5, May 2011, pp. 1396–1408.
[11] J. Zeng, C. Yu., Q. Qi., Z. Yan., Y. Ni., B. Zhang, S. Chen,
and F. F. Wu., “A novel hysteresis current control for active
power filter with constant frequency”, Electric Power
Systems Research, Vol. 68, 2004, pp. 75–82.
[12] M. Milosevic, “Hysteresis current control in three-phase
voltage source inverter”,Zurich: Technical Report, 2003.
[13] S. Srikanthan, M. Mishra, and R. Rao, “Improved hysteresis
current control of three-level inverter for distribution static
compensator application”, IET Power Electronics, Vol. 2,
No. 5, 2009, pp. 517–526.
[14] S. K. Sahu, D. D. Neema, and T. V. Dixit, “Indirect vector
control of induction motor using ANN estimator and ANFIS
controller,” International Journal of Computer Applications,
Vol. 66, No. 14, March 2013.
[15] M. N. Uddin, T. S. Radwan, and M. A. Rahman,
“Performances of fuzzy-logic-based indirect vector control
for induction motor drive,” IEEE Transaction on Industry
Applications, Vol. 38, No. 5, Sept./Oct 2002, pp. 1219–
1225.
[16] C. C. Lee, Fuzzy Logic in Control Systems: Fuzzy Logic
controller Part 1 and Part 2. New Delhi: IEEE Press, 1990.
[17] Z. Zhao, M. Tomizuka, and S. Isaka, “Fuzzy gain
scheduling of PID controllers,” IEEE Transactions on
Systems, Man and Cybernetics, Vol. 23, No. 5, Sept./Oct
1993, pp. 1392–1398.
[18] Y. Bekakra and D. B. Attous, “Sensorless speed based on
MARS with tuning of PI speed controller in FOC of
induction motor drive using PSO,” World Academy of
Science, Engineering and Technology, Vol. 60, 2011, pp.
1550–1555.
[19] R. Eberhart. and Y. Shi, “Comparing inertial weights and
constriction factor in particle swarm optimization,”
Proceeding of the International Congress on Evaluationing
Computation, 2000, pp. 84–88.
[20] E. AP, Fundamentals of Computational Swarm Intelligence.
John Wiley Sons, 2005.
[21] R. Perez and K. Behdinan, “Particle swarm approach for
structural design optimization”, Computers and Structures,
Vol. 85, 2007, pp. 1579–88.
[22] V. Donescu, D. Neacsu, G. Griva, and F. Profumo, “A
systematic design method for fuzzy controller for brushless
dc motor drives,” Proc. of the 27th. IEEE Annual Power
Electronics Specialists Conference, 1996, pp. 689– 694.
[23] F. D. S. Cardoso, J. F. Martins, and V. F. Pires, “A
comparative study of a PI, neural network and fuzzy genetic
approach controllers for an ac drive,” IEEE 5thInternational
Workshop on Advanced Motion Control AMC, Coimbra,
1998, pp. 375–380.
| 9cs.NE
|
MISO: An intermediate language to express
parallel and dependable programs
Alcides Fonseca, Raul Barbosa
arXiv:1608.06171v1 [cs.DC] 22 Aug 2016
CISUC, Department of Informatics Engineering
University of Coimbra
P-3030 290, Coimbra, Portugal
Email: {amaf, rbarbosa}@dei.uc.pt
Abstract—One way to write fast programs is to explore the
potential parallelism and take advantage of the high number
of cores available in microprocessors. This can be achieved by
manually specifying which code executes on which thread, by
using compiler parallelization hints (such as OpenMP or Cilk),
or by using a parallel programming language (such as X10,
Chapel or Æminium). Regardless of the approach, all of these
programs are compiled to an intermediate lower-level language
that is sequential, thus preventing the backend compiler from
optimizing the program and observing its parallel nature.
This paper presents MISO, an intermediate language that
expresses the parallel nature of programs and that can be
targeted by front-end compilers. The language defines ‘cells’,
which are composed by a state and a transition function from
one state to the next. This language can express both sequential
and parallel programs, and provides information for a backendcompiler to generate efficient parallel programs. Moreover, MISO
can be used to automatically add redundancy to a program, by
replicating the state or by taking advantage of different processor
cores, in order to provide fault tolerance for programs running
on unreliable hardware.
I. I NTRODUCTION
In recent years, processor architectures have evolved towards processors with multiple CPU cores, rather than increasing the operating frequency. In order to improve the performance of programs, developers have to design (or redesign)
programs to execute code in parallel. Writing parallel code
can be done through the usage of libraries, such as Pthreads,
or with special parallel programming-languages. In any of the
cases, the parallel code is translated to a sequential intermediate language. Regardless of being compiled or interpreted, the
intermediate language can be used to perform optimizations at
the lower-level code. These optimizations are unaware of the
parallel nature of the program and cannot take into account.
Intermediate languages handle a program as a list of operations, organized in blocks expected to be executed sequentially.
Parallel programs are also considered to be sequential, and
the parallel nature of a program is hidden behind system calls
for which the compiler does not know the semantics. This is
true for LLVM, B3 and Swift intermediate languages, which
are compiled, as well for virtual machine bytecode languages,
such as Java bytecode, .NET Intermediate Language, Python
bytecode and others.
Since these intermediate languages are shared between
several high-level front-end programming languages, they are
the focus of optimization. Parallelization is one type of optimization that can result in very significant speedups for large
programs. Polly [1] is a tool that performs automatic parallelization of loops at the LLVM-IR level. Universal Translation
Library [2] is another approach that takes LLVM-IR, translates
to its own representation of the parallel program and then
generates the resulting LLVM-IR. Both these approaches use
internal representations that are not intended to be a target of
compilation.
SPIRE [3] has tackled this issue by extending LLVM with
detach, attach and barrier constructs that allow asynchronous
execution. This simple extension makes LLVM able to represent parallel languages such as OpenCL, Cilk, OpenMP, X10
and Chapel, while maintaining some lower-level semantics.
Despite this effort, there are some other parallel semantics
not being represented, such as parallel for-loops and task
dependencies, among others. The limited semantics of SPIRE
prevent some optimizations that could have been done otherwise.
In this paper, we present MISO, an intermediate language
for representation of parallel programs, supporting both Single
Instruction Multiple Data (SIMD) and Multiple Instructions
Multiple Data (MIMD). MISO can be used for performing
optimization at the parallelization level, as well as for improving the dependability of programs running on unreliable
hardware.
II. A MISO P RIMER
MISO stands for Multiple-Input Single-Output, which describes the semantics of each MISO block. MISO is organized
in ‘cells’, of which there can be several instances. A cell has
two components: the state definition and the state transition
function.
Listing 1 shows the example code of a program that
progressively blends one image into another. The ImageBlend
cell is defined as having three memory slots: r, g and b, all
integers (the state). That cell also has a transition function
that performs a weighted sum of the value of each color in
the previous state, and the same color from the second image
in the previous state as well. The saved values are stored in
a new state that will be feed into the next transition. In the
bottom of the program, we can see the instantiation of several
cells, one for each pixel of the two images. The source code
for StaticImage is omitted, as it is similar to ImageBlend with
an empty transition function.
particle strikes, that must be corrected in order for programs
to produce correct results. We introduce redundancy in MISO
in order to tolerate hardware errors.
Listing 1. An example of a MISO program
In MISO, all state transitions are isolated from one another.
Since the state of one cell is only written by one state-transition
c e l l I m ag eBlen d {
function, replication of operations can be achieved through
var r : In t = 0;
the same means as parallelization. Given that each cell has
var g : In t = 0;
its own memory, the memory contents may be duplicated and
var b : In t = 0;
the state-transition function can operate on both replicas. This
replication can be applied in parallel in different CPU cores or
transition {
processors. In NUMA architectures, the state duplication can
r = . 9 9 ∗ r + . 0 1 ∗ im ag e2 ( t h i s . p o s ) . r ; be achieved even in different memories. Identifying a soft error
g = . 9 9 ∗ g + . 0 1 ∗ im ag e2 ( t h i s . p o s ) . g ; can be done by comparing the two new states. If there is a
b = . 9 9 ∗ b + . 0 1 ∗ im ag e2 ( t h i s . p o s ) . b ; mismatch, a third equal transition should be executed to decide
}
between the two possible outcomes. By identifying MISO cells
}
that are frequently erroneous, it is possible to detect permanent
failures in hardware requiring maintenance.
im ag e1 = new I m ag eBlen d ( 3 0 0 ∗ 2 0 0 )
Replication is managed by the runtime environment, and
im ag e2 = new S t a t i c I m a g e ( 3 0 0 ∗ 2 0 0 )
therefore the same MISO program may be executed with
One aspect of MISO is that loading input and output data different levels of redundancy. In some circumstances a given
can be performed by the runtime. In this case, the program piece of code may be crucial for a specific application, thereby
could load the two images from file and could output all requiring full duplication and comparison of MISO cells, while
intermediate states to screen, resulting in a video animation the same piece of code may play a secondary role in a
non-critical application and require no replication. Selective
transition from one image to another.
The state transition moves each cell from one state to the replication of key cells may also be applied by the runtime,
next. In order to achieve this safely, there can be only writes in order to balance the fault tolerance and the overhead.
to the current state, or local variables. Reads can be performed
V. C ONCLUSION
from the previous state of either the current cell or any other
This paper presented MISO, an intermediate language that
cell. This semantic restriction allows for further optimizations.
can represent programs with both task and data parallelism
implicit in its semantics. By isolating state and state transiIII. PARALLELIZATION OF MISO
One of the main challenges for automatic parallelization is tions, compiler tools can reason about the program in order to
the extraction of data-flow information, in order to understand perform optimizations in terms of parallelization. Furthermore,
what computations can influence further operations, in order to the state isolation of MISO cells allows for replication in
establish dependencies and guarantee that the parallel program different hardware processors and memories, in order to detect
has the same semantics of the sequential version. MISO de- and recover from transient hardware faults.
The MISO toolset is currently under development, and a
scribes those dependencies explicitly in the transition function.
prototype
is available at https://github.com/alcides/miso-dsl,
Cells that have dependencies can be synchronously executed,
featuring
a
sequential and parallel runtime execution system.
while cells without any direct or indirect dependency can be
The
toolset
is open source and open to improvements and
executed in parallel, removing the need for a global barrier
suggestions
from
the community.
per transition step.
Furthermore, both task and data parallelism can be described
in MISO. Task parallelism (Multiple Instructions, Multiple
Data) can be represented using different types of cell and
Data parallelism (Single Instruction, Multiple Data) can be
expressed by having several instances of the same cell. Data
parallelism in MISO can be directly translated to GPU languages, such as CUDA or OpenCL.
IV. D EPENDABILITY
OF
MISO
Integrated circuits are continually evolving toward higher
density and greater number of transistors, as well as smaller
gates and lower operating voltages. Although this trend vastly
increases the performance of microprocessors, it also increases
the rate at which soft errors occur. Soft errors are incorrect
states in the hardware, caused by transient events such as
ACKNOWLEDGMENT
The first author was supported by the Portuguese National
Foundation for Science and Technology (FCT) through a
Doctoral Grant (SFRH/BD/84448/2012).
R EFERENCES
[1] T. Grosser, H. Zheng, R. Aloor, A. Simbürger, A. Größlinger, and L.N. Pouchet, “Polly-polyhedral optimization in llvm,” in Proceedings of
the First International Workshop on Polyhedral Compilation Techniques
(IMPACT), vol. 2011, 2011.
[2] A. Y. Drozdov, S. Novikov, V. Vladislavlev, E. Kochetkov, and P. Ilin,
“Program auto parallelizer and vectorizer implemented on the basis of
the universal translation library and llvm technology,” Programming and
Computer Software, vol. 40, no. 3, pp. 128–138, 2014.
[3] D. Khaldi, P. Jouvelot, F. Irigoin, and C. Ancourt, “Spire: A methodology
for sequential to parallel intermediate representation extension,” in 17th
Workshop on Compilers for Parallel Computing (CPC 2013), 2012.
| 6cs.PL
|
ON TOPOLOGIES ON THE GROUP (Zp )N
arXiv:1610.00046v1 [math.GR] 30 Sep 2016
I. K. BABENKO AND S. A. BOGATYI
Abstract. It is proved that, on any Abelian group of infinite cardinality m, there exist
m
precisely 22 nonequivalent bounded Hausdorff group topologies. Under the continuum
hypothesis, the number of nonequivalent compact and locally compact Hausdorff group
topologies on the group (Zp )N is determined.
The systematic study of Abelian topological groups was initiated in Pontryagin’s paper [18] and van Kampen’s paper [9] (see also [1]). In these papers, an exceptionally
important relationship between an Abelian topological group and the topological group
of its characters was revealed. It is not an exaggeration to say that this relationship is one
of the most important tools for studying Abelian topological groups. This relationship
manifests itself most pronouncedly for compact and locally compact groups.
Topological groups close to compact and locally compact ones are, respectively, totally
bounded and locally bounded Abelian topological groups. This type of topologies on
Abelian groups was studied by A. Weil [21], who showed that an Abelian topological
group is totally bounded if and only if it is embedded in a compact Abelian group as a
dense subgroup and that such a group is locally bounded if and only if it is embedded
in a locally compact group as a dense subgroup. Another deep characteristic property
of locally bounded and totally bounded topological, not necessarily Abelian, groups were
found by A.D. Alexandroff [2]. He showed in particular that a topological group G is
locally bounded if and only if G admits a left invariant generalized measure, see [2] for
more details.
Among all bounded group topologies on a given Abelian group there is a maximal
topology; the Weil completion of a group with this topology coincides with the Bohr
compactification of this group and can be continuously mapped onto any compact group
extension.
The next important step in the study of Abelian topological groups was made by
Kakutani [7], who showed that the cardinality of an infinite compact Abelian group G
and the cardinality of the discrete group G∗ of its characters are related by
∗
|G| = 2|G | .
In the 1950s, interest in the study of general properties of Abelian topological groups
motivated the interest in the study of the relationship between the algebraic structure of
a given group and the possible topologies consistent with its group structure. This topic
includes, in particular, the problem of describing Abelian groups admitting a compact
topology, which was posed by Kaplansky in [8]. This problem was completely solved by
Hulanicki in [14, 15].
This work was partially supported by the Russian Foundation for Basic Research (project no. 16-0100357A), and by ANR-Finsler.
1
2
I. K. BABENKO AND S. A. BOGATYI
We mention two results concerning the possible compact topologies on a given Abelian
group. In the very early 1950s, Kulikov [11] showed that the cardinality of the family
of all compact Abelian groups of cardinality 2m , where m is an infinite cardinal, equals
2m . In the light of this result the following example of Fuchs [12] looks quite surprising.
He constructed an Abelian group G of cardinality 2m admitting infinitely many different
compact topologies. Moreover, the cardinality of the set of pairwise non-homeomorphic
compact invariant topologies on G equals 2m .
It is well known that, on any Abelian group G, there are as many nonequivalent Hausdorff group topologies as possible. To be more precise, the cardinality of the set of
|G|
nonequivalent Hausdorff group topologies on G equals 22 . This result was obtained
by Podewski [17], who studied the problem in its most general setting. Namely, given
any set with a finite or countable system of algebraic structures, Podewski found necessary and sufficient conditions on these algebraic structures ensuring the existence of
a nondiscrete Hausdorff topology consistent with all of them [17]. Moreover, Podewski
showed that these conditions ensure the existence of the maximal possible number of
pairwise non-equivalent Hausdorff topologies which agree with all algebraic structures.
For Abelian groups and fields, Podewski’s conditions always hold; details can be found
in [17]. In 1970, Arnautov constructed an example of an infinite commutative ring such
that any Hausdorff ring topology on this ring is discrete [4]. In the 1979s–1980s, examples
of infinite non-Abelian groups admitting only discrete Hausdorff group topologies were
given [13], [16], [20].
Fuchs’ example shows that properties of various topologies that an abstract Abelian
group G carries directly depend on the algebraic properties of the group G.
In our view, the influence of the algebraic structure of an Abelian group on the structure
of the set of Hausdorff group topologies on this group has been studied insufficiently. All
topologies considered in the following are assumed to be Hausdorff.
This paper studies the structure of the set of possible topologies on the well known
Abelian group
(1)
G = (Zp )N ,
where Zp = Z/pZ for a prime p. Henceforth for two sets X and Y the set X Y is understood
in usual sense of set theory. As customary in topology, we denote the prime field with p
elements by the same symbol Zp .
Note that there are four canonical, in a certain sense, topologies on G. The group G
has the natural compact topology determined by the structure of the direct product (1);
we denote this topology by τc . The second topology is the discrete topology τd on G. The
character group of (G, τc ), which we denote by (G, τc )∗ = ⊕ℵ0 Zp , is countable and discrete.
Thus, we obtain two topological groups, (G, τc )×(G, τc )∗ and (G, τc )×(G, τd ). Since these
two groups are algebraically isomorphic to G, it follows that, in fact, we obtain two more
topologies on G, which are surely different (they have different weights); we denote these
topologies by τ0 and τ1 , respectively. Obviously, these two topologies are locally compact.
The following theorem is one of the main results of this paper.
Theorem 0.1. For the Abelian group G = (Zp )N , the following assertions hold.
2ℵ0
1. The set of all Hausdorff group topologies on G contains a subset of cardinality 22
consisting of pairwise non-homeomorphic totally bounded topologies.
ON TOPOLOGIES ON THE GROUP (Zp )N
3
2. Under the continuum hypothesis, τc , τd , τ0 , and τ1 are the only locally compact Hausdorff group topology on G. Only one of them, τc , is compact.
2ℵ0
As mentioned above, the general result of Podewski [17] affirms that there are 22
pairwise non-homeomorphic group topologies on G. The first assertion of the theorem
shows that the set of all group topologies on G contains a subset of the same cardinality
consisting of totally bounded pairwise non-homeomorphic topologies on G.
Although totally bounded topologies are very close to compact, under the continuum
hypothesis, only four of these numerous topologies are locally compact, and only one of
them is compact. This fact contrasts with Fuchs’ result in [12] cited above.
Theorem 0.1 follows from more general statements proved below. Thus, its first assertion is, in fact, a general property of Abelian groups and follows directly from Theorem 1.3
proved in Section 1.
Compact and locally compact topologies are studied in Section 2 in a more general
situation. However, in Section 2, we essentially use the algebraic structure of G, namely,
the fact that each of its element has order p. The second assertion of the theorem follows
directly from Corollary 2.4 and, according to this corollary, is equivalent to the continuum
hypothesis. Note that the negation of the continuum hypothesis drastically changes the
situation with locally compact topologies; see Proposition 2.3.
Section 1.1 considers the set of all totally bounded topologies on the group G. This
subset, which has the largest possible cardinality in the set of all topologies on G, has a
number of interesting structures. For example, we can single out lattices of topologies.
This, in turn, allows us to construct chains of decreasing totally bounded topologies
on G. Moreover, there exist chains of maximum length, which have maximal or minimal
elements. These structures are studied in Section 1.1.
One of our main tools for studying Abelian topological groups is, as usual, the character
theory founded by Pontryagin in [18]. This theory has been developed most fully for locally
compact Abelian groups, for which Pontryagin’s duality theorem is valid (see [19]). Among
other classes of topological groups whose connection with their character groups remains
strong enough are the already mentioned classes of totally bounded and locally bounded
groups. The relationship between these groups and their character groups was studied by
Comfort and Ross [5].
For other classes of Abelian topological groups, connection with character groups is
not so pronounced; nevertheless, character groups remain an important tool for studying
them.
In conclusion, we mention that our interest in this topic was aroused by a series of
questions asked to us by R. I. Grigorchuk. We thank him for numerous fruitful discussions.
We are also grateful to O. V. Sipacheva for useful conversations.
1. The Structure of the Set of All Topologies
Q
Let
m
be
a
finite
or
infinite
cardinal.
In
what
follows,
we
use
the
notation
Q
P m Zp =
(Zp )α , where the cardinality of the index set is |A| = m; accordingly,
m Zp =
α∈A
⊕ (Zp )α .
α∈A
4
I. K. BABENKO AND S. A. BOGATYI
Given another cardinal k, we set
(2)
Fk,m (p) =
Y
Zp ×
X
Zp .
m
k
In the case where p fixed or is clear from the context, we denote groups of the form (2)
by Fk,m .
Endowing the first factor with the compact product topology and the second with the
discrete topology, we turn Fk,m into an Abelian topological group. In what follows, we
assume that at least one of the cardinals k and m is infinite and the finite cardinals can
take only the value 0. The easy-to-prove properties of this Abelian topological group are
collected in the following proposition.
Proposition 1.1. The topological group Fk,m is locally compact; it has cardinality
max{2k , m} and weight max{k, m}. The group of characters of Fk,m is isomorphic to
the group Fm,k . For different pairs of cardinals (k, m), the topological spaces Fk,m are
nonhomeomorphic.
Below we give general facts about topologies on Abelian groups generated by subgroups
of the character group.
We denote the entire group of characters of an Abelian group A, i.e., the group of
all possible homomorphisms of A to the circle S 1 = {eit }, by A⋆ . If A is endowed with
a group topology, then we denote the group of topological characters, i.e., continuous
homomorphisms, of the Abelian topological group A to the circle by A∗ . Note that A∗ is
always a subgroup of A⋆ ; if the topology of A is discrete, then these two groups coincide.
As usual, a subset X ⊂ A⋆ is said to separate vectors if, for any different v, w ∈ A,
there exists an h ∈ X such that h(v) 6= h(w). Any set X ⊂ A⋆ generates a diagonal map
(3)
∆X : A −→ (S 1 )X .
A set X separates vectors if and only if the diagonal map (3) is a monomorphism. Note
that X separates vectors if and only if so does the subgroup hXi generated by this set.
As is known [19], for any nonzero element v ∈ A, there exists a character φv ∈ A⋆ such
that φv (v) 6= 1. Uniting such characters over all elements v ∈ A, we obtain a set X ⊂ A⋆
separating the elements of A. Let H = hXi be the subgroup generated by this set; as
mentioned above, H separates points as well. If |A| = m, where m is an infinite cardinal,
then, obviously, we have |H| ≤ m. Note also that
diagonal map ∆H implements an
Q the
1
isomorphism of the group A onto its image in
(S )h .
h∈H
Given a subgroup H ⊂ A⋆ , we denote the minimal topology on A with respect to which
all maps in H are continuous by τH . This topology coincides with the minimal topology
with respect to which the diagonal map (3) is continuous (the image is assumed to be
endowed with the product topology). Obviously, τH is Hausdorff if and only if H separates
points. Recall that a topology τ on a group G is said to be totally bounded if, for any
neighborhood U of zero, there exists a finite set {0, g1, . . . , gk } of elements of G such that
the family {U, g1 + U, . . . , gk + U} covers the entire group.
Topologies determined by various subgroups H ⊂ A⋆ were studied by Comfort and
Ross [5]; we collect the results of [5] needed later on in the following theorem.
ON TOPOLOGIES ON THE GROUP (Zp )N
5
Theorem 1.2. [5] 1. For any subgroup H ⊂ A⋆ , the topology τH is totally bounded,
and for each totally bounded topology τ on A, there exists a subgroup H ⊂ A⋆ such that
τ = τH .
2. The identity map id : (A, τH2 ) −→ (A, τH1 ) is continuous if and only if H1 ⊂ H2 .
Now, we prove a general statement on the number of totally bounded topologies on an
Abelian group.
Theorem 1.3. Let A be an Abelian group of cardinality |A| = m, where m is an infinite
m
cardinal. Then, on A, there are precisely 22 totally bounded Hausdorff group topologies.
m
Among them there are 22 pairwise nonhomeomorphic topologies.
Proof. Let A⋆ be the group of all characters of A; by Kakutani’s theorem, we have |A⋆ | =
2m . Fix a subgroup H ✁ A⋆ separating the points of characters. As mentioned above,
this subgroup can always be chosen so that |H| ≤ m. Let B = A⋆ /H; then |B| = 2m .
Given a subgroup C ✁ B, we set C̄ = p−1 (C), where p : A −→ B is the natural
projection. Since H ✁ C̄, it follows that the topology τC̄ on A is Hausdorff. Thus, to
each subgroup of B we have assigned a totally bounded Hausdorff group topology on A.
These topologies are different for different subgroups. Indeed, according to Theorem 1.2,
the identity map is not a homeomorphism with respect to topologies corresponding to
different subgroups. In turn, Lemma 1.4 implies that the cardinality of the set of such
m
topologies is 22 .
Finally, let us count nonequivalent topologies among those obtained above. By nonequivalent topologies we mean topologies that cannot be mapped to each other by a homeomorphism, not necessarily consistent with the group structure.
Since |A| = m, it follows that the set of all self-maps of A has cardinality mm = 2m
(as a set). It follows that any topology on A is equivalent to at most 2m other topologies
m
on this set. Thus, we obtain a partition of 22 different totally bounded group topologies
into classes of equivalent topologies, each of which has cardinality at most 2m . Therefore,
m
the number of classes is precisely 22 , and topologies from different classes are pairwise
nonequivalent. This completes the proof of the theorem.
Lemma 1.4. Let m be an infinite cardinal, and let B be an Abelian group of cardinality
m
2m . Then the cardinality of the set of pairwise different subgroups in B equals 22 .
Proof. The upper bound for the cardinality of the set of subgroups in B is obvious and
coincides with the cardinality of the set of all subsets in B. Let us prove the lower bound.
Let T = Tors(B) be the torsion subgroup of B; then B ′ = B/T is torsion-free. The
cardinality of at least one of the groups T and B ′ equals 2m , and the presence of the
required number of pairwise different subgroups in T or in B ′ implies the presence of the
same number of subgroups in B. Thus, it suffices to prove the lemma for periodic groups
and torsion-free groups separately. Consider two cases.
1. Suppose that the group B is periodic. Since B decomposes into the direct sum of its
primary components [10], it follows that at least one primary component has cardinality
2m , and it suffices to prove the required assertion for the case of a primary p-group, where
p is a prime.
Suppose that B is a p-group. Consider the subgroup B(p) consisting of all elements
of order p. Obviously, |B(p)| = 2m , so that it suffices to prove the required assertion
for the group B(p). In this case, B(p) is a Zp -vector space, and it has a Hamel basis of
6
I. K. BABENKO AND S. A. BOGATYI
cardinality 2m . Different subsets of this basis generate different subgroups of B(p) and,
therefore, of B. This completes the proof in the case of periodic groups.
2. Finally, suppose that B is torsion-free. Since |B| = 2m , it follows that the (Prüfer)
rank of B equals 2m . Thus, there is a free Z-module of rank 2m in B. As above, different
subsets in a family of generators of such a module generate different subgroups in the
module and, therefore, in B.
1.1. The Structure of the Family of Totally Bounded Topologies. In what follows,
we essentially use the structure of the groups Fk,m (p) and, in particular, the fact that these
group can be treated as Zp -vector spaces. The characters of such groups are simply the
elements of the dual space. In what follows, we use this dual understanding of characters
without mention.
Consider the case where the group A is a Zp -vector space of dimension equal to an infinite cardinal m. In this case, Theorem 1.3 is substantially simpler and has a transparent
combinatorial interpretation. The subgroup H ✁ A⋆ of characters separating points is a
subspace of the same dimension m. Next, we choose a Hamel basis B in H and extend it
to a Hamel basis B⋆ in the entire space A⋆ . Let M = B⋆ \ B; obviously, we have |M| = 2m .
Each set N ⊂ M now determines a subgroup in A being the linear span of the system
of vectors {N ∪ B}. The set of pairwise different subgroups thus defined has cardinality
m
22 and can be used in Theorem 1.3 for the special case where A is a Zp -vector space.
The Hausdorff group topologies on A thus obtained have an additional structure: on this
set of topologies, there arise lattices.
The construction described above shows that, for the set M = B⋆ \ B of cardinality 2m ,
the lattice (2M ; ⊂, ∪, ∩) of subsets is embedded in the lattice of totally bounded linearly
invariant topologies on the space A. Therefore, the structures on the lattice (2M ; ⊂, ∪, ∩)
can be transferred to topologies. Thus, any linearly ordered set M generates a chain of
length 2m of one-to-one continuous maps. In the family of topologies under consideration,
there is the minimal topology τmin = τHB and the maximal topology τmax = τA⋆ .
Thus, on a vector space A of dimension m, there are chains of length 2m of one-to-one
continuous maps, and there are no longer chains. Such chains carry various combinatorial
structures, which correspond to different types of ordering of the vectors in the Hamel
basis.
m
m
Moreover, the existence of 22 independent subsets in M implies the existence of 22
totally bounded linearly invariant Hausdorff topologies on A such that the identity map
is not continuous with respect to any pair of these topologies.
Proposition 1.5. Let A be a vector space of infinite dimension m. Then any separating
subspace H ⊂ V ⋆ contains a separating subspace of dimension at most m.
Proof. Indeed, given any pair of different vectors in the space A, it suffices to take a linear
functional on H separating them. The dimension of the linear span of such functionals
does not exceed m2 = m.
Proposition 1.6. Let τ be a totally bounded linearly invariant Hausdorff topology on a
vector space A of infinite dimension m. Then there is a set M of cardinality 2m such
that, on the set of totally bounded linearly invariant Hausdorff topologies on A, there is
a lattice isomorphic to (2M ; ⊂, ∪, ∩). Moreover, the initial topology τ is either minimal
ON TOPOLOGIES ON THE GROUP (Zp )N
7
or maximal in this lattice. In particular, any totally bounded linearly invariant Hausdorff
topology on A can be included in a linear chain of length 2m of different topologies.
Proof. Consider the space H of linear functionals generating the given topology. If the
dimension of H is less than 2m , then we take the Hamel basis B in H, extend it to a
Hamel basis B⋆ in A⋆ , and set M = B⋆ \ B̃. In this case, (2M ; ⊂, ∪, ∩) is the required
lattice, and the given topology is minimal in this lattice.
Suppose that the dimension of H equals 2m . According to Proposition 1.5, H contains
a separating subspace Q ⊂ H of dimension m. We take a Hamel basis B in Q and extend
it to a Hamel basis B̃ in H. As above, we set M = B̃ \ B. The desired lattice is again
(2M ; ⊂, ∪, ∩), and the topology determined by H is maximal in it.
Q
Proposition 1.7. Let k be an infinite cardinal, and let A = Fk,0 = k Zp . Then the
subspace H0 ⊂ A⋆ generated by the coordinate projections of A is a minimal separating
space.
Proof. Obviously, the subspace H0 itself is separating. Suppose that there exists a separating subspace H ⊂ H0 . Consider the one-to-one continuous map generated by the
embeddings (A, τH0 ) −→ (A, τH ). The topology of τH0 coincides with the Tychonoff product topology on A and, therefore, is compact. Since a continuous one-to-one map of a
compact space to any Hausdorff space is a homeomorphism, the separating subspace H
cannot be proper.
The following result shows that there are no other minimal separating subspaces.
Theorem 1.8. For a Zp -vector space A, a subspace H ⊂ A⋆ is a minimal separating
space if and only if the natural map e : A −→ H ⋆ is an isomorphism.
Proof. Obviously, the natural map e is a monomorphism if and only if H ⊂ A⋆ is a
separating subspace.
Suppose that e is an isomorphism and H1 ⊂ H is a proper subspace. Since the map
p : H ⋆ −→ H1⋆ of the dual spaces has nontrivial kernel, it follows that so does the natural
map e1 = p ◦ e in H1⋆ . Therefore, H1 does not separate points, i.e., H is minimal.
Now, let us suppose that H is a minimal separating space and show that e is an
isomorphism. Take a nonzero element φ ∈ H ⋆ and let ker φ = H1 ⊂ H. By minimality, H1
is not separating; hence there exists a nonzero vector v ∈ A such that e(v)(x) = x(v) = 0
for all x ∈ H1 . Since H separates points, it follows that there exists a y ∈ H for which
y(v) 6= 0. Without loss of generality, we can assume that y(v) = 1.
The subspace H1 has codimension 1, because this is the kernel of a linear functional.
Therefore, H1 and y generate H; in particular, φ(y) 6= 0. Finally, for any vector z ∈ H,
we have z = x + ty, where x ∈ H1 . This implies
φ(z) = tφ(y) = tφ(y)y(v) = ty(φ(y)v) = (x + ty)(φ(y)v) = e(φ(y)v)(z).
Thus, φ = e(φ(y)v), so that e is an epimorphism and, by the remark at the beginning of
the proof, an isomorphism.
Corollary 1.9. Let A be a Zp -vector space for which H ⊂ A⋆ is a minimal
separating
Q
subspace of infinite
dimension
k.
Then
there
exists
an
isomorphism
i
:
Z
−→
A such
p
k
Q
⋆
⋆
that
i
(H)
⊂
(
Z
)
coincides
with
the
subspace
generated
by
the
basis
projections
of
k p
Q
k Zp .
8
I. K. BABENKO AND S. A. BOGATYI
Proof.PIndeed, it suffices to choose
Q a Hamel basis in H; this determines an isomorphism
⋆
H ≃ k Zp , whence A ≃ H ≃ k Zp .
We define the logarithm of a cardinal m by
Ln(m) = {n : m = 2n }.
Obviously, k ∈ Ln(2k ) for any cardinal k. Moreover, the set Ln(ℵ0 ) is empty by Cantor’s
theorem. Finally, the generalized continuum hypothesis (2ℵα = ℵα+1 ) implies Ln(2k ) =
{k}.
Corollary 1.10. If the dimension m of a vector space A satisfies the condition Ln(m) =
∅, then A⋆ contains no minimal separating subspace.
2. Locally Compact Topologies
Q In this section, we analyze the set of locally compact topologies arising on the group
n Zp , where n is an infinite cardinal.
Let G be an infinite compact Abelian group; then, according to Kakutani’s theorem
[7], we have |G| = 2m , where m = |G∗ |. Since the character group G∗ is discrete, it
follows that if
Qany element of G has order p, then the group G is isomorphic to one of the
groups G ≃ n Zp , where n ∈ Ln(|G|),
Q with the natural product topology. For different
n ∈ Ln(|G|), the topological groups
n Zp are nonhomeomorphic, because, according to
Q
Proposition 1.1, the weight of n Zp equals n. Finally, note that under the generalized
continuum
hypothesis there exists only one compact topology on each direct product
Q
n Zp . Thus, the following assertion is valid.
Proposition 2.1. If n is an infinite cardinal, then
Q the cardinality of the set of compact
Hausdorff group topologies on the group G ≃ n Zp equals |Ln(|G|)|. Moreover, each
n′ ∈ Ln(|G|) corresponds to precisely one compact topology, and its weight equals n′ .
P
Theorem 2.2. Let τ be a locally compact Hausdorff topology on the group G ≃ m Zp .
Then the group (G, τ ) is topologically isomorphic to Fk,n , where the cardinals k and n
satisfy the condition max{2k , n} = m.
Proof. All elements of the group G and, therefore, all elements of its character group G∗
have order p. It follows that the group (G, τ ) is totally disconnected.
According to van Dantzig’s theorem [6], there exists a compact open subgroup H ⊂ G.
Let p : G −→ N = G/H be the natural projection, and let i : H −→P
G be an embedding.
Since H is open, it follows that N is discrete and, therefore,
N ≃ n Zp , where n is a
Q
cardinal. According to Proposition 2.1, we have H ≃ k Zp for an appropriate cardinal
k (naturally, the direct product is endowed with the compact product topology).
Next, choosing a Hamel basis in the Zp -vector space N, we take a splitting q : N −→ G
of the projection p. The homomorphism q is automatically continuous, and its image is
discrete in G.
Finally, consider the homomorphism i×q : H×N −→ G defined by i×q(h, n) = i(h) +
q(n). Obviously, i × q is an algebraic isomorphism. Since H is open, it follows that this
is a local homeomorphism between H×N and G, which, in turn, implies that i×q is an
isomorphism of topological groups.
Thus, we conclude that (G, τ ) is isomorphic to Fk,n as a topological group. The equality
max{2k , n} = m is now obvious.
ON TOPOLOGIES ON THE GROUP (Zp )N
9
The P
following proposition shows what happens to locally compact topologies on the
group n Zp , where n = ℵ1 or 2ℵ0 , under the negation of the continuum hypothesis.
Proposition 2.3.PIf ℵ1 < 2ℵ0 , then there is only one (discrete) locally compact topology on the group ℵ1 Zp . Under the same assumption, there
P exist at least five pairwise
nonhomeomorphic locally compact topologies on the group 2ℵ0 Zp .
P
Proof. According to Theorem 2.2, the group ℵ1 Zp with a locally compact topology is
isomorphic to a group Fk,n . Taking into account the inequality ℵ1 < 2ℵ0 , we see that the
only possible choice of the cardinals is k = 0 and n = ℵ1 , i.e., the topology is discrete.
ℵ0
The groups F
Pk,n are locally compact and, under the assumption ℵ1 < 2 , algebraically
isomorphic to 2ℵ0 Zp for the following pairs (k, n) of cardinals:
(2ℵ0 , 0), (2ℵ0 , ℵ0 ), (2ℵ0 , ℵ1 ), (2ℵ0 , ℵ2 ), (0, 2ℵ0 ).
Here the first topology is compact and the last one is discrete. All of these topologies are
pairwise nonhomeomorphic by virtue of Proposition 1.1.
Theorem 2.2 and Proposition 2.3 have the following immediate corollary.
Corollary 2.4. The following conditions are equivalent:
ℵ0
(1) the continuum
P hypothesis holds: ℵ1 = 2 ;
(2) the group Pℵ1 Zp admits a compact group topology;
(3) the group Q ℵ1 Zp admits a nondiscrete locally compact group topology;
(4) the group Qℵ0 Zp admits at most four locally compact group topologies;
(5) the group ℵ0 Zp admits precisely four locally compact group topologies.
In conclusion, we mention that the four locally compact topological groups in condition
(5) are precisely as follows:
Fℵ0 ,0 (this is a compact topology of weight ℵ0 );
Fℵ0 ,ℵ0 (a noncompact nondiscrete topology of weight ℵ0 );
Fℵ0 ,2ℵ0 (a noncompact nondiscrete topology of weight 2ℵ0 );
F0,2ℵ0 (the discrete topology of weight 2ℵ0 ).
This list completes the proof of the second assertion of Theorem 0.1 stated in the
introduction.
References
[1] J. W. Alexander, On the characters of discrete Abelian groups, Ann. of Math., 1934, Vol. 35, No. 2,
389–395.
[2] A.D. Alexandroff, On groups with an invariant measure. C. R. (Doklady) Acad. Sci. URSS (N.S.),
1942, Vol. 34, No. 1, 59.
[3] H. Anzai and Sh. Kakutani, Bohr compactifications of a locally compact Abelian group, Proc. Imp.
Acad. Tokyo, 1943, Vol. 19, No. 8, 476–480.
[4] V. I. Arnautov, Non-discrete topologizability of infinite commutative rings, Dokl. Academy of Sciences SSSR, 1970, Vol. 111, No. 8, 476–480.
[5] W. W. Comfort and K. A. Ross, Topologies induced by groups of characters, Fundamenta Math.,
1964, Vol. 55, 283–291.
[6] D. van Dantzig, Zur topologischen Algebra I, Math. Ann., 1933, Vol. 107, No. 1, 587–626.
[7] Sh. Kakutani, On cardinal numbers related with a compact Abelian group, Proc. Imp. Acad. Tokyo,
1943, Vol. 19, 366–372.
10
I. K. BABENKO AND S. A. BOGATYI
[8] I. Kaplansky, Infinite Abelian Groups (Univ. of Michigan Press, Ann Arbor, 1954).
[9] E. R. van Kampen, Locally bicompact Abelian groups and their character groups, Annals of Math.,
1935, Vol. 36, No. 2, 448–463.
[10] M. I. Kargapolov and Yu. I. Merzlyakov, Foundations of Group Theory (Nauka, Moscow, 1972) [in
Russian].
[11] L. Ya. Kulikov, Generalized primary groups. I–II, Tr. Mosk. Mat. O-va, 1952, Vol. 1, 247–326; 1953,
Vol. 2, 85–167.
[12] L. Fuchs, On character groups of discrete Abelian groups, Acta Math. Hungarica, 1959, Vol. 10,
Nos. 1–2, 133–140.
[13] G. Hesse, Zur Topologisierbarkeit von Gruppen, Dissertation (Univ. of Hannover, Hannover, 1979)
[in German].
[14] A. Hulanicki, Algebraic characterization of Abelian divisible groups which admit compact topologies, Fundamenta Math., 1957, Vol. 44, 192–197.
[15] A. Hulanicki, Algebraic structure of compact Abelian groups, Bull. Acad. Polon. Sci., 1958, Vol. 6,
71–73.
[16] A. Yu. Ol’shanskii, A remark on a countable nontopologizable group, Vestnik Mosk. Univ. Mat.
Mekh., 1980, No. 3, 103.
[17] K.-P. Podewski, Topologisierung algebraischer Strukturen, Rev. Roumaine Math. Pures Appl., 1977,
Vol. 22, No. 9, 1283–1290.
[18] L. Pontrjagin, Theory of topological commutative groups, Ann. of Math., 1934, Vol. 35, 361–388.
[19] L.S. Pontryagin, Continuous Groups (Nauka, Moscow, 1973) [in Russian].
[20] S. Shelah, On a problem of Kurosh, Jonsson groups, and applications, in Word Problems II, Ed.
by S. I. Adian, W. W. Boone, and G. Higman (North-Holland, Amsterdam, 1980), pp. 373–394.
[21] A. Weil, Sur les espaces à structure uniforme et sur la topologie générale, (Hermann, Paris, 1937)
[in French].
Université de Montpellier CNRS UMR 5149, Institut Montpelliérain Alexander Grothendieck, Place Eugène Bataillon, Bât. 9, CC051, 34095 Montpellier CEDEX 5, France;, Mechanics and Mathematics Faculty, M. V. Lomonosov Moscow State University, Moscow
119992, Russia
E-mail address: [email protected]; [email protected]
| 4math.GR
|
A generalization of the Log-Lindley distribution – its properties and applications
S. Chakraborty1*, S. H. Ong2 and C. M. Ng2
1
2
Department of Statistics, Dibrugarh University, India
Institute of Mathematical Sciences, University of Malaya, Malaysia
* Corresponding author. Email: [email protected]
(Version II-21st October 2017)
Abstract
An extension of the two-parameter Log-Lindley distribution of Gomez et al. (2014) with
support in (0, 1) is proposed. Its important properties like cumulative distribution
function, moments, survival function, hazard rate function, Shannon entropy, stochastic n
ordering and convexity (concavity) conditions are derived. An application in distorted
premium principal is outlined and parameter estimation by method of maximum
likelihood is also presented. We also consider use of a re-parameterized form of the
proposed distribution in regression modeling for bounded responses by considering a
modeling of real life data in comparison with beta regression and log-Lindley regression
models.
1. Introduction
In statistical research many attempts have been made to introduce alternative to the
classical beta distribution (see Gomez et al., 2014 for details). But most of these
distributions involved special functions in their formulations barring the Kumaraswamy
distribution (Jones, 2012).
Distribution with support (0, 1) and having a compact
cumulative distribution function (cdf) can be used as a distortion function to define a
premium principle. Recently, a new pdf with bounded support in (0,1) was introduced by
Gomez et al. (2014), as an alternative to the classical beta distribution by suitable
transformation of a particular case of the generalized Lindley distribution proposed by
Zakerzadeh and Dolati (2010). The new distribution called Log-Lindley (LL) distribution
Preprint Version II
1
has compact expressions for the moments as well as the cdf. Gomez et al. (2014) studied
its important properties relevant to the insurance and inventory management applications.
The new model is also shown to be appropriate regression model to model bounded
responses as an alternative to the beta regression model. The LL distribution of Gomez et
al. (2014) is probably the latest in a series of proposals as alternatives to the classical beta
distribution. Unlike its predecessors the main attraction of the LL distribution is that it
has nice compact forms for the probability density function (pdf), cdf and moments,
which do not involve any special functions.
The LL distribution of Gomez et al. (2014) is given by
Pdf: f ( x; , )
2
( log x ) x 1 , 0 x 1, 0, 0
1
(1)
Cdf: F ( x; , )
x [1 ( log x)]
1
(2)
Moments: E ( X r ; , )
2 1 (r )
, r ,2,1,1, 2,
1 (r ) 2
(3)
In this paper we attempt to extend the LL distribution to provide another distribution
in (0, 1) having compact expressions for the pdf, cdf and moments and study its
properties and applications.
2. Extended Log-Lindley distribution LL p ( , )
The LL distribution of Gomez et al., (2014) was obtained from a two-parameter
particular case of the three-parameter generalized Lindley (GL) distribution of
Zakerzadeh and Dolati (2010) which has the pdf
2 ( x) 1 ( x)e x
f ( x)
, 0 x 1, 0, 0 .
( ) ( 1)
Gomez et al. (2014) first substituted 1 and then employed the transformation
X log(1 / Y ) to obtain their LL distribution in (1).
In this section, we extend their proposal by employing the same transformation to
the GL distribution directly to derive what we call the generalized Log-Lindley
2
distribution. This proposed three-parameter distribution can be seen as an extension of
the LL distribution of Gomez et al. (2014) with an additional parameter ‘p’. We denote
the new distribution by LL p ( , ) . The pdf and cdf of the proposed distribution are given
respectively by
f ( x; , , p)
2 p
( log x ) p ( log x ) x 1 ,
(1 p)[1 p ]
0 x 1, 0, 0, p 0
(4)
and
F ( x; , , p )
[ x (1 p ) Ei( p, log x )] ( log x)1 p
(1 p ) [1 p ]
(5)
where Ei(n, z ) e zt / t n dt z n1(1 n, z ) is the exponential integral and (1 n, z ) is
z
the upper incomplete gamma function. In particular, the recurrence relations
n Ei(n 1, z ) e z zEi(n, z ) and
d
Ei(n, z ) zEi (n 1, z )
dz
can be utilized in conjunction with Ei(0, z ) e z / z to obtain expressions for higher
integral values on n. Moreover, as z , Ei(n, z ) e z / z (see http:// mathworld.
wolfram.com/ En-Function.html for more results).
Alternatively, we can write the cdf in (5) in terms of upper incomplete gamma function as
x ( log x )1 p (1 p )(1 p, log x )
.
F ( x; , , p )
(1 p ) [1 p ]
p
Further using the result ( p 1, y ) p! e y y k / k! for integer p, the cdf can be written
k 0
in a compact from devoid of any special function as
p
x ( log x)1 p (1 p ) p! x ( log x) k / k!
F ( x; , , p)
k 0
(1 p ) p!
3
.
Important particular cases of LL p ( , ) :
i.
for p 0 , we get back the LL distribution of Gomez et al. (2014) in (1).
ii.
for p 0, 1 , as , LL p ( , ) Uniform (0, 1) .
iii.
for p 1 , we get new LL1 ( , ) distribution as:
Pdf: f ( x; , )
3
( log x)( log x) x 1 ,
[2 ]
Cdf: F ( x; , )
x [2 log x (2 log x )]
2
3 [(r ) 2]
Moment: E ( X ; , )
,
(r ) 3 [2 ]
r
iv.
0 x 1, 0, 0
r ,2,1, 1, 2,
(7)
(8)
(9)
for p 2 , we get new LL2 ( , ) distribution as
Pdf: f ( x; , )
4
( log x) 2 ( log x ) x 1 ,
6 2
Cdf: F ( x; , )
x [6 2 log x {6 2 log x (3 log x)}]
6 2
Moment: E ( X r ; , )
4 [ 2 ( r ) 6]
,
2(r ) 4 (3 )
0 x 1, 0, 0
r ,2,1, 1, 2,
(10)
(11)
(12)
2.1 Shape of LL p ( , ) distributions
Here we have drawn some illustrative pdf of LL p ( , ) distributions for different choices
of parameters , and index p to study their shapes. It is observed that the distribution
can be positively or negatively skewed and symmetrical. The pdf can also be increasing
or decreasing.
4
(a)
(b)
(c)
(d)
(e)
(f)
Figure 1. Pdf plots of LL p ( , ) when p = 0 (red), 1 (green), 3(blue), 5 (brown), 7
(light green) for (a) 20, 0.001 (b) 5, 0.1 (c) 20, 5 (d)
2, 0.1 (e) 0.5, 10 (f) 0.9, 0
5
2.2 Moment
Here we derive general result for moments of LL p ( , ) distribution and obtain its
particular cases.
2 p ((r ) (1 p ))
E ( X ; , )
, r ,2,1,1, 2,
(r ) 2 p [1 p ]
r
In particular, E ( X ; , )
1
2 p
1 p (1 )
.
1 p
(13)
(14)
2.3 Mode
The mode plays an important role in the usefulness of a distribution. For the LL p ( , )
distribution the mode occurs at
exp[{1 p (1 ) 4 p ( 1) (1 p (1 ))2 } / 2(1 )] .
2.4 Survival and Hazard Rate functions
S ( x; , , p )
r ( x; , , p )
(1 p ){(1 p) ( log x)1 p Ei ( p, log x)} x ( log x)1 p
(1 p) [1 p ]
(1 p ){ (1 p) (1 p, log x)} x ( log x )1 p
(1 p )[1 p ]
2 ( log x ) p ( log x ) x 1
(1 p ){(1 p) ( log x )1 p Ei( p, log x )} x ( log x)1 p
2 ( log x)( log x ) p
(1 p ){(1 p) (1 p, log x)} x ( log x)1 p
(15)
It can be easily checked that for p 0 , the above results reduce to those of the LL0 ( , )
distribution of Gomez et al. (2014). When p is an integer the expressions can be written
p
in compact form using the result ( p 1, y ) p! e y y k / k!.
k 0
6
Some illustrative plots of the hazard function of LL p ( , ) distributions for different
choices of parameters , and index p are presented in Figure 2, which reveals that the
hazard function can be increasing and bath-tub shaped.
Figure 2. Hazard rate function plots of LL p ( , ) when p = 0 (red), 1 (green), 3 (blue), 5
(brown), 7 (light green) for (a) 0.9, 2 (b) 2, 2 (c) 5, 2
2.5 Shannon Entropy of LL p ( , )
Here we derive the Shannon entropy of LL p ( , ) in terms of that of LL0 ( , ) .
H p ( X ) E[ log f ( X ; , , p)]
2 p
E log
( log X ) p ( log X ) X 1
(1 p)[1 p ]
2 p
log
pE log (log(1 / X ) E[log{( log X ) X 1}]
(1 p)[1 p ]
2 p
log
pE log (log(1 / X )
(1 p)[1 p ]
2
2
E log
( log X ) X 1 log
1
1
(1 p )(1 p)
log
pE[log (log(1 / X )] H 0 ( X )
(1 ) p
(16)
It can be checked that for p 0 , H p ( X ) reduces to H 0 ( X ) , the Shannon entropy of
LL0 ( , ) which is the Log-Lindley (see proposition 2 in page 51 of Gomez et al., 2014).
7
For integral values of index parameter p, exact expression for Elog (log(1 / X )] can be
obtained using Mathematica as follows.
p 1, E log {log(1 / X )}
3 (2 )( log )
2
p 2, E log {log(1 / X )}
11 3 2(3 )( log )
6 2
p 3, E log {log(1 / X )}
50 11 6(4 )( log )
24 6
p 4, E log {log(1 / X )}
274 50 24(5 )( log )
,…
120 24
Thus, for p 1 , Shannon Entropy for LL1 ( , ) using (16) is given by
(2 ) 3 (2 )( log )
H1 ( X ) log
H0( X ) .
2
2
(1 )
2.6 Stochastic Ordering
We will consider the likelihood ratio (LR), hazard rate (HR) and stochastic (ST)
orderings for LL p ( , ) random variables in this section.
Theorem1. Let X 1 and X 2 be random variables following LL p1 (1 , 1 ) and LL p2 ( 2 , 2 )
distributions, respectively. If 1 2 , 1 2 and p2 p1 then X1 LR X 2 .
Proof: Consider the ratio
2 p
f ( x; 2 , 2 , p2 ) 2 2 (1 p1 )[1 p1 11 ]
( log x) p2 p1 h( x )
f ( x; 1 , 1 , p1 ) 12 p1 (1 p2 )[1 p2 2 2 ]
where h( x)
(17)
2 log x 2 1
x
. Gomez-Deniz et al. (2014) has shown that the function
1 log x
h(x) is non-decreasing for x (0,1) if 1 2 , 1 2 . Here, if p2 p1 , it is clear that
( log x ) p2 p1 is non-decreasing for x (0,1) . This implies that if 1 2 , 1 2 and
p2 p1 , then the ratio in (17) is non-decreasing for x (0,1) and hence, X1 LR X 2 .
8
The result on LR ordering leads to the following (Gomez-Deniz et al., 2014):
X 1 LR X 2 X 1 HR X 2 X 1 ST X 2 . Therefore, similar results as in Corollary 1 of
Gomez-Deniz et al. (2014) can be shown for the new log-Lindley distribution as follows.
Corollary 1. Let X1 and
X 2 be random variables following LL p1 (1 , 1 ) and
LL p2 ( 2 , 2 ) distributions, respectively. If 1 2 , 1 2 and p2 p1 , then
a. the moments, E[ X 1k ] E[ X 2k ] for all k 0 ;
b. the hazard rates, r1 ( x) r2 ( x) for all x (0,1) .
2.7 Convexity and Concavity for Generalized Log-Lindley Distribution
For brevity, we suppress the cdf notation for LL p , distribution in (5) as F (x) in this
section.
Theorem 2: If 0 1 , p 0 and 0 , then F (x) is concave for x (0,1) . Hence,
for 0 1 , F (x) is also log-concave for x (0,1) since F ( x) 0 .
Proof: If 0 1 , then ( log x) p , ( log x) and ( x ) 1 are decreasing in x (0,1) for
p 0 and 0 . This implies that the pdf
F ( x) f ( x; , , p )
2 p
( log x) p ( log x ) x 1 ,
(1 p )[1 p ]
is decreasing in x (0,1) . Thus, F (x) is concave for x (0,1) .
Since F ( x) 0 , concavity implies F (x) is also log-concave for x (0,1) .
Theorem 3. The function F (x) is not convex or concave for x (0,1) for any 1 ,
p 0 and 0 .
Proof: For any 1 , p 0 and 0 , consider the second order derivative of F (x)
given by
F ( x )
2 p
p 1
p
x 1 ( log x) p1 ( 1)(log x) 2
log x
(1 p )[1 p ]
1
1
9
In order for
F (x)
to be convex,
F (x)
must be
0
or that the term
p 1
p
2
log x
(log x )
0 for all x (0,1) . However, when x 1 ,
1
1
p 1
p
p
2
0.
log x
(log x )
1
1
1
Thus, F (x) is not convex for x (0,1) for any 1 , p 0 and 0 .
When x 0 ,
p 1
p 1
p
p
1
2
2
(log
x
)
log
x
(log
x
)
1
.
1
1
log x
( 1)(log x) 2
This implies that F (x) cannot be 0 for all x (0,1) . Therefore, F (x) is not concave
for x (0,1) for any 1 , p 0 and 0 .
Note: F (x) is convex for x (0,1) only for p 0 and ( 1) 1 as shown in GomezDeniz et al. (2014).
3. Application to insurance premium loading
Theorem 4. If F ( x; , ) of LL p ( , ) given by (5) is concave, then 1 F (1 x; , ) is
a convex function from (0, 1) to (0, 1) for 0 1 and p 0, 0 .
Proof: It has been shown in Theorem 2 that the cdf F ( x; , ) of LL p ( , ) is concave
for 0 1 and p 0, 0 . Hence, 1 F (1 x; , ) is a convex function from (0, 1) to
(0, 1).
Remark: F ( x; , ) can be used as a distortion function to distort survival function (sf) of
a given random variable as stated in the corollary below.
Corollary 2. If X is the risk with sf G (x ) and let Z be a distorted random variable with sf
H , ( x ) F [G ( x); , ] for 0 1 and p 0, 0 . Let E[Z ]
F [G ( x); , ] dx
0
10
P , ( X ) be the distorted premium and Pn ( X ) [G ( x)]n dx , 0 n 1 be the
0
proportional hazard premium (see Wang, 1995) of a insurance product respectively. Then
P , ( X ) is a premium principle such that
i.
Pn ( X ) P , ( X ) max( X ) , for all n
ii.
P , (aX b) aP , ( X ) b ,
iii. if G1 ( X 1 ) and G1 ( X 2 ) are sf of two non negative risk random variables X 1 and
X 2 with G1 ( X 1 ) G2 ( X 2 ) , that is, X 1 precedes X 2 under first stochastic
dominance then P , ( X 1 ) P , ( X 2 ) ,
X 1 precedes
iv. if
X 2 under second stochastic dominance, that is, if
G ( x ) dx G ( x ) dx
1
1
1
x
2
2
2
for all x 0 then P , ( X 1 ) P , ( X 2 ) .
x
Proof: From theorem 2 above we know that F ( x; , ) is concave for x (0,1) when
0 1 , p 0 and 0 . Also being a cdf, it is an increasing function of x with
F (0; , ) 0 and F (1; , ) 1 .
The results (ii), (iii) and (iv) follow immediately from definition 6 of distortion premium
principle and subsequent properties thereof in Wang (1996).
For the result (i) from Wang (1996) it follows that P , ( X ) max( X ) .
Now we provide a proof of Pn ( X ) P , ( X ) . We first prove that for any p, F x; , , p
x .
Case I: When p is an integer, we get
p
x
F x; , , p
1 p
1 p
( log x )
(1 p ) p! x
( log x)
k 0
(1 p ) p!
11
k
/ k!
p
( log x ) k
( log x )1 p
x 1
x .
k
!
(
1
p
)
p
!
k 1
Case II: For real p, we apply the integral representation (https://en.wikipedia.org/wiki/
Incomplete_gamma_function; see section “Evaluation Formulae”)
(1 p, y ) e
y
y
1 p
yu
e 1 u
p
du
0
p
e y y1 p e yu u du e y e t t p 11dt e y (1 p )
0
0
Hence,
x
F x; , , p
( log x)1 p (1 p )(1 p, log x)
(1 p ) [1 p ]
x ( log x)1 p (1 p )e log x (1 p)
( log x)1 p
x 1
x
(1 p) [1 p ]
(1 p) [1 p ]
Therefore for any p, F x; , , p x . It follows that
F ( x; , ) x x F ( x; , ) for x (0,1)
(G ( x )) F (G ( x ); , ) since 0 G ( x ) 1 for all x .
(G ( x )) F (G ( x ); , ) for all x .
Now, for all x , we have
(G ( x )) n F (G ( x ); , ) when 0 n 1
0
0
(G ( x )) n dx F (G ( x); , ))dx Pn ( X ) P , ( X ) .
Note: This new distorted premium principle is a tradeoff between the proportional hazard
premium
and
maximal
loss.
For
n 1,
Pn ( X ) E ( X )
and
Pn ( X ) max( X ) (see Wang, 1995).
4. Parameter Estimation: Maximum Likelihood Estimation
The likelihood function for a random sample of size n from the LL p , is
n( 2 p )
L
(1 p)(1 p )n
n
log xi
i 1
12
p
n
( log x ) x
i
i 1
1
.
for
n 0,
The log-likelihood function is then given by
n
l log L n(2 p) log n log (1 p ) n log(1 p ) p log( log xi )
i 1
n
n
log( log xi ) ( 1) log xi .
i 1
i 1
The first and second order derivatives of the log-likelihood function are:
n
l n(2 p)
n
log xi
1 p i 1
n
l
n
1
1 p i 1 log xi
n
l
/ (1 p)
n
n log n
log[log[1 / xi ]]
p
(1 p ) 1 p i 1
2l n ( 2 p )
n2
n[(2 p)(1 p ) 2 2 2 ]
2
2
(1 p ) 2
2 (1 p ) 2
n
2l
n 2
1
2
2
2
(1 p )
i 1 ( log xi )
2
/ (1 p )
2l
// (1 p )
n
n
p 2
(1 p)
(1 p )
2l
n
n
2
(1 p ) 1 p
2l
n
n
p (1 p ) 2
2l
n
p
(1 p ) 2
For information matrix we obtain the following result
13
r
p
n
p 1
1
2 p e
p
E
( ) (r , ) .
2
i 1 ( log xi ) (1 p)[1 p ] r 0 r
For p 0 , this reduces to the result given in Gomez et al. (2014).
The information matrix is given by
n[( 2 p)(1 p ) 2 2 2 ]
2 (1 p ) 2
n
(1 p ) 2
n
1 p
p p 1
2 p e
r ( ) p (r , )
(1 p)[1 p ] r 0
r
n 2
(1 p )
2
//
(1 p)
n
(1 p )
n
n
(1 p ) 2
n 2
(1 p ) 2
/ (1 p)
n
(1 p)
2
This matrix can be inverted to get asymptotic variance-covariance matrix for the
maximum likelihood estimates.
5. Numerical Applications
In this section, for the purpose of application with the LL p , regression model, reparameterization of the distribution is required. The LL p , will be re-parameterized
in terms of its mean, in (14) together with two other new parameters and such
that the parameters in (4) are replaced by:
(2 ) 2 2 2 4 (1 )
log
1 p
,p
,
2(1 )(1 )
log(1 )
1
The new parameters are such that 0 < < 1, > 0, 1 1 . The LL p ( , ) will
now be denoted in terms of the new parameters as LL ( , ) . When γ = 1, the reparameterized log-Lindley distribution of Gomez et al. (2014) is obtained.
Suppose that a random sample Y1 , Y2 , , Yn of size n is obtained from the
LL , distribution. For a set of k covariates, the logit link for the LL ( , )
regression model gives the mean for each Yi as
14
i
where
xTi 1, xi1 , , xik
are
exp( x Ti β )
1 exp( x Ti β)
the
, i 1, 2, , n
covariates
with
corresponding
coefficients
β 0 , 1 , , k .
Here, the beta, LL and LL ( , ) regression models are applied on the data set
originally used in Schmit and Roth (1990) and considered in Gomez-Deniz et al. (2014)
about the cost effectiveness of risk management (measured in percentages) in relation to
exposure to certain property and casualty losses, adjusted by several other variables such
as size of assets and industry risk. Description on the data set may be found in GomezDeniz et al. (2014). We take the response variable to be Y = FIRMCOST/100. Six other
variables (covariates) are ASSUME (X1), CAP (X2), SIZELOG (X3), INDCOST (X4),
CENTRAL (X5) and SOPH (X6). Similar to the analysis on the same data set by Jodra and
Jimenez-Gamero (2016), which used a different re-parameterization of the Log-Lindley
distribution, we investigated the cases of y and 1 – y with and without covariates.
Analysis was performed using the R packages. Beta regression is performed using
the package betareg, while LL and LL ( , ) regressions are performed through
optimization using the function mle2 under the package bbmle. The log-likelihood values
and parameter estimates for the models considered, with and without covariates, are
presented in Table 1.
15
Table 1. Log-likelihood values and parameter estimates for beta, Log-Lindley and
generalized Log-Lindley models, with and without covariates
Y
1–Y
Models
Log-likelihood
Estimates
Log-likelihood
Estimates
a = 0.6125
76.1175
a = 3.7979
(a) Without covariates
Beta(a,b)
76.1175
b = 3.7979
Log-Lindley,
76.6042
LL
Generalized
= 0.03427
b = 0.61252
69.0196
= 0.6907
83.2511
Log-Lindley,
= 0.3824
= 4.16×103
= 5.9076
69.0195
= 2.66×103
= 1.2694
= 5.9077
p = 1.7819
p = 1.0×10-6
LL p ( , )
(b) With covariates and logit link for regression
Beta ( , )
Log-Lindley,
LL1 ( , )
87.7230
83.7575
0 = 1.8880
87.7230
0 = -1.8880
1 = -0.01214
1 = 0.01214
2 = 0.1780
2 = -0.1780
3 = -0.5115
3 = 0.5115
4 = 1.2363
4 = -1.2363
5 = -0.01216
5 = 0.01216
6 = -0.003721
6 = 0.003721
= 6.331
= 6.331
0 = 1.6767
1 = -7.57×10-3
16
96.7689
0 = -2.7200
1 = 0.03208
Generalized
91.9525
2 = 0.08903
2 = -0.7378
3 = -0.4281
3 = 0.7024
4 = 0.9687
4 = -3.7854
5 = -0.02318
5 = 6.86×10-3
6 = 2.91×10-4
6 = 0.03593
= 0.03488
= 67.212
0 = 0.2175
96.8252
0 = -2.7149
Log-Lindley,
LL ( , )
1 = 0.004907
1 = 0.03221
2 = -0.1335
2 = -0.7508
3 = -0.2644
3 = 0.7049
4 = 0.4451
4 = -3.7945
5 = -0.07726
5 = 0.002473
6 = 0.005268
6 = 0.03570
= 0.1755
= 65.0001
γ = 2.6735
γ = 1.0021
In terms of the log-likelihood values, it is clear from the results in Table 1 that the
generalized Log-Lindley model fits the data best with or without covariates for the
response variable y. It is also the best model for the regression of 1 – Y, while the beta
model is the best for this case without covariates. For the case of modeling 1 – Y without
covariates, it is seen that the estimates for the parameter p of the LL p , model
approaches 0 and hence, approaches the results for the LL distribution. Similarly for the
regression of 1 – Y with covariates, the estimates of the parameter γ for the LL ,
model approaches 1 and hence, the results approach that of the LL regression model.
17
6. Conclusion
A generalized Log-Lindley distribution has been investigated with important
properties such as the pdf, cdf, survival function and hazard rate function derived.
Application to a real life data set shows the relevance of the newly proposed distribution
in regression modeling. The proposed distribution nests the Log Lindley distribution
proposed by Gomez-Deniz et al. (2014) and offers compact expression for cdf and
moments.
References
1. G´omez−D´eniz, E, Sordo, M.A. and Calder´in−Ojeda, E (2014). The LogLindley distribution as an alternative to the beta regression model with
applications in insurance, Insurance: Mathematics and Economics, 54, 49-57.
2. Jodra, P. and Jimenez-Gamero, M.D. (2016). A note on the Log-Lindley
distribution, Insurance: Mathematics and Economics, 71, 189-194.
3. Jones, M.C. (2009). Kumaraswamy’s distribution: a beta-type distribution with
some tractability advantages, Statistical Methodology, 6, 70-81.
4. Nadaraja, S. (2005) Exponentiated Beta Distribution -, Computers and
Mathematics with Application, 49, 1029-1035
5. Schmit, J.T., Roth, K. (1990). Cost effectiveness of risk managements practice.
The Journal of Risk and Insurance 57 (3), 455–470.
6. Wang, S. (1995). Insurance pricing and increased limits rate making by
proportional hazards transformations, Insurance: Mathematics and Economics, 17,
43-54.
7. Wang, S. (1996). Premium calculation by transforming the layer premium
density, Astin Bulletin, 26 (1), 71-92.
8. Zakerzadeh, H., Dolati, A., 2010. Generalized Lindley distribution. Journal of
Mathematical Extension 3 (2), 13–25.
18
| 10math.ST
|
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
1
VLSI Designs for Joint Channel Estimation and
Data Detection in Large SIMO Wireless Systems
arXiv:1709.07860v1 [cs.IT] 22 Sep 2017
Oscar Castañeda, Tom Goldstein, and Christoph Studer
Abstract—Channel estimation errors have a critical impact on
the reliability of wireless communication systems. While virtually
all existing wireless receivers separate channel estimation from
data detection, it is well known that joint channel estimation
and data detection (JED) significantly outperforms conventional
methods at the cost of high computational complexity. In this
paper, we propose a novel JED algorithm and corresponding
VLSI designs for large single-input multiple-output (SIMO)
wireless systems that use constant-modulus constellations. The
proposed algorithm is referred to as PRojection Onto conveX hull
(PrOX) and relies on biconvex relaxation (BCR), which enables us
to efficiently compute an approximate solution of the maximumlikelihood JED problem. Since BCR solves a biconvex problem
via alternating optimization, we provide a theoretical convergence
analysis for PrOX. We design a scalable, high-throughput VLSI
architecture that uses a linear array of processing elements
to minimize hardware complexity. We develop corresponding
field-programmable gate array (FPGA) and application-specific
integrated circuit (ASIC) designs, and we demonstrate that PrOX
significantly outperforms the only other existing JED design in
terms of throughput, hardware-efficiency, and energy-efficiency.
Index Terms—FPGA and ASIC designs, joint channel estimation and data detection (JED), large single-input multiple-output
(SIMO) wireless systems, biconvex relaxation (BCR).
I. I NTRODUCTION
IRELESS communication with a large number of
antennas at the base-station (BS) will play a major
role in fifth-generation (5G) systems. By equipping the BS
with hundreds or thousands of antenna elements, large wireless
systems enable fine-grained beamforming and, hence, improved
spectral-efficiency within each cell compared to traditional
communication systems that use a small number of BS
antennas [2]–[7]. All these advantages come at significantly
increased signal-processing complexity at the BS [8], which
necessitates the development of high-performance transceiver
algorithms that scale well to large BS array sizes and can be
implemented in very-large scale integration (VLSI) circuits at
low costs and in an energy-efficient manner.
W
A. The Importance of Accurate Channel State Information
Due to the fine-grained nature of beamforming in such
large wireless systems, the acquisition of accurate channel
O. Castañeda and C. Studer are with the School of Electrical and Computer
Engineering, Cornell University, Ithaca, NY; e-mail: [email protected],
[email protected]; web: http://vip.ece.cornell.edu
T. Goldstein is with the Department of Computer Science, University of
Maryland, College Park, MD; e-mail: [email protected]
A short version of this paper summarizing the PrOX FPGA design has been
presented at the IEEE Int. Symp. on Circuits and Systems (ISCAS) 2017 [1].
The MATLAB simulator for PrOX used in this paper is available on GitHub:
https://github.com/VIP-Group/PrOX.
state information (CSI) at the BS is critical for reliable, highthroughput data transmission. In particular, accurate channel
state information is not only required in the uplink (to
coherently detect data transmitted from the users to the BS),
but also required for beamforming or precoding in the downlink
(to precisely focus the transmit energy towards the users
and to mitigate multi-user interference). However, the fading
nature of wireless channels as well as pilot contamination,
i.e., channel training that may be contaminated by pilots or
data transmission of users communicating in adjacent cells [2],
render the acquisition of accurate CSI without a significant
channel-training overhead a difficult task.
Most proposed large wireless systems rely on time-division
duplexing (TDD) and perform pilot-based channel training
during the uplink phase [3], [5]. It is, however, well known
that the quality of CSI can be improved significantly by
joint channel estimation and data detection (JED), which is
capable of approaching the performance of idealistic systems
with perfect CSI [9], [10]. While a few JED algorithms have
been proposed for traditional, small-scale wireless systems [9]–
[16], their computational complexity is typically high and not
much is known about their efficacy for systems with hundreds
or thousands of antennas. Moreover, with the exception of
the recently proposed VLSI designs in [17], no hardware
implementations for JED have been described in the literature.
B. Contributions
In this paper, we propose a novel, computationally efficient
and near-optimal JED algorithm for large SIMO wireless systems, and we develop corresponding very-large scale integrated
(VLSI) designs. Our contributions are summarized as follows:
• We use biconvex relaxation (BCR) [18] to develop PRojection Onto conveX hull (PrOX), a novel algorithm that
achieves near-optimal JED performance at low complexity.
• We theoretically analyze the convergence of PrOX, which
solves a biconvex problem via alternating optimization.
• We introduce an approximation that significantly reduces
the preprocessing complexity of PrOX at virtually no loss
in terms of error-rate performance.
• We provide simulation results that showcase the robustness
of PrOX to a broad range of system parameters.
• We develop a scalable VLSI architecture for PrOX that
uses a linear array of processing elements (PEs) to achieve
high-throughput at low hardware complexity.
• We show reference FPGA and ASIC implementation
results, and compare our designs to that of the recentlyreported JED implementations in [17].
2
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
Our results demonstrate that PrOX provides near-optimal JED at
significantly lower complexity than existing reference designs.
C. Related Relevant Results
theoretical convergence analysis. Section IV details the VLSI
architecture for PrOX. Section V shows error-rate performance
and VLSI implementation results, and compares PrOX to
existing FPGA and ASIC designs. We conclude in Section VI.
While a considerable number of algorithms and VLSI designs
II. S YSTEM M ODEL AND ML-O PTIMAL JED
have been proposed for small- and large-scale multi-antenna
We now introduce the considered SIMO system model and
wireless systems that separate channel estimation and data
develop
the associated ML-JED problem.
detection (see, e.g., [8], [17], [19] and the references therein),
only a few of results have been proposed for JED. Spheredecoding (SD) algorithms have been proposed to perform A. System Model
exact maximum-likelihood (ML) JED in SIMO and MIMO
We study a (potentially large) SIMO wireless uplink system
systems that use a small number of time slots [9]–[12], [20],
in which a single-antenna user transmits data over K + 1
[21]. Unfortunately, the complexity of SD methods quickly
time slots to a BS with B antennas. We consider the standard
becomes prohibitive for larger dimensional problems [22], [23]
block-fading, narrow-band1 channel model with the following
and approximate linear methods, which are widely used for
input-output relation [9]–[11], [20]:
coherent data detection in massive MIMO systems [8], cannot
be used for JED (see Section II-B for the reasons). Very recently,
Y = hsH + N.
(1)
a handful of approximate JED algorithms have been proposed
B×(K+1)
contains the B-dimensional
for large wireless systems [17], [24]–[26]. To the best of our Here, the matrix Y ∈ C
knowledge, reference [17] describes the only VLSI design of a receive vectors for the K + B1 time slots, i.e., Y =
, . . . , yK+1 ] with yk ∈ C and k = 1, 2, . . . , K + 1,
JED algorithm reported in the open literature. While this design [y1 , y2B
(assumed to
achieves near-ML-JED performance, the algorithm relies on h ∈ C is the unknown SIMO channel vector K+1
contains
semidefinite relaxation (SDR), which lifts the dimensionality of remain constant over the K +1 time slots), s ∈ O
the
transmitted
data
symbols
for
all
K
+
1
time
slots,
and
the problem to the square of the number of time slots causing
B×(K+1)
N
∈
C
models
i.i.d.
complex
circularly-symmetric
high hardware complexity. In contrast to all these results, PrOX
avoids lifting while achieving near-ML-JED performance and Gaussian noise with variance N0 per entry. In the remainder
can be implemented efficiently in VLSI, even for scenarios of the paper, we assume constant-modulus constellations, i.e.,
|s| = σ for all s ∈ O and a fixed σ > 0.
that operate with a large number of time slots.
Another line of related work investigates data detection Remark 1. The SIMO case is relevant in many rural dealgorithms that are robust to imperfect CSI [27], [28]. The ployment scenarios in which only a few users are active
idea is to statistically model channel estimation errors and at a time [30]. The assumption of having constant-modulus
to solve suitably adapted data detection problems. The re- constellations limits our results to low-rate modulation schemes,
sulting algorithms, however, (i) do not achieve the error-rate such as BPSK and PSK constellations. While the (multi-user)
performance of JED as they are not taking into account all MIMO case and higher-order QAM modulation schemes would
the received information (i.e., from training and data symbols be of interest in more general deployment scenarios, the
received over multiple time slots) and (ii) do not improve the associated ML-JED problem is significantly more challenging
channel estimate itself—the latter is critical in TDD systems to solve, and requires different, more complex, and complicated
that perform beamforming in the downlink using acquired CSI algorithms; see, e.g., [12], [21] for more details—we are
in the uplink [2]–[5]. In contrast, we propose JED algorithms planning to address this scenario in the future.
that (i) approach optimal error-rate performance as they jointly
process all received information and (ii) generate improved
B. Joint Channel Estimation and Data Detection (JED)
channel estimates that can be used for beamforming.
Let h be a deterministic—but unknown—channel vector
with unknown prior statistics. Then, one can formulate the
D. Notation
following ML-JED problem [11]:
Lowercase and uppercase boldface letters stand for column
JED
ŝ , ĥ = arg min
Y − hsH F .
(2)
vectors and matrices, respectively. For the matrix A, the
K+1
B
s∈O
, h∈C
Hermitian is AH and the kth row and `th column entry is Ak,` .
For the vector a, the kth entry is ak . The Euclidean norm of a, This problem aims at simultaneously finding an estimate for
the Frobenius norm, and the spectral norm of A are denoted the transmitted data vector s and the channel vector h. We
by kak2 , kAkF , and kAk, respectively. The real and imaginary emphasize that there is a phase ambiguity between both output
vectors of ML-JED in (2). Specifically, if ŝJED ejφ ∈ OK+1 for
parts of the vector a are <(a) and =(a), respectively.
φ ∈ [0, 2π), then ĥejφ is also a valid solution to (2). In order to
E. Paper Outline
The rest of the paper is organized as follows. Section II
describes the system model as well as ML-optimal JED.
Section III introduces the PrOX algorithm and provides a
1 An extension to wideband systems that use OFDM [29] is straightforward
as the flat-fading model in (1) remains to be valid per active sub-carrier. An
extension of our methods to channels with multi-user interference or wideband
channels with inter-symbol interference is, however, not straightforward; see
also Remark 1 for more details.
O. CASTAÑEDA ET AL.
3
avoid this phase ambiguity, one can either transmit information
as phase differences among the entries in the vector s, which
is known as differential signaling [24], or fix the phase of at
least one entry in the transmit vector s. As in [17], we set the
first entry of the transmit vector to a fixed constellation point
š ∈ O and exploit this knowledge at the receiver.
where α > 0 is a suitably chosen regularization parameter. To
ensure that the problem in (4) is convex in the vector q, the
parameter α must be larger than the maximum eigenvalue of
the Gram matrix G = YH Y, i.e., α > kYH Yk.
We next relax the finite-alphabet constraint to the convex
hull CO of the set O given by [34]
|O|
|O|
Remark 2. While this approach resembles that of conventional,
X
X
pilot-based data transmission, we emphasize that ML-JED
CO =
αi si | (αi ∈ R+ , ∀i) ∧
αi = 1 ,
is fundamentally different. In traditional, pilot-based data
i=1
i=1
transmission, a small number of known training symbols are
used to generate channel estimates, which are then used during with si , i = 1, . . . , |O|. For example, the convex hull for BPSK
the detection of the data symbols. In contrast, ML-JED uses all constellations CBPSK is the line along the real axis in [−1, +1];
received symbols, i.e., pilots and data symbols, to improve the the convex hull CQPSK for QPSK is the square with the four
quality of CSI. We also note that ML-JED as in (2) jointly solves constellation points as corners. By relaxing O to the convex
for the transmitted data vector s and the channel vector h; this hull CO , we arrive at the following relaxed ML-JED problem:
is in contrast to JED methods that alternate between channel
minimize − kYqk22 + αkq − sk22 .
K+1
estimation and data detection (see, e.g., [13]–[16]) and for
s∈CO
,q∈CK+1
which ML-JED optimality can, in general, not be guaranteed.
For the above problem, it is likely to obtain a solution that is
Since we assumed the entries in s to be constant-modulus, inside the convex hull. We are, however, interested in finding
the ML-JED problem in (2) can be rewritten in the following, solutions that lie at the boundary of the convex hull CO as
more compact form [11]:
the original constellation points in O lie at the boundary. To
this end, we include a norm constraint that promotes large
ŝJED = arg max kYsk2 ,
(3)
values
in s, with the goal of pushing the entries of s towards
K+1
s∈O
the boundary of the convex hull. This leads to the final BCR
and the associated channel estimate is given by ĥ = formulation of the ML-JED problem:
YŝJED /kŝJED k22 . We note that the optimization problem in (3)
ŝBCR =
arg min
−kYqk22 + αkq − sk22 − βksk22 . (5)
resembles the famous MaxCut problem that is known to be NPK+1
K+1
s∈CO ,q∈C
hard [31]. Indeed, general problems of the form (3) are known
to be NP-hard with respect to randomized reductions, even Here, β is a suitably chosen algorithm parameter that satisfies
when only approximate solutions are sought [32]. Nevertheless, β < α, which ensures that the optimization problem is biconvex
for a small number of time slots K + 1, the problem in (3) in the vectors s and q. In words, for a fixed vector q, the
can be solved exactly and at low average complexity using problem in (5) is convex in s, and vice versa.
sphere decoding (SD) methods [11]. For a large number of
time slots, however, SD is known to entail prohibitively high
complexity [33]. Furthermore, linear approximations are useless B. Alternating Optimization
for JED, since the entries of s would grow without bound when
We solve the BCR problem in (5) using alternating minirelaxing the finite-alphabet constraint s ∈ OK+1 to the complex mization, i.e., we keep one of the vectors fixed while solving
numbers s ∈ CK+1 . As a consequence, the design of non-linear for the other. The resulting iterative procedure is given by
and low-complexity algorithms is necessary to enable JED for
q(t) = arg min −kYqk22 + αkq − s(t−1) k22
(6)
practical SIMO systems that use a large number of time slots.
q∈CK+1
(t)
s
III. P ROX: PROJECTION O NTO CONVE X HULL
= arg min αkq(t) − sk22 − βksk22 ,
(7)
K+1
s∈CO
We now develop a novel algorithm to approximately solve the where t = 1, 2, . . . , t
max is the iteration counter.
ML-JED problem in (3) using biconvex relaxation (BCR) [18],
We initialize the above algorithm with s(0) = š(G1,1 )−1 g1c ,
a recent framework to solve large semidefinite programs in
where g1c is the first column of G, and š ∈ O is the fixed
computer vision. The resulting algorithm is referred to as PRosymbol transmitted in the first time slot and known at the
jection Onto conveX hull (PrOX), requires low computational
receiver.
complexity, and achieves near-ML-JED performance.
Since the problem in (5) is biconvex in s and q, both
of the above steps are convex and can be solved optimally.
Furthermore, each of these optimization problems turns out
A. Biconvex Relaxation (BCR) of the ML-JED Problem
to have a closed form solution. Concretely, we obtain the
We start from (3) and include a regularization term that
following two-step PrOX algorithm:
K+1
forces the transmit vector s ∈ O
to be close to a copy
q ∈ CK+1 that is relaxed to the complex numbers as follows:
q(t) = (I − α−1 G)−1 s(t−1)
(8)
minimize
s∈O K+1 ,q∈CK+1
− kYqk22 + αkq − sk22 ,
(4)
s(t) = proxC K+1 (θq(t) ),
O
(9)
4
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
α
> 0, I is the identity matrix, and the proximal
where θ = α−β
operator [35] is defined as
always lie on the boundary of the convex hull CO ; a proof is
given in Appendix B.
proxC K+1 (v) = arg min kv − sk22 .
Theorem 2. Let the conditions of Theorem 1 hold and further
assume α > β > 0. Then, any non-zero stationary point of (5)
lies along the boundary of the set CO .
O
(10)
K+1
s∈CO
This proximal operator is the element-wise orthogonal projection of the vector v = θq(t) onto the convex hull CO around the
constellation set O. For BPSK, the proximal operator in (10)
is given by proxCBPSK (vb ) = max {min {<(vb ), +1} , −1}, i.e.,
the projection onto the real line in [−1, +1]; for QPSK, (10)
corresponds to applying the same operation, independently, to
both <(vb ) and =(vb ). In the case of (9), we have v = θq(t)
with θ > 0 and hence, the proximal operator in (10) projects a
scaled version of q(t) onto the convex hull of the constellation
set CO —this is why we dub our algorithm PRojection Onto
conveX hull (PrOX).
C. Convergence Theory of PrOX
Remark 3. Theorem 1 and Theorem 2 do not guarantee PrOX
to find a global minimizer to the ML-JED problem (3), but
rather stationary (or saddle) points that lie on the boundary of
the convex hull—although we often observe global minimizers
in practice (cf. Section V-A). The development of stronger
results that provide conditions for which PrOX finds the optimal
solution is challenging and left for future work.
D. Simplifying the Preprocessing Stage
For a large number of time slots K + 1, computing the
b = (I − α−1 G)−1 in (8) results in high
matrix inverse G
preprocessing complexity. We now propose an approximate
method that substantially reduces the preprocessing complexity
for PrOX at a negligible loss in terms of error-rate performance.
Let kGk < α for any consistent matrix norm k · k. Then,
we have the following Neumann series expansion [36]:
We now show that the PrOX algorithm is well behaved,
even though it solves a biconvex problem using alternating
optimization. More specifically, we begin by establishing that
the iterates q(t) and s(t) generated by PrOX converge to
stationary points of the objective function in (5), provided that
∞
X
the algorithm parameters α and β are chosen appropriately.
(α−1 G)k .
(11)
(I − α−1 G)−1 =
We then show that these stationary points lie on the boundary
k=0
K+1
of the convex hull CO
, which implies that our biconvex As put forward in [8] for data detection in massive MIMO
relaxation is reasonably tight. In what follows, we denote the systems, one can approximate a matrix inverse by truncating
objective function of PrOX in (5) as
the series expansion to the first two terms, which is
( 1
β
K+1
α
2
2
2
− 2 kYqk2 + 2 kq − sk2 − 2 ksk2 , if s ∈ CO
(I − α−1 G)−1 ≈ I + α−1 G.
(12)
f (q, s) =
∞, otherwise.
Evidently, the expression on the right-hand side does not require
1
Note that the factors of 2 do not affect the solution of PrOX. the computation of a matrix inverse, which significantly reduces
At an optimal solution, the sub-gradients of f with respect the preprocessing complexity. More specifically, the matrix
to both s and q should be zero. Because s(t) is computed by to be inverted in the left-hand side of (12) is of dimension
minimizing f with respect to s(t) , we already know that the (K + 1) × (K + 1). The complexity (in terms of the number
sub-gradient of f with respect to s(t) contains zero. We can of real-valued multiplications) of an explicit matrix inversion
thus measure the optimality of the iterates by examining only using, e.g., the Cholesky factorization scales with (K + 1)3
∇q f, the gradient of f with respect to q. We now show that and exhibits stringent data dependencies when implemented
this gradient vanishes for a large number of iterations t; the in VLSI; see [8] for more details on the computational
complexity of Cholesky-based matrix inversion and a VLSI
proof is given in Appendix A.
design. Hence, for a large number of time slots K, the
Theorem 1. Suppose that the parameters α and β satisfy approximate preprocessing method in (12) yields significant
α > kGk and α > β. Then, the PrOX algorithm in (8) and (9) savings in terms of hardware complexity.
converges in the gradient sense, i.e.,
We have the following result that bounds the error of the
(t) (t)
approximation
in (12); the proof is given in Appendix C.
lim ∇q f (q , s ) = 0.
t→∞
Furthermore, any limit point of the sequence of iterates is a
stationary point.
Theorem 3. Let α > kGk. Then, the error of the approximation in (12) is bounded by
kα−1 Gk2
The PrOX algorithm is founded on the idea of replacing the
k(I − α−1 G)−1 − (I + α−1 G)k ≤
.
1 − kα−1 Gk
discrete constellation O with its convex hull CO . This results
in a biconvex problem that is easily solved without resorting
This result implies that if kGk is smaller than α, then
to expensive discrete programming or lifting-based methods. the approximation error will be small. In other words, the
However, in using a biconvex relaxation, there is a risk of approximation error will depend on the parameter α, which we
finding a minimizer that lies somewhere in the interior of the can tune in practice for optimal performance. In Section V-A,
convex hull CO , far away from any of the constellation points. we will show that PrOX with this approximation results in
We now provide conditions for which this situation does not excellent error-rate performance while significantly reducing
b
happen. The following theorem shows that minimizers of (5) the complexity compared to using the original matrix G.
O. CASTAÑEDA ET AL.
5
E. Hardware-Friendly Variant of PrOX
As a last step, we now modify the PrOX algorithm in (8)
and (9) to make it more hardware friendly. Since the BS is
assumed to know the first entry of s, there is no need to apply
the algorithm to this particular entry. Consequently, we force
(t)
the entry s1 to be š at the end of each iteration.
b exhibits, in general, a large dynamic range
The matrix G
for different system configurations and channel realizations.
To facilitate fixed-point design, we divide all of its elements
by a constant γ > 0, so that the entries of the resulting matrix
are close to one in absolute value. This scaling procedure
requires us to introduce a new vector q̃ = γ −1 q, resulting in
the following modified PrOX algorithm:
b (t−1)
q̃(t) = γ −1 q(t) = γ −1 Gs
(t)
s
(13)
(t)
(t)
= proxC K+1 (θγ q̃ ) = proxC K+1 (%q̃ ),
O
O
(14)
where % = θγ. With these two modifications, we arrive at the
hardware-friendly version of PrOX summarized as follows:
Algorithm 1 (Hardware-Friendly PrOX). Fix the algorithm parameters % > 0 and α > kGk. Precompute the
b using one of the following expressions:
matrix G
b = γ −1 (I − α−1 G)−1
G
b = γ −1 (I + α−1 G),
G
(15)
(16)
architecture consists of a linear array of N = K + 1 processing
elements (PEs), each one dedicated to computing an entry of
the vectors s and q̃. To execute Algorithm 1, each PE has three
b
key components: (i) a G-matrix
memory, (ii) a complex-valued
multiply-accumulate (MAC) unit, and (iii) a projection unit (see
b
the right side of Figure 1). The G-matrix
memory comprises
two memories to store the real and imaginary parts of ĝkr , i.e.,
b We assume that G,
b which is
the kth row of the matrix G.
the result of either (15) (for PrOX) or (16) (for APrOX), was
computed and loaded in the memories during a preprocessing
step. The MAC unit of the kth PE is used to sequentially
compute the kth row of the matrix-vector product (MVP) in
(t)
Step (17) of Algorithm 1, resulting in q̃k (see Section IV-B
for the details). Finally, the projection unit of the kth PE
computes the kth entry of Step (18) of Algorithm 1, i.e., it
(t)
(t)
(t)
uses q̃k to acquire sk . Since s1 = š, the first PE does
not need the previous three key components. Instead, PE 1
only contains two multiplexers and flip-flops that store the
predefined constellation point š in order to implement Step (19)
of Algorithm 1. The implementation of all other PEs is identical
and uses the aforementioned three key components. For these
PEs, the hard-output estimates in Step (20) are extracted directly
from the sign bits at the outputs of the projection units.
B. Matrix-Vector Product (MVP)
Before explaining the operation details of the proposed VLSI
architecture, we focus on the main computation of PrOX: the
MVP operation in Step (17) of Algorithm 1. A straightforward
b (t−1) using a linear array of
way
for computing q̃(t) = Gs
(t)
(t−1)
b
q̃ = Gs
(17)
PEs is depicted in Figure 2(a) for N = K + 1 = 3. This
s(t) = proxC K+1 (%q̃(t) )
(18)
architecture computes q̃(t) sequentially and on a column-byO
PN
c (t−1)
(t)
column basis, i.e., it evaluates q̃(t) =
in a
k=1 ĝk sk
s1 = š
(19)
sequential manner over k = 1, . . . , N clock cycles. Here, ĝkc
b While such an approach is
At the end of iteration tmax , compute hard-output estimates
represents the kth column of G.
of the transmitted signals as follows:
conceptually simple, it requires a centralized vector memory
which suffers from a high fan-out at its output (highlighted
(t )
ŝk = arg min |sk max − s|, k = 1, 2, . . . , K + 1. (20)
with red color in Figure 2(a)). More concretely, the vector
s∈O
memory output must be connected to N = K + 1 MAC units,
eventually becoming the critical path for systems with a large
In what follows, we will use PrOX to refer to Algonumber of time slots K +1. While wire pipelining at the output
rithm 1 with exact preprocessing as in (15) and approximate
of the vector memory could be used to mitigate the fan-out, we
PrOX (APrOX) to refer to Algorithm 1 with approximate
propose an alternative solution that avoids this issue altogether.
preprocessing as in (16). We also note that α > 0 and % > 0
Consider the MVP architecture depicted in Figure 2(b). In
are algorithm parameters that can be tuned via numerical
this architecture, we replace the centralized vector memory
simulations to improve the error-rate performance.
by a set of registers: each register is associated with a PE
and contains an entry of the vector s(t−1) . Furthermore, the
IV. VLSI A RCHITECTURE
registers are chained together in a cyclic manner, forming a
We now propose a VLSI architecture for PrOX, which shift register that enables all entries of s(t−1) to reach all the
exhibits high regularity and enables us to achieve high PEs in a round-robin fashion. Then, each PE only needs to
throughput at low hardware complexity.
access the correct entry of ĝkr to ensure that the accumulator of
their MAC unit contains the correct MVP result after N clock
cycles. By following this approach, each register only drives
A. Architecture Overview
the next register in the cycle and one of the MAC unit inputs,
Figure 1 shows the VLSI architecture for PrOX in a
regardless of N , hereby completely avoiding the fanout issue.
system that uses QPSK2 modulation. At a high level, our
We henceforth call this architecture input-cyclic MVP, which
builds the core component of the PrOX architecture shown in
2 For BPSK modulation, the VLSI architecture shown in Figure 1 can be
simplified as data-path elements used to compute = (s) can be removed.
Figure 1 and detailed in the next subsection.
and initialize s(0) = š(G1,1 )−1 g1c . Then, for every
iteration t = 1, 2, . . . , tmax , compute the following steps:
6
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
Complex-valued MAC unit
Projection unit
PE
X
Memory
PE
+
+
-
+
+
+
+
+
X
Memory
X
PE
X
PE
Fig. 1. VLSI architecture of PrOX. Left: linear array of K + 1 processing elements (PEs), which enables us to achieve high throughput at low complexity.
Right: architecture details of the kth PE, which mainly consists of a complex-valued multiply-accumulate (MAC) unit and a projection unit.
b
We note that the reading pattern of the G-matrix
memory
for each PE corresponds to simply accessing the entries of ĝkr
in order. However, in contrast to the conventional architecture
in Figure 2(a), reading can start from any element and wraps
around after reaching the end of ĝkr . Instead of using address
generation logic required to implement this behavior, we store
b
a cyclically-permuted version of ĝkr in the G-matrix
memory
of each PE. By doing so, each PE simply needs to access
b
its G-matrix
memory in regular order starting from the first
address, as it would be done in the conventional architecture
depicted in Figure 2(a).
C. Operation Details of the PrOX Architecture
To implement the input-cyclic MVP architecture, the entries
b matrix for the kth PE are stored in the following
of the G
b k,k ; the
way: The first address of the PE memory contains G
b
second address contains Gk,k+1 , and so on. For example, in
the case shown in Figure 2(b), the second PE would have its
b
b 2,2 , G
b 2,3 , G
b 2,1 }.
G-matrix
entries organized as follows: {G
b k,k
In the first clock cycle, the kth PE has access to both G
(t−1)
(t−1)
b
and sk
, so it can compute the Gk,k sk
product, and store
the result in its accumulator. At the same time, this PE passes
its current s(t−1) entry to the subsequent PE in the chain, i.e.,
the (k − 1)th PE. Passing the current s(t−1) entry to the next
PE will be performed in all PEs during all the clock cycles of
the MVP computation. An example of this first clock cycle is
b 2,2 s2
shown in Figure 2(b), where the second PE computes G
and passes the s2 value to the first PE, so that the first PE has
access to s2 in the second clock cycle.
(t−1)
In the second clock cycle, the kth PE receives sk+1 from
the (k + 1)th PE, the previous PE in the array. As a result, the
b k,k+1 s(t−1) and accumulate it with the
kth PE can compute G
k+1
previous product. For example, in the second clock cycle of
Figure 2(b), the second PE receives s3 from the third PE and
b 2,3 s3 and add it with G
b 2,2 s2 .
is hence able to compute G
(t−1)
In the third clock cycle, the kth PE receives sk+2 from the
(t−1)
(k + 1)th PE. Note that, while the sk+2 value was initially in
the (k+2)th PE, the (k+2)th PE previously passed this value to
the (k + 1)th PE. Then, the kth PE has the necessary operands
to continue executing its complex-valued MAC operation. To
clarify this aspect, let us examine the third clock cycle in
b 2,1 and s1 to finish
Figure 2(b): the second PE has access to G
its computation. While the second PE received s1 from the
third PE, the third PE received s1 from the first PE in the
previous clock cycle.
This example illustrates how the exchange of values between
PEs enables each entry of s(t−1) to circulate through the linear
array of PEs. In the case of the example in Figure 2(b), a
complete circulation of all values in the shift registers takes
three clock cycles. In general, after K + 1 clock cycles, all PEs
have had access to all the entries of s(t−1) and used them to
compute their respective entry of q̃(t) . During this procedure,
the first PE (which is implemented differently than the other
PEs) executes the same procedure: it passes š to the (K + 1)th
PE during the first clock cycle, and in the subsequent clock
cycles sends the s(t−1) entry received from the second PE to
the (K + 1)th PE. As the MAC units contain three pipeline
stages, two clock cycles are required to flush the pipeline.
Hence, the MVP operation in Step (17) of Algorithm 1 is
computed in only K + 3 clock cycles.
We now describe the operation of the projection unit
shown on the right side of Figure 1, which implements
s(t) = proxC K+1 (%q̃(t) ). For QPSK, the projection unit of
O
the kth PE consists of two identical modules: one module
(t)
takes q̄ = <(q̃k ) as its input, while the other module takes
(t)
q̄ = =(q̃k ) as its input. For BPSK, only the module with an
(t)
input of q̄ = <(q̃k ) is required. The function of each module
is to compute %q̄ and to clip its value in case it exceeds an
absolute value of 1. In other words, each module determines
if %q̄ is smaller than −1, larger than +1, or in-between these
numbers, and outputs −1, +1, or %q̄, respectively. As % > 0,
to determine if %q̄ ≥ +1, one can also check q̄ ≥ +1/% or,
equivalently, if q̄ − 1/% ≥ 0, which can be extracted from the
sign bit of q̄ − 1/%. Analogously, to determine if %q̄ < −1, one
X
+
X
+
+
X
X
+
+
Vector Memory
Vector Memory
+
X
X
+
+
+
+
+
+
Vector Memory
X
Cycle 3
Cycle 2
7
Cycle 1
O. CASTAÑEDA ET AL.
X
X
+
+
+
+
+
+
X
+
Cycle 3
Cycle 2
Cycle 1
(a) Conventional column-by-column matrix-vector product
X
+
X
+
+
X
+
X
+
X
+
+
X
+
X
+
+
+
+
+
+
+
X
+
+
(b) Proposed input-cyclic matrix-vector product for PrOX
Fig. 2. Processing details of the matrix-vector product (MVP) operation: (a) conventional approach that proceeds on a column-by-column basis and leads to
high memory fan-out; (b) proposed input-cyclic method that reduces fan-out.
can also check q̄ + 1/% < 0, which can be extracted from the
sign bit of q̄ + 1/%. As a result, each module takes its input q̄
and computes q̄ − 1/% and q̄ + 1/%. Using the sign bits of these
two results, the module can select if the output will be −1,
+1, or %q̄. The quantity %q̄ is computed in parallel with the
calculation of q̄ + 1/% and q̄ − 1/%, which reduces the critical
path. By restricting the parameter % to be a power of two, the
multiplication can be carried out with inexpensive arithmetic
shifts. The projection unit uses one clock cycle to complete
its operation, but it can start only after the K + 3 clock cycles
used by the complex MAC unit. After the projection unit has
(t)
finished, the new sk will be available at the inputs of the
complex-valued MAC unit of the kth PE, ready to start a new
iteration. Consequently, each PrOX iteration requires a total
number of K + 4 clock cycles.
D. Implementation Details
We now provide the remaining implementation details,
including the fixed-point parameters as well as memory
considerations for our FPGA and ASIC designs.
1) Fixed-Point Parameters: To minimize area and power
consumption, and to maximize the throughput, we deploy fixedpoint arithmetic in our design. Specifically, our architecture
uses 6 bit signed fixed-point values for representing the s(t)
b 12 bit
entries, with 3 fraction bits. For the elements of G,
signed fixed-point values are used, with 11 fraction bits. While
18 bit values are generated at the outputs of the multipliers in
the complex-valued MAC unit, only 15 bits are used in the
subsequent adders and accumulators (of which 11 are fraction
bits). The adder and subtractor that receive the outputs from
the multipliers do not saturate, but rather wrap-around which
reduces area and delay. In contrast, the accumulator adders
saturate. The outputs of the complex-valued MAC unit are also
15 bit signed fixed-point values with 11 fraction bits. In the
projection unit, the −1/% and +1/% quantities are represented
using 12 bit signed fixed-point numbers with 11 fraction bits,
and are added to the outputs of the complex MAC unit using
15 bit saturating adders. The outputs of these adders are used
to determine the output of the projection unit. Furthermore,
for all the considered system configurations, % > 1. As % is
restricted to be a power of two, a multiplication with % is
8
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
implemented by an arithmetic left shift. The number of shifts i.e., the downlink channel is the transpose of the uplink channel.
is encoded with a 4 bit number. Only the 6 most significant Hence, we can use the estimated channel acquired in the
bits of the shifted value are taken into the multiplexer of the uplink for MRC-based beamforming in the downlink. We
projection unit, as the output of this multiplexer will become the observe similar trends as for the uplink shown in Figure 3. In
next iterate s(t+1) . We note that these fixed-point parameters particular, PrOX, APrOX, and TASER all achieve near-MLare sufficient to achieve near-floating-point performance; see JED performance. MRC and MRC that uses the estimated
Figures 3 and 4 in Section V-A for corresponding symbol data vector sMRC to compute a re-trained channel estimate
error-rate performance results.
according to ĥ = YŝMRC /kŝMRC k22 (referred to as “RT”) is
b
2) Memory Considerations: For the FPGA designs, the G- able to approach our algorithms by 2 dB and 4 dB, respectively.
matrix memory is implemented using look-up tables (LUTs) Hence, we conclude that JED also significantly improves the
as distributed RAM, i.e., no block RAMs have been used. For reliability of data transmission in the downlink.
b
3) Performance vs. Complexity Trade-offs: Figure 5 shows
the ASIC designs, the G-matrix
memory is implemented using
latch arrays as described in [37]; these memories are built from the trade-offs between the ASIC throughput of PrOX and
standard cells and simplify automated design synthesis, without AProX, and the minimum signal-to-noise ratio (SNR) required
to achieve 1% SER in the uplink for SIMO systems with
a significant area penalty compared to SRAM macrocells.
K = 16 time slots and a varying number of BS antennas.
As
a reference, we include the performance of MRC with
V. R ESULTS
perfect
CSIR and that of the optimal ML-JED detector. We
We now show error-rate performance results as well as FPGA
observed
that by increasing the number of PrOX and APrOX
and ASIC implementation results. We also compare our design
iterations
tmax , the mean squared error (MSE) of the channel
to the only other existing JED design reported recently in [17].
estimate is reduced monotonically and the desired SER is
achieved at a lower SNR at the expense of reduced throughput.
A. Error-Rate Performance
Clearly, PrOX outperforms APrOX for a small number of
1) Uplink Data Detection: Figure 3 shows uplink symbol iterations; this implies that the preprocessing approximation
error-rate (SER) simulation results for PrOX and APrOX, both proposed in Section III-D entails a small performance loss
running tmax = 5 iterations. The simulation results are obtained when a small number of iterations is used. Note that for BPSK
from 50, 000 Monte-Carlo trials in a B = 16 BS antenna modulation, tmax = 2 PrOX iterations are typically sufficient
SIMO system with K = 16 time slots. We consider both to reach near-ML-JED performance at an ASIC throughput of
an i.i.d. flat Rayleigh block-fading channel model as well as 338 Mb/s per PrOX instance. For QPSK, 3 PrOX iterations are
a line-of-sight (LoS) channel with a spherical wave model sufficient at an ASIC throughput of 451 Mb/s.
and a linear BS antenna array with λ/2-wavelength spacing
Remark 4. For all provided simulation results, we have
as described in [38]. As reference points, we also include
assumed perfect synchronization and a transmission free of
the performance of maximum ratio combining (MRC)-based
impairments or pilot contamination3 . Hence, the provided
data detection with both perfect receive-side CSI (denoted by
simulation results may not be representative for other system
“CSIR”) and with conventional channel estimation, optimal MLconfigurations or more realistic communication scenarios. The
JED detection using the sphere-decoding algorithm put forward
MATLAB simulator for PrOX used in this paper is available on
in [12], and the recently proposed TASER (short for Triangular
GitHub: https:// github.com/ VIP-Group/ PrOX; this enables the
Approximate SEmidefinite Relaxation) algorithm [17] running
interested readers to investigate such aspects in more detail.
for 10 iterations. Note that MRC with perfect CSIR is optimal
for these scenarios. However, the assumption of perfect CSIR
cannot be realized in practice and one must resort to channel B. FPGA and ASIC Implementation Results
estimation (CHEST), which entails a performance loss of
To demonstrate the efficacy of PrOX, we now provide FPGA
more than 4 dB at 1% SER. PrOX and AProX achieve similar implementation results on a Xilinx Virtex-7 XC7VX690T and
error rate performance. Furthermore, both of our algorithms ASIC implementation results in a 40 nm CMOS technology.
significantly outperform MRC with CHEST and approach We implemented for different array sizes N ∈ {5, 9, 17, 33} in
near-ML-JED performance but, in stark contrast to ML-JED Verilog on register transfer level (RTL) and hence, our FPGA
data detection, at very low computational complexity. Our and ASIC designs support near-ML-JED for K ∈ {4, 8, 16, 32}
algorithms also outperform TASER, especially for QPSK time slots, respectively. We also provide a comparison to the
modulation; see Section V-B for a hardware comparison. We only other existing FPGA and ASIC designs for JED proposed
also show the fixed-point performance of the VLSI designs of in the literature [17].
PrOX. The markers of PrOX and APrOX correspond to the
1) FPGA Implementation Results: Our FPGA implemenfixed-point performance of golden models that exactly match tation results were obtained using the Xilinx Vivado Design
the outputs of our VLSI designs, whereas the curves correspond Suite, and are summarized in Table I. As expected, the resource
to MATLAB floating-point performance—evidently, our fixed- utilization scales nearly linearly with the array size N . For the
point VLSI designs exhibit virtually no implementation loss.
3 A straightforward model for pilot contamination would be to add inde2) Downlink Beamforming: Figure 4 shows downlink SER
pendent Gaussian noise to the received signals—such a model would simply
simulation results for PrOX and APrOX with the same system reduce the operating SNR. Without more accurate models, however, it is not
and algorithm parameters. We assume channel reciprocity [5], clear what the SNR reduction would be in a practical cellular system.
O. CASTAÑEDA ET AL.
9
100
10−1
10−2
10−3
−12
−10
−8
−6
−4
−2
Average SNR per receive antenna [dB]
10−1
10−2
10−3
−10
0
100
MRC (CSIR)
MRC
ML-JED
TASER
PrOX
APrOX
Uncoded symbol error rate (SER)
MRC (CSIR)
MRC
ML-JED
TASER
PrOX
APrOX
Uncoded symbol error rate (SER)
Uncoded symbol error rate (SER)
100
(a) BPSK, Rayleigh fading
−8 −6 −4 −2
0
2
4
Average SNR per receive antenna [dB]
10−1
10−2
10−3
−10
6
MRC (CSIR)
MRC
ML-JED
TASER
PrOX
APrOX
(b) QPSK, Rayleigh fading
−8 −6 −4 −2
0
2
4
Average SNR per receive antenna [dB]
6
(c) QPSK, LoS
Fig. 3. Uncoded uplink symbol-error rate (SER) for a SIMO uplink with B = 16 BS antennas and transmission over K = 16 time slots. PrOX and APrOX
achieve near-optimal SER performance (close to that of MRC detection with perfect CSIR) and exhibit a performance similar to ML-JED. At a SER of 0.1%,
TASER [17] entails a 1 dB SNR loss for QPSK, while MRC with conventional CHEST suffers more than 3 dB SNR loss for BPSK and QPSK. For PrOX and
APrOX, the curves show the floating-point performance whereas the markers show the fixed-point performance of our golden models.
100
10−1
10−2
10−3
−12
−10
−8
−6
−4
−2
Average SNR per receive antenna [dB]
10−1
10−2
10−3
−10
0
100
MRC (CSIR)
MRC
MRC (RT)
ML-JED
TASER
PrOX
APrOX
Uncoded symbol error rate (SER)
MRC (CSIR)
MRC
MRC (RT)
ML-JED
TASER
PrOX
APrOX
Uncoded symbol error rate (SER)
Uncoded symbol error rate (SER)
100
(a) BPSK, Rayleigh fading
−8 −6 −4 −2
0
2
4
Average SNR per receive antenna [dB]
10−1
10−2
10−3
−10
6
MRC (CSIR)
MRC
MRC (RT)
ML-JED
TASER
PrOX
APrOX
(b) QPSK, Rayleigh fading
−8 −6 −4 −2
0
2
4
Average SNR per receive antenna [dB]
6
(c) QPSK, LoS
Fig. 4. Uncoded downlink symbol-error rate (SER) for a SIMO uplink with B = 16 and K = 16. Beamforming with the channel estimate from PrOX, APrOX
and TASER [17] achieve near-optimal SER performance (close to that of perfect CSIR) and a performance similar to that of ML-JED. At a SER of 0.1%, MRC
with conventional CHEST suffers a 4 dB SNR loss for QPSK; retraining the channel with the estimated uplink data vector can reduce this loss, as in MRC
(RT). For PrOX and APrOX, the curves show the floating-point performance whereas the markers show the fixed-point performance of our golden models.
500
400
300
PrOX
B = 16
B = 32
B = 64
B = 128
APrOX
B = 16
B = 32
B = 64
B = 128
1
1 11
1
1
1,400
1 1
1,200
200
100
1,000
800
600
PrOX
B = 16
B = 32
B = 64
B = 128
APrOX
B = 16
B = 32
B = 64
B = 128
1
1
1 1
1
1
1,400
1
1,200
ASIC throughput [Mb/s]
600
ASIC throughput [Mb/s]
ASIC throughput [Mb/s]
700
400
200
0
−18 −16 −14 −12 −10 −8 −6 −4
Min. SNR [dB] required to achieve 1% SER
(a) BPSK, Rayleigh fading
−2
0
−14
1,000
800
600
PrOX
B = 16
B = 32
B = 64
B = 128
APrOX
B = 16
B = 32
B = 64
B = 128
1
1 1
1 1
1
1 1
400
200
−12 −10 −8
−6
−4
−2
Min. SNR [dB] required to achieve 1% SER
(b) QPSK, Rayleigh fading
0
0
−14
−12 −10 −8
−6
−4
−2
Min. SNR [dB] required to achieve 1% SER
0
(c) QPSK, LoS
Fig. 5. ASIC throughput vs. performance trade-off of PrOX and APrOX for K = 16 time slots. Vertical solid lines represent MRC with CSIR; vertical dashed
lines represent ML-JED performance. The numbers next to the markers correspond to the number of iterations tmax . The throughput was obtained from post
place-and-route ASIC implementation results. We can see that PrOX requires a smaller number of iterations than APrOX to achieve ML-JED performance.
10
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
TABLE I
FPGA IMPLEMENTATION RESULTS OF P ROX ON A X ILINX V IRTEX -7
XC7VX690T FOR DIFFERENT P ROX ARRAY SIZES
Array size (N = K +1)
Time slots (K = N −1)
Slices
LUTs
FFs
DSP48s
Max. clock freq. [MHz]
Min. latency [cycles]
Max. throughputa [Mb/s]
Power estimateb [W]
a Assuming
b Statistical
N =5
K =4
N =9
K =8
N = 17
K = 16
N = 33
K = 32
327
990
515
16
358
8
358
0.45
658
1 991
989
32
341
12
454
0.47
1 491
4 818
1 953
64
297
20
475
0.79
3 018
9 861
3 857
128
240
36
426
1.14
(a) N = 5
(b) N = 9
QPSK modulation; for BPSK, the throughput is halved.
power estimation at max. clock freq. and 1.0 V supply voltage.
(d) N = 33
(c) N = 17
TABLE II
C OMPARISON OF JED FPGA DESIGNS FOR A QPSK, B = 128, K = 8
LARGE -SIMO SYSTEM ON A X ILINX V IRTEX -7 XC7VX690T
Detection algorithm
Error-rate performance
Iterations tmax
PrOX
Near ML-JED
3
TASER [17]
Near ML-JED
3
Slices
LUTs
FFs
DSP48s
Clock frequency [MHz]
Latency [clock cycles]
Throughput [Mb/s]
Power estimatea [W]
658 (0.61 %)
1 991 (0.46 %)
989 (0.11 %)
32 (0.89 %)
341
12
151
0.47
4 350 (4.02 %)
13 779 (3.18 %)
6 857 (0.79 %)
168 (4.67 %)
225
72
50
1.30
75 841
3.09
3 629
26.0
Throughput/LUTs
Energy/bit [nJ/b]
a Statistical
power estimation at max. clock freq. and 1.0 V supply voltage.
N ∈ {5, 9, 17} arrays, the critical path of PrOX is in the PEs’
projection unit; for the largest N = 33 array, the critical path
is in the distribution of the control signals.
Table II compares PrOX with TASER [17], which have been
implemented on the same FPGA for a SIMO system with
B = 128 BS antennas and communication through K = 8
time slots. PrOX requires significantly fewer resources and
lower power than TASER, while achieving a substantially
higher throughput. This makes our design superior to TASER
in terms of both hardware-efficiency (measured in throughput
per FPGA LUTs) and energy per bit. Concretely, PrOX is 20×
more hardware-efficient and 8× more energy-efficient than
TASER for the considered scenario.
2) ASIC Implementation Results: Our ASIC post placeand-route implementation results summarized in Table III
were obtained using Synopsys DC and IC compilers with
a 40 nm CMOS standard-cell technology. Figure 6 shows the
corresponding ASIC layouts. For the N ∈ {5, 9} arrays, the
critical path of PrOX is in the PEs’ projection unit; for N = 17,
the critical path is in the multipliers of the MAC unit and, for
the largest N = 33 array, the critical path is in the address
generation logic and readout circuitry of the ĝkr memories.
Table IV compares PrOX with TASER [17], which have
Fig. 6. Layouts of the PrOX ASIC designs for array sizes N ∈ {5, 9, 17, 33}.
The PEs of each design were colored differently. One can observe a nearly
linear increase in silicon area depending on the array size N = K + 1.
TABLE III
ASIC IMPLEMENTATION RESULTS OF P ROX IN 40 NM CMOS FOR
DIFFERENT ARRAY SIZES
Array size (N = K +1)
Time slots (K = N −1)
N =5
K =4
N =9
K =8
N = 17
K = 16
N = 33
K = 32
Core area [µm2 ]
Core density [%]
Cell area [GE]
Max. clock freq. [MHz]
Min. latency [cycles]
Max. throughput [Mb/s]
Power estimatea [mW]
26 419
77.24
28 922
976
8
976
10
53 361
77.39
58 522
942
12
1 256
18
132 903
72.96
137 428
846
20
1 354
38
303 722
76.38
328 753
695
36
1 235
58
a Post
place-and-route power estimation at max. clock freq. and 1.1 V.
TABLE IV
C OMPARISON OF JED ASIC S FOR A QPSK, B = 128, K = 8
LARGE -SIMO SYSTEM IN 40 NM CMOS
Detection algorithm
Error-rate performance
Iterations tmax
PrOX
Near ML-JED
3
TASER [17]
Near ML-JED
3
Core area [µm2 ]
Core density [%]
Cell area [GE]
Max. clock freq. [MHz]
Latency [clock cycles]
Throughput [Mb/s]
Power estimatea [mW]
53 361
77.39
58 522
942
12
418
18
482 677
68.89
471 238
560
24
125
87
7 153
279
43
697
Throughput/cell
area [b/(s×GE)]
Energy/bit [pJ/b]
a Post
place-and-route power estimation at max. clock freq. and 1.1 V.
O. CASTAÑEDA ET AL.
11
been implemented on the same 40 nm CMOS technology for a constellation sets [12], [21], they are either computationally
SIMO system with B = 128 and K = 8. Similar to the complex or exhibit a considerable loss with respect to the
FPGA case, the PrOX ASIC designs require significantly optimal ML-JED performance. The development of efficient
fewer resources and lower power than the TASER ASIC JED algorithms that can be implemented as high-throughput
designs, while achieving a substantially higher throughput. VLSI designs for large (multi-user) MIMO systems with
Concretely, PrOX is 25× more hardware-efficient (in terms of arbitrary constellations is part of ongoing work.
the throughput per cell area) and 16× more energy-efficient
than TASER for the considered scenario.
A PPENDIX A
We also note that PrOX exhibits a similar (for BPSK) or
P ROOF OF T HEOREM 1
better (for QPSK) SER performance than TASER, while using
We need a closed form expression for the gradient so that
fewer algorithm iterations; see Figures 3 and 4 for the associated we can say something about its convergence. We have
SER results. However, while PrOX is only suitable for JED in
∂q f (q(t) , s(t) ) = −YT Yq(t) + α(q(t) − s(t) ).
large SIMO systems, TASER is also able to perform near-ML
data detection in coherent massive multi-user MIMO systems From the optimality condition for the minimization step (6),
that use conventional pilot-based channel training [17]. Hence, we have
the error-rate performance and hardware complexity advantages
0 = −YT Yq(t) + α(q(t) − s(t−1) )
of PrOX over TASER come at the cost of reduced flexibility.
= −YT Yq(t) + α(q(t) − s(t) ) + α(s(t) − s(t−1) )
= ∂q f (q(t) , s(t) ) + α(s(t) − s(t−1) ),
Remark 5. While a plethora of VLSI designs exist for data detection in coherent small-scale and massive (multi-user) MIMO
and thus
systems that separate channel estimation from data detection
(see [8], [19], [22], [23], [39]–[46] and the references therein),
∂q f (q(t) , s(t) ) = α(s(t−1) − s(t) ).
(21)
TASER is the only other SIMO-JED VLSI design that has been
Now that we have a simple representation for the gradient,
described in the literature [17]. Furthermore, low-complexity
we can bound its norm. Note that f is strongly convex in q
linear methods that are commonly used for conventional (multiwith parameter α − kYT Yk, and so
user) MIMO data detection cannot be used in the JED scenario
(as briefly discussed in Section II-B). Hence, we limited our
α − kYT Yk
kq − qs k22
(22)
f
(q,
s)
−
f
(q
,
s)
≥
s
comparisons in Table II and Table IV to PrOX and TASER.
2
VI. C ONCLUSIONS
where qs = arg minq f (q, s). Likewise, f is strongly convex
in s with parameter α − β, and so
α−β
f (q, s) − f (q, sq ) ≥
ks − sq k22
(23)
We have proposed a novel joint channel estimation and data
2
detection (JED) algorithm, referred to as PRojection Onto
where sq = arg mins f (q, s).
conveX hull (PrOX). Our algorithm builds upon biconvex
If we choose q = q(t−1) and s = s(t−1) in (22), then
relaxation (BCR) [18], which enables near-maximum-likelihood
qs = q(t) and we have
(ML) JED performance at low computational complexity for
large SIMO systems with constant-modulus constellations. We
f (q(t−1) , s(t−1) ) − f (q(t) , s(t−1) )
have provided theoretical convergence guarantees for PrOX
α − kYT Yk (t−1)
and introduced an approximation that significantly reduces the
≥
kq
− q(t) k22 . (24)
2
preprocessing complexity. To demonstrate the effectiveness
of our algorithm, we have proposed a VLSI architecture Similarly, from (23) we get
that consists of a linear array of processing elements (PEs).
α − β (t−1)
f (q(t) , s(t−1) ) − f (q(t) , s(t) ) ≥
ks
− s(t) k22 .
Our architecture enables high-throughput near-ML-JED at low
2
silicon area and power consumption. We have provided FPGA Adding these two inequalities yields
and ASIC implementation results, which demonstrate that
(t−1) (t−1)
,s
) − f (q(t) , s(t) )
PrOX significantly outperforms the only other existing JED f (q
design [17] in terms of hardware- and energy-efficiency as well
α − β (t−1)
α − kYT Yk (t−1)
kq
− q(t) k22 +
ks
− s(t) k22 .
≥
as error-rate performance. More generally, PrOX constitutes a
2
2
first step towards hardware accelerators that are able to find Summing from t = 1 to t = T, and observing that the
approximate solutions to problems that resemble the famous summation on the left telescopes, we get
MaxCut problem in a hardware-efficient manner.
f (q(0) , s(0) ) − f (q(T ) , s(T ) )
(25)
There are many avenues for future work. An extension of
T
PrOX to compute soft-outputs in the form of log-likelihood raT
X α − kY Yk
≥
kq(t−1) − q(t) k22
(26)
tios (LLRs) is part of ongoing work. Furthermore, a theoretical
2
t=1
error-rate performance analysis of PrOX is a challenging open
α − β (t−1)
research problem. Finally, while some algorithms exist for JED
+
ks
− s(t) k22 .
in the MIMO case and for general (non-constant modulus)
2
12
TO APPEAR IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I: REGULAR PAPERS
Because the iterates remain bounded and f is continuous, the
left-hand side of (25) is bounded from above. We can conclude
that ks(t−1) − s(t) k → 0, and from (21), we see that the
residuals converge as well.
Finally, because the sub-gradients of f depend continuously
on s and q, any limit point of the sequence of iterates must
have zero sub-gradients, and is thus a stationary point.
A PPENDIX B
P ROOF OF T HEOREM 2
Let (q∗ , s∗ ) denote a non-zero stationary point of f that
lies in the interior of CO . Such a point satisfies the first-order
conditions
∗
T
∗
∗
−Y Yq + α(q − s ) = 0
∗
∗
(28)
α−β ∗
α s .
Plugging this result
α(s − q ) − βs = 0.
From (28), we see that q∗ =
into (27) yields
−YT Y
(27)
∗
α−β ∗
s + (α − β)s∗ − αs∗ = 0,
α
which re-arranges to
YT Ys∗ =
−αβ ∗
s .
α−β
Since s∗ is non-zero, this shows that s∗ is an eigenvector of
YT Y with negative eigenvalue. This is impossible because the
Gram matrix YT Y is positive semi-definite. It follows that
(q∗ , s∗ ) cannot satisfy the first-order conditions (27) and (28),
and thus, cannot lie in the interior of CO .
A PPENDIX C
P ROOF OF T HEOREM 3
Since kGk < α, we have the Neumann series expansion [36]
in (11). Hence, we can bound the approximation error from
above as follows:
k(I − α−1 G)−1 − (I + α−1 G)k =
(a)
∞
X
(c)
k=2
∞
X
≤
=
k=0
(d)
=
(b)
(α−1 G)k ≤
α−1 G
k
∞
X
∞
X
(α−1 G)k
k=2
α−1 G
k
k=2
− 1 − kα−1 Gk
−1
1
Gk2
(e) kα
− 1 − kα−1 Gk =
.
−1
1 − kα Gk
1 − kα−1 Gk
Here, step (a) follows from the triangle inequality, step (b)
from the fact that the spectral norm is a consistent matrix norm,
step (c) is a result of basic arithmetic manipulations, step (d)
follows from geometric series expansions, and step (e) is the
same expression in simplified form. We note that the proof
continues to hold for any consistent matrix norm.
ACKNOWLEDGMENTS
The authors would like to thank S. Jacobsson and G. Durisi
for discussions on line-of-sight channels. We also thank C. Jeon
and G. Mirza for discussions on the input-cyclic MVP engine.
The work of O. Castañeda and C. Studer was supported by
Xilinx, Inc. and by the US National Science Foundation (NSF)
under grants ECCS-1408006, CCF-1535897, CAREER CCF1652065, and CNS-1717559. The work of T. Goldstein was
supported by the US NSF under grant CCF-1535902 and by the
US Office of Naval Research under grant N00014-17-1-2078.
R EFERENCES
[1] O. Castañeda, T. Goldstein, and C. Studer, “FPGA design of lowcomplexity joint channel estimation and data detection for large SIMO
wireless systems,” in Proc. IEEE Int. Symp. Circuits Syst. (ISCAS), May
2017, pp. 128–131.
[2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers
of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 11,
pp. 3590–3600, Nov. 2010.
[3] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors,
and F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60,
Jan. 2013.
[4] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
[5] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO
for next generation wireless systems,” IEEE Commun. Mag., vol. 52,
no. 2, pp. 186–195, Feb. 2014.
[6] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong,
and J. C. Zhang, “What will 5G be?” IEEE J. Sel. Areas Commun.,
vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[7] L. Lu, G. Li, A. Swindlehurst, A. Ashikhmin, and R. Zhang, “An overview
of massive MIMO: Benefits and challenges,” IEEE J. Sel. Topics Signal
Process., vol. 8, no. 5, pp. 742–758, Oct. 2014.
[8] M. Wu, B. Yin, G. Wang, C. Dick, J. R. Cavallaro, and C. Studer,
“Large-scale MIMO detection for 3GPP LTE: algorithms and FPGA
implementations,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp.
916–929, Oct. 2014.
[9] P. Stoica and G. Ganesan, “Space–time block codes: Trained, blind, and
semi-blind detection,” Elsevier Dig. Signal Process., vol. 13, no. 1, pp.
93–105, Jan. 2003.
[10] H. Vikalo, B. Hassibi, and P. Stoica, “Efficient joint maximum-likelihood
channel estimation and signal detection,” IEEE Trans. Wireless Commun.,
vol. 5, no. 7, pp. 1838–1845, Jul. 2006.
[11] H. A. J. Alshamary, M. F. Anjum, T. Al-Naffouri, A. Zaib, and W. Xu,
“Optimal non-coherent data detection for massive SIMO wireless systems
with general constellations: A polynomial complexity solution,” arXiv
preprint: 1507.02319, Jul. 2015.
[12] W. Xu, M. Stojnic, and B. Hassibi, “On exact maximum-likelihood
detection for non-coherent MIMO wireless systems: A branch-estimatebound optimization framework,” in Proc. IEEE Int. Symp. Inf. Theory
(ISIT), Jul. 2008, pp. 2017–2021.
[13] T.-H. Pham, Y.-C. Liang, and A. Nallanathan, “A joint channel estimation
and data detection receiver for multiuser MIMO IFDMA systems,” IEEE
Trans. Commun., vol. 57, no. 6, pp. 1857–1865, Jun. 2010.
[14] R. Prasad, C. R. Murthy, and B. D. Rao, “Joint channel estimation and
data detection in MIMO-OFDM systems: A sparse Bayesian learning
approach,” IEEE Trans. Sig. Proc., vol. 63, no. 20, pp. 5369–5382, Oct.
2015.
[15] E. Kofidis, C. Chatzichristos, and A. L. F. de Almeida, “Joint channel
estimation / data detection in MIMO-FBMC/OQAM systems - a tensorbased approach,” arXiv preprint: 1609.09661, Sep. 2016.
[16] C.-K. Wen, C.-J. Wang, S. Jin, K.-K. Wong, and P. Ting, “Bayes-optimal
joint channel-and-data estimation for massive MIMO with low-precision
ADCs,” IEEE Trans. Sig. Proc., vol. 64, no. 10, pp. 2541–2556, Dec.
2016.
[17] O. Castañeda, T. Goldstein, and C. Studer, “Data detection in large
multi-antenna wireless systems via approximate semidefinite relaxation,”
IEEE Trans. Circuits Syst. I, no. 99, pp. 2334–2346, Nov. 2016.
O. CASTAÑEDA ET AL.
[18] S. Shah, A. Kumar, C. Castillo, D. Jacobs, C. Studer, and T. Goldstein,
“Biconvex relaxation for semidefinite programming in computer vision,”
European Conf. Comput. Vision (ECCV), pp. 717–735, Sep. 2016.
[19] S. Yang and L. Hanzo, “Fifty years of MIMO detection: The road to
large-scale MIMOs,” IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp.
1941–1988, Sep. 2015.
[20] P. Stoica, H. Vikalo, and B. Hassibi, “Joint maximum-likelihood channel
estimation and signal detection for SIMO channels,” in Proc. IEEE Int.
Conf. Acoust., Speech, Signal Process. (ICASSP), vol. 4, May 2003, pp.
13–16.
[21] H. A. J. Alshamary and W. Xu, “Efficient optimal joint channel estimation
and data detection for massive MIMO systems,” in Proc. IEEE Int. Symp.
Inf. Theory (ISIT), Jul. 2016, pp. 875–879.
[22] A. Burg, M. Borgmann, M. Wenk, M. Zellweger, W. Fichtner, and
H. Bölcskei, “VLSI implementation of MIMO detection using the sphere
decoding algorithm,” IEEE J. Solid-State Circuits, vol. 40, no. 7, pp.
1566–1577, Jul. 2005.
[23] C. Studer, M. Wenk, and A. Burg, “VLSI implementation of hard-and
soft-output sphere decoding for wide-band MIMO systems,” in VLSI-SoC:
Forward-Looking Trends in IC and Systems Design. Springer, 2010, pp.
128–154.
[24] A. Schenk and R. F. H. Fischer, “Noncoherent detection in massive
MIMO systems,” in Int. ITG Workshop on Smart Antennas (WSA), Apr.
2013, pp. 1–8.
[25] G. Yammine and R. F. Fischer, “Soft-decision decoding in noncoherent
massive MIMO systems,” in Int. ITG Workshop on Smart Antennas
(WSA), Mar. 2016, pp. 1–7.
[26] J. Feng, H. Gao, T. Wang, T. Lv, and W. Guo, “A noncoherent differential
transmission scheme for multiuser massive MIMO systems,” in Proc.
IEEE Wireless Commun. Networking Conf. (WCNC), Mar. 2017, pp. 1–6.
[27] B. S. Thian and A. Goldsmith, “Decoding for MIMO systems with
imperfect channel state information,” in Proc. IEEE Global Telecommun.
Conf. (GLOBECOM), Dec. 2010, pp. 1–6.
[28] R. Ghods, C. Jeon, G. Mirza, A. Maleki, and C. Studer, “Optimally-tuned
nonparametric linear equalization for massive MU-MIMO systems,” in
Proc. IEEE Int. Symp. Inf. Theory (ISIT); arXiv preprint: 1705.02985,
Jun. 2017.
[29] R. Prasad, OFDM for Wireless Communications Systems. Norwood,
MA, USA: Artech House, Inc., 2004.
[30] “3GPP TS 36.211 Evolved Universal Terrestrial Radio Access (E-UTRA)
physical channels and modulation (release 8),” 3rd Generation Partnership
Project.
[31] D. Steinberg, “Computation of matrix norms with applications to robust
optimization,” Master’s thesis, Technion, Israel, 2005.
13
[32] M. Ajtai, “The shortest vector problem in L2 is NP-hard for randomized
reductions,” in Proc. ACM Symp. Theory Comput. ACM, 1998, pp.
10–19.
[33] D. Seethaler, J. Jaldén, C. Studer, and H. Bölcskei, “On the complexity
distribution of sphere decoding,” IEEE Trans. Inf. Theory, vol. 57, no. 9,
pp. 5754–5768, Sep. 2011.
[34] M. Wu, C. Dick, J. R. Cavallaro, and C. Studer, “High-throughput
data detection for massive MU-MIMO-OFDM using coordinate descent,”
IEEE Trans. Circuits Syst. I, no. 99, pp. 2357–2367, Nov. 2016.
[35] N. Parikh and S. Boyd, “Proximal algorithms,” Foundations and Trends
Optimization, vol. 1, no. 3, pp. 123–231, Jan. 2013.
[36] G. Stewart, Matrix Algorithms: Basic decompositions, 1998.
[37] P. Meinerzhagen, C. Roth, and A. Burg, “Towards generic low-power
area-efficient standard cell based memory architectures,” in Proc. IEEE
Int. Midwest Symp. Circuits Syst. (MWSCAS), Aug. 2010, pp. 129–132.
[38] D. Tse and P. Viswanath, Fundamentals of wireless communication.
Cambridge University Press, 2005.
[39] K. Wong, C. Tsui, R. Cheng, and W. Mow, “A VLSI architecture of a
K-best lattice decoding algorithm for MIMO channels,” in Proc. IEEE
Int. Symp. Circuits Syst. (ISCAS), vol. 3, May 2002, pp. 273–276.
[40] C. Studer, A. Burg, and H. Bölcskei, “Soft-output sphere decoding:
Algorithms and VLSI implementation,” IEEE J. Sel. Areas Commun.,
vol. 26, no. 2, pp. 290–300, Feb. 2008.
[41] C. H. Liao, T. P. Wang, and T. D. Chiueh, “A 74.8 mW soft-output
detector IC for 8×8 spatial-multiplexing MIMO communications,” IEEE
J. Solid-State Circuits, vol. 45, no. 2, pp. 411–421, Feb. 2010.
[42] C. H. Yang, T. H. Yu, and D. Marković, “A 5.8 mW 3GPP-LTE compliant
8×8 MIMO sphere decoder chip with soft-outputs,” in Symp. VLSI
Circuits, Jun. 2010, pp. 209–210.
[43] E. M. Witte, F. Borlenghi, G. Ascheid, R. Leupers, and H. Meyr, “A
scalable VLSI architecture for soft-input soft-output single tree-search
sphere decoding,” IEEE Trans. Circuits and Syst. II, vol. 57, no. 9, pp.
706–710, Sep. 2010.
[44] C.-F. Liao, J.-Y. Wang, and Y.-H. Huang, “A 3.1 Gb/s 8×8 sorting
reduced K-Best detector with lattice reduction and QR decomposition,”
IEEE Trans. VLSI Syst., vol. 22, no. 12, pp. 2675–2688, Feb. 2014.
[45] C. Senning, L. Bruderer, J. Hunziker, and A. Burg, “A lattice reductionaided MIMO channel equalizer in 90 nm CMOS achieving 720 Mb/s,”
IEEE Trans. Circuits Syst. I, vol. 61, no. 6, pp. 1860–1871, Jun. 2014.
[46] B. Yin, M. Wu, J. Cavallaro, and C. Studer, “VLSI Design of Large-Scale
Soft-Output MIMO Detection Using Conjugate Gradients,” in Proc. IEEE
Int. Conf. Circuits Syst. (ISCAS), May 2015, pp. 1498–1501.
| 7cs.IT
|
Analysis of Feature Detector and Descriptor Combinations with a Localization
Experiment for Various Performance Metrics
Ertugrul BAYRAKTAR*, Pınar BOYRAZ
Graduate School of Science Engineering and Technology Mechatronics Engineering Department,
Istanbul Technical University, [email protected]
Mechanical Engineering Department, Istanbul Technical University, Inonu Cd. No: 65, 34437
Beyoglu Istanbul, Turkey, [email protected]
Abstract
The purpose of this study is to give a detailed performance comparison about the feature
detector and descriptor methods, particularly when their various combinations are used for
image matching. As the case study, the localization experiments of a mobile robot in an
indoor environment are given. In these experiments, 3090 query images and 127 dataset
images are used. This study includes five methods for feature detectors such as features from
accelerated segment test (FAST), oriented FAST and rotated binary robust independent
elementary features (BRIEF) (ORB), speeded-up robust features (SURF), scale invariant
feature transform (SIFT), binary robust invariant scalable keypoints (BRISK), and five other
methods for feature descriptors which are BRIEF, BRISK, SIFT, SURF, and ORB. These
methods are used in 23 different combinations and it was possible to obtain meaningful and
consistent comparison results using some performance criteria defined in this study. All of
these methods are used independently and separately from each other as being feature detector
or descriptor. The performance analysis shows the discriminative power of various
combinations of detector and descriptor methods. The analysis is completed using five
parameters such as (i) accuracy, (ii) time, (iii) angle difference between keypoints, (iv)
number of correct matches, and (v) distance between correctly matched keypoints. In a range
of 60°, covering five rotational pose points for our system, “FAST-SURF” combination gave
the best results with the lowest distance and angle difference values and highest number of
matched keypoints. The combination “SIFT-SURF” is obtained as the most accurate
combination with 98.41% of correct classification rate. The fastest algorithm is achieved with
“ORB-BRIEF” combination with a total running time 21303.30 seconds in order to match 560
images captured during the motion with 127 dataset images.
1. Introduction
Achieving beneficial information from visual sensors for indoor robots is crucial. In order to
get this type of information capable of representing the real world with minimum loss of
details, robots use various computer vision algorithms under the names of object detection,
segmentation, and recognition. These algorithms work by matching and obtaining structural
or inferred information about objects. Then, relating these separate low-level information sets,
the algorithm constructs a framework in order to obtain semantic information. In this context,
semantic information is necessary to get robots, machines, or digital systems to make sense
from numerical data such as understanding what is happening in a scene or what the context
of a speech/conversation could be. Consequently, semantic information requires more
*: Corresponding Author
computational effort and a deeper knowledge representation in comparison to low-level
computing such as basic object detection and recognition algorithms.
Image matching has a wide range of applications in real world such as object and face
recognition. In addition, image matching is the main operation for obtaining semantic
information. On the other hand, image matching is still a challenging problem for real-time
applications because of the amount of the data to be processed. One can briefly summarize the
process of image matching as follows; 1) Constructing an appropriate feature database for the
desired application, 2) Streaming live/recorded video or loading images to the system, 3)
Computing the features of frame grabbed/captured from streamed live/recorded video or
loaded images with the database, 4) Comparing the features, 5) Decision making about the
quality of the matches such as accurately matched features.
Furthermore, our work includes design of an algorithm which may be applied to known
indoor environments to obtain semantic information. In order to develop and test this
algorithm,
we
created
an
image
database
(https://web.itu.edu.tr/bayraktare/Visual_Indoor_Dataset.rar) consisting 3090 images of an
office environment and we specified various office objects within this database to match these
objects with query images. We have selected 127 images which contain only one object in
each image without any occlusions as database images to be compared by query images. In
addition, these 127 images are taken as the ground-truth for our performance analysis. In
order to accurately find the location of the robot, we consider the 3 height levels in a range of
300. This means the localization algorithm may give six possible results for the same point. By
using the output of the localization process, it is possible to determine the location of the robot
within a cube which contains the 6 possible coordinates as boundary points. For this reason,
scale and rotation performance of these methods are crucial for our localization algorithm.
This paper has been organized as follows; previous studies about feature detector-descriptor
algorithms and performance evaluation of these methods are given in section 2, in 3rd section
the datasets and the methods used in this study are explained, performance results are shown
in 4th section, and finally, in section 5 we discuss the results.
2. Related Work
In literature, there is a wide range of studies based on feature detector-descriptor
combinations of which some compare the feature detectors and feature descriptor methods;
some of them argue the best detector-descriptor combinations, and a few of them focus on
their performance in the recognition of objects. A feature in an image can be defined in a
specific 2-dimensional structure that is composed of a detector and a descriptor. In this
structure, the detector finds the repeatable interest points, and the descriptor is a distinctive
specification that is obtained by computing each detected feature which can be matched
between different images.
It is commonly accepted that SIFT [1], SURF [2], BRISK [3], ORB [6] methods used in this
work consist of similar content and it may be given in 4 steps as follows; 1) Scale-space
representation, 2) Key-point localization, 3) Orientation assignment, and 4) Key-point
descriptor as given in [5]. In other words, one can summarize the first three of these steps as
detector, and the last one as descriptor. Besides, some of these methods may include both the
detector and descriptor, while some are distinctly known as detector or descriptor. As an
example, FAST [7, 8] is a detector method and BRIEF [4] is a descriptor method. In the step
of scale-space representation, SIFT applies a series of Difference-of-Gaussian (DoG) filters
for multiple scales, and in this way we may get DoG filtered and down-sampled versions of
the original image. SIFT descriptor is composed of a 4x4 array of gradient orientation
histogram weighed by the gradient magnitude. On the other hand, scale-space representation
in SURF is based on sums of 2D Haar wavelets using integral images that approximate
Gaussian derivatives given by [2]. SURF detector approximates the determinant of Hessian
matrix which will give a local maximum. SURF descriptor is a 64-dimensional vector which
is obtained by summing the Haar Wavelet coefficients over 4x4 pixels around the key-point.
As described in [7] and [8] FAST, a detection method that is actually used to detect corners,
uses a circle consisting of 16 pixels around candidate corner pixels to classify whether that
point is a corner or not by comparing these 16 pixels’ brightness with the intensity of
candidate pixel including a threshold. BRIEF is a binary descriptor based on pairwise
intensity comparison. The detector part of the BRISK is given at [3] as computing FAST
score, which is computed at each octave and intra-octave separately, across scale space and
obtained continuous maximum across scales by calculating the sub-pixel maximum across
patch at pixel level non-maximal suppression. BRISK descriptor contains concatenated
brightness results tests with a binary string and it is rotation and scale invariant apart from
BRIEF. Descriptor part of ORB is a method that is similar to BRIEF with rotation and scale
invariances. Detector part of ORB applies FAST detector in a Gaussian pyramid.
The system proposed in [5] demonstrates the performance results of 4 descriptors [SIFT,
SURF, BRISK, and FREAK] and looks for the best matching results in terms of detection
accuracy and speed in the context of detecting the abandoned objects in real-time. This
method is very sensitive to disturbances and the robustness is not completely ensured in the
experimental setup. In addition to this, the camera system movement of the work is very
limited such as linear back and forth motion. Another study, which investigates the detector
and descriptor methods for photogrammetric applications [9], compares 5 interest point
detectors with respect to correct detected corners, their localizations, the density of detected
points/regions, but it is clear that the number of methods is very limited and performance
analysis does not cover the time cost for these methods. Similar to a part of our work, [10]
analyzes the different detector-descriptor combinations with 7 detectors and 2 descriptors with
the aim of finding the best combination. They used a dataset that includes 60 scenes from 119
positions with 19 different illumination conditions. They conclude as a result of their
experiments that DOG or MSER detector with a SIFT or DAISY descriptor as the best
combination by using a performance measure computed from area under receiver operating
characteristic (ROC) curve which depends on a proportion of the distance values between the
best matched features.
As a matter of course, previous studies have used different measuring scales and criteria in
order to evaluate the performance outputs of different feature detector-descriptor
combinations. In [11], although it does not consider time performance, the performance
evaluation is realized counting the number of correctly detected interest points and their
locations with comparing the density and relative orientations of these points with stereo pairs
for 5 interest point detectors (Förstner, Heitger, Susan, Harris, and Hessian) and two region
detectors/descriptors (Harris-Affine and SIFT). Another study, [12], considers the occlusions
and realized using a moving camera gives the performance comparison results for 4
descriptors (SURF, SIFT, BRISK, and FREAK). This paper uses accuracy and speed as
performance criteria, but it is very sensitive to disturbances, and the robustness is provided by
limiting the movement of the robot just by a linear back-and-forth motion at the experimental
setup. Although it tells that the real-time detection, normalized cross correlation, which scans
all the frame which makes this method slow rotation-variant, is applied to the binary image
that means loss of information in the image used as image comparison method. [13] compares
only binary descriptors (SIFT, FAST, BRIEF, BRISK, ORB, and FREAK) and their
combinations without giving any details about these combinations’ total performance and
compliance of different detectors and descriptors. The performance evaluation results are
obtained by matching the given images and pixel based distance values of the corresponding
points. In this study, SIFT is assumed to be ground-truth, but ground-truth would be obtained
by direct matching of images more precisely. Another local descriptor comparison is given in
[14] for only 4 methods. These 4 methods are based on Harris-Affine detector and they are
compared with regard to the complexity of the compared methods’ individual parameters and
usage areas with detection rate with respect to false positive rate as evaluation criterion based
on the calculation of the area under Receiver Operating Characteristics (ROC) curves. In
addition to the comparison of local detectors and descriptors, [15] investigates the
performance of in terms of two properties; robustness and distinctiveness using a unified
framework. The framework is composed of two steps; the first step is detector evaluation
criteria, which takes into account the accuracy of localization under different conditions, and
the repeatability score for 6 detectors (Harris-Affine, Hessian-Affine, Maximally stable
extremal regions, Intensity based regions, Geometry based regions, Salient regions). The
second step is descriptor evaluation criterion that considers the distinctiveness, which is
measured by ROC of the number of correct matches with respect to the number of
corresponding regions against false positive rate with a distance threshold for 6 descriptors
(SIFT, steerable filters, differential invariants, complex filters, moment invariants, crosscorrelation). On the other hand, [16] gives the effects of different detectors (SIFT, SURF,
BRISK, ORB, FAST, GFTT, STAR) and descriptors (SIFT, SURF, BRISK, ORB, BRIEF,
FREAK) on RGB-D SLAM (Simultaneous Localization and Mapping) methods. The
performance evaluations of these methods are investigated in terms of accuracy by measuring
empirically the distance of matched objects and time to process per frame. The detectordescriptor combinations of this study are limited and only BRIEF and FREAK descriptors are
combined with other detectors. In an addition to the comparison of feature detectors and
descriptors performance on visual SLAM, [18] compares the effects of 6 feature descriptors
(BRIEF, BRISK, FREAK, ORB, SIFT, SURF) on graph-based VSLAM algorithm according
to localization accuracy and motion speed of the camera for real-time performance using two
different datasets as being used as two different motion scenario. Even though, both the
number of performance parameters and compared descriptors are limited, and there are no
detectors compared, similarly to our accuracy and time performance evaluation results given
in Table 1, the results are clearly show that the accuracy of SIFT and the speed of BRIEF are
the best options. In the context of image search and fine-grained classification, [17] provides
the performance assessment results in terms of accuracy and time by proposing new
approaches such as modifying Harris, Hessian and DoG detectors to extract dense patches,
changing the scale of selecting these patches’ edges, and filter selection is made by their own
method to locate patches. Although this paper studies on the developing the accuracy and
process time, it is underlined that not to target high repeatability. Besides, [19] examines the
JPEG compression’s effects on different feature detector-descriptor combinations. They used
VLBenchmarks [21] framework to test their 60 combinations composed of 10 detectors with 6
descriptors, which the dependency to this framework may harm some important properties
about performance evaluation results such as repeatability, accuracy, speed and extracted
feature numbers. Performance evaluation parameter of this study is very limited with mAP
(mean average precision) Scores, which is computed by taking the average of 55 same-level
compressed query images with JPEG with a quality range from 4 to 20 after applying 3
deblocking methods (spatial domain, frequency domain, hybrid filtering), by comparing the
number of feature detector-descriptor combinations and the feature extraction speed is not
considered as a performance evaluation criterion.
Similar to [15], [20] gives the comparison of performance evaluation results about affine
covariant region detectors using a structured and textured scenes under different conditions
such as viewpoints changes, scale changes, illumination and blurring variations. This
comprehensive study assesses the performance in terms of measuring the repeatability by the
detector’s performance on determining the corresponding scene region, and the accuracy with
respect to regions’ shape, scale, and localization. In addition, also, the distinctiveness of the
detected regions is evaluated as another parameter for performance to emphasize the
discriminative power of these methods.
Our study is one of the most comprehensive comparisons available because we give 23
different feature detector-descriptor performance evaluation results in terms of accuracy and
speed in a real-world scenario including localization experiment. Additionally, we present 19
different feature detector-descriptor comparisons regarding the number of correct matches,
mean angle difference between key-points, and minimum distance metrics. Although, as
mentioned above, the literature studies mostly focused on accuracy and time performance
evaluations, this work investigates the time consumed to match two features, time
performance from start to finish of the comparison of all query images with all dataset images
for different detector-descriptor methods combinations, accuracy values of these
combinations, and the relative relations of different detector-descriptor combinations with
respect to correct matches, mean angle difference between key-points, and minimum distance
between key-points.
3. Data and Methodology
In this section we give details about the dataset and query images which we created for this
study. Query image dataset is created by grabbing indoor images with the dimensions of
555x480 pixels using a low-cost CMOS camera from a laboratory environment by rotating the
camera by 15o with an apparatus at predetermined points for 3 different heights. Besides, the
dataset that contains the template images to be matched with the query images which are
chosen from the query image dataset. Figure 1.a shows the image grabbing process and the
hypothetical volumetric location cube, and the objects in the template images are given in
Figure 1. b. A collage of query images grabbed from two different points is displayed in
Figure 1. c.
In addition, before obtaining performance results of a combination of methods between query
and template images, two elimination steps are evaluated in order to prevent obtaining trivial
results. After a satisfactory matching score is obtained then the robot location is determined
approximately in a hypothetical cube.
In the first step, we used FAST method to detect the key-points of the query image and
determined a hysteresis threshold that provides us to eliminate images which have less keypoints than the lower threshold or have much key-points than the upper threshold. In the
second step, we used histogram comparison results to eliminate query images before matching
process. Thus we have same number of matching results for all combinations. After these
steps we achieve the performance results at matching process of the algorithm.
Figure 1 a. Demonstration of grabbing images from the environment and the localization of the mobile robot in
a hypothetical location cube. b. A part of template image dataset. c. A sequence of query image dataset which
grabbed from 2 points.
In order to visualize experimental localization results, two different tools are used. Figure 2.a
demonstrates the mobile robot’s path during its flight in the office room and Figure 2.b
displays the same path in a virtual reality environment. The red symbols indicate the
recognized location of the robot which means a matched query image with a template image
and it can be seen from the Figure 1. a that there are much more recognized points at the last
stage of the movement than the beginning because of the speed of the robot is faster in the
beginning and then slows down.
Figure 2 Localization experiment results. a. Demonstration of the location of the mobile robot in 3D
coordinates. b. Path followed by mobile robot using a different visualization tool.
4. Performance Analysis
This section demonstrates the performance results using different parameters such as time for
total computation, number of matches per second, and accuracy given in Table 1 and mean
key-point angle differences, number of correct of matches, distance metrics of matches, given
in Figure 3.
In order to complete Table 1, we compared all query images grabbed from the office room
with template images chosen from query images. Thus, we observed the parameter changes
whether an image pair (query and template) is related according to its position/rotation or
these query and template images are irrelevant. Before the comparison section of the program,
we eliminate some of the query images using a hysteresis threshold with respect to the
number of key-points which are computed using FAST method. The first column of Table 1,
total running time of the algorithm means the time passed until the robot finishes the same
number of image comparison process using different combinations on the same path between
560 query images and 127 template images. The last column is the time passed to match
correctly per features extracted by different feature-detector-descriptor combinations. It is
clear from Table 1 that the fastest combination consists of ORB for both key-point detector
and key-point descriptor. On the other hand, minimum number of correct matches per second
belongs to the combination of BRISKSIFT; conversely, maximum number of correct matches
per second belongs to the combination of FASTBRIEF. Furthermore, SURFSIFT combination
has the biggest running time. In an addition, if the SIFT is used as the key-point descriptor,
time for total computation is lower than all the other combinations for each key-point detector
methods. It can be easily estimated that the accuracy for the 127 template images are %100
for all combinations because the template images are chosen from the query images and these
are exactly the same.
%𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
∑(𝑇𝑟𝑢𝑒 𝑝𝑜𝑖𝑠𝑖𝑡𝑖𝑣𝑒𝑠 + 𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠)
127 + 0
∗ 100 =
∗ 100 = %100
∑(𝑇𝑜𝑡𝑎𝑙 𝐶𝑎𝑠𝑒𝑠)
127
Therefore, giving these accuracy rates is trivial. However, we formulated a different accuracy
term that includes the 3 height level for 5 different pose angles from -300 to 300 according to
the current pose angle and height, and if the histogram comparison value is higher than the
threshold value (0.9) of that comparison result, it is accepted that this comparison is made for
the same point at different pose angles. This measure of accuracy works because the
algorithm is designed so that if any match is achieved in this range in terms of given
parameters, the robot location detected at that point is acceptable. Furthermore, in Table 1 the
accuracy rates are given for explained situation and SIFTSURF is ascertained as the most
accurate combination with % 98.41. Moreover, as one can see that excluding BRISK method
as key-point detector combinations, the best accuracy rates are achieved when the SURF
method is the key-point descriptor.
Table 1 Performance Results for All Combinations.
KEYPOINT
DETECTOR
TYPE
ORB
SURF
SIFT
FAST
BRISK
Parameters Utilized in Performance Analysis
KEYPOINT DESCRIPTOR
TYPE
Total Time of the
Algorithm (sec)
Accuracy
GroundTruth
Number of Correct
Matches per Second
BRIEF
21303.299
%62.83
127*3*5
1457.011
BRISK
23461.675
%74.28
127*3*5
956.158
SIFT
97603.701
%72.28
127*3*5
318.012
SURF
79391.059
%97.90
127*3*5
390.965
ORB
21330.015
%63.62
127*3*5
1455.186
BRIEF
32277.702
%62.36
127*3*5
1568.788
BRISK
35133.182
%63.52
127*3*5
1275.433
SIFT
196487.667
%68.82
127*3*5
280.432
SURF
79135.063
%89.54
127*3*5
696.295
ORB
35074.987
%63.78
127*3*5
1414.503
BRIEF
30938.270
%62.73
127*3*5
879.923
BRISK
32422.058
%64.67
127*3*5
920.443
SIFT
45919.045
%62.31
127*3*5
698.415
SURF
35319.080
%98.41
127*3*5
908.024
BRIEF
23355.701
%62.52
127*3*5
2736.879
BRISK
33752.854
%63.20
127*3*5
454.341
SIFT
56154.363
%72.44
127*3*5
1373.775
SURF
37357.531
%88.30
127*3*5
2065.004
ORB
22734.255
%62.62
127*3*5
275.125
BRIEF
20517.530
%64.62
127*3*5
320.6933
SIFT
34284.954
%86.61
127*3*5
202.3933
SURF
26043.988
%80.32
127*3*5
266.4356
ORB
23335.765
%69.76
127*3*5
289.8432
In order to make our comparison more comprehensive, we give the performance results for
different pose cases using the query and template image matches within an rotational pose
range of [−300 , 300 ] for 5 cases in Figure 3. From Figure 3.a, it is clear that the all methods
are scattered in a wide range in 3-dimensional space, and also FASTSURF combination gives
the best results for all rotations excluding the comparison given for the same images with
respect to number of correct matches, mean of angle of differences between correctly matched
key-points and minimum distance metrics. If we limit our comparison parameters to two of
recent parameters for these comparison results by taking the application priorities into
account, the best combination may change. For instance, if the minimum distance metrics and
the number of correct matches are important for an application, then for 150 and 300 there are
4 best results (SURFSURF, SURFBRIEF, FASTORB, and SURFBRISK). It is possible to
make such analysis for any kind of priorities from the given performance evaluation
parameters.
If the comparison is given for the same image, the combinations align on a straight line at
different points. In this case, SURFSURF and SURFSIFT combinations both give the best
results, and SIFTBRIEF, ORBBRISK, SIFTSURF, SIFTBRISK, and SIFTSIFT are very close
to each other giving the worst results. As a matter of course, only the number of correct
matches change between 220 matches and 990 matches, other parameters are the same for all
combinations.
Whether the number of correct matches is changing between around 200 and 1000 for all
rotation cases, minimum distance metrics and mean of angle difference values between keypoints vary. For positive signed rotations (150 , 300 ), minimum distance metrics values change
between 0 and 170 pixels. On the other hand, for negative signed rotations (−150, −300 ),
minimum distance metrics values change between 0 and 300 pixels. Furthermore, average of
angle difference values between correctly matched key-points vary from −300 to 300 for
negative signed rotations, and from −100 to 400 for positive signed rotations.
Consequently, we may obtain the worst combination results with respect to present
parameters by calculating the geometrical distance using the relative geometric positions of
all combinations from the figures to the best combination. For the rotations of 300 , 150 , and
−300 , ORBSIFT, for the same image comparison SIFTBRIEF, and for −150 , SURFSIFT
combinations are obtained as the worst combinations with respect to given comparison
parameters.
Figure 3 Performance results for the different angle cases. These comparison results are given for the points
that are rotated from -300 to +300 according to the current point which query image and template image is the
same with respect to number of correct matches, minimum distance metrics and the average of angle difference
values between key-points. Lower the average of angle difference values between key-points, and minimum
distance metrics, and higher the number of correct matches is desired for the best result. In summary for all
subfigures displayed in here is the longer to the central point, which indicates the best method, the worse the
performance. a. Comparison of the template image with +𝟑𝟎𝟎 rotated query image from the same environment
and height. b. Comparison of the template image with +𝟏𝟓𝟎 rotated query image from the same environment
and height c. Comparison of the template image with −𝟑𝟎𝟎 rotated query image from the same environment and
height d. Comparison of the template image with −𝟏𝟓𝟎 rotated query image from the same environment and
height e. Comparison of the template image with the same image as query image.
5. Conclusion and Future Work
Developing good feature detectors and descriptors is still a challenging research topic.
Independent from the development of hardware, especially for embedded and/or onboard
applications, and in any autonomous system, the parameters such as computation time,
robustness, repeatability, and accuracy are crucial in the context of soft algorithm
development of feature detectors and descriptors. It has been a usual path to create a new
feature detector-descriptor method by combining previous detector and descriptor methods in
an effective algorithm. Our study provides more comprehensive information with more
performance evaluation parameters (in terms of accuracy and temporal costs such as the
running time from start to end, time per matched features) wider coverage of methods (i.e. 23
combinations). In addition to these commonly used metrics, we give the performance results
of 19 combinations with respect to new metrics, such as mean keypoint angle differences,
number of correct matches, and the distance metrics of matches.
It is clear from our results that there are trade-offs between different parameters and
performance criteria when different feature detector-descriptor combinations are analyzed. If
a wide rotation range is desired to be matched, then the algorithm finds weak features and
matches are not realized because of predefined threshold values. On the other hand, if a
narrow rotation range is desired, the algorithm finds too many features to match and this
causes the increase at total loop time because of an increase at number of correct matches per
unit time. The lowest total running time belongs to the combination of BRISK-BRIEF with
64.62% accuracy rate and 320.6933 correct matches per second. The highest running time is
for the SURF-SIFT combination with 68.82% accuracy rate and 280.432 correct matches per
second. Additionally, SIFT-SURF combination has 98.41% accuracy rate in a total of
35319.080 seconds with 908.024 correct matches per second. Furthermore, FAST-SURF
combination is the method to give the best results for the angular rotations of 30°, 15°, -15°,
and -30° for comparisons. Moreover, for the case of the comparison of the same images,
SURF-SURF and SURF-SIFT combinations give the best results with respect to number of
correct matches, mean of angle of differences between correctly matched keypoints and
minimum distance metrics.
Future research in the computer vision, perception and robotics areas can benefit from the
results provided in this study. From the investigated methods, the individual priorities of
different applications can easily be reflected in combinations as the best option. To be more
specific, our results can improve the applications in the fields of object recognition by
image/object matching in different conditions, and in visual SLAM field by extracting robust
features with its generic outcomes.
In our future work, using the results of this study, an experiment in real-time on our ongoing
humanoid project UMAY [21] with a wider object database will be performed to get the
semantic information of unknown indoor environments. Since the recognition of the
environment is crucial for the robot in terms of executing or interpreting the given tasks in a
dynamical way, object recognition algorithms here can provide an optimized method to start
with processing the visual information. As a consequence, this framework can be considered
to form a basis for future applications involving extracting of visual semantic cues.
Acknowledgements
The research was conducted at the Mechatronics Research and Education Center at Istanbul
Technical University.
References
[1] David G. Lowe, "Distinctive image features from scale-invariant keypoints," International Journal of Computer Vision,
60, 2 (2004), pp. 91-110.
[2] H. Bay, A. Ess, T. Tuytelaars, L. van Gool, "Speeded-up Robust Features (SURF)", Computer Vision and Image
Understanding (CVIU), Vol. 110, No. 3, pp. 346--359, 2008
[3] Leutenegger, S. ; Autonomous Syst. Lab., ETH Zurich, Zurich, Switzerland ; Chli, M. ; Siegwart, R.Y., “BRISK: Binary
Robust invariant scalable keypoints”, Computer Vision (ICCV), 2011 IEEE International Conference on, pp. 2548-2555,
Barcelona, 6-13 Nov. 2011.
[4] Calonder, M., Lepetit, V., Strecha, C., Fua, P., “BRIEF: Binary Robust Independent Elementary Features”, 11th European
Conference on Computer Vision, pp 778-792, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV
[5] Kucharczak, F., Da Silva, A., F., Thomaz., L., A., Carvalho, G., Da Silva, E., A., B., Netto, S., L., "Comparison and
Optimization of Image Descriptors for Real-Time Detection of Abandoned Objects."
[6] Rublee, E., Rabaut, V., Konolige, K., Bradski, G., “ORB: an efficient alternative to SIFT or SURF”, In Proc. of the
IEEE Intl. Conf. On Computer Vision (ICCV), volume 13, 2011.
[7] Rosten, E., Drummond, T., “Fusing points and lines for high performance tracking”, Computer Vision, 2005. ICCV 2005.
Tenth IEEE International Conference on (Volume:2 ), pp 1508-1515 Vol. 2, Beijing, 17-21 Oct. 2005.
[8] Rosten, E., Drummond, T., “Machine learning for high-speed corner detection”, 9th European Conference on Computer
Vision Proceedings, Part I, pp 430-443, Graz, Austria, May 7-13, 2006.
[9] Remondino, F., “Detectors and descriptors for photogrammetric applications.”, International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences, 36(3): 49–54, 2006.
[10] Dahl, L., A., Aanæs, H., Pedersen, K., S., “Finding the Best Feature Detector-Descriptor Combination”. International
Conference on 3D Imaging, Modeling, Processing, Visualization and Transmission (3DIMPVT), 16-19 may 2011
[11] Remondino, Fabio. "Detectors and descriptors for photogrammetric applications." International Archives of
Photogrammetry, Remote Sensing and Spatial Information Sciences 36.3 (2006): 49-54.
[12] Kucharczak, Florentin, et al. "Comparison and Optimization of Image Descriptors for Real-Time Detection of
Abandoned Objects."
[13] Figat, Jan, Tomasz Kornuta, and Włodzimierz Kasprzak. "Performance evaluation of binary descriptors of local
features." Computer Vision and Graphics. Springer International Publishing, 2014. 187-194.
[14] Mikolajczyk, Krystian, and Cordelia Schmid. "A performance evaluation of local descriptors." Pattern Analysis and
Machine Intelligence, IEEE Transactions on 27.10 (2005): 1615-1630.
[15] Mikolajczyk, Krystian, and Cordelia Schmid. "Comparison of Affine-Invariant Local Detectors and Descriptors”, 12th
European Signal Processing Conference, Vienna, pp. 1729-1732, 2004.
[16] Guclu, Oguzhan, and Ahmet Burak Can. "A Comparison of Feature Detectors and Descriptors in RGB-D SLAM
Methods." Image Analysis and Recognition. Springer International Publishing, pp. 297-305, 2015.
[17] Iscen, Ahmet, et al. "A comparison of dense region detectors for image search and fine-grained classification." Image
Processing, IEEE Transactions on (Volume:24, Issue: 8), pp. 2369-2381, 2015.
[18] Hartmann, Jean-Michel, Jan Helge Klussendorff, and Erik Maehle. "A comparison of feature descriptors for visual
SLAM." Mobile Robots (ECMR), 2013 European Conference on. IEEE, 2013.
[19] Chao, J., Al-Nuaimi, A., Schroth, G., Steinbach, E., “ Performance comparison of various feature detector-descriptor
combinations for content-based image retrieval with JPEG-encoded query images”., Multimedia Signal Processing
(MMSP), 2013 IEEE 15th International Workshop on, pp. 029-034, Pula, 2013.
[20] Mikolajczyk K, Tuytelaars T, Schmid C, et al., “A Comparison of Affine Region Detectors”, International Journal of
Computer Vision, Vol:65, ISSN:0920-5691, Pages:43-72, 2005.
[21] Lenc, K., Gulshan, V., Vedaldi, A. “VLBenchmarks, http://www.vlfeat.org/benchmarks/”., 2012.
| 1cs.CV
|
arXiv:1710.07792v2 [math.ST] 17 Nov 2017
Cointegrated Density-Valued Linear Processes
Won-Ki Seo
Department of Economics, University of California, San Diego
November 20, 2017
Abstract
In data rich environments we may sometimes deal with time series
that are probability density-function valued, such as observations of
cross-sectional income distributions over time. To apply the methods of functional time series analysis to such observations, we should
first embed them in a linear space in which the essential properties
of densities are preserved under addition and scalar multiplication.
Bayes Hilbert spaces provide one way to achieve this embedding. In
this paper we investigate the use of Bayes Hilbert spaces to model
cointegrated density-valued linear processes. We develop an I(1) representation theory for cointegrated linear processes in a Bayes Hilbert
space, and adapt existing statistical procedures for estimating the corresponding attractor space to a Bayes Hilbert space setting. We revisit
empirical applications involving earnings and wage densities to illustrate the utility of our approach.
I thank Brendan K. Beare for a lot of discussion at all stages of this paper. I am
also grateful to Dakyung Seong and Lam Nguyen for useful discussion on section 6 of this
paper.
1
1
Introduction
While the subject of time series analysis has traditionally dealt with time
series taking values in finite dimensional Euclidean space, a recent literature
on functional time series analysis deals with time series taking values in an
infinite dimensional Banach or Hilbert space. Each observation of such a time
series is a functional object; for example, it could be a square-integrable function, continuous function, or probability density function. Bosq (2000) gives
a rigorous theoretical treatment of linear processes in Banach and Hilbert
spaces. Horváth and Kokoszka (2012) discuss statistical aspects of functional
data and time series analysis, and provide various empirical applications.
Granger (1981) introduced the notion of cointegration as a way to model
long-run equilibrium relationships between real-valued time series. A recent
paper by Chang et al. (2016) is the first to consider the possibility of cointegration in a functional time series setting. The authors consider a time
series of probability densities taking values in the space of square-integrable
real functions on a compact interval K, denoted by L2 (K), and provide a
notion of cointegration adapted to this space. They develop associated statistical methods based on functional principal component analysis (FPCA),
and provide empirical applications to time series of earnings and stock return
densities.
Beare (2017) notes that a technical complication arises in the framework
developed by Chang et al. (2016): the nonnegativity property of probability
densities is incompatible with the type of nonstationarity exhibited by integrated time series, and this incompatibility cannot be resolved by a simple
demeaning of densities. Nontrivial examples of cointegrated density-valued
processes in L2 (K) therefore do not exist. Even in models of stationary probability density-valued time series, Petersen and Müller (2016) have observed
that it is generally inadvisable to treat such time series as taking values in
the subset of L2 (K) consisting of probability densities, as this subset does
not form a linear subspace of L2 (K). It is clear that an arbitrary linear
combination of densities is not a proper density; only a convex combination
of densities is a proper one.
The primary purpose of this paper is to show that the technical complications just discussed can be resolved by viewing probability density-valued
observations as elements of a Bayes Hilbert space. Such spaces were introduced by Egozcue et al. (2006) and developed further by van den Boogaart
et al. (2014). They are constructed in such a way that different elements
2
of the space correspond to different probability densities, and the essential
properties of probability densities are preserved under addition and scalar
multiplication. The notion of cointegration may easily be adapted to this
space as in Beare et al. (2017), who extend the framework developed by
Chang et al. (2016) to arbitrary complex separable Hilbert spaces.
A secondary purpose of the paper is to study the behavior of densityvalued autoregressive processes of order p (AR(p) processes) taking values
in a Bayes Hilbert space. We provide conditions under which an AR(p)
law of motion has a stationary solution or I(d) solution for some positive
integer d, and a necessary and sufficient condition for such an I(d) solution
to in fact be I(1). These results are closely related to the so-called GrangerJohansen representation theorem and its generalization to a possibly infinite
dimensional Hilbert space setting by Beare et al. (2017) and Beare and Seo
(2017).
A third contribution of the paper is the provision of statistical methods
to estimate the attractor space (to be defined later) for an I(1) process in a
Bayes Hilbert space. These methods are based on the functional unit root
test proposed by Chang et al. (2016). We illustrate their usefulness with
empirical applications to time series of cross-sectional densities of individual
earnings and wages.
The remainder of the paper is organized as follows. In Section 2, we review some background material on Bayes Hilbert spaces and other essential
mathematical concepts. In Section 3, we explain how Bayes Hilbert spaces
can provide a useful setting for models of cointegrated density-valued time
series. Results on I(d) representations of cointegrated AR(p) processes in
Bayes Hilbert space are provided in Section 4. Statistical tools for studying cointegrated density-valued time series are introduced in Section 5, and
illustrated with empirical applications in Section 6.
2
Preliminaries
Here we briefly review essential background for the study of cointegrated
density-valued linear processes, and fix standard notation and terminology.
3
2.1
Bayes Hilbert spaces
Bayes Hilbert spaces provide the setting for our treatment of cointegrated
density-valued linear processes. The discussion provided here will omit some
details to conserve space. The reader is referred to Egozcue et al. (2006) and
van den Boogaart et al. (2014) for a rigorous introduction to the subject.
Let (M, A) be a measurable space and P be the set of σ-finite positive
real-valued measures defined on it. For any measure λ, called a reference
measure, define
M(λ) := {ν ∈ P : ν(A) = 0 ⇔ λ(A) = 0,
∀A ∈ A}.
That is, M(λ) is the collection of σ-finite positive measures absolutely continuous with respect to λ. Due to the Radon-Nikodym theorem, any element
ν in M(λ) may be identified with the λ-density dν/dλ, so hereafter we always
regard an element in M(λ) as its λ-density.
We define an equivalence relation ' on M(λ) as follows. For λ-densities
f, g ∈ M(λ), the equivalence f ' g holds if there is some a > 0 such that
Z
Z
gdλ for all A ∈ A.
f dλ = a ·
A
A
The above says that if two measures ν1 and ν2 are proportional in the sense
that ν1 (A) = a · ν2 (A) for all A ∈ A, their λ-densities are equivalent under
'. The collection of '-equivalence classes of λ-densities, denoted by B(λ),
are the elements of a Bayes space associated with λ. We define two vector
operations (⊕, ), called perturbation and powering, as follows. For a ∈ R
and f, g ∈ B(λ), define
Z
f ⊕g '
f gdλ,
Z
a
f'
f a dλ.
Negative perturbation is defined as f g = f ⊕ (−1 g). B(λ) equipped
with the two vector operations (⊕, ) is a vector space.
We hereafter assume that λ is a finite measure and let B 2 (λ) be the
collection of λ-densities that have square-integrable logarithms, i.e.
Z
2
2
B (λ) := f ∈ B(λ) :
|log f | dλ < ∞ .
4
It turns out that B 2 (λ) is a vector subspace of B(λ). Define the inner product
h·, ·iB 2 (λ) in the following way: for f, g ∈ B 2 (λ),
Z
Z
Z
log f − log f dλ
log g − log f dλ dλ.
hf, giB 2 (λ) =
The following result is obtained in van den Boogaart et al. (2014).
Lemma 2.1. B 2 (λ) equipped with (⊕, ) and h·, ·iB 2 (λ) is a separable Hilbert
space.
Define L2 (λ) as the set of measurable square-integrable functions f : Ω →
R. An element of L2 (λ) is regarded as a class of square-integrable functions
that are λ-almost surely equivalent. L2 (λ) is assumed to be
R equipped with
the inner product h·, ·iL2 (λ) defined as follows: hf˜, g̃iL2 (λ) = f˜g̃dλ for f˜, g̃ ∈
L2 (λ). Define
Z
2
2
L (λ) = f˜ ∈ L (λ) : f˜ = 0 ,
which is a closed subspace of L2 (λ). One useful fact for the study of densityvalued functional objects is that there exists an isometric isomorphism between B 2 (λ) and L2 (λ).
Lemma 2.2. The following hold:
(a) B 2 (λ) is isometrically isomorphic to L2 (λ).
(b) clr : B 2 (λ) → L2 (λ) defined as
Z
clr(f ) = log f −
log f dλ,
f ∈ B 2 (λ),
is an isometrical isomorphism. Moreover, clr−1 is given by
clr−1 (f˜) ' exp(f˜),
f˜ ∈ L2 (λ).
See van den Boogaart et al. (2014) for more details. The map clr is called
the centered log-ratio (clr) transformation. Since clr is a unitary linear map,
the following are naturally implied:
clr(f ⊕ g) = clr(f ) + clr(g),
hf, giB 2 (λ)
clr(a f ) = a · clr(f ),
Z
= hclr(f ), clr(g)iL2 (λ) = clr(f ) clr(g)dλ.
5
Let λ be the uniform measure defined on a compact interval K. The Bayes
Hilbert space associated with λ, B 2 (λ), was considered in Egozcue et al.
(2006), and it may be understood as the space of probability densities with
support K. Note that B 2 (λ) is the space of equivalence classes of positive
functions defined on K with square-integrable logarithms with respect to
the
R uniform measure λ. Within an equivalence class, the integral constraint
f dλ = 1 singles out the representative element as a probability density
K
function. Then for f, g ∈ B 2 (λ), two operations ⊕ and are be defined as
f (x)g(x)
,
f
(y)g(y)dy
K
f ⊕ g(x) = R
a
f (x) = R
f (x)a
.
a dy
f
(y)
K
If we denote the indicator function on K by 1K , then the '-equivalence
class associated with 1K is the neutral element of perturbation. Naturally,
1 ∈ R is the neutral element of powering. Perturbation and powering are
constructed such that the results of f ⊕ g and a f are always densities even
if one or both of f and g are not proper densities. If one regards an element
in B 2 (λ) as a probability density, perturbation can be interpreted as a Bayes
updating. This is why B 2 (λ) is called a Bayes (Hilbert) space in the previous
literature.
The fact that the resulting density from any arbitrary linear combination
of two elements of B 2 (λ) is a proper density is crucial to the approach taken
in this paper. In fact, the point of Beare (2017)’s comments on Chang et al.
(2016) is rooted in the fact that linear combinations of probability densities
in L2 (K) are not necessarily probability densities. It will be shown in the
subsequent sections that the Bayes Hilbert space is not just a good candidate on which to define (cointegrated) density-valued processes, but also its
Hilbert space structure makes it possible to use well-developed statistical
techniques in Hilbert spaces without further technicality.
We will consider a time series of densities with common support K.
Therefore, λ will be assumed to be the uniform measure on K and densityvalued linear processes will be defined in the associated Bayes Hilbert space
B 2 (λ). Allowing the reference measure to be another measure with compact
support is trivial. However, generalization to a reference measure with unbounded support is nontrivial and we will not consider this case. When the
reference measure has unbounded support, the space of probability measures
as a subset of B 2 (λ) is not a subspace, as the following example demonstrates.
6
Example 2.1. Let λ be the standard normal measure defined on the real
line and B 2 (λ) be the associated Bayes Hilbert space. Let ν be the normal
2
Rmeasure 2with mean 0 and variance 2. Then f := dν/dλ ' exp(x /4), so
| log f | dλ < ∞ since the fourth moment of the standard normal measure
is finite. However, note that f itself represents an infinite measure; i.e. it is
not integrable. So in this case, an element of B 2 (λ) need not be a proper
density.
Aitchison (1982) introduced the space B 2 (λ) equipped with (⊕, ) and
h·, ·iB 2 (λ) in the special case where λ is the discrete uniform measure supported on a finite set K. The term Aitchison geometry is used to connote
the geometric properties of this Bayes Hilbert space. The term generalized
Aitchison geometry may be preferred when referring to Bayes Hilbert spaces
constructed with general reference measures.
2.2
Bounded linear operators on Bayes Hilbert space
Here we briefly summarize essential concepts on bounded linear operators.
The reader is referred to Bosq (2000) and Conway (1994) for a detailed
introduction.
Let H be B 2 (λ) or its complexification BC2 (λ), defined as
BC2 (λ) = {f ⊕ ig : f, g ∈ B 2 (λ)},
where i denotes the imaginary unit. BC2 (λ) is understood to be equipped
with the inner product hf ⊕ ig, f 0 ⊕ ig 0 iBC2 (λ) = hf, f 0 iB 2 (λ) − ihf, g 0 iB 2 (λ) +
ihg, f 0 iB 2 (λ) + hg, g 0 iB 2 (λ) .
Let B(H) be the space of bounded linear operators, equipped with the
usual operator norm
kAkB(H) = sup kAf kH ,
A ∈ B(H).
kf kH ≤ 1
A ∈ B(H) is said to be compact if for two orthonormal bases (uj , j ∈ N)
and (vj , j ∈ N) of H, it can be written as
Af =
∞
M
γj hf, uj iH
j=1
for some sequence (γj , j ∈ N) tending to zero.
7
vj ,
Given A ∈ B(H), we define two fundamental subspaces, the range and
kernel of A, as follows.
ker A := {f ∈ H : Af = λ},
ran A := {Af : f ∈ H}.
The dimension of ker A is called the nullity of A, and the dimension of ran A
is called the rank of A.
For A ∈ B(H), there exists an operator A∗ ∈ B(H), called the adjoint of
A, that is uniquely determined by the following property
hAf, giH = hf, A∗ giH ,
for all f, g ∈ H.
Given a subset V ⊂ H, the orthogonal complement of V is denoted by
V ⊥ = {g ∈ H : hf, giH = 0 for all f ∈ V }.
The closure of V , defined as the union of V and its limit points, is denoted
by cl(V ). It turns out that the following relationship holds (see e.g. Conway,
1994, pp. 35-36)
ker A = (ran A∗ )⊥
and (ker A)⊥ = cl(ran A∗ ).
(2.1)
A ∈ B(H) is said to be positive semidefinite if hAf, f iH ≥ 0 for all f ∈ H.
If the inequality is strict for all f 6= λ, A is said to be positive definite.
Given a subspace V ∈ H, A|V denotes the restriction of an operator
A ∈ B(H), i.e. A|V : V → H.
Let K = L2 (λ) if H = B 2 (λ), and let K = L2C (λ) if H = BC2 (λ), where
L2C (λ) is the complexification of L2 (λ) similarly defined as BC2 (λ). Let clrC
denote the complexification of clr, i.e. clrC (f ⊕ ig) = clr(f ) + i · clr(g). Then
clrC is an isometric isomorphism between BC2 (λ) and L2C (λ). Since clr or clrC
is an isometric isomorphism between H and K, any bounded linear operator
in B(H) can be understood as the corresponding element in B(K). This
property will be used repeatedly in subsequent sections.
2.3
Random elements of Bayes Hilbert space
Here we summarize some essential concepts relating to random elements of
a Bayes Hilbert space. See Bosq (2000) for a detailed discussion of random
elements of Banach and Hilbert spaces.
8
Let (Ω, F, P ) be the underlying probability triple and let H = B 2 (λ) or
which is understood to be equipped with its Borel σ-field. An Hvalued random variable is a measurable map X : Ω → H. Let k · kH be the
norm defined on H. We say that X is integrable if EkXkH < ∞. If X is
integrable, there exists a unique element, denoted by EX, such that
BC2 (λ),
EhX, f iH = hEX, f iH ,
for all f ∈ H.
EX is called the expectation of X.
If EkXk2H < ∞ then X is said to be square-integrable. Let L2H denote
the space of square-integrable random variables X with EX = λ, equipped
with the norm
1/2
kXkL2H = EkXk2H
.
It turns out that L2H is a Banach space. For X ∈ L2H , the Cauchy-Schwarz
inequality implies that Xhf, XiH is integrable for all f ∈ H. Define the
operator CX ∈ B(H) by
CX (f ) := E (Xhf, XiH ) .
CX is called the covariance operator of X.
2.4
Operator pencils
Let H be a complex separable Hilbert space and U be an open connected set
in the complex plane C. An operator-valued map Q : U → B(H) is called
an operator pencil. In this paper, H = BC2 (λ) or L2C (λ) is considered.
For notational convenience, we put H = L2C (λ). An operator pencil Q(z)
is holomorphic on an open connected set D if and only if
Q(1) (z0 ) := lim
z→z0
Q(z) − Q(z0 )
z − z0
exists in the norm of B(H) for each z0 ∈ D. A well known fact is that when
Q(z) is holomorphic at z = z0 , it allows the power series expansion
Q(z) =
∞
X
Qk (z − z0 )k
k=0
for some sequence (Qk , k ∈ N) in B(H).
The set σ(Q) := {z ∈ U : Q(z) is not invertible} is called the spectrum
of Q, which turns out to be closed (Markus, 2012, p. 56).
9
3
Cointegrated density-valued linear processes
Hereafter the following is always assumed without explicitly stating it.
Assumption (R). λ is the uniform measure on a compact interval K ⊂ R.
Under the above assumption, B 2 (λ) is the Bayes Hilbert space studied in
Egozcue et al. (2006). The assumption is not essential but convenient. In
fact, extending the subsequent results of this paper to other measures supported on K, such as the truncated normal measure, can be easily done with
trivial modifications.
First we define linear processes in B 2 (λ), called Bayes linear processes.
Due to the Hilbert space structure, linear processes in B 2 (λ) may be easily
defined and studied according to the work of Bosq (2000, 2007). As seen in
Section 2, a Bayes linear process may be understood as a linear process of
probability densities with compact support K.
Then, we introduce I(1) processes in B 2 (λ) based on the previous work of
Beare et al. (2017) who provided a notion of cointegrated linear processes in
an arbitrary Hilbert space. Of course, it can be understood as an I(1) process
of probability densities with support K. We further investigate cointegration
in B 2 (λ).
3.1
Linear processes in B 2 (λ)
Suppose that we have a time series of densities f = (ft , t ≥ t0 ) with compact
t
for each t, where νt is a probability
support K ⊂ R. We assume that ft := dν
dµ
measure for each t and µ is the Lebesgue measure defined on R. Since
ft has compact support K, it is absolutely continuous with respect to λ.
This implies that the corresponding λ-density of the underlying probability
measure νt is µ(K)−1 ft . Thus, the unit integral constraint singles out ft
itself from the equivalence class containing ft . The time series of µ-densities
f = (ft , t ≥ t0 ) may therefore be regarded as a stochastic process taking
values in B 2 (λ).
Let ε = (εt , t ∈ Z) be an independent and identically distributed (iid)
sequence in L2B 2 (λ) and (Ak , k ≥ 0) be a sequence in B(B 2 (λ)) satisfying
P∞
2
k=0 kAk kB(B 2 (λ)) < ∞. For fixed t0 ∈ Z ∪ {−∞}, the sequence f = (ft , t ≥
10
t0 ) defined as
ft =
∞
M
Ak εt−k
(3.1)
k=0
is convergent in L2B 2 (λ) (Bosq, 2000, Lemma 7.1). We call the sequence f =
(ft , t ≥ t0 ) a Bayes linear process. Bayes linear processes are necessarily
stationary.
P∞
A Bayes linear process,
L∞ (3.1), is said to be standard2 if k=0 kAk kB(B 2 (λ)) <
∞. In this case, A := k=0 Ak is convergent in B(B (λ)). As in Beare et al.
(2017), we define the long-run covariance operator Λ ∈ B(B 2 (λ)) for a standard linear process as follows
Λ := ACε0 A∗ .
Since clr is an isometric isomorphism, it follows that the clr image of a
(standard) linear process in B 2 (λ) is a (standard) linear process in L2 (λ).
Denote by g̃ ∈ L2 (λ) the clr image of g ∈ B 2 (λ). Then the L2 (λ)-linear
process paired with (3.1) is
f˜t =
∞
X
Ãk ε̃t−k ,
k=0
where f˜t = clr(ft ), Ãk = clr ◦Ak ◦ clr−1 and ε̃ = (ε˜t , t ∈ Z) is an iid sequence
in L2 (λ).
3.2
I(1) processes in B 2 (λ) and cointegration
Chang et al. (2016) were the first to study I(1) processes taking values in
an infinite dimensional Hilbert space, specifically the space L2 (λ). Subsequently, Beare et al. (2017) considered I(1) processes in an arbitrary separable complex Hilbert space. In this section we specialize the latter setting
to the Bayes Hilbert space B 2 (λ). Let the time series of random densities
of interest f = (ft , t ≥ 0) be a sequence in B 2 (λ). Denote the time series
of first differences ∆f = (∆ft , t ≥ 1) with ∆ft = ft ft−1 . We say that
f = (ft , t ≥ 0) is I(1) if ∆ft satisfies
∆ft =
∞
M
k=0
11
Nk εt−k
(3.2)
for all t ≥ 1, where ε = (εt , t ∈ Z) is an iid sequence in L2B 2 (λ) , and (Nk , k ≥ 0)
P∞
2
is
k=0 kkNk kB(B 2 (λ)) < ∞ and N :=
La∞ sequence in B(B (λ)) satisfying
k=0 Nk 6= 0.
From Chang et al. (2016) and Beare et al. (2017), it can be shown that
the above I(1) sequence allows the so-called Beveridge-Nelson decomposition,
ft = (f0
ν0 ) ⊕ N ξ t ⊕ νt ,
t≥0
Lt
L∞
L
where ξt =
( ∞
s=1 εs and νt =
k=0 Nk εt−k with Nk = −1
j=k+1 Nj ).
This means that ft is obtained by combining three different components: an
initial condition f0 ν0 , a random walk component N ξt , and a stationary
component νt .
Given an I(1) sequence f = (ft , t ≥ 0), we define the cointegrating space
associated with f , denoted by C(f, B 2 (λ)), to be the set
C(f, B 2 (λ)) := g ∈ B 2 (λ) : hft , giB 2 (λ) , t ≥ 0 is stationary for some f0 .
Moreover, we call A(f, B 2 (λ)) := C(f, B 2 (λ))⊥ the attractor space. These
terms to indicate such subspaces are commonly used in the literature on
cointegration, see e.g. Johansen (1996).
Define the long-run variance operator of ∆f as Λ := N Cε0 N ∗ . If Cε0 is
positive definite, it must be the case that
ker Λ = ker N ∗ ,
and further Beare et al. (2017) showed that the cointegrating space is equal
to ker N ∗ .
The cointegrating space (or attractor space) can be identified by analyzing
its clr-image in L2 (λ), which may be convenient in practice. Suppose that
the clr image of (3.2) is
∆f˜t =
∞
X
Ñk ε̃t−k ,
k=0
where ∆f˜t = f˜t − f˜t−1 and Ñk = clr ◦Nk ◦ clr−1 . The Beveridge-Nelson
decomposition is
f˜t = (f˜0 − ν̃0 ) + Ñ ξ˜t + ν˜t ,
12
t≥0
where Ñ = clr ◦N ◦ clr−1 . We also define the long-run variance operator
of ∆f˜ as Λ̃ = Ñ Cε̃0 Ñ ∗ , where Cε̃0 is the covariance operator of ε̃0 . The
following result shows that we can fully identify the cointegrating space in
B 2 (λ) associated with f from the cointegrating space in L2 (λ) associated
with f˜, the clr image of f .
Proposition 3.1. If Cε0 is positive definite, then C(f, B 2 (λ)) = clr−1 ker Ñ ∗
Proof. Proposition 3.1 in Beare et al. (2017) implies that C(f, B 2 (λ)) =
ker N ∗ . Note that clr∗ = clr−1 since for g̃ := clr(g),
hf, clr−1 g̃iB 2 (λ) = hclr(f ), g̃iL2 (λ) .
This holds for any arbitrary g ∈ B 2 (λ). By the uniqueness of the adjoint
map, clr∗ = clr−1 . From these results, it is clear that
Ñ ∗ = clr ◦N ∗ ◦ clr−1 .
Since clr and clr−1 are bijective, it follows that ker N ∗ = clr−1 ker Ñ ∗ .
Chang et al. (2016) considered a cointegrated density-valued process with
values in L2 (λ). However, Beare (2017) pointed out that the attractor space
under their assumptions is always trivial, i.e. A(f, L2 (λ)) = {0}. This problem occurs because an I(1) process of probability densities in L2 (λ) cannot
satisfy the nonnegativity constraints required for each realization to be a
proper density; the space of probability densities with compact support K
is strictly smaller than L2 (λ) and not a subspace. Due to the geometry of
the Bayes Hilbert space, a cointegrated linear process in B 2 (λ) does not suffer from this problem. Furthermore, it is clear from the earlier discussion
of Bayes Hilbert spaces in Section 2.1 that the space of probability densities supported on K is isomorphic to L2 (λ), which is just a slightly smaller
subspace than L2 (λ) itself.
4
Cointegrated AR(p) processes in B 2(λ)
In this section, we study AR(p) processes with values in B 2 (λ) and obtain a
version of the Granger-Johansen representation theorem.
Suppose that a sequence of λ-densities (ft , t ≥ −p + 1) in B 2 (λ) evolves
according to
ft = A1 ft−1 ⊕ · · · ⊕ Ap ft−p ⊕ εt
13
(4.1)
for t ≥ 1, where ε = (εt , t ∈ Z) is an iid sequence in L2B 2 (λ) and A1 , . . . , Ap ∈
B(B 2 (λ)). Equation (4.1) may be understood as an autoregressive law of
motion for density-valued processes in B 2 (λ). The corresponding clr image
of (4.1) with values in L2 (λ) is given by
f˜t = Ã1 f˜t−1 + · · · + Ãp f˜t−p + ε˜t
(4.2)
for t ≥ 1, where Ãk = clr ◦Ak ◦ clr−1 . It is clear that Ãk ∈ B(L2 (λ)), and
that Ãk is compact if and only if Ak is compact. Let idL2 (λ) be the identity
operator on L2 (λ), and for z ∈ C define the operator pencil
Ã(z) = idL2 (λ) −z Ã1 − · · · − z p Ãp .
For each z ∈ C, Ã(z) is a bounded linear operator on L2C (λ), the complexification of L2 (λ). The operator pencil Ã(z) is polynomial in z; see Markus
(2012) for a detailed discussion of polynomial operator pencils. Let idB 2 (λ)
be the identity operator on B 2 (λ), and define
A(z) = idB 2 (λ) (z
A1 )
···
(z p
Ap ).
(4.3)
For each z, A(z) is a bounded linear operator on BC2 (λ). One can easily show
that A(z) = clr−1
C ◦Ã(z) ◦ clrC , so σ(A) = σ(Ã) holds.
To further analyze (4.1), we employ the following mutually exclusive assumptions.
Assumption (S). If z ∈ σ(A) then |z| > 1.
Let Dr be the open disk centered at zero with radius r > 0. Since σ(A)
is closed, Assumption (S) implies that there exists η > 0 such that A(z) is
invertible on D1+η .
Assumption (N).
(i) A1 , . . . , Ap ∈ B(B 2 (λ)) are compact operators.
(ii) If z ∈ σ(A) then |z| > 1 or z = 1. Moreover, 1 ∈ σ(A).
(iii) If P ∈ B(B 2 (λ)) is a projection on ran A(1), then the map
(idB 2 (λ) P)A(1) (1)|ker A(1) : ker A(1) → ran(idB 2 (λ) P)
is invertible.
14
Under Assumption (N)-(i), A(z) is an index-zero Fredholm operator for each
z ∈ C. This implies that ran A(1) is closed. (N)-(i) is not restrictive in practice since it is in fact common to assume the compactness of autoregressive
operators. (N)-(ii) implies the existence of η > 0 such that A(z) is invertible
everywhere on D1+η except at z = 1. The role of (N)-(iii) will be discussed
later.
The AR(p) law of motion (4.1) behaves very differently depending on
whether A(z) satisfies Assumption (S) or (N). In the former case it generates
a stationary process which may be represented as a standard Bayes linear
process.
Proposition 4.1. Suppose that Assumption (S) is satisfied. Then the AR(p)
law of motion (4.1) admits a unique stationary solution in L2B 2 (λ) given by
ft =
∞
M
Nk εt−k
(4.4)
k=0
for all t ≥ 1. Moreover, Nk is given by
Nk =
1
k!
N (k) (0)
and N (z) := A(z)−1 is holomorphic on D1+η for some η > 0.
Proof of Proposition 4.1. Consider the clr image of (4.2) in L2 (λ) and let
H = L2 (λ). Using the Markovian representation in Bosq (2000, p. 128), the
AR(p) law of motion in H can be re-expressed as an AR(1) law of motion in
Hp by writing
Ft = AFt−1 + Et ,
where
f˜t
f˜
t−1
Ft = . ,
..
f˜t−p+1
ε˜t
0
Et = .. ,
.
0
Ã1 Ã2 · · · Ãp−1 Ãp
id
0
0
H 0 ···
A= .
.. . .
..
.. .
.
.
.
.
.
.
0
0 · · · idH 0
15
(4.5)
Define A(z) := idHp −Az, i.e.
idH −Ã1 z −Ã2 z −Ã3 z −Ã4 z
···
−Ãp z
− id z
idH
0
0
···
0
H
0
− idH z idH
0
···
0
A(z) =
.
.
.
.
..
.
..
..
..
..
..
.
0
0
0
0
− idH z idH
A11 (z) A12 (z)
=:
,
A21 (z) A22 (z)
where one can easily verify that A22 (z) is invertible. Define the Schur com−1
plement of A22 (z) as A+
11 (z) := A11 (z) − A12 (z)A22 A21 (z). From a little
algebra, it can be easily verified that A+
11 (z) = Ã(z). Since A22 (z) is invert+
ible, A(z) is invertible if and only if A11 is invertible. Therefore, it follows
that σ(A) = σ(Ã) ⊂ {z : |z| > 1}. Define
s(A) := {z ∈ C : A − z idHp is invertible}
and
ρ(A) := sup{|z| : z ∈ σ(A)},
the spectrum and spectral radius of the bounded linear operator A. From
the construction of the operator pencil A(z) and Assumption (S), it is clear
that
ρ(A) := sup{|z| : z ∈ s(A)} < 1.
(4.6)
In view of Gelfand’s formula (Conway, 1994, Proposition VII.3.8), (4.6) implies that there exists j such that
kAj kB(Hp ) < aj ,
a ∈ (0, 1).
(4.7)
From a well known result on the inverse of Banach-valued functions , A(z)−1
is given as
A(z)−1 = idH +Az + A2 z 2 + · · ·
which is convergent in D1+η , and indeed holomorphic on D1+η , for some η > 0
because of (4.7). This shows that the AR(1) law of motion in (4.5) may be
represented as
Ft =
∞
X
Ak Et−k ,
k=0
16
t ∈ Z.
Let Π : Hp → H be the coordinate projection given as Π(x1 , . . . , xp ) = x1
and let Π∗ : H → Hp be its adjoint. Then one can easily verify that
f˜t =
∞
X
ΠAk Π∗ ε̃t−k ,
t ∈ Z.
k=0
−1
Simple algebra yields ΠA(z)−1 Π∗ = A+
= Ã(z)−1 , and it is clear that
11 (z)
ΠAk Π∗ is the kth coefficient in the Taylor series of Ã(z)−1 around z = 0.
Holomorphicity of A(z)−1 on D1+η is inherited from the holomorphicity of
Ã(z)−1 on D1+η , which is implied by the holomorphicity of A(z)−1 established
above and the relation Ã(z)−1 = ΠA(z)−1 Π∗ .
P
Remark 4.1. Norm-summability, ∞
k=0 kÑk kB(B 2 (λ)) < ∞, is a natural consequence of holomorphicity of A(z)−1 on D1+η for some η > 0. This shows
that (4.4) is a standard Bayes linear process.
Under Assumption (N), a process in B 2 (λ) satisfying the AR(p) law of
motion (4.1) is no longer stationary. Loosely speaking, it corresponds to an
I(d) process for d ≥ 1.
Proposition 4.2. Under Assumption (N)-(i, ii), the AR(p) law of motion
does not admit a stationary solution. Instead, given a suitable initialization
f−d+1 , . . . , f0 , it is satisfied by a sequence in L2B 2 (λ) whose dth difference is
given by
∞
M
d
Nk εt−k
∆ ft =
k=0
for all t ≥ 1. Nk is defined by
Nk =
1
k!
N (k) (0)
and N (z) := (z − 1)d A(z)−1 is holomorphic on D1+η for some η > 0.
If Assumption (N)-(iii) also holds, then d = 1 and
lim (z − 1)
z→1
A(z)−1 =
−1
idB 2 (λ) P A(1) (1)|ker A(1)
idB 2 (λ) P , (4.8)
which does not depend on the choice of the projection P on ran A(1).
17
Proof. For notational convenience, we work with Ã(z) and set H = L2 (λ).
Assumption (N)-(i,ii) implies that 1 is an isolated singularity of Ã(z). From
Lemma 8.1, we know that there exist finite dimensional projections Q1 , . . . , Qd
such that, for z in a punctured neighborhood of 1,
Ã(z) = (P1 + (1 − z)Q1 ) · · · (Pd + (1 − z)Qd )G(z),
where Pi = idH −Qi for i = 1, . . . , d, and G(z) is a holomorphic function that
is invertible at z = 1. Note that for i = 1, . . . , d
(Pi + (1 − z)Qi )−1 = (Pi + (1 − z)−1 Qi ).
This shows that in a punctured neighborhood of z = 1,
Ã(z)−1 = G(z)−1 (Pd + (1 − z)−1 Qd ) · · · (P1 + (1 − z)−1 Q1 ).
(4.9)
Therefore, Ã(z)−1 is meromorphic and the lowest order of the Laurent series
is −d. Therefore Ñ (z) := (1 − z)d Ã(z)−1 can be continued holomorphically
over z = 1. By applying the linear filter induced by Ñ (z) to (4.1), we have
for all t ≥ d + 1
∆ f˜t =
d
∞
X
Ñk ε̃t−k ,
k=0
Ñk =
1 (k)
Ñ (0)
k!
(4.10)
Since (∆d f˜t , t ≥ d + 1) is stationary, it may be extended for t ≥ 1 in a
natural way and f−d+1 , . . . , f0 may be chosen accordingly. This proves the
first statement.
Additionally under Assumption (N)-(iii), Lemma 8.2 implies that d = 1.
The only remaining thing is to verify the residue formula (4.8). From (4.9)
we know that
lim (1 − z)Ã(z)−1 = G(1)−1 (idH −P1 ).
z→1
Define
A := G(1)−1 |ran(idH −P1 ) ,
B := G(1)|ker Ã(1) .
From Lemma 8.3(a,b) we know that B = −(idH −P1 )Ã(1) (1)|ker Ã(1) and
ran B ⊂ ran(idH −P1 ). Since B as a map from ker Ã(1) to ran(idH −P1 )
18
is invertible under Assumption (N)-(iii), it is clear that AB = idker Ã(1) . This
shows that
Ñ−1 := lim(1 − z)Ã(z)−1 = A(idH −P1 ) = B−1 (idH −P1 ).
z→1
Under Lemma 8.1, we may choose P1 to be any projection on ran Ã(1), so
the expression for Ñ−1 just obtained cannot depend on the particular choice
of projection.
Under Assumption (N)-(i,ii), Proposition 4.2 says that ∆d f = (∆d ft , t ≥ 1)
is a Bayes linear process under a suitable initial condition. In this sense, we
say that the AR(p) law of motion (4.1) is I(d) for some d ≥ 1. The most
common case in practice may be I(1), and such an I(1) solution is guaranteed
under an extra condition, Assumption (N)-(iii).
Remark 4.2. When H = Rn , Johansen (1991) provided a necessary and sufficient condition on the autoregressive matrix polynomial and its first derivative at one for an AR(p) law of motion to be I(1). It is commonly called the
I(1) condition. Assumption (N)-(iii) plays the same role for an AR(p) law
of motion in B 2 (λ). Beare et al. (2017) provided an I(1) condition for an
AR(p) law of motion in an arbitrary complex Hilbert space, and Assumption
(N)-(iii) is a reformulation of it.
Remark 4.3. When p = 1, an I(1) solution can be guaranteed without
requiring compactness of the autoregressive operator. See Theorem 4.1 in
Beare et al. (2017).
Under Assumption (N)-(i,ii,iii), one natural consequence of Proposition
4.2 is the following Beveridge-Nelson decomposition: for some f0 , ν0 ∈ B 2 (λ),
ft = (f0
ν0 ) ⊕ N (1)ξt ⊕ νt ,
t≥0
(4.11)
Lt
L∞
where ξt =
s=1 εs , νt =
k=0 Nk εt−k and N (1) is a finite rank operator
which can be explicitly obtained from the residue formula (4.8). If ε = (εt , t ∈
Z) has a positive definite covariance operator, one can easily identify the
cointegrating and attractor spaces of B 2 (λ) from the decomposition (4.11).
Proposition 4.3. Assume that Cε0 is positive definite and (N)-(i, ii, iii) hold.
Then C(f, B 2 (λ)) = [ker A(1)]⊥ .
19
Proof. We only provide an informal argument. For a more detailed and
rigorous proof, refer to Beare et al. (2017, Proposition 3.1). From (4.11), for
any g ∈ B 2 (λ) we have
hft , giB 2 (λ) = hf0
ν0 , giB 2 (λ) + hN (1)ξt , giB 2 (λ) + hνt , giB 2 (λ) .
By employing the initial condition f0 = ν0 , we can make the first term
vanish. Moreover, the inner product process (hνt , giB 2 (λ) , t ≥ 0) is stationary
because (νt , t ≥ 0) is stationary. The variance of hN (1)ξt , giB 2 (λ) increases in
t. Thus, hN (1)ξt , giB 2 (λ) must vanish in order for hft , giB 2 (λ) to be stationary.
Given that ε has a positive definite covariance operator, for hN (1)ξt , giB 2 (λ)
to vanish we require g ∈ ker N (1)∗ . From the closed form solution for N (1)
given in (4.8) combined with (2.1) , it can be easily deduced that ker N (1)∗ =
[ker A(1)]⊥ .
Naturally, the attractor space A(f, B 2 (λ)) = C(f, B 2 (λ))⊥ = ker A(1), which
is finite dimensional under Assumption (N).
Example 4.1 (A numerical example).
Let λ be the uniform measure on [−3, 3] and consider the Bayes Hilbert
space associated with λ. Let g be the standard normal measure truncated
on [−3, 3]. We consider the following AR(1) law of motion
g) = Ψ(ft−1 g) ⊕ εt ,
∞
M
Ψ(·) =
λj h·, ej iB 2 (λ) ej ,
(ft
j=1
where λj = 21−j . Let (uj , j ∈ N) be the Fourier basis for L2 (λ), and let e1 be
the Cauchy probability density with location zero and scale 0.25. We obtain
an orthonormal basis (ej , j ∈ N) for B 2 (λ) by applying the Gram-Schmidt
process to the basis (e1 , clr−1 u2 , clr−1 u3 , . . .). If εt has a positive definite
covariance operator, this specification makes the attractor space equal to
A(f, B 2 (λ)) = {g ⊕ (a
e1 ) : a ∈ R}.
(4.12)
Figure 4.1 shows typical elements in the attractor space. An element g ⊕
(a e1 ) shows higher concentration around 0 when a > 0, or is bimodal
when a < 0. To obtain a simulated sequence, I generated a sequence of
independent standard Brownian bridges (Bt , t ≥ 1), and set
εt = clr−1 {0.3 · Pm (λ) B t },
20
where Pm (λ) denotes orthogonal projection on theR space of mth-order poly1
nomials in L2 (λ), and B t (u) = Bt ((u + 3)/6) − 6 0 Bt (r)dr for u ∈ [−3, 3].
A consequence of using the projection Pm to construct the innovation εt is
that the attractor space may not be exactly equal to that in (4.12), since the
covariance operator of εt is not positive definite. However, this is a convenient way to force the simulated probability densities to be smooth, and by
choosing large m the attractor space should approximate that in (4.12). I
set f1 = g and generated a sequence which is shown in Figure 4.2. We can
easily see that ft floats around the attractor space in the long run.
Figure 4.1: Typical elements in the attractor
g ⊕ e1
g
e1
Figure 4.2: Realization of (ft , t ≥ 1)
t=1
t = 10
t = 30
t = 100
t = 300
t = 500
t = 800
t = 1000
21
5
Statistical inference
In this section we provide a statistical procedure to estimate the attractor
space, which is assumed here to be finite dimensional. Chang et al. (2016)
provided statistical methods based on FPCA for a cointegrated densityvalued linear process with values in L2 (λ). Even though Beare (2017) commented that their time series cannot accommodate a nontrivial attractor
space, the statistical procedure they provided is useful since we can apply it
to the time series of clr-images f˜ = (f˜t , t ≥ 1) with values in L2 (λ). Using
their procedure, the attractor space of L2 (λ) can be estimated, and then it
is easy to obtain the attractor space of B 2 (λ) using the clr−1 transformation.
We therefore do not have to develop a new statistical technique adapted to
Bayes Hilbert spaces.
Assumption (T1).
(i) For some finite rankPÑ ∈ B(L2 (λ)) and a sequence (Ñk , k ≥ 0) in
B(L2 (λ)) such that ∞
k=0 kkÑk kL2 (λ) < ∞, we have
f˜t = f˜0 − ν̃0 + Ñ ξ˜t + ν̃t ,
where ν̃t =
P∞
k=0
t ≥ 0.
Ñk ε̃t−k .
(ii) (ε̃t , t ∈ Z) is an iid sequence with E ε̃t = 0, positive definite covariance
operator Cε̃0 , and Ekε̃t kpL2 (λ) < ∞ for some p ≥ 4.
In the above assumption, it is explicitly required that the attractor space
A(f˜, L2 (λ)) is finite dimensional. An example of such an I(1) process is
an AR(p) process satisfying Assumption (N), which we considered in the
previous section. Assume that dim(A(f˜, L2 (λ))) = r > 0 and let (vi , i ∈ N)
be the orthonormal basis of L2 (λ) such that
A(f˜, L2 (λ)) = span(v1 , . . . , vr ),
C(f˜, L2 (λ)) = span(vr+1 , vr+2 , . . .).
The probability density functions (ft , t = 1, . . . , T ) are not directly observed and must be estimated from the data. We assume that there are n
cross-sectional observations that are available to estimate ft at each time
period t. One possibility would be to obtain estimated densities (fˆ1 , . . . , fˆT )
22
from a standard nonparametric kernel method. However, since we would
ˆ
Rbe applying the procedure of Chang et al. (2016) to the clr-images log ft −
ˆ
log ft dλ, this naive approach could lead to unacceptable bias if the densities are close to zero at the boundary of their support. Possibly, a better
approach is to estimate the log-densities directly. Methods for doing so include the maximum penalized likelihood method of Silverman (1982), the
spline-based method of O’Sullivan (1988) and the local likelihood method of
Loader (1996). However the clr-images are estimated, we assume that they
satisfy the following high level condition.
Assumption (T2). The estimated clr-images (f˜tE , t = 1, . . . , T ) satisfy
(i) sup1≤t≤T kf˜tE − f˜t kL2 (λ) = Op (1),
(ii) T −1
PT
t=1
kf˜tE − f˜t kL2 (λ) →p 0.
The testing procedure is based on FPCA of the empirical covariance operator
!
!
T
T
T
X
X
X
T
−1
E
−1
E
E
−1
E
Ṽ = T
f˜t − T
f˜t ⊗ f˜t − T
f˜t .
t=1
t=1
t=1
Let (ζiT , viT ), i ∈ N, be the eigenpairs of the empirical covariance operator
Ṽ T , where ζiT is assumed to be decreasing as i gets larger. Let PA be the
orthogonal projection on ran Ñ and define PC = idL2 (λ) −PA . Ṽ T can be
written as
T
T
T
T
Ṽ T = ṼAA
+ ṼAC
+ ṼCA
+ ṼCC
,
T
T
T
T
where ṼAA
= PA Ṽ T PA , ṼAC
= PA Ṽ T PC , ṼCA
= PC Ṽ T PA , and ṼCC
=
PC Ṽ T PC . Under Assumptions (T1) and (T2), it can be shown that
T
kB(L2 (λ)) = Op (T −1 ).
kT −1 Ṽ T − T −1 ṼAA
T
That is, the difference between T −1 Ṽ T and T −1 ṼAA
is a norm-vanishing opT
erator. Noting that the eigenfunctions of ṼAA are in A(f˜, L2 (λ)), it follows
that the eigenfunctions associated with the r leading eigenvalues are a consistent estimator of a spanning set of A(f˜, L2 (λ)). Formally, it can be shown
(Chang et al., 2016, Proposition 3.2 and Section 4) that
kP̂A − PA kL2 (λ) = Op (T −1 ),
23
(5.1)
where P̂A is the projection operator based on the eigenfunctions. An alternative formulation of (5.1) is
span(v1T , . . . , vrT ) →p A(f˜, L2 (λ)).
(5.2)
Needless to say, C(f˜, L2 (λ)) can be estimated as the orthogonal complement,
[span(v1T , . . . , vrT )]⊥ . These results show that if r is known, then it is easy to
estimate the attractor space, A(f˜, L2 (λ)), since the r leading eigenfunctions
of Ṽ T asymptotically span A(f˜, L2 (λ)).
In practice, the dimension of A(f˜, L2 (λ)) is typically unknown. To determine the dimension, we may consider testing the null hypothesis
2
˜
H0 : dim A(f , L (λ)) = R,
(5.3)
against the alternative
H1 : dim A(f˜, L2 (λ)) ≤ R − 1.
(5.4)
For some positive integer Rmax ≥ 1, successive tests of the null hypothesis
for R = Rmax , Rmax − 1, . . . , 1 can determine the dimension of A(f˜, L2 (λ)).
For example, if the null is rejected in favor of the alternative
for R > r, but
2
˜
not rejected for R = r , we can conclude that dim(A f , L (λ)) = r. A
feasible test statistic and its limiting distribution under the null hypothesis
are given by Chang et al. (2016). Let
*
+
*
+ !0
T
T
X
X
ẑt =
v T , f˜E − T −1
f˜E
, . . . , v T , f˜E − T −1
f˜E
1
t
t
t=1
R
H
t
t
t=1
H
for t =P
1, . . . , T , and define ẐT = (ẑ1 , . . . , ẑT )0 . Further let Q̂TR = ẐT0 ẐT and
Σ̂TR = |k|≤` w` (k)Γ̂(k), where Γ̂(k) is the sample autocovariance of ẑt and
w` (k) is a bounded weight function1 . The proposed test statistic is
τRT = T −2 ζmin (Q̂TR , Σ̂TR )
where ζmin (Q̂TR , Σ̂TR ) denotes the smallest generalized eigenvalue of Q̂TR with
respect to Σ̂TR .
1
See Chang et al. (2016) for further details. The truncation parameter ` may be chosen
as in Andrews (1991). In the empirical applications reported in Section 6 of this paper,
w` (·) is chosen to be the quadratic spectral kernel.
24
Proposition 5.1 (Theorem 4.3 in Chang et al., 2016). If Assumptions (T1)
and (T2) are satisfied then, under the null hypothesis H0 in (5.3), we have
Z
0
T
W R (r)W R (r) dr, IR ,
(5.5)
τR →d ζmin
R
where I is the R × R identity matrix, W R (r) = WR (r) − WR (s)ds, and WR
is an R × 1 vector of independent standard Brownian motions. On the other
hand, under the alternative hypothesis H1 in (5.4),
τRT →p 0.
R
That is, τRT converges in law to the smallest eigenvalue of W R (r)W R (r)0 dr
under the null hypothesis, and vanishes under the alternative. The critical
values can be easily obtained from a large number of simulated sample paths
of W R . The critical values are presented in Table 5.1.2
Table 5.1: Critical values of τRT
R
1%
5%
10%
1
2
3
4
5
6
7
0.0248 0.0163 0.0123 0.0100 0.0084 0.0073 0.0065
0.0365 0.0215 0.0156 0.0122 0.0101 0.0086 0.0075
0.0459 0.0254 0.0177 0.0136 0.0111 0.0094 0.0081
Remark 5.1. The case when f˜t is observable at each t may be easily dealt
with. Test statistics can be calculated based on true observations f˜ = (f˜t , t =
1, . . . , T ). The limit distribution is slightly different from (5.5); it is given by
Z
0
ζmin
WR (r)WR (r) dr, IR .
See Chang et al. (2016) for more details.
6
Empirical application
6.1
Example 1: Cross-sectional densities of earnings
In this section, we revisit the application of FPCA to cross-sectional densities of individual weekly earnings undertaken by Chang et al. (2016). The
2
The reported critical values for R = 1, . . . , 5 are taken from Table 1 in Chang et al.
(2016). The values for R = 6 and 7 are newly calculated for reference in the next section.
25
cross-sectional observations that are used to estimate densities are obtained
at monthly frequency from the Current Population Survey (CPS) database,
running from January 1990 to March 2017 (327 months in total). Since individual earnings in each month are reported in current dollars, I adjusted them
all to January 1990 prices.3 I excluded earnings below the 2.5th percentile
and above the 97.5th percentile since there are many near-zero4 earnings and
top-coded earnings in the raw dataset.5 Weekly earnings are censored from
above, with the threshold for censoring changing partway through the sample: the top-coded nominal earning is $1923 before January 1998, and $2885
afterward. This difference introduces significant heterogeneity; for example,
the support of the earnings distribution changes greatly after January 1998
since the number of observations contained in (1923, 2885) is always zero
before this month, but nonzero afterward. However, after dropping the top
2.5% of observations, the range of earnings is quite stable over time. Moreover, after dropping the bottom 2.5% of observations, all near-zero earnings
are excluded and the smallest earnings become reasonably sized. After applying this truncation, the number of observations for each month ranges
from 11760 to 15489. The last thing we need to notice is that the CPS individual earnings data are collected from a monthly survey of individuals, with
each individual assigned their own design weight. Therefore, any estimates
constructed from the dataset should take design weights into account. To
estimate log-densities, I used the local likelihood method of Loader (2006)
which can easily accommodate design weights. The procedure is described in
detail in the Appendix. Figure 6.2 shows the time series of density estimates
and their clr-images.
To investigate the dimension of the attractor space, I calculated the test
statistics described in the previous section. Figure 6.1 displays the test statistics for R = 1, . . . , 7 as well as a scree plot of the eigenvalues. From the scree
plot it is apparent that Rmax can be set to around 5. Referring back to Table
5.1, we see that R = 2 is rejected even at the 1% level, but R = 1 is not
rejected at the 5% level. We tentatively conclude that the dimension of the
2
attractor space is one. Viewed as a subspace of L (λ), the estimated attrac3
Monthly CPI data are obtained from the Federal Reserve Economic Data (FRED).
The dataset contains a lot of abnormal values of nominal earnings near zero. For
example, there are total 9188 observations corresponding to weekly earnings less than
$0.01 over the whole span.
5
This symmetrical trimming was considered in, for example, Autor et al. (2008) for
analysis of the CPS wage data.
4
26
tor space is the span of the leading eigenfunction v̂1 , displayed in Figure 6.1.
Alternatively, we may regard the attractor space to be the span of clr−1 (v̂1 ),
a subspace of B 2 (λ).
To describe the attractor space, I orthogonally projected the estimated
densities upon the estimated cointegrating space, obtaining the projected
series (fˆtC , t = 1, . . . , T ). I positively and negatively perturbed the sample
mean f¯TC of (fˆtC , t = 1, . . . , T ), called the stationary mean of f , in the direction associated with the leading eigenfunction clr−1 v̂1 . More specifically, the
positive perturbation is f¯TC ⊕ (ζ̂1 clr−1 (v̂1 )) and the negative perturbation
is f¯TC (ζ̂1 clr−1 (v̂1 )), where ζ̂1 is the eigenvalue associated with v̂1 . Figure
6.3 shows the results.
Figure 6.1: Test statistics — individual earnings
R
τRT
1
0.03638
2
0.01253
3
0.00908
4
0.00590
Scree plot of eigenvalues
6.2
5
0.00537
6
0.00506
7
0.00472
Leading eigenfunction
Example 2 : cross-sectional densities of wages
Now we consider cross-sectional densities of hourly wages. The cross-sectional
observations are obtained at monthly frequency from the CPS database, and
the span is the same as in the previous example. Wages are deflated using
CPI data and measured in January 1990 dollars. There are many near-zero
values and top-coded values in the raw data set, so I excluded earnings below
the 1.25th percentile and above the 98.75th percentile; unlike the earnings
27
Figure 6.2: Density estimates and clr-images - earnings
Time series of density estimates
Time series of clr-images
Figure 6.3: Description of the attractor - earnings
data set, the top-coded values are constant over time, so only a small truncation is required to get rid of abnormal or extreme values. The number
of observations for each month ranges from 7348 to 9584 after truncation.
Figure 6.5 shows the time series of density estimates and their clr-images.
Test statistics are calculated as before and reported in Figure 6.4 with
28
other graphs. From the scree plot, we may set Rmax to around 5. From
successive tests we conclude that the dimension of the attractor space is 1.6
Thus, the estimated attractor space, viewed as a subspace of L2 (λ), is the
span of the eigenfunction associated with the first leading eigenvalue. Viewed
as a subspace of B 2 (λ), it is the span of the clr−1 -image of this eigenfunction.
Figure 6.6 provides a visual impression of the attractor space, similar to
Figure 6.3 above. In this case I perturbed the stationary mean f¯TC in the
direction of clr−1 v̂1 by ±2ζ̂1 , so as to more clearly emphasize the difference
between the two perturbations.
7
Concluding remarks
In this paper we have investigated cointegrated linear processes with values
in a Bayes Hilbert space of densities. Autoregressive density-valued processes
were also studied and a version of the Granger-Johansen representation theorem is provided. We showed that the statistical methods developed by Chang
et al. (2016) can be used to estimate the attractor space associated with a
cointegrated linear process in a Bayes Hilbert space.
6
On the other hand, if we set Rmax equal to 6 (resp. 7) then our sequential testing
procedure indicates an attractor space dimension of 6 (resp. 7), which seems large.
Figure 6.4: Test statistics - wages
R
τRT
1
0.05174
2
0.01181
3
0.01081
4
0.01076
Scree plot of eigenvalues
5
0.00982
6
0.00977
7
0.00966
Leading eigenfunction
29
Figure 6.5: Density estimates and clr-images - wages
Time series of density estimates
Time series of clr-images
Figure 6.6: Description of the attractor - wages
30
References
Aitchison, J. (1982). The statistical analysis of compositional data. Journal
of the Royal Statistical Society. Series B (Methodological), 44(2):139–177.
Andrews, D. W. K. (1991). Heteroskedasticity and autocorrelation consistent
covariance matrix estimation. Econometrica, 59(3):817–858.
Autor, D. H., Katz, L. F., and Kearney, M. S. (2008). Trends in u.s. wage
inequality: Revising the revisionists. Review of Economics and Statistics,
90(2):300–323.
Beare, B. K. (2017). The Chang-Kim-Park model of cointegrated densityvalued time series cannot accommodate a stochastic trend. Econ Journal
Watch, 14(2):133 – 137.
Beare, B. K., Seo, J., and Seo, W.-K. (2017). Cointegrated linear processes
in hilbert space. Journal of Time Series Analysis, 38(6):1010–1027.
Beare, B. K. and Seo, W.-K. (2017). Representation of I(1) autoregressive
Hilbertian processes. ArXiv e-print, arXiv:1701.08149v1 [math.ST].
Bosq, D. (2000). Linear Processes in Function Spaces. Springer-Verlag New
York.
Bosq, D. (2007). General linear processes in Hilbert spaces and prediction.
Journal of Statistical Planning and Inference, 137(3):879 – 894. Special
Issue on Nonparametric Statistics and Related Topics: In honor of M.L.
Puri.
Chang, Y., Kim, C. S., and Park, J. Y. (2016). Nonstationarity in time series
of state densities. Journal of Econometrics, 192(1):152 – 167.
Conway, J. B. (1994). A Course in Functional Analysis. Springer.
Egozcue, J. J., Dı́az-Barrero, J. L., and Pawlowsky-Glahn, V. (2006). Hilbert
space of probability density functions based on Aitchison geometry. Acta
Mathematica Sinica, 22(4):1175–1182.
Granger, C. W. J. (1981). Some properties of time series data and their use
in econometric model specification. Journal of Econometrics, 16(1):121 –
130.
31
Horváth, L. and Kokoszka, P. (2012). Inference for Functional Data with
Applications. Springer-Verlag GmbH.
Howland, J. S. (1971). Simple poles of operator-valued functions. Journal of
Mathematical Analysis and Applications, 36(1):12 – 21.
Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models. Econometrica, 59(6):1551–
1580.
Johansen, S. (1996). Likelihood-Based Inference in Cointegrated Vector Autoregressive Models (Advanced Texts in Econometrics). Oxford University
Press.
Loader, C. R. (1996). Local likelihood density estimation. Annals of Statistics, 24(4):1602–1618.
Loader, C. R. (2006). Local Regression and Likelihood. Springer New York.
Markus, A. S. (2012). Introduction to the Spectral Theory of Polynomial
Operator Pencils (Translations of Mathematical Monographs). American
Mathematical Society.
O’Sullivan, F. (1988). Fast computation of fully automated log-density and
log-hazard estimators. SIAM Journal on Scientific and Statistical Computing, 9(2):363–379.
Petersen, A. and Müller, H.-G. (2016). Functional data analysis for density functions by transformation to a Hilbert space. Annals of Statistics,
44(1):183–218.
Silverman, B. W. (1982). On the estimation of a probability density function by the maximum penalized likelihood method. Annals of Statistics,
10(3):795–810.
van den Boogaart, K. G., Egozcue, J. J., and Pawlowsky-Glahn, V. (2014).
Bayes Hilbert spaces. Australian & New Zealand Journal of Statistics,
56(2):171–194.
32
8
8.1
Appendix
Useful lemmas
Lemma 8.1 (Theorem 1.4 in Howland (1971)). Let B(H) be the space of
bounded linear operators H → H for a separable complex Hilbert space H and
let A(z) = idH −C(z) where C(z) is an analytic family of compact operators.
If z0 ∈ σ(A) is an isolated element, there exist finite dimensional projections
Q1 , . . . , Qd such that
A(z) = (P1 + (z0 − z)Q1 ) · · · (Pd + (z0 − z)Qd )G(z),
where G(z) is analytic, G(z0 ) is invertible, and Pi = idH −Qi . We may
choose P1 to be any projection on ran A(z0 ).
Lemma 8.2. Under the conditions of Lemma 8.1,
(idH −P1 )A(1) (z0 )|ker A(z0 ) : ker A(z0 ) → ran(idH −P1 )
is invertible if and only if A(z)−1 has a simple pole at z = z0 .
Proof. From Corollary 3.5 in Howland (1971) A(z)−1 has a simple pole at
z = z0 if and only if (i) A(1) (z0 ) is injective on ker A(z0 ) and (ii) ran A(z0 ) ∩
A(1) (z0 ) ker A(z0 ) = {0}. Since A(z) = idH −C(z) and C(z) is a compact
family of operators, A(z) is an index-zero Fredholm family of operators. This
implies that dim(ker A(z0 )) = dim(ran(idH −P1 )). From this, one can easily
show that the invertibility of (idH −P1 )A(1) (z0 )|ker A(z0 ) is another equivalent
condition to (i) and (ii).
Lemma 8.3. Let d = 1 in Lemma 8.1. Then,
(a) G(z0 ) = A(z0 ) − (idH −P1 )A(1) (z0 ),
(b) ran G(z0 ) |ker A(z0 ) = ran(idH −P1 ),
Proof. We may easily deduce (b) from (a) and Lemma 8.2, so we only show
(a). We know that G(z) is holomorphic at z = z0 from Lemma 8.1, so
(1)
G(z) = G(z0 ) − G (z0 )(z0 − z) +
∞
X
G(j) (z0 )
j=2
33
j!
(z − z0 )j .
Therefore,
A(z) = [P1 + (idH −P1 )(z0 − z)][G(z0 ) − G(1) (z0 )(z0 − z) + H1 (z)],
(8.1)
where H1 (z) denotes the remainder of the Taylor series of G(z). Since A(z)
is holomorphic at z = z0 , we have
A(z) = A(z0 ) − A(1) (z0 )(z0 − z) + H2 (z),
(8.2)
where H2 (z) is the remainder of the Taylor series of A(z). Collecting terms
associated with the same powers from (8.1) and (8.2) we obtain
P1 G(z0 ) = A(z0 )
(8.3)
(idH −P1 )G(z0 ) = −A(1) (z0 ) + P1 G(1) (z0 ).
(8.4)
The left-hand side of the second equation is invariant under (idH −P1 ), implying that
(idH −P1 )G(z0 ) = −(idH −P1 )A(1) (z0 ).
(8.5)
Combining (8.3) and (8.5), we obtain (a).
8.2
Log-density estimation
In Section 6, the estimated log-density ĝt is obtained from the following
procedure. The reader is referred to Loader (1996, 2006) for more details.
Given
P survey responses X1 , . . . , Xn with design weights w1 , . . . , wn such
that i=1 wi = n, consider the weighted log-likelihood
Z
n
X
L(ft ) =
wi log(ft (Xi )) − n
ft (u)du − 1 .
i=1
Let K be the support of ft . Under some local smoothness assumptions, we
can consider a localized version of the log-likelihood and log ft (u) can be
locally approximated by a polynomial function, so follows.
Lp (ft )(x) =
n
X
i=1
wi W
Xi − x
h
Q(Xi − x; αt )
Z
−n
W
34
u−x
h
exp(Q(u − x; αt ))du (8.6)
where W is a suitable kernel function, h is a bandwidth which assumed to
be fixed, and Q(u; αt ) is polynomial in u with coefficients αt .
(1 − |u|3 )3 and Q(u; α) = α0,t + α1,t u. For fixed x ∈ K, let
I set W(u) = 70
81
(α̂0,t , α̂1,t ) be the maximizer of (8.6). Then the local likelihood log-density
estimate is given by
ĝt (x) = α̂0,t
The procedure is repeated for a fine grid of points, and then ĝt may be
obtained from an interpolation method described in (Loader, 2006, Chapter
12). Needless to say, the above estimation procedure depends on h, a fixed
−1/5
bandwidth parameter. Let bt := [qt (99) − qt (1)] nt
where qt (a) is the a-th
percentile of observations at time t and nt is the number of observations.
Within the range [0.75 bt , 1.25 bt ], bandwidth h is set to the minimizer of a
generalized BIC criterion.7
7
A generalized version of AIC as a diagnostic for the local likelihood method is given
in Loader (2006). A suitable generalization of BIC is obtained by changing the penalty
term in an obvious way.
35
| 10math.ST
|
arXiv:1108.4455v2 [math.AC] 8 Jul 2012
Local cohomology properties of direct summands
Luis Núñez-Betancourt
July 10, 2012
Abstract
In this article, we prove that if R → S is a homomorphism of
Noetherian rings that splits, then for every i ≥ 0 and ideal I ⊂ R,
i
AssR HIi (R) is finite when AssS HIS
(S) is finite. In addition, if S is a
Cohen-Macaulay ring that is finitely generated as an R-module, such
i
that all the Bass numbers of HIS
(S), as an S-module, are finite, then
i
all the Bass numbers of HI (R), as an R-module, are finite. Moreover, we show these results for a larger class a functors introduced by
Lyubeznik [6]. As a consequence, we exhibit a Gorenstein F -regular
UFD of positive characteristic that is not a direct summand, not even
a pure subring, of any regular ring.
1
Introduction
Throughout this article all rings are commutative Noetherian with unity.
Let R denote a ring. If M is an R-module and I ⊂ R is an ideal, we
denote the i-th local cohomology of M with support in I by HIi (M ). If I is
generated by the elements f1 , . . . , fℓ ∈ R, these cohomology groups can be
computed by the Čech complex,
0 → M → ⊕j Mfj → . . . → Mf1 ···fℓ → 0.
The structure of these modules has been widely studied by several authors. Among the results obtained, one encounters the following finiteness
properties for certain regular rings:
(1)
the set of associated primes of HIi (R) is finite;
(2)
the Bass numbers of HIi (R) are finite;
(3)
inj.dimHIi (R) ≤ dimSuppHIi (R).
1
Huneke and Sharp proved those properties for characteristic p > 0 [4].
Lyubeznik showed them for regular local rings of equal characteristic zero
and finitely generated regular algebras over a field of characteristic zero [6].
These properties have been proved for a larger family of functors introduced by Lyubeznik [6]. If Z ⊂ Spec(R) is a closed subset and M is an
R-module, we denote by HZi (R) the i-th local cohomology module of M
with support in Z. We notice that HZi (R) = HIi (R), where Z = V(I) =
{P ∈ Spec(R) : I ⊂ P }. For two closed subsets of Spec(R), Z1 ⊂ Z2 , there
is a long exact sequence of functors
. . . → HZi 1 → HZi 2 → HZi 1 /Z2 → . . .
(1)
We denote by T any functor of the form T = T1 ◦· · · ◦Tt , where every functor
Tj is either HZi for some closed subset Z of Spec(R) or the kernel, image
or cokernel of some morphism in the previous long exact sequence for some
closed subsets Z1 , Z2 of Spec(R).
Our aim in this manuscript is to prove the finiteness properties (1) and
(2) for direct summands. We need to make some observations before we
are able to state our theorems precisely. Let R → S be a homomorphism
of Noetherian rings. For an ideal I ⊂ R, we have two functors associated
i (−) : S-mod → S-mod, which
with it, HIi (−) : R-mod → R-mod and HIS
are naturally isomorphic when we restrict them to S-modules. Moreover,
for two ideals of R, I2 ⊂ I1 , the natural morphism HIi1 (−) → HIi2 (−) is the
same as the natural morphism HIi1 S (−) → HIi2 S (−) when we restrict the
functors to S-modules. Thus, their kernel, cokernel and image are naturally
isomorphic as S-modules. Hence, every functor T for R is a functor of the
same type for S when we restrict it to S-modules.
As per the previous discussion, for an S-module, M , we will make no
distinction in the notation or meaning of T(M ) whether it is induced by
ideals of R or their extensions to S and, therefore, by the corresponding
closed subsets of their respective spectra. Now, we are ready to state our
main results.
Theorem 1.1. Let R → S be a homomorphism of Noetherian rings that
splits. Suppose that AssS T(S) is finite for a functor T induced by extension
of ideals of R. Then, AssR T(R) is finite. In particular, AssR HIi (R) is finite
i (S) is finite.
for every ideal I ⊂ R, if AssS HIS
Theorem 1.2. Let R → S be a homomorphism of Noetherian rings that
splits. Suppose that S is a Cohen-Macaulay ring such that all the Bass
numbers of T(S), as an S-module, are finite for a functor T induced by
2
extension of ideals of R. If S is a finitely generated R-module, then all the
Bass numbers of T(R), as an R-module, are finite. In particular, for every
ideal I ⊂ R the Bass numbers of HIi (R) are finite, if the Bass numbers of
i (S) are finite.
HIS
The first theorem holds when S is a polynomial ring over a field and R
is the invariant ring of an action of a linearly reductive group over S. It also
holds when R ⊂ K[x1 , . . . , xn ] is an integrally closed ring that is finitely
generated as a K-algebra by monomials. This is because such a ring is a
direct summand of a possibly different polynomial ring (cf. Proposition 1
and Lemma 1 in [2]).
We would like to mention another case in which an inclusion splits. This
is when R → S is a module finite extension of rings containing a field
of characteristic zero such that S has finite projective dimension as an Rmodule. Moreover, such a splitting exists when Koh’s conjecture holds (cf.
[5, 9, 1]). Therefore, if Koh’s conjecture applies to R → S and T(S) has
finite associated primes or finite Bass numbers, so does T(R).
We point out that property (3) does not hold for direct summands
of regular rings, even in the finite extension case. A counterexample is
R = K[x3 , x2 y, xy 2 , y 3 ] ⊂ S = K[x, y], where S is the polynomial ring
in two variables with coefficients in a field K. The splitting of the inclusion is the map θ : S → R defined in the monomials by θ(xα y β ) = xα y β
if α + β ∈ 3Z and as zero otherwise. We have that the dimension of
2
Supp(H(x
3 ,x2 y,xy 2 ,y 3 ) (R)) is zero, but it is not an injective module, because
R is not a Gorenstein ring, since R/(x3 , y 3 )R has a two dimensional socle.
The manuscript is organized as follows. In section 2, we prove Theorem
1.1, and we show some consequences. In particular, we exhibit a Gorenstein
F -regular UFD of positive characteristic that is not a direct summand, not
even a pure subring, of any regular ring. In section 3, we give a proof for
Theorem 1.2.
2
Associated Primes
Lemma 2.1. Let R → S be an injective homomorphism of Noetherian rings,
and let M be an S-module. Then, AssR M ⊂ {Q ∩ R : Q ∈ AssS M }.
Proof. Let P ∈ AssR M and u ∈ M be such that AnnR u = P . We have that
(AnnS u) ∩ R = P . Let Q1 , . . . , Qt denote the minimal primes of AnnS u. We
obtain that
p
√
P = P = AnnS u ∩ R = (∩j Qj ) ∩ R = ∩j (Qj ∩ R),
3
so, there exists a Qj such that P = Qj ∩ R. Since Qj is a minimal prime for
AnnS u, we have that Qj ∈ AssS M and the result follows.
Definition 2.1. We say that a homomorphism of Noetherian rings R → S
is pure if M = M ⊗R R → M ⊗R S is injective for every R-module M. We
also say that R is a pure subring of S.
Proposition 2.1 (Cor. 6.6 in [3]). Suppose that R → S is a pure homeomorphism of Noetherian rings and that G is a complex of R-modules. Then,
the induced map j : H i (G) → H i (G ⊗R S) is injective.
Proposition 2.2. Let R → S be a pure homomorphism of Noetherian rings.
Suppose that AssS HIi (R) is finite for some ideal I ⊂ R and i ≥ 0. Then,
i (S) is finite.
AssS HIS
i (S) is injective by Proposition 2.1, Ass H i (R) ⊂
Proof. Since HIi (R) → HIS
R I
i
AssR HIS (S) and the result follows by Lemma 2.1.
Proof of Theorem 1.1. The splitting between R and S makes T(R) into a direct summand of T(S); in particular, T(R) ⊂ T(S). Therefore, AssR T(R) ⊂
AssR T(S) and the result follows by Lemma 2.1.
If R is a ring containing a field of characteristic p > 0, Theorem 1.1 gives
a method for showing that R is not a direct summand of a regular ring. We
used this method to prove that there exists a Gorenstein strongly F -regular
UFD of characteristic p > 0 that is not a direct summand of any regular
ring.
Theorem 2.1 (Thm. 5.4 in [8]). Let K be a field, and consider the hypersurface
K[r, s, t, u, v, w, x, y, z]
R=
.
(su2 x2 + sv 2 y 2 + tuxvy + rw2 z 2 )
Then, R is a unique factorization domain for which the local cohomology
3
module H(x,y,z)
(R) has infinitely many associated prime ideals. This is preserved if R is replaced by the localization at its homogeneous maximal ideal.
The hypersurface R has rational singularities if K has characteristic zero,
and it is F -regular if K has positive characteristic.
Corollary 2.1. Let R be as in the previous theorem taking K of positive
characteristic. Then, R is a Gorenstein F -regular UFD that is not a pure
subring of any regular ring. In particular, R is not direct summand of any
regular ring.
4
3
Proof. Since H(x,y,z)
(R) has infinitely many associated prime ideals, it cannot be a direct summand or pure subring of a regular ring by Theorem 1.1,
Proposition 2.1 and finiteness properties of regular rings of positive characteristic (cf. [7]).
Theorem 2.2 (Thm. 1 in [10]). Assume that S = K[x1 , . . . , xn ] is a polynomial ring in n variables over a field K of characteristic
p > 0. SupP
pose that I = (fP
,
.
.
.
,
f
)
is
an
ideal
of
S
such
that
deg
f
1
s
i < n. Then
i
i
dim S/Q ≥ n − i deg fi for all Q ∈ AssS HI (S).
Corollary 2.2. Let S = K[x1 , . . . , xn ] be a polynomial ring in n variables
over a field K of characteristic p > 0. Let R → S be a homomorphism of
NoetherianPrings that splits. Suppose that I = (f1 , . . . , fs ) is an ideal of R
such that i deg(fi )P
< dimR. If S is a finitely generated R-module, then
dim R/P ≥ dim R − i deg fi for all P ∈ AssR HIi (R).
Proof. Since HIi (−) commutes with direct sum of R-modules, we have that
a splitting of R ֒→ S over R induces an splitting of HIi (R) ֒→ HIi (S) over
R. Then, by Lemma 2.1, for any P ∈ AssR HIi (R) ⊂ AssR HIi (S) there exists
Q ∈PAssR HIi (S) such that P = Q ∩ R and then dimR/P = dimS/Q >
n − i deg fi , and the result follows.
3
Bass Numbers
Lemma 3.1. Let (R, m, K) be a local ring and M be an R-module. Then,
the following are equivalent:
a) dimK (ExtjR (K, M )) is finite for all j ≥ 0;
b) length(ExtjR (N, M )) is finite for every finite length module N for all
j ≥ 0;
c) there exists one module N of finite length such that length(ExtjR (N, M ))
is finite for all j ≥ 0.
Proof. a) ⇒ b): Our proof will be by induction on h = length(N ). If h = 1,
then N = K, and the proof follows from our assumption. We will assume
that the statement is true for h and prove it when length(N ) = h + 1. In
this case, there is a short exact sequence 0 → K → N → N ′ → 0, where N ′
has length h. From the induced long exact sequence
j−1
. . . → ExtR
(N ′ , M ) → ExtjR (K, M ) → ExtjR (N, M ) → . . . ,
5
we see that length(ExtiR (N, M )) is finite for all i ≥ 0.
b) ⇒ c): Clear.
c) ⇒ a): We will prove the contrapositive. Let j be the minimum nonnegative integer such that dimK (ExtjR (K, M )) is infinite. We claim that
length(ExtiR (N, M )) < ∞ for i < j and length(ExtjR (N, M )) = ∞ for any
module N of finite length. Our proof will be by induction on h = length(N ).
If h = 1, then N = K and it follows from our choice of j. We will assume
that this is true for h and prove it when length(N ) = h + 1. We have a short
exact sequence 0 → K → N → N ′ → 0, where N ′ has length h. From the
induced long exact sequence
j−1
. . . → ExtR
(N ′ , M ) → ExtjR (K, M ) → ExtjR (N, M ) → . . . ,
we have that length(ExtiR (N, M )) < ∞ for i < j and that the map
j−1
ExtjR (K, M )/Im(ExtR
(N ′ , M )) → ExtjR (N, M ) is injective. Therefore,
length(ExtjR (N, M )) = ∞.
Lemma 3.2. Let R → S be a pure homomorphism of Noetherian rings.
Assume that S is a Cohen-Macaulay ring. If S is finitely generated as an
R-module, then R is a Cohen-Macaulay ring.
Proof. Let P ⊂ R be a prime ideal. Let x = x1 , . . . , xd denote a system of parameters of RP , where d = dim(RP ). It suffices to show that
Hi (K(x; RP )) = 0 for i 6= 0, where K is the Koszul complex with respect
to x. We notice that the natural inclusion RP → SP is a pure homeomorphism of rings. This induces an injective morphism of R-modules
Hi (K(x; RP )) → Hi (K(x; SP )) by Proposition 2.1. Thus, it is enough to
show that Hi (K(x; SP )) = 0 for i 6= 0. Since SP is a module finite extension
of RP , we have that every maximal ideal Q ⊂ SP contracts to P RP and x is
a system of parameters for SQ . Then, Hi (K(x; SQ )) = 0 for i 6= 0 and every
maximal ideal Q ⊂ SP . Hence, Hi (K(x; SP )) = 0 for i 6= 0 and the result
follows.
Proposition 3.1. Let R → S be a homomorphism of Noetherian rings that
splits. Assume that S is a Cohen-Macaulay ring and S is finitely generated
as an R-module. Let N be an R-module and M be an S-module. Let N → M
be a morphism of R-modules that splits. If all the Bass numbers of M , as
an S-module, are finite, then all the Bass numbers of N , as an R-module,
are finite.
6
Proof. Since N ֒→ M splits, we have that ExtiRP (RP /P RP , NP ) is a direct
summand of ExtiRP (RP /P RP , MP ), so, we may assume that N = M .
Let P be a fixed prime ideal of R and let KP denote RP /P RP . Since
we want to show that dimKP (ExtiRP (KP , MP )) is finite, we may assume
without loss of generality that R is local and P is its maximal ideal. Let
x = x1 , . . . , xn be a system of parameters for R. Since R is Cohen-Macaulay
by Lemma 3.2, we have that the Koszul complex, KR (x), is a free resolution for R/I, where I = (x1 , . . . , xn ). We also have that for every maximal ideal Q ⊂ S lying over P , x is a system of parameters of SQ because dimR = dimSQ and SQ /ISQ is a zero dimensional ring. From the
Cohen-Macaulayness of S and the previous fact, we have that the Koszul
complex KS (x) is a free resolution for S/IS. Therefore, ExtiR (R/I, M ) =
H i (HomR (KR (x), M )) = H i (HomS (KS (x), M )) = ExtiS (S/IS, M ). Since
ExtiS (S/IS, M ) = ⊕Q ExtiSQ (SQ /ISQ , MQ ) has finite length as an S-module
by Lemma 3.1, we have that ExtiR (R/I, M ) has finite length as an R-module
because S is finitely generated. Then, we have that dimKP (ExtiR (KP , M ))
is finite by Lemma 3.1.
Proof of Theorem 1.2. The splitting between R and S induces a splitting
between T(R) ֒→ T(S). The rest follows from Proposition 3.1.
Acknowledgments
I would like to thank my advisor Mel Hochster for his valuable comments
and suggestions. I also wish to thank Juan Felipe Perez-Vallejo for carefully
reading this manuscript. I am grateful to the referee for her or his comments.
Thanks are also due to the National Council of Science and Technology of
Mexico by its support through grant 210916.
References
[1] R. Flórez, J.D. Vélez, Failure of splitting from module-finite extension rings. Beitrage Algebra Geom. 41 (2000), no. 2, 345-357.
[2] M. Hochster, Rings of invariants of tori, Cohen-Macaulay rings
generated by monomials, and polytopes, Ann. of Math. (2) 96
(1972), 318337.
7
[3] M. Hochster, J.L. Roberts, Rings of invariants of reductive groups
acting on regular rings are Cohen-Macaulay, Advances in Mathematics 13, 115–175 (1974).
[4] C. Huneke, R. Sharp, Bass numbers of local cohomology modules,
Transactions of the American Mathematical Society, Vol. 339, No.
2, October 1993), pp. 765–779.
[5] J. H. Koh, The direct summand conjecture and behavior of codimension in graded extensions, Thesis, University of Michigan,
1983.
[6] G. Lyubeznik, Finiteness properties of local cohomology modules
(an application of D-modules to commutative algebra), Invent.
Math. 113 (1993), no. 1, 41–55.
[7] G. Lyubeznik, F -modules: applications to local cohomology and
D-modules in characteristic p > 0, J. Reine Angew. Math. 491
(1997), 65–130.
[8] A. Singh, I. Swanson, Associated primes of local cohomology modules and of Frobenius powers, Int. Math. Res. Not. 2004, no. 33,
1703–1733.
[9] J.D. Vélez, Splitting results in module-finite extension rings and
Koh’s conjecture, J. Algebra 172 (1995), no. 2, 454-469.
[10] Y. Zhang, A property of local cohomology modules over polynomial
rings, Proc. Amer. Math. Soc. 139 (2011), 125–128.
Department of Mathematics, University of Michigan, Ann Arbor, MI 48109–1043, USA.
Email address: [email protected]
8
| 0math.AC
|
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Sang-Woo Lee Yu-Jung Heo Byoung-Tak Zhang
arXiv:1802.03881v1 [cs.CV] 12 Feb 2018
Abstract
Goal-oriented dialogue has been paid attention
for its numerous applications in artificial intelligence. To solve this task, deep learning and reinforcement learning have recently been applied.
However, these approaches struggle to find a competent recurrent neural questioner, owing to the
complexity of learning a series of sentences. Motivated by theory of mind, we propose “Answerer in
Questioner’s Mind” (AQM), a novel algorithm for
goal-oriented dialogue. With AQM, a questioner
asks and infers based on an approximated probabilistic model of the answerer. The questioner
figures out the answerer’s intent via selecting a
plausible question by explicitly calculating the
information gain of the candidate intentions and
possible answers to each question. We test our
framework on two goal-oriented visual dialogue
tasks: “MNIST Counting Dialog” and “GuessWhat?!.” In our experiments, AQM outperforms
comparative algorithms and makes human-like dialogue. We further use AQM as a tool for analyzing the mechanism of deep reinforcement learning
approach and discuss the future direction of practical goal-oriented neural dialogue systems.
1. Introduction
Goal-oriented dialogue is a classical artificial intelligence
problem including digital personal assistants, order-byphone tools, and online customer service centers. Goaloriented dialogue occurs when a questioner asks an actionoriented question and an answerer responds with the intent
of letting the questioner know a correct action to take. Significant research on goal-oriented dialogue has tackled this
problem (Lemon et al., 2006; Williams & Young, 2007),
though a good solution has not yet been provided (Bordes
& Weston, 2017).
Motivated by the achievement of neural chit-chat dialogue
School of Computer Science and Engineering, Seoul National University. Correspondence to: Byoung-Tak Zhang
<[email protected]>.
Copyright 2018 by the authors.
Figure 1. Illustration of an AQM algorithm for goal-oriented visual
dialogue. AQM makes a decision tree on the image for asking
efficient questions.
research (Vinyals & Le, 2015), recent studies on goaloriented dialogues have utilized deep learning, using massive data to train their neural networks in for self-play
environments. Many researchers attempted to solve goaloriented dialogue tasks by using deep supervised learning
(deep SL) approach (Wen et al., 2016) based on seq2seq
models (Cho et al., 2014) or deep reinforcement learning
(deep RL) approach utilizing rewards obtained from the result of the dialogue (Zhao & Eskenazi, 2016; Li et al., 2017).
However, these methods struggle to find a correct RNN
model that uses back-propagation, owing to the complexity
of learning a series of sentences. These algorithms tend to
generate redundant sentences, making generated dialogues
inefficient (Kim et al., 2017; Das et al., 2017b).
Our idea to deal with goal-oriented dialogue is to introduce
opponent modeling. In this approach, an agent considers
what the opponent will respond by using an explicit approximated model of the opponent. Our study is motivated by
theory of mind, the ability to attribute mental states to others and to understand how our mental states are different
(Premack & Woodruff, 1978). If one wishes to efficiently
convey information to an opponent, it is best to converse in
a way that maximizes the opponent’s understanding (Bruner,
1981). For our method, we consider the mind to be beyond
a part of mental states (e.g., belief, intent, knowledge). The
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
mind is the probabilistic distribution of the opponent model
itself.
We propose an “Answerer in Questioner’s Mind’’ (AQM)
algorithm to allow an agent to ask appropriate consecutive questions during goal-oriented dialogue (Figure 1).
AQM’s questioner explicitly possesses an approximated
model of the answerer. The questioner utilizes the approximated model to calculate the information gain of the candidate answerer’s intentions and answers for each question. In
the experiment, AQM’s question generator extracts proper
questions from training data, not really generates.
We test AQM mainly on goal-oriented visual dialogue tasks.
There have been several kinds of visual-language tasks including image captioning (Vinyals et al., 2015) and visual
question answering (VQA) (Antol et al., 2015; Mao et al.,
2016), and recent research goes further to propose multiturn visual dialogue tasks (Kim et al., 2017). Our main
experiment is conducted on “GuessWhat?!”, a cooperative
two-player guessing game on the image (Figure 2). AQM
achieves an accuracy of 63.63% in 3 turns and 78.72% in 10
turns, outperforming deep SL (46.8% in 4.1 turns) (de Vries
et al., 2017) and deep RL (52.3% in 5 turns) (Strub et al.,
2017) algorithms. Though we demonstrate the performance
of our models in visual dialogue tasks, our approach can
be directly applied to general goal-oriented dialogue where
there is a non-visual context.
In the discussion section, aside from the experimental success of AQM, we leverage AQM as a tool for analyzing
deep RL approach on goal-oriented dialogue tasks from the
perspective of theory of mind. According to our argument,
training two agents to make plausible dialogues via rewards
during self-play is not adaptable to a service scenario. To enable an agent to converse with a human, the opponent agent
in self-play should model a human as much as possible. We
prove that AQM and RL have a similar objective function,
and that the objective function of RL can be replaced by a
direct objective function for each question, instead of reward
assigned at the end of dialogue. Through these series of arguments, we discuss future directions for building practical
goal-oriented dialogue systems.
2. Previous Works
2.1. Theory of Mind Approach
In this section, we introduce a series of studies considering
opponent’s uncertainty or the opponent model itself explicitly. We refer to these methods as theory of mind approach.
Opponent Modeling Studies on opponent modeling have
treated simple games with a multi-agent environment where
an agent competed with the other (Hernandez-Leal &
Kaisers, 2017; Foerster et al., 2017). In the study of Fo-
Figure 2. An example of the GuessWhat?! game. A green mask
highlights the correct object.
erster et al., the agent has the model of the opponent and
updates it assuming the opponent will be updated by gradient descent with RL. They argued modeling opponent could
be applied to track the non-stationary behavior of an opponent agent. Their model outperformed classical RL methods
in simple games, such as tic-tac-toe and rock-paper-scissors.
On the other hand, AQM applied opponent modeling to a
cooperative multi-agent setting. We believe that opponent
modeling can also be applied to dialogue systems in which
agents are partially cooperative and partially competitive.
Obverter A recent study applied the obverter technique
(Batali, 1998), motivated by theory of mind, to an imagedescription-match classification task to study language emergence (Choi et al., 2018). In the experiments of Choi et al.,
one agent transmitted one sentence for describing artificial
images to the opponent agent. In their study, the obverter
technique can be understood as that an agent play both questioner and answerer, maximizing the consistency between
visual and language modules. Their experimental results
showed that their obverter technique generated a word corresponding to a specific object (e.g. ‘bbbbbbb{b,d}’ for a blue
box). They argued their method could be an alternative to
RL-based language emergence systems. Compared to their
model, however, AQM uses real images, creates multi-turn
dialogue, and can be used for general goal-oriented dialogue
tasks.
Information Gain AQM’s question-generator optimizes
information gain using an approximated opponent model.
However, the concept of utilizing information gain in a dialogue task is not new for a classical rule-based approach.
Polifroni and Walker used information gain to build a dialogue system for restaurant recommendations (Polifroni &
Walker, 2006). They made a decision tree using information gain and asked a series of informative questions about
restaurant preferences. Rothe et al. applied a similar method
to generate questions on a simple Battleship game experiment (Rothe et al., 2017). It is noticeable that they used
pre-defined logic to generate questions with information
gain criteria to make novel (i.e., not shown in the training
dataset) and human-like questions. Unlike these previous
studies, AQM makes a new decision tree for every new context (see the decision tree in Figure 1). AQM also considers
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
uncertainty by using a multi-modal deep learning module,
and does not require hand-made or domain-specific rules.
Theory of AI’s Mind Chandrasekaran et al. argued that
human-machine collaboration could be improved when humans understand the properties of behavior (e.g., strengths,
weakness, or quirks) of the opponent machine (Chandrasekaran et al., 2017). They referred to this kind of human
understanding as “theory of AI’s mind.’’ In their human
experiment on visual question answering, a team consisting
of a human-questioner and a machine-answerer performed
better when the human knew whether the opponent could
answer correctly.
2.2. Visually Grounded Dialogue Task
GuessWhat?! GuessWhat?! is a cooperative two-player
guessing game proposed by De Vries et al. (Figure 2)
(de Vries et al., 2017). GuessWhat?! has received attention in the field of deep learning and artificial intelligence
as a testbed for research on the interplay of computer vision and dialogue systems. The goal is to locate a hidden
object in a rich image scene by asking a sequence of questions. One participant, called “Oracle,” or “Answerer” in
our paper, is randomly assigned an object in the image. The
other participant, “Questioner,” guesses the object assigned
to the answerer. The questioner asks a series of questions,
for which the answerer responds as “yes,” “no,” and “n/a.”
When the questioner decides to guess the correct object, a
list of candidate objects is then revealed. A win occurs when
the questioner picks the correct object. The GuessWhat?!
dataset contains 66,537 MSCOCO images (Lin et al., 2014),
155,280 games, and 831,889 question-answer pairs.
Visual Dialog In Visual Dialog (Das et al., 2017a), two
agents also communicate with questioning and answering
about the given MSCOCO image. Unlike GuessWhat?!, an
answer can be a sentence and there is no restriction for the
Visual Dialogue answerer. Das et al. used this dataset to
make a goal-oriented dialogue task, where the questioner
guesses a target image from 9,627 candidates in the test
dataset (Das et al., 2017b). The dataset includes a true caption of each image achieving percentile ranks around 90%.
In their self-play experiments, adding information via a dialogue improved the percentile ranks to around 93%, where
the questioner and answerer were trained with deep SL and
deep RL methods. This means that the models predict the
correct image to be more exact than 93% of the rest images
in the test dataset.
3. Answerer in Questioner’s Mind (AQM)
Preliminary In our experimental setting, two machine players, a questioner and an answerer, communicate via natural
dialogue. Specifically, there exists a class c, which is an
Algorithm 1 AQM’s Question-Generator
p̂(c) ∼ p̂0 (c)-model
p̃(at |c, qt , a1:t−1 , q1:t−1 ) ∼ p̃(a|c, q)-model
Q ← Q-sampler
for t = 1:T do
˜ At ; qt , a1:t−1 , q1:t−1 ] in Eq. 2
qt ← argmaxqt ∈Q I[C,
Get at from the answerer
Update p̂(c|a1:t , q1:t ) ∝ p̃(at |c; qt , a1:t−1 , q1:t−1 ) ·
p̂(c|a1:t−1 , q1:t−1 )
end for
answerer’s intention or a goal-action, the questioner should
perform. The answerer knows the class c, whereas the questioner does not. The goal of the dialogue for the questioner
is to find the correct class c by asking a series of questions
to the answerer. The answerer responds the correct answer
to given question.
We treat C, Qt , and At as random variables of class, t-th
question, and t-th answer, respectively. c, qt , and at becomes
their single instance. In a restaurant scenario example, qt
can be “Would you like to order?” or “What can I do for
you?” at can be “Two coffees, please.” or “What’s the password for Wi-Fi?” c can then be “Receive the order of two
hot Americanos.” or “Let the customer know the Wi-Fi
password.”
AQM’s claim In our problem setting, the answerer requires
one module, an answer-generator, and the questioner needs
two modules, a question-generator and a guesser. The objective function of three modules is as follows.
Answer-generator: argmaxat p(at |c, qt , a1:t−1 , q1:t−1 )
Guesser: argmaxc p(c|a1:t , q1:t )
Question-generator: argmaxqt I[C, At ; qt , a1:t−1 , q1:t−1 ]
I is information gain or mutual information of the class
C and the current answer At , where the previous history
(a1:t−1 , q1:t−1 ) and a current question qt are given (See
Equation 2). Note that maximizing information gain can be
understood as minimizing the conditional entropy of class
C, given a current answer At .
Answer-generator
For the answerer’s answergenerator module, a neural network is trained by
minimizing cross-entropy over the answer distribution
p(at |c, qt , a1:t−1 , q1:t−1 ), like the deep SL approach. On
the other hand, the questioner’s question-generator and
guesser module possess the approximated answer distribution p̃(at |c, qt , a1:t−1 , q1:t−1 ), or simply the likelihood p̃.
The likelihood p̃ can be obtained by learning training data
as does in the answer-generator, or by distilling from an
answer-generator module directly (Hinton et al., 2015). If
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Figure 3. Illustration of the AQM question-generator module.
at is a sentence, as in Visual Dialog (Das et al., 2017b), the
probability is extracted from the multiplication of the word
probability of RNN.
Guesser For the questioner’s guesser module, the posterior
of class c, p̂(c), is calculated based on the history (a1:t , q1:t ),
the likelihood p, and the prior of class c, p̂0 (c).
p̂(c|a1:t , q1:t ) ∝ p̂0 (c)
t
Y
p̃(aj |c, qj , a1:j−1 , q1:j−1 ) (1)
j
We use a term of likelihood as p̃, prior as p̂0 , and posterior as p̂ from the perspective that the questioner classifies class c. When the answerer’s answer distribution
p(at |c, qt , a1:t−1 , q1:t−1 ) is fixed, the questioner achieves
an ideal performance when the likelihood p̃ is the same as
p(at |c, qt , a1:t−1 , q1:t−1 ).
Question-generator AQM’s question-generator module selects qt∗ , which has a maximum information gain
˜ At ; qt , a1:t−1 , q1:t−1 ], or simply I.
˜ To calculate inforI[C,
˜
mation gain I, the question-generator module uses the likelihood p̃ and the posterior p̂.
& Fergus, 2014). AQM is explainable because it is clear
how a question is selected. Asking a series of questions
in AQM corresponds to constructing an efficient decision
tree classifier. We further introduce two properties of AQM
related to explainability.
AQM’s Property 1. The performance of AQM’s questioner is optimal, where the likelihood p̃ is equivalent to the
answering distribution of the opponent p. For the guesser
module, the performance of the guesser with the posterior
p̂ is optimal when p̃ is p. The performance of questiongenerator also increases as the p̃ becomes more similar to
the opponent, when the p̂ is fixed.
AQM’s Property 2. The contexts of history questioner
requires are the posterior p̂ and the history itself (a1:t , q1:t )
used as input for the likelihood p̃. In comparative deep learning methods, hidden neurons in RNN are expected to track
the context of history; though it has not been known yet
what kinds of information should be tracked. If the question
to be asked is independent from the previous questions, the
only context AQM should track is the posterior p̂. In this
case, the posterior p̂ in the yellow box of Figure 3 corresponds to the hidden vector of the RNN in the comparative
dialogue studies.
4. Experiments
˜ At ; qt , a1:t−1 , q1:t−1 ]
qt∗ = argmax I[C,
qt ∈Q
= argmax
qt ∈Q
XX
at
p̂(c|a1:t−1 , q1:t−1 )·
c
p̃(at |c, qt , a1:t−1 , q1:t−1 )
p̃0 (at |qt , a1:t−1 , q1:t−1 )
(2)
P
0
where p̃ (at |qt , a1:t−1 , q1:t−1 ) =
c p̂(c|a1:t−1 , q1:t−1 ) ·
p̃(at |c, qt , a1:t−1 , q1:t−1 ). Q is the set of the candidate questions.
p̃(at |c, qt , a1:t−1 , q1:t−1 ) ln
Algorithm 1 and Figure 3 explain the question-generator
module procedure. The question-generator requires the p̂0 model for the prior, the p̃(a|c, q)-model for the likelihood,
and the Q-sampler for the set of candidate questions.
Explainable AQM Explainable machine learning has recently received much attention (Ribeiro et al., 2016; Zeiler
4.1. MNIST Counting Dialog
To clearly explain the mechanism of AQM, we introduce
an MNIST Counting Dialog task, which is a toy goaloriented visual dialogue problem, illustrated in Figure 4.
This task utilizes the concept of the MNIST Dialog dataset
suggested by Seo et al. (Seo et al., 2017). Like MNIST
Dialog, each image in MNIST Counting Dialog contains
16 digits, each having four randomly assigned properties:
color = {red, blue, green, purple, brown}, bgcolor = {cyan,
yellow, white, silver, salmon}, number = {0, 1, · · · , 9}, and
style = {flat,stroke}. Unlike MNIST Dialog, MNIST Counting Dialog only asks counting questions. This type does not
require the information of previous questions and answers.
The goal of the MNIST Counting Dialog task is to inform
the questioner to pick the correct image among 10K candidate images via questioning and answering. In other words,
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Figure 4. Illustration of MNIST Counting Dialog, a simplified version of MNIST Dialog (Seo et al., 2017).
class c is an index of the true target image (1 ∼ 10,000).
In the example of Figure 4, asking about the number of 1
digits or 6 digits classifies a target image perfectly if the
recognition accuracy on each digit in the image is 100%.
Asking about the number of 0 digits does not help classify,
because all images have one zero. If recognition accuracy
is less than 100%, asking about the number of 1 digits is
better than asking about the number of 6 digits, because the
variance of the number of 1 digits is larger than that of the 6
digits.
For the MNIST Counting Dialog task, we do not use neural
networks for modeling the questioner or answerer. The questioner’s answering model is count-based for each property
value and is trained for 30K training data. To add randomness or uncertainty to this task, we set the ratio of recognition accuracy of each property to that of the answerer. If the
recognition accuracy on the digit decreases, the answering
accuracy on the image is decreased much. If the recognition
accuracy of color is 85%, the answering accuracy is 46.6%,
on average. 22 questions about red to stroke are used for the
candidate questions.
Figure 5 shows that AQM nearly always chooses the true target image from 10K candidates in six turns if the recognition
accuracy is 100%. However, AQM also chooses correctly
with a probability of 51%, 39%, and 31% (test accuracy)
in six turns, when the recognition accuracy (Acc in the legend) is 95%, 90%, 85%, respectively. “Random” denotes a
questioner with a random question-generator module and
the AQM’s guesser module.
4.2. GuessWhat?!
p̂0 (c)-model for the Prior The questioner does not know
the list of candidate objects while asking questions. This
makes the GuessWhat?! task difficult, although the number
of candidates is around 8. We use YOLO9000, a real-time
object detection algorithm, to estimate the set of candidate
objects (Redmon & Farhadi, 2016). The prior p̂0 (c) is set to
Figure 5. Test accuracy from the MNIST Counting Dialog task.
Acc is the average of the randomly assigned recognition accuracy.
1/N , where N is the number of extracted objects.
p̃(a|q, c)-model for the Likelihood We use the answerer
model from previous GuessWhat?! research (de Vries et al.,
2017). The inputs of the answer-generator module consist
of a VGG16 feature of a given context image, a VGG16
feature of the cropped object in the context image, spatial
and categorical information of the target object, and the
question qt at time step t. Our answer-generator module
assumes the answer distribution is independent from the
history (a1:t−1 , q1:t−1 ).
p̃(at |c, qt , a1:t−1 , q1:t−1 ) ∝ p̃00 (at |c, qt )
(3)
We use two strategy to make the questioner’s answer
distribution p̃(at |c, qt , a1:t−1 , q1:t−1 ) approximate the answerer’s answer distribution p(at |c, qt , a1:t−1 , q1:t−1 ). The
first is “indA,” in which p and p̃ is trained separately for the
training data. The second is “depA,” in which p̃ is trained
for the answer from p. The question and the image is also
sampled from the training data.
Q-sampler for the Candidate Question Set We compare two Q-samplers. The first is “randQ,’’ which samples
questions randomly from the training data. The second is
“countQ,’’ which causes every other question from the set Q
to be less dependent on the other. countQ checks the dependency of two questions with the following rule: the probability that two consecutive questions’
P answers are same
cannot exceed 95%. In other words, a p̃† (ai = a|qi , aj =
a, qj ) < 0.95, where p̃† (ai |qi , aj , qj ) is derived from the
count of a pair of answers for two questions in the training
data. p̃ made by indA is used for countQ. We set the size of
Q to 200.
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Table 1. Test accuracies from the GuessWhat?! task.
Model
Baseline
Deep SL (de Vries et al., 2017)
Deep RL (Strub et al., 2017)
Random-randQ-indA (5-q)
Random-randQ-depA (5-q)
AQM-randQ-indA (5-q)
AQM-countQ-indA (5-q)
AQM-countQ-depA (5-q)
AQM-countQ-depA (10-q)
Human (Strub et al., 2017)
Figure 6. Test accuracy from the GuessWhat?! task. Previous
works do not report the performance change with an increase
in the number of turns.
Experimental Results Figure 6 and Table 1 shows the
experimental results. Figure 7 illustrates the generated dialogue. Our best algorithm, AQM-countQ-depA, achieved
63.63% in three turns, outperforming deep SL and deep
RL algorithms. By allowing ten questions, the algorithms
achieved 78.72% and reached near-human performance. If
answerer’s answer distribution p is directly used for the questioner’s answer distribution p̃ (i.e., p̃ = p), AQM-countQ
achieved 63.76% in three turns and 81.96% in ten turns.
depA remarkably improved the score in self-play but did not
increased the quality of the generated dialogue significantly.
On the other hand, countQ did not boost the score much but
increased the quality of the generated dialogue. The comparative deep SL method used the question-generator with
hierarchical recurrent encoder-decoder (Serban et al., 2015),
achieving an accuracy of 46.8% in five turns (de Vries et al.,
2017). However, Random-randQ-depA achieved 46.36%
in five turns, which is a competitive result to the deep SL
model. The comparative deep RL method applied reinforcement learning on long short-term memory, achieving 52.3%
in about 4.1 turns (Strub et al., 2017).
5. Discussion
Many researchers have studied dialogue systems for cooperative games using RL, to increase the score in self-play
environments (Lazaridou et al., 2016; Kottur et al., 2017;
Mordatch & Abbeel, 2017). In the discussion section, we
discuss various issues in RL research on goal-oriented dialogue, mainly based on the case of goal-oriented visual
dialogue studies (de Vries et al., 2017; Das et al., 2017b).
Our primary argument is that considering the opponent’s
Accuracy
16.04
46.8
52.3
42.48 (± 0.84)
46.36 (± 0.91)
65.66 (± 0.55)
66.73 (± 0.76)
72.89 (± 0.70)
78.72 (± 0.54)
84.4
mind in implementing an agent is important.
Section 5.1 discusses three objectives of goal-oriented visual
dialogue research: score in self-play, service, and language
emergence. Section 5.2 introduces three properties of RL
algorithms in goal-oriented dialogue tasks using AQM as a
tool. Section 5.3 discusses the direction of future works to
make the questioner agent applicable for service.
5.1. Objective Function of Goal-Oriented Visual
Dialogue
Score In goal-oriented visual dialogue research, score is
used as one of the main measurements of dialogue efficiency.
However, a high score can be achieved via optimization over
p(c|a1:T , q1:T ), which is the objective function of RL. Particularly, the agent can achieve a high score if the probability
distribution of both the questioner and answerer is not bound
to a human’s distribution, even when human-like dialogue is
generated. For example, in GuessWhat?!, Han et al. showed
that pre-defined questions about location can provide an
accuracy of 94.34% in five turns (Han et al., 2017). Their
methods divided an image evenly into three parts using two
vertical or horizontal lines for three cases of answer {yes,
no, n/a}. A natural language-based protocol can also be created by using size, color, category, or other major properties
of the object.
Service One of the ultimate objectives of goal-oriented
dialogue research is to create an agent that can be used in a
real service (Bordes & Weston, 2017). However, successful
reports have been limited. Chattopadhyay et al. reported
the human-machine performance of the deep RL method
in a study on Visual Dialog (Das et al., 2017b; Chattopadhyay et al., 2017). In their method, questioner and answerer
were both fine-tuned by RL. Thus, the answerer’s answering distribution differed from the training data. This RL
method also used the objective function of deep SL as the
regularizer, conserving generated dialogue as human-like.
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Figure 7. Generated dialogues from our algorithm and the comparative algorithms. The tested games are sampled from the selected results
of previous papers.
Nevertheless, the RL algorithm deteriorated the score in
the game with a human, compared to deep SL. The authors
also assessed six measures of generated dialogue quality:
accuracy, consistency, image understanding, detail, question
understanding, and fluency. However, the human subjects
reported that the deep RL algorithm performed worse than
the deep SL algorithm for all measures, except “detail.”
Language Emergence Plenty of research has recently
been published on language emergence with RL in a multiagent environment. Some studied artificial (i.e., non-natural)
language (Evtimova et al., 2017; Lazaridou et al., 2018),
whereas others attempted to improve the quality of generated natural dialogue. One of the best progress found in the
latter study is the improvement on the quality of a series of
questions in a multi-turn VQA. When deep RL is applied,
the questioner generates fewer redundant questions than
deep SL (Strub et al., 2017; Das et al., 2017b). It can be
understood that the questioner in these methods are optimized by both p(qt |a1:t−1 , q1:t−1 ) and p(c|a1:T , q1:T ). The
questioners are improved more than the answerers, because
the answerer gets the objective function directly for each
answer (i.e., VQA), whereas the questioner does not.
5.2. Analyzing RL via AQM
RL’s Property 1. AQM and RL approaches share
same objective function. Two algorithms have the same
objective function with the assumption that qt only considers the performance of the current turn. The assumption is
used in the second line of the following equation.
argmax max ln p(c|a1:T , q1:T )
qt
qt−
≈ argmax ln p(c|a1:t , q1:t )
qt
p(at |c, qt , a1:t−1 , q1:t−1 )
= argmax E ln
p(at |qt , a1:t−1 , q1:t−1 )
qt
(4)
= argmax I[C, At ; qt , a1:t−1 , q1:t−1 ]
qt
In the third line, at ∼ p(at |c, qt , a1:t−1 , q1:t−1 ), c ∼
p(c|a1:t−1 , q1:t−1 ), and Bayes rule is used. The assumption in the second line can be alleviated via multi-step
AQM, which uses I[C, At:t+k ; qt:t+k , a1:t−1 , q1:t−1 ] as the
objective function of the question-generator module. In the
multi-step AQM, the optimal question cannot be selected
in a greedy way, unlike AQM. The multi-step AQM needs
to search in a tree structure; a Monte Carlo tree search
(Coulom, 2006) can be used to find a reasonable solution.
The RL and AQM question-generator are closely related,
as is the discriminative-generative pair of classifiers (Ng
& Jordan, 2002). AQM’s question-generator and guesser
module explicitly have a likelihood p̃, whereas the RL’s
modules do not have explicitly. The properties of AQM,
for example, including optimal conditions and sentence dynamics (AQM’s property 1 and 2), can be extended to RL.
The complexity of RL’s question-generator can be decomposed to tracking class posteriors p(c|a1:t−1 , q1:t−1 ) and
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
history (a1:t−1 , q1:t−1 ) for multi-turn question answering.
For human-like learning, the context for language generation
p(qt |a1:t−1 , q1:t−1 ) is also required; Q-sampler corresponds
to this context.
RL’s Property 2. Optimizing both questioner and answerer with rewards makes the agent’s performance with
human worse. This is true, even when the process improves
a score during self-play or uses tricks to maintain a humanlike language generation. The property of self-play in a
cooperative goal-oriented dialogue task is different from the
case of AlphaGo, which defeated a human Go champion
(Silver et al., 2017). For example, in GuessWhat?!, the reversed response of the answerer (e.g., “no” for “yes”) may
preserve the score in the self-play, but it makes the score in
the human-machine game near 0%.
According to AQM’s Property 1, for the play of a machine
questioner and a human answerer, the performance is optimal only if the approximated answerer’s distribution p̃ of
AQM’s questioner is the same as the human’s answering
distribution. According to RL’s Property 1, RL and AQM
shares the optimality condition about the distribution of
the opponent. Fine-tuning an answerer agent with a reward
makes the distribution of the agent different from a human’s.
Therefore, fine-tuning both agents decreases performance in
service situations. The experiment with a human, studied by
Chattopadhyay et al. explained in Section 5.1, empirically
demonstrates our claim.
RL’s Property 3. An alternative objective function
exists, which is directly applicable to each question. A reward can only be applied when one game is finished. If the
goal of training is to make language emergence itself or to
make an agent for service, two agents can communicate with
more than just question-answering for back-propagation, including sharing attention for the image (Kim et al., 2017).
Cross-entropy for the guesser of each round can be considered to replace reward. Information gain can also be used
not only for AQM but also for alternative objectives of backpropagation.
5.3. Future Works
RL with Theory of Mind RL methods can be enhanced in
a service scenario by considering the answering distribution
of human. It is advantageous for the machine questioner to
ask questions for which the human answer is predictable
(Chandrasekaran et al., 2017). In other words, a question
having a high VQA accuracy is preferred. The model uncertainty of the questioner can also be measured and utilized with recent studies on Bayesian neural networks and
uncertainty measures (Kendall & Gal, 2017). Because the
questioner has the initiative of dialogue, the questioner does
not need to necessarily learn the entire distribution of human conservation. The question, which the questioner uses
frequently in self-play, can be asked more to a human. Then,
the obtained question-answer pairs can be used for improving the answerer, like in active learning.
Opponent Modeling in the Test-Phase In the perspective
of making an agent for service, RL in self-play can be understood as opponent modeling in the training-phase; An
opponent model in self-play is just an approximated model
of the true opponent, a human. On the other hand, AQM
can be understood as opponent modeling in the test-phase,
especially when the approximated model of the agent is the
opponent model itself in self-play.
Opponent modeling during the test-phase can make it easier to handle a non-stationary environment (Foerster et al.,
2017) or a multi-domain problem. The AQM models can
also be easily trained because research on training an answerer model on VQA tasks has been more progressed than
training an RNN for the questioner. However, during testphase opponent modeling, it is difficult to back-propagate in
an end-to-end way. Thus, it requires additional techniques.
For example, in GuessWhat, if the Q-sampler in our experiment is replaced with seq2seq trained by a deep SL method,
AQM would generate a question by optimizing both sentence probability from the seq2seq model and information
gain, because the Q-sampler in AQM corresponds to regularizing with p(qt |a1:t−1 , q1:t−1 ) in the deep RL approach.
6. Concluding Remarks
Our contributions are two-fold. First, we proposed the AQM,
a practical and explainable goal-oriented dialogue system.
The algorithm is a scaled-up version of theory of mind methods for a realistic problem. During the self-play experiment,
AQM outperformed comparative deep SL and RL methods.
Second, we used AQM as a tool for analyzing deep RL research. We argued that the AQM and RL approach shared
the same objective function. From our discussion: 1) optimizing both agents through reward during self-play makes
the service performance worse; and 2) the reward in RL can
be replaced with an alternative objective function directly
applicable to each question.
The limitation of AQM is that the question-generator retrieves a question from the training data, not generating a
new question. One of the future directions would include
generating questions from the perspective discussed in this
paper. However, just generating novel questions can be produced with a simple rule-based program (Rothe et al., 2017)
or replacing Q-sampler to a seq2seq model. Our primary
interest is to make an agent that can talk with a human for
a practical service. In our future work, we will make a new
algorithm that has strengths of both theory of mind and deep
RL approach.
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Acknowledgements
The authors would like to thank Jin-Hwa Kim, Cheolho Han,
Wooyoung Kang, Jaehyun Jun, Christina Baek, Hanock
Kwak, Marco Baroni, and Jung-Woo Ha for helpful comments and editing.
References
Antol, Stanislaw, Agrawal, Aishwarya, Lu, Jiasen, Mitchell,
Margaret, Batra, Dhruv, Lawrence Zitnick, C, and Parikh,
Devi. Vqa: Visual question answering. In Proceedings of
the IEEE International Conference on Computer Vision,
pp. 2425–2433, 2015.
Batali, John. Computational simulations of the emergence
of grammar. Approaches to the evolution of language:
Social and cognitive bases, 405:426, 1998.
Bordes, Antoine and Weston, Jason. Learning end-to-end
goal-oriented dialog. In ICLR, 2017.
Bruner, Jerome S. Intention in the structure of action and
interaction. Advances in infancy research, 1981.
Chandrasekaran, Arjun, Yadav, Deshraj, Chattopadhyay,
Prithvijit, Prabhu, Viraj, and Parikh, Devi. It takes two
to tango: Towards theory of ai’s mind. arXiv preprint
arXiv:1704.00717, 2017.
Chattopadhyay, Prithvijit, Yadav, Deshraj, Prabhu, Viraj,
Chandrasekaran, Arjun, Das, Abhishek, Lee, Stefan, Batra, Dhruv, and Parikh, Devi. Evaluating visual conversational agents via cooperative human-ai games. arXiv
preprint arXiv:1708.05122, 2017.
Cho, Kyunghyun, van Merrienboer, Bart, Gulcehre, Caglar,
Bahdanau, Dzmitry, Bougares, Fethi, Schwenk, Holger,
and Bengio, Yoshua. Learning phrase representations
using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical
Methods in Natural Language Processing (EMNLP), pp.
1724–1734, 2014.
Choi, Edward, Lazaridou, Angeliki, and Freitas, Nando de.
multi-agent compositional communication learning from
raw visual input. In ICLR, 2018.
Coulom, Rémi. Efficient selectivity and backup operators
in monte-carlo tree search. In International conference
on computers and games, pp. 72–83. Springer, 2006.
Das, Abhishek, Kottur, Satwik, Gupta, Khushi, Singh, Avi,
Yadav, Deshraj, Moura, José MF, Parikh, Devi, and Batra,
Dhruv. Visual dialog. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017a.
Das, Abhishek, Kottur, Satwik, Moura, José MF, Lee, Stefan,
and Batra, Dhruv. Learning cooperative visual dialog
agents with deep reinforcement learning. arXiv preprint
arXiv:1703.06585, 2017b.
de Vries, Harm, Strub, Florian, Chandar, Sarath, Pietquin,
Olivier, Larochelle, Hugo, and Courville, Aaron. Guesswhat?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.
Evtimova, Katrina, Drozdov, Andrew, Kiela, Douwe, and
Cho, Kyunghyun. Emergent language in a multimodal, multi-step referential game. arXiv preprint
arXiv:1705.10369, 2017.
Foerster, Jakob N, Chen, Richard Y, Al-Shedivat, Maruan,
Whiteson, Shimon, Abbeel, Pieter, and Mordatch, Igor.
Learning with opponent-learning awareness. arXiv
preprint arXiv:1709.04326, 2017.
Han, Cheolho, Lee, Sang-Woo, Heo, Yujung, Kang, Wooyoung, Jun, Jaehyun, and Zhang, Byoung-Tak. Criteria for
human-compatible ai in two-player vision-language tasks.
In 2017 IJCAI Workshop on Linguistic and Cognitive
Approaches to Dialogue Agents, 2017.
Hernandez-Leal, Pablo and Kaisers, Michael. Learning
against sequential opponents in repeated stochastic games.
In The 3rd Multi-disciplinary Conference on Reinforcement Learning and Decision Making, Ann Arbor, 2017.
Hinton, Geoffrey, Vinyals, Oriol, and Dean, Jeff. Distilling the knowledge in a neural network. arXiv preprint
arXiv:1503.02531, 2015.
Kendall, Alex and Gal, Yarin. What Uncertainties Do We
Need in Bayesian Deep Learning for Computer Vision?
In Advances in neural information processing systems,
2017.
Kim, Jin-Hwa, Parikh, Devi, Batra, Dhruv, Zhang, ByoungTak, and Tian, Yuandong. Codraw: Visual dialog for
collaborative drawing. arXiv preprint arXiv:1712.05558,
2017.
Kottur, Satwik, Moura, José, Lee, Stefan, and Batra, Dhruv.
Natural language does not emerge ‘naturally’in multiagent dialog. In Proceedings of the 2017 Conference on
Empirical Methods in Natural Language Processing, pp.
2962–2967, 2017.
Lazaridou, Angeliki, Peysakhovich, Alexander, and Baroni,
Marco. Multi-agent cooperation and the emergence of
(natural) language. In ICLR, 2016.
Answerer in Questioner’s Mind for Goal-Oriented Visual Dialogue
Lazaridou, Angeliki, Hermann, Karl Moritz Hermann,
Tuyls, Karl, and Clark, Stephen. Emergence of linguistic communication from referential games with symbolic
and pixel input. In ICLR, 2018.
Rothe, Anselm, Lake, Brenden M, and Gureckis, Todd.
Question asking as program generation. In Advances in
Neural Information Processing Systems, pp. 1046–1055,
2017.
Lemon, Oliver, Georgila, Kallirroi, Henderson, James, and
Stuttle, Matthew. An isu dialogue system exhibiting reinforcement learning of dialogue policies: generic slotfilling in the talk in-car system. In Proceedings of the
Eleventh Conference of the European Chapter of the Association for Computational Linguistics: Posters & Demonstrations, pp. 119–122. Association for Computational
Linguistics, 2006.
Seo, Paul Hongsuck, Lehrmann, Andreas, Han, Bohyung,
and Sigal, Leonid. Visual reference resolution using
attention memory for visual dialog. In Advances in neural
information processing systems, 2017.
Li, Xuijun, Chen, Yun-Nung, Li, Lihong, and Gao, Jianfeng. End-to-end task-completion neural dialogue systems. arXiv preprint arXiv:1703.01008, 2017.
Silver, David, Schrittwieser, Julian, Simonyan, Karen,
Antonoglou, Ioannis, Huang, Aja, Guez, Arthur, Hubert,
Thomas, Baker, Lucas, Lai, Matthew, Bolton, Adrian,
et al. Mastering the game of go without human knowledge. Nature, 550(7676):354, 2017.
Lin, Tsung-Yi, Maire, Michael, Belongie, Serge, Hays,
James, Perona, Pietro, Ramanan, Deva, Dollár, Piotr, and
Zitnick, C Lawrence. Microsoft coco: Common objects
in context. In European conference on computer vision,
pp. 740–755. Springer, 2014.
Mao, Junhua, Huang, Jonathan, Toshev, Alexander, Camburu, Oana, Yuille, Alan L, and Murphy, Kevin. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 11–20, 2016.
Mordatch, Igor and Abbeel, Pieter. Emergence of grounded
compositional language in multi-agent populations. arXiv
preprint arXiv:1703.04908, 2017.
Serban, Iulian V, Sordoni, Alessandro, Bengio, Yoshua,
Courville, Aaron, and Pineau, Joelle. Hierarchical neural
network generative models for movie dialogues. arXiv
preprint arXiv:1507.04808, 2015.
Strub, Florian, de Vries, Harm, Mary, Jeremie, Piot, Bilal,
Courville, Aaron, and Pietquin, Olivier. End-to-end optimization of goal-driven and visually grounded dialogue
systems. arXiv preprint arXiv:1703.05423, 2017.
Vinyals, Oriol and Le, Quoc. A neural conversational model.
In ICML deep learning workshop, 2015.
Vinyals, Oriol, Toshev, Alexander, Bengio, Samy, and Erhan, Dumitru. Show and tell: A neural image caption
generator. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pp. 3156–3164,
2015.
Ng, Andrew Y and Jordan, Michael I. On discriminative
vs. generative classifiers: A comparison of logistic regression and naive bayes. In Advances in neural information
processing systems, pp. 841–848, 2002.
Wen, Tsung-Hsien, Vandyke, David, Mrksic, Nikola, Gasic, Milica, Rojas-Barahona, Lina M, Su, Pei-Hao, Ultes,
Stefan, and Young, Steve. A network-based end-to-end
trainable task-oriented dialogue system. arXiv preprint
arXiv:1604.04562, 2016.
Polifroni, Joseph and Walker, Marilyn. Learning database
content for spoken dialogue system design. In 5th International Conference on Language Resources and Evaluation (LREC), 2006.
Williams, Jason D and Young, Steve. Partially observable
markov decision processes for spoken dialog systems.
Computer Speech & Language, 21(2):393–422, 2007.
Premack, David and Woodruff, Guy. Does the chimpanzee
have a theory of mind? Behavioral and brain sciences, 1
(4):515–526, 1978.
Redmon, Joseph and Farhadi, Ali. Yolo9000: better, faster,
stronger. arXiv preprint arXiv:1612.08242, 2016.
Ribeiro, Marco Tulio, Singh, Sameer, and Guestrin, Carlos.
Why should i trust you?: Explaining the predictions of
any classifier. In Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and
Data Mining, pp. 1135–1144. ACM, 2016.
Zeiler, Matthew D and Fergus, Rob. Visualizing and understanding convolutional networks. In European conference
on computer vision, pp. 818–833. Springer, 2014.
Zhao, Tiancheng and Eskenazi, Maxine. Towards endto-end learning for dialog state tracking and management using deep reinforcement learning. arXiv preprint
arXiv:1606.02560, 2016.
| 1cs.CV
|
Finite Uniform Bisimulations for Linear Systems with Finite
Input Alphabets
arXiv:1510.04209v1 [math.OC] 14 Oct 2015
Donglei Fan and Danielle C. Tarraf∗
Abstract
We consider a class of systems over finite alphabets, namely discrete-time systems with linear dynamics and a finite input alphabet. We formulate a notion of finite uniform bisimulation, and motivate and
propose a notion of regular finite uniform bisimulation. We derive sufficient conditions for the existence
of finite uniform bisimulations, and propose and analyze algorithms to compute finite uniform bisimulations when the sufficient conditions are satisfied. We investigate the necessary conditions, and conclude
with a set of illustrative examples.
Index Terms— finite uniform bisimulations, systems over finite alphabets, abstractions.
1
Introduction
The past decade has witnessed much interest in finite state approximations of hybrid systems for both analysis
and control design. Multiple complementary approaches have been developed and used [8, 10, 16, 23, 24]
including qualitative models and l-complete approximations [11–13], approximating automata [2–4], exact
or approximate bisimulation and simulation abstractions [6, 18], and ρ/µ approximations [20–22].
In particular, the existence of “equivalent" finite state representations of systems with infinite memory
has been a problem of academic interest, and continues to be so. Typically, this equivalence is captured by
the existence of a bisimulation relation between the original system and the finite memory model, a concept
originally introduced in the context of concurrent processes [15]. For instance in [9], the authors show that
if a hybrid transition system is O-minimal, then it has a finite bisimulation quotient. In [1], by interpreting
the trajectories of linear systems as O-minimal language structures, the authors present instances of linear
systems which admit finite bisimulation quotients. In [26], the authors provide an algorithm for finding finite
bisimulations for piecewise affine systems, and show that it can be applied to linear systems in a dead-lock free
manner. In [19], the authors show that certain controllable linear systems admit finite bisimulations. In [6]
the authors propose an algorithm, based on polyhedral Lyapunov functions, to generate finite bisimulations
for switched linear systems with stable subsystems.
Concurrently, bisimulation has been explored in a more traditional systems setting. Particularly in
[25], the author discusses the connection between bisimulation relations and classical notions of state space
equivalence and equality of external behavior in systems theory. Specifically, he shows that a bisimulation
relation between two linear time-invariant (LTI) systems exists if and only if their transfer matrices are
identical.
In this paper, we revisit the question of existence of finite state equivalent models (a preliminary version
of this work appeared in [5]), focusing on a special class of systems over finite alphabets, namely systems
with linear dynamics and finite input alphabets. This class of systems provides potential models of simple
practical systems where the actuation involves multi-level switching. For this class of systems:
1. We formalize the notion of finite uniform bisimulation and discuss its connections to the existing
literature, and we propose a new notion of regular finite uniform bisimulation.
2. We derive sufficient conditions for the existence of finite uniform bisimulations.
∗ The authors are with the Department of Electrical & Computer Engineering Department at Johns Hopkins University,
Baltimore, MD, 21218 ([email protected], [email protected]).
1
3. We propose and analyze constructive algorithms for computing finite uniform bisimulations when the
sufficient conditions are satisfied.
4. We explore the question of necessity and derive a set of necessary conditions, highlighting the relevance
of the regularity property of finite uniform bisimulations.
5. We provide a set of examples, thereby illustrating how existence of finite uniform bisimulations can be
exploited to derive an “equivalent" deterministic finite state machine model for the system.
Paper Organization: We formulate the notion of finite uniform bisimulation and propose that of regular
finite uniform bisimulation in Section 2. We describe the class of systems of interest, pose our problem, and
place our work in the context of existing literature in Section 3. We state our main analytical results in
Section 4 and present the corresponding constructive algorithms in Section 5. We present a full derivation
of the analytical results and an analysis of our constructive algorithms in Section 6. We present a set of
illustrative examples in Section 7 and conclude with directions for future work in Section 8.
Notation: We briefly summarize our notation here and, for completeness, we provide a review of all
relevant concepts in the Appendix. We use N to denote the non-negative integers, Z+ to denote the positive
integers, Q to denote the rationals, R to denote the reals, and C to denote the complex numbers. For a set
A ⊂ Rn , we use |A| to denote its cardinality (which could be infinite), diam(A) to denote its diameter, Ac
to denote its complement, cl(A) to denote its closure, int(A) to denote its interior, and ∂A to denote its
boundary. A pair W and V of disjoint, nonempty, open sets in Rn is a disconnection of A if W ∩ A =
6 ∅,
V ∩ A 6= ∅ and A ⊂ W ∪ V, and we say A is not connected if there is a disconnection of A.
For
v ∈ Rn , we use kvk1 to denote its 1-norm. We use Br (v) to denote the open ball centered at v with
radius r. For a square matrix A, we use kAk1 to denote its 1-induced norm and ρ(A) to denote its spectral
radius. For sets S, R ⊂ Rn , a matrix H ∈ Rn×n and a vector v ∈ Rn , we use HS to denote the set
{z ∈ Rn |z = Hx, for some x ∈ S}, use v + S to denote the set {z ∈ Rn |z = v + x, for some x ∈ S}, and use
S + R to denote the set {x + r|x ∈ S, r ∈ R}. We use ∼ to denote an equivalence relation, x ∼ y to denote
that x is equivalent to y, x y to denote that x is not equivalent to y, and [x] to denote the equivalence
class of x. For completeness, we provide detailed definitions of all our notation in the Appendix.
2
2.1
Finite Uniform Bisimulations
Proposed Notions
We begin by defining the notion of finite uniform bisimulation, which is simply an equivalence relation that
satisfies certain desired properties:
Definition 1. Consider a discrete-time system
xt+1 = f (xt , ut )
(1)
where t ∈ N is the time index, xt ∈ Rn is the state, ut ∈ U is the input, f : Rn × U → Rn is given, and
input alphabet U represents the collection of possible values of the input. Given a set S ⊂ Rn , we say
an equivalence relation ∼⊂ S × S is a finite uniform bisimulation on S if the following two conditions are
satisfied:
(i) For any x, x0 ∈ S and any u ∈ U, if x ∼ x0 , then
f (x, u) ∼ f (x0 , u)
(2)
(ii) For x ∈ S with [x] = {y ∈ S|y ∼ x}, we have
1 < |{[x]|x ∈ S}| < ∞
(3)
Essentially (2) requires that each equivalence class transition into another equivalence class under any
input, and (3) requires that there be a finite number of equivalence classes while avoiding the trivial instance
of a single equivalence class.
We define a finite uniform bisimulation to be regular if the equivalence classes have a specific topological
structure:
2
Definition 2. Given a finite uniform bisimulation ∼ on S of system (1), we say ∼ is regular if for all x ∈ S,
[x] = {y ∈ S|y ∼ x} consists of open sets in Rn and possibly their boundary points.
We are interested in regular finite uniform bisimulations because we wish to avoid certain “pathological"
finite uniform bisimulations, as will become clear when we discuss the necessary conditions for the existence
of finite uniform bisimulations in Section 4.2.
2.2
Deterministic Finite State Bisimulation Models
Given a finite uniform bisimulation ∼ on S of system (1), it is straightforward to construct a deterministic
finite state machine (DFM) that is bisimilar to the original system when the latter is restricted to evolve on
S. Indeed:
Definition 3. Given a system (1) denoted by P and a finite uniform bisimulation ∼ on S of P , consider the
DFM P̂ defined by
qt+1 = f∼ (qt , ut ),
(4)
where t ∈ N is the time index, qt ∈ Q is the state, ut ∈ U is the input, Q = {[x]|x ∈ S} (essentially Q is
the finite quotient set of S under equivalence relation ∼), U is the input alphabet of system (1), and state
transition function f∼ : Q × U → Q is defined as
f∼ (q, u) = [f (x, u)], ∀ x ∈ q.
(5)
We say that P̂ is uniformly bisimilar to P .
Note that since ∼ is a finite uniform bisimulation, it follows from (2) that f∼ is well-defined.
3
Problem Setup and Formulation
3.1
Systems of Interest & Problem Statement
We first introduce the specific class of systems (1) that we will study in this paper. Consider a discrete-time
dynamical system described by
xt+1 = Axt + But ,
(6)
where t ∈ N is the time index, xt ∈ Rn is the state, ut ∈ U is the input, and A ∈ Rn×n and B ∈ Rn×m are
given. We enforce that the input ut can only take finitely many values in U ⊂ Rm (that is, |U| < ∞).
For this class of systems, we are interested in questions of existence and construction of finite uniform
bisimulations on a subset S of the state space Rn . Particularly, in order for the bisimulation relation to yield
a meaningful “equivalent" DFM, we require the set S be an invariant set of the system:
N
Definition 4. A set S ⊂ Rn is an invariant set of system (6) if for any input sequence {ut }∞
t=0 ∈ U
x0 ∈ S ⇒ xt ∈ S, for all t ∈ N.
(7)
We are now ready to state the first problem of interest:
Problem 1. Given system (6), under what conditions on A, B, U does there exist a finite uniform bisimulation
∼ on some invariant set S of system (6)?
When Problem 1 has an affirmative answer, another set of problems naturally follows:
Problem 2. Given a system (6) that admits a finite uniform bisimulation on some invariant set S, under
what conditions on A, B, U can an arbitrarily large number of equivalence classes be generated by a finite
uniform bisimulation?
Note that we seek (and propose) both analytical and constructive, algorithmic solutions to the above
problems.
3
3.2
Comparison with Existing Work on Finite Bisimulations
Before presenting our main results, we briefly discuss the similarities and differences between the current
problem of interest and some of the previous developments on finite bisimulations:
• Our definition of finite uniform bisimulation is stronger than that of finite bisimulation used in some
of the literature, of which we pick [19] as a representative paper. In particular in that setting, the
definition requires that if two states are bisimilar (x ∼ y) and x transitions to x0 under input u, then
there exists an input u0 such that y transition to y 0 under u0 and y 0 ∼ y. Note that u and u0 need not
be the same, and thus a finite bisimulation as in [19] is not necessarily a finite uniform bisimulation.
We will use Example 1 in Section 7 to illustrate this difference.
• Our definition of finite uniform bisimulation is in accordance with the definitions of finite bisimulation
introduced in [6, 9]. However, the sufficient conditions for existence of finite bisimulations derived
in [9] concern linear vector fields, and as such correspond to special cases of (6) where B is the zero
matrix, whereas the present contribution addresses the more general case where B is nonzero. Likewise,
the dynamics of the system of interest in [6] are different, as the authors study systems of the form
xt+1 = Aσ(t) xt , where σ(t) is the switching signal and is considered to be the input.
• Finally, the finite input alphabet setup is unique in the literature, in contrast to typically studied
setups where the input signal takes arbitrary instantaneous values in Euclidean space, or else the input
signal is of certain form such as polynomial, exponential or sinusoidal as in [1].
4
4.1
Statement of Main Results
Sufficient Conditions and Construction
We begin by defining a set that will be useful for formulating a sufficient condition for the existence of finite
uniform bisimulations.
Definition 5. Given system (6), define set A as
A = {α ∈ Rn |α =
t
X
At−τ Buτ , u(·) ∈ U, t ∈ N}.
(8)
τ =0
Essentially, A is the collection of forced responses of system (6). Now, we are ready to propose a sufficient
condition for the existence of finite uniform bisimulations on some invariant subset of the state space.
Theorem 1. Given system (6) with 0 ∈ U, assume that A has all eigenvalues within the unit disc. If cl(A)
is not connected, then there exists a finite uniform bisimulation on a subset of Rn that is an invariant set of
system (6).
When the conditions in Theorem 1 are satisfied, we also propose an algorithm for computing finite uniform
bisimulations. To keep the flow of presentation, we present this algorithm in the following section (Algorithm
1 in Section 5). We show that Algorithm 1 is guaranteed to generate a finite uniform bisimulation when the
sufficient condition is satisfied.
Theorem 2. Given system (6), and let the hypothesis in Theorem 1 hold, then Algorithm 1 terminates, and
returns X1 , X2 such that X1 , X2 afford a finite uniform bisimulation on an invariant set S, namely S = X1 ∪X2 ,
of system (6).
Next, we continue to study Problem 2. It turns out that additional assumptions are needed to guarantee
the existence of arbitrarily many equivalence classes, as we shall see in Section 7 Example 2. In order to
describe such conditions, we first define a relevant collection of subsets of the state space Rn : Given system
(6), let U = {u1 , u2 , . . . , uq } for q ∈ Z+ and define sets {Sj1 }qj=1 as follows
Sj1 = Buj + cl(AA),
j = 1, 2, . . . , q.
(9)
We can now propose a sufficient condition for the existence of an arbitrarily large number of equivalence
classes.
4
Theorem 3. Given system (6) with 0 ∈ U and |U| > 1, assume that A has all eigenvalues within the unit disc.
If A is invertible, and {Sj1 }qj=1 (9) are disjoint, then for any z ∈ Z+ there is a finite uniform bisimulation ∼
of system (6) such that the number of equivalence classes associated with ∼ is greater than z.
We also propose an algorithm, which is an extension of Algorithm 1, to compute many equivalence classes.
Corollary 1. Given system (6), and let the hypothesis in Theorem 3 hold, then for any z ∈ Z+ , Algorithm
2 (see Section 5) terminates, and returns a finite uniform bisimulation ∼ that has more than z equivalence
classes.
Remark 1. These equivalence classes computed by Algorithm 2 can also be made arbitrarily fine, that is to
say, the diameter of each equivalence class can be made arbitrarily small (see Section 6.3).
4.2
Necessary Conditions for the Existence of Finite Uniform Bisimulations
Next, we investigate necessary conditions for the existence of finite uniform bisimulations. We quickly realize
that system (6) may admit “pathological" finite uniform bisimulations: If A, B, U have entries in Q, then the
partition Qn and Rn \ Qn affords a finite uniform bisimulation of system (6). This motivates us to study
regular finite uniform bisimulations. We propose a necessary condition for the existence of regular finite
uniform bisimulations.
Theorem 4. Given system (6) with 0 ∈ U. If ∼ is a regular finite uniform bisimulation on an invariant set
S of system (6), 0 ∈ int([0]), and [0] is bounded, then ρ(A) ≤ 1.
Remark 2. Theorem 4 states that under certain assumptions, there do not exist regular finite uniform
bisimulations for Schur unstable systems (6). This justifies why we study Schur stable systems in Theorem
1.
We point out that the condition “[0] is bounded" in Theorem 4 cannot be dropped (see Example 3 in
Section 7). However, the condition “[0] is bounded" in Theorem 4 can be dropped for scalar systems, where
we restrict our attention to instances of (6) described by
xt+1 = axt + but
(10)
where xt ∈ R, ut ∈ U, and a, b ∈ R. U is a finite subset of R.
Corollary 2. Given system (10) with 0 ∈ U. If ∼ is a regular finite uniform bisimulation on an invariant set
S of system (10) and 0 ∈ int([0]), then |a| ≤ 1.
5
Constructive Algorithms
First, we present an algorithm for computing finite uniform bisimulations when the conditions in Theorem
1 are satisfied.
We begin by introducing the notation of binary partitions of the finite input set U with |U| > 1: A pair
(U1 , U2 ) is a binary partition of U if U1 , U2 are nonempty, disjoint subsets of U, and U1 ∪U2 = U. The order of
U1 , U2 is not relevant: (U1 , U2 ) is the same as (U2 , U1 ). Since U is a finite set, there are finitely many distinct
(i)
(i)
binary partitions of U. We use {(U1 , U2 ) : i = 1, . . . , r} to denote the collection of all binary partitions of
q!
1
2
q−1
represents the quantity “q choose
U. Here r = (Cq + Cq + · · · + Cq )/2, where q = |U|, and Cqj = j!(q−j)!
j". Now we are ready to present the following algorithm to compute finite uniform bisimulations of system
(6).
5
Algorithm 1 Computing a Finite Uniform Bisimulation
17:
18:
19:
20:
Input: Matrix A, B, set U
Compute: h = max{kBuk1 : u ∈ U}
Choose: such that 0 < < 1 − ρ(A).
Compute: Matrix T , invertible, such that kT −1 AT k1 ≤ ρ(A) + .
(i)
(i)
Compute: All binary partitions of U: (U1 , U2 ), i = 1, . . . , r.
−1
k1 kT k1
Compute: κ = 2kT
1−ρ(A)−
k ← 1.
loop
Compute: lk = hkAk k1
i ← 1.
while i ≤ r do
(i)
(i)
Compute: C1 = {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 ∈ U1 , u2 , . . . , uk ∈ U}
(i)
(i)
C2 = {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 ∈ U2 , u2 , . . . , uk ∈ U}
(i)
(i)
(i)
Compute: dk = min{kα − βk1 : α ∈ C1 , β ∈ C2 }
(i)
if dk ≥ κlk then
ĩ ← i, k̃ ← k.
Exit the loop
end if
i ← i + 1.
end while
k ← k + 1.
end loop
21:
Compute: S = {x ∈ Rn : kT −1 xk1 <
22:
23:
Compute: X1 = C1 + S, X2 = C2 + S
Return: X1 , X2
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
(ĩ)
(ĩ)
d
k̃
2kT k1 }
(ĩ)
Remark 3. In the preceding algorithm, one approach to compute matrix T involves Schur’s triangularization
of matrix A (pp. 79, [7]). We refer interested readers to [7] on the specifics of computing matrix T such that
kT −1 AT k1 ≤ ρ(A) + is satisfied.
Remark 4. Here we explain why Algorithm 1 returns two equivalence classes. We first point out that if
the conditions in Theorem 1 are satisfied, the number of equivalence classes generated by a finite uniform
bisimulation could be greater than two, which is the case in Example 4 in Section 7. However, for certain
systems (see Example 2 in Section 7), two, and only two equivalence classes can be generated based on the
analytical result stated in Theorem 1. Therefore Algorithm 1 returns two equivalence classes, since it is
capable of computing finite uniform bisimulations for any system that satisfies the conditions in Theorem 1.
As we shall see next, we propose another algorithm in case more equivalence classes are desired.
Next, we present a second algorithm, which is an extended version of Algorithm 1, to generate an
arbitrarily large number of equivalence classes when the conditions in Theorem 3 are satisfied.
6
Algorithm 2 Computing a Finite Uniform Bisimulation with Many Equivalence Classes
1:
2:
3:
4:
5:
6:
7:
8:
Input: Matrix A, B, set U = {u(1) , u(2) , . . . , u(q) }, integer z: Lower bound of the number of equivalence
classes.
Compute: h = max{kBuk1 : u ∈ U}
Choose: such that 0 < < 1 − ρ(A).
Compute: Matrix T , invertible, such that kT −1 AT k1 ≤ ρ(A) + .
k1 kT −1 k1
Compute: κ = 2kT
1−ρ(A)−
k ← 1.
loop
Compute: lk = hkAk k1
(k)
Compute: C1 = {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 = u(1) , u2 , . . . , uk ∈ U}
(k)
= {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 = u(2) , u2 , . . . , uk ∈ U}
C2
..
.
(k)
Cq
9:
10:
11:
12:
13:
14:
15:
16:
= {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 = u(q) , u2 , . . . , uk ∈ U}
(i)
(i)
Compute: dk = min{kα − βk1 : α ∈ Cv , β ∈ Cw , w 6= v, 1 ≤ w, v ≤ q}
if dk ≥ κlk then
k̃ ← k.
Exit the loop
end if
k ← k + 1.
end loop
d
Compute: S = {x ∈ Rn : kT −1 xk1 < 2kTk̃k1 }
(k̃)
(k̃)
(k̃)
Compute: X¯1 = C1 + S, X¯2 = C2 + S, . . . , X¯q = Cq + S
η+1
Choose: η ∈ Z+ such that q
> z.
η
19: Compute: An enumeration {u1 , u2 , . . . , uq η } of the set U η , where uj = (u1j , . . . , uj ).
20: Compute: Xk = Bu1 + ABu2 + · · · + Aη−1 Buη + Aη X̄i , 1 ≤ k ≤ q η+1 , where (u1 , . . . , uη ) = uj for
some 1 ≤ j ≤ q η , and 1 ≤ i ≤ q.
21: Return: X1 , . . . , Xq η+1
17:
18:
As we shall see in the derivation of Theorem 3 in Section 6, we claim that the sets X1 , . . . , Xqη+1 returned
η+1
by Algorithm 2 afford a finite uniform bisimulation on ∪qk=1 Xk of system (6).
6
6.1
Derivation of Main Results
Derivation of Theorem 1
We first introduce several Lemmas which will be instrumental in this derivation of Theorem 1.
Lemma 1. Given system (6), if matrix A has all eigenvalues within the unit disc, then cl(A) is compact.
P∞
Proof. If A ∈ Rn×n has all eigenvalues within the unit disc, then τ =0 kAτ k1 converges (pp. 298, [7]). Since
U is finite, max{kBuk1 : u ∈ U} is also finite. Combining these two facts, and applying triangle inequality,
we conclude that A is bounded and therefore cl(A) is bounded. Since cl(A) is closed and bounded in Rn ,
cl(A) is compact.
Next, we study the structure of set A as defined in (8). By the definition of A and 0 ∈ U, and recall (9),
we have
q
[
Sj1 = cl(A).
(11)
j=1
7
Generally, for any k ∈ Z+ , let {u1 , u2 , . . . , uqk } be an enumeration of the set U k , where uj = (u1j , . . . , ukj ),
k
u1j , . . . , ukj ∈ U, we define sets {Sjk }qj=1 as follows
Sjk = Bu1j + ABu2j + · · · + Ak−1 Bukj + cl(Ak A),
j = 1, 2, . . . , q k .
(12)
We also have
k
q
[
Sjk = cl(A).
(13)
j=1
Now we introduce the following Lemma.
Lemma 2. Given system (6), assume that A has all eigenvalues within the unit disc. If open sets W and V
∗
is a disconnection of cl(A), then there exists k ∗ ∈ Z+ such that for all j ∈ {1, . . . , q k },
∗
Sjk ∩ W 6= ∅
∗
Sjk ⊂ W
⇒
(14)
Proof. We show this Lemma by contradiction. We first assume that for all k ∈ Z+ , there is j(k) ∈ {1, . . . , q k }
k
k
k
k
∩W =
6 ∅ and Sj(k)
∩V =
6 ∅. For each k, choose wk ∈ Sj(k)
∩ W and vk ∈ Sj(k)
∩ V. Then
such that Sj(k)
∞
we have constructed two sequences {wk }∞
and
{v
}
.
k k=1
k=1
∞
Since {wk }∞
k=1 ⊂ cl(A), {vk }k=1 ⊂ cl(A) and cl(A) is compact (by Lemma 1), there exists a subsequence
∞
{wkl }∞
l=1 that converges to a point in cl(A). Similarly, there also exists a subsequence of {vkl }l=1 that
∞
converges to a point in cl(A). By relabeling, we have found two sequences {wkp }∞
and
{v
}
such
that
kp p=1
p=1
lim wkp = w,
p→∞
lim vkp = v
and
(15)
p→∞
where w, v ∈ cl(A).
By the construction of Sjk (12), we see that for any j, diam(Sjk ) ≤ kAk k1 diam(A). Since A has all
eigenvalues within the unit disc, lim Ak = 0n×n (pp.298, [7]). By boundedness of set A, diam(A) is finite.
k→∞
k
k
p
p
Therefore diam(Sjk ) goes to 0 as k tends to infinity. Note that wkp ∈ Sj(k
and vkp ∈ Sj(k
, and kp ≥ p,
p)
p)
therefore limp→∞ kwkp − vkp k1 = 0. Combine with (15), we have limp→∞ wkp = limp→∞ vkp = w, where
w ∈ cl(A). Without loss of generality, let w ∈ W. Since W is open, there exist > 0 such that the open
ball B (w) ⊂ W. Since W ∩ V = ∅, {vkp }∞
p=1 ∩ B (w) = ∅. Therefore kvkp − wk1 ≥ for all p. This is a
contradiction with limp→∞ vkp = w. Therefore (14) holds.
Next, we introduce another Lemma which is based on Lemma 2.
Lemma 3. Given system (6), assume that A has all eigenvalues within the unit disc. If open sets W and V
is a disconnection of cl(A), then there exist open sets W 0 and V 0 in Rn such that the pair W 0 and V 0 is also
a disconnection of cl(A), and for all j ∈ {1, . . . , q}
Sj1 ∩ W 0 6= ∅
Sj1 ⊂ W 0
⇒
(16)
Proof. By Lemma 2 , (14) holds, and we only need to consider the case when k ∗ ≥ 2.
Define a function f : Rn → Rn as :f (x) = Ax. Clearly f is continuous. For any set S, use f −1 (S) to
denote the set f −1 (S) = {x ∈ Rn |f (x) ∈ S}.
∗
k∗
For {Sjk }qj=1 as constructed in (12), let u be an element of U, then define an index set J as
∗
J = {j ∈ {1, . . . , q k } : u1j = u},
then |J | = q k
∗
−1
. Define sets
∗
∗
S̃jk = −Bu + Sjk , j ∈ J .
∗
∗
(17)
For any j ∈ J , by (14), either S̃jk ⊂ −Bu + W or S̃jk ⊂ −Bu + V. Write W 0 = f −1 (−Bu + W) and
∗
∗
V 0 = f −1 (−Bu + V), then either f −1 (S̃jk ) ⊂ W 0 or f −1 (S̃jk ) ⊂ V 0 .
8
For each j ∈ J , by (12), (17), and the compactness of cl(A), we have
∗
S̃jk = A(Bu2j + · · · + Ak
∗
for some (u2j , . . . , ukj ) ∈ U k
that
∗
−1
∗
−2
∗
Bukj + cl(Ak
∗
−1
A))
. Consequently, we can determine one and only one j 0 ∈ {1, . . . , q k
∗
Sjk0
−1
∗
⊂ f −1 (S̃jk ).
We also observe that
[
∗
∗
−1
} such
(18)
(u2j , u3j , . . . , ukj ) = U k
∗
−1
.
(19)
j∈J
Recall (13), (18), we see that
∗
cl(A) =
−1
q k[
∗
Sjk0
−1
⊂
j 0 =1
[
∗
f −1 (S̃jk ) ⊂ W 0 ∪ V 0 .
(20)
j∈J
It is clear that W 0 and V 0 are disjoint open sets. Therefore (14) holds for k ∗ − 1 and W 0 , V 0 . Repeat this
argument k ∗ − 1 times, we conclude that (16) holds.
Finally, we provide the proof of Theorem 1.
Proof. (of Theorem 1) Since cl(A) is not connected, let W and V be a disconnection of cl(A). Then by
Lemma 3, (16) holds. We propose an equivalence relation on A. Since A is an invariant set of system (6),
the proof is complete if we can show that this equivalence relation satisfies (2) and (3).
Given open sets W 0 and V 0 that satisfy (16), let X1 = A ∩ W 0 and X2 = A ∩ V 0 . Define an equivalence
relation ∼ as
x ∼ x0 ⇔ x ∈ Xi and x0 ∈ Xi for some i ∈ {1, 2}.
For any x, x0 ∈ A, any uj ∈ U, if x ∼ x0 , then Ax + Buj ∈ Sj1 and Ax0 + Buj ∈ Sj1 . By (16), we see that
Ax + Buj ∼ Ax0 + Buj . Therefore (2) is satisfied. Since 1 < 2 < ∞, (3) is also satisfied. This completes the
proof.
6.2
Derivation of Theorem 2
In this section, we derive Theorem 2. We first show that Algorithm 1 terminates, and then show that the
equivalence classes X1 , X2 returned by Algorithm 1 afford a finite uniform bisimulation on X1 ∪ X2 .
Proof. (of Theorem 2) Given system (6), since matrix A has all eigenvalues within the unit disc, and cl(A)
is not connected, by Lemma 3, there is a disconnection of cl(A), W and V, such that for all j ∈ {1, . . . , q}
Sj1 ∩ W 6= ∅
⇒
Sj1 ⊂ W
(21)
where q = |U|. Let U1∗ = {uj ∈ U|Sj1 ∩ W 6= ∅}, and U2∗ = U \ U1∗ . Recall (11), we see that U1∗ is nonempty,
otherwise cl(A) ∩ W = ∅, which contradicts with W and V being a disconnection of cl(A). U2∗ is also
nonempty, otherwise cl(A) ⊂ W, then cl(A) ∩ V = ∅, which draws a contradiction. We also observe that
|U| > 1, otherwise U = 0 by assumption, and cl(A) = 0 is connected. Therefore the binary partitions of U
are well-defined. Since U1∗ and U2∗ are nonempty, disjoint subsets of U, and U1∗ ∪ U2∗ = U, there is a binary
(i∗ )
(i∗ )
partition of U, (U1 , U2 ), such that
(i∗ )
(U1
(i∗ )
, U2
) = (U1∗ , U2∗ )
where i∗ is an integer between 1 and r.
Since for any k ∈ Z+ ,
(i)
(i)
(i)
dk = min{kα − βk1 : α ∈ C1 , β ∈ C2 },
9
(22)
(23)
(i∗ )
we claim that dk
(23) is uniformly bounded away from zero, that is: There exists d > 0 such that
(i∗ )
dk
≥ d,
for all
To see this claim, we define two sets G1 , G2 by
[
G1 =
Sj1 ,
k ∈ Z+ .
[
G2 =
j∈U1∗
(24)
Sj1 .
(25)
j∈U2∗
By the definition of U1∗ , we see that G1 ⊂ W. Recall (11) and that W and V is a disconnection of cl(A),
we see that G2 ⊂ V. Because V and W are disjoint, G1 and G2 are also disjoint. Since G1 is a finite union
of closed sets, G1 is closed. By Lemma 1, cl(A) is bounded, and therefore G1 is bounded. We see that G1 is
closed, bounded, and therefore compact. Similarly, G2 is also compact. By an observation in analysis: The
distance between two disjoint compact sets is positive (pp. 18, [17]), we have
d = inf{kα − βk1 : α ∈ G1 , β ∈ G2 } > 0.
Since
(i)
(i)
(i)
(i)
(26)
C1 = {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 ∈ U1 , u2 , . . . , uk ∈ U}
C2 = {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 ∈ U2 , u2 , . . . , uk ∈ U}
(27)
and recall (9), (22), and (25), we observe that: For all k ∈ Z+ ,
(i∗ )
C1
(i∗ )
⊂ G1 ,
C2
⊂ G2 .
(28)
∗
(i )
Recall (23), we have dk ≥ d > 0 for all k ∈ Z+ .
Since matrix A is Schur stable, we see that lk = hkAk k1 → 0 as k → ∞. Consequently, there exists
∗
k ∈ Z+ such that
2kT k1 kT −1 k1
(i∗ )
lk∗ .
dk∗ ≥ κlk∗ =
1 − ρ(A) −
Now we see that the loop in Algorithm 1 terminates, and returns two sets X1 , X2 :
(ĩ)
X1 = C1 + S,
(29)
(ĩ)
X2 = C2 + S.
For the second part of this derivation, we show that X1 ∪ X2 is an invariant set of system (6), and that
X1 , X2 afford a finite uniform bisimulation on X1 ∪ X2 .
For any x ∈ X1 ∪ X2 , by (27) and (29), there exist (u1 , . . . , uk̃ ) ∈ U k̃ and s ∈ S such that
x = Bu1 + ABu2 + · · · + Ak̃−1 Buk̃ + s.
(30)
Ax + Bu = (Bu + ABu1 + · · · + Ak̃−1 Buk̃−1 ) + (Ak̃ Buk̃ + As).
(31)
Then for any u ∈ U,
Recall kT −1 AT k1 ≤ ρ(A) + , κ =
2kT k1 kT −1 k1
1−ρ(A)− ,
and h = max{kBuk1 : u ∈ U}, we observe that
kT −1 (Ak̃ Buk̃ + As)k1
≤ kT −1 Ak̃ Buk̃ k1 + kT −1 Ask1
≤ kT −1 k1 kAk̃ k1 kBuk̃ k1 + k(T −1 AT )T −1 sk1
≤ kT −1 k1 lk̃ + k(T −1 AT )k1 kT −1 sk1
(ĩ)
< kT −1 k1
d
1 − ρ(A) − (ĩ)
dk̃ + (ρ(A) + ) k̃
−1
2kT k1 kT k1
2kT k1
(ĩ)
=
dk̃
2kT k1
.
10
Therefore (Ak̃ Buk̃ + As) ∈ S. By (27), we observe that (Bu + ABu1 + · · · + Ak̃−1 Buk̃−1 ) ∈ C1ĩ ∪ C2ĩ , therefore,
we have
Ax + Bu ∈ X1 ∪ X2 .
(32)
We conclude that X1 ∪ X2 is an invariant set of system (6).
Next, we show X1 ∩ X2 = ∅. We show by contradiction: Assume z ∈ X1 ∩ X2 , then by (29), there exist
c1 ∈ C1ĩ , c2 ∈ C2ĩ , s1 ∈ S, and s2 ∈ S such that z = c1 + s1 , and z = c2 + s2 , and recall S = {x ∈ Rn :
(ĩ)
kT −1 xk1 < dk̃ /(2kT k1 )}, we have
kc1 − c2 k1 ≤ kc1 − zk1 + kz − c2 k1
= ks1 k1 + ks2 k1
= kT (T −1 s1 )k1 + kT (T −1 s2 )k1
≤ kT k1 (kT −1 s1 k1 + kT −1 s2 k1 )
(ĩ)
< dk̃ .
(ĩ)
But by (23), we have kc1 − c2 k1 ≥ dk̃ , which draws a contradiction. Therefore X1 ∩ X2 = ∅.
Now we are ready to define an equivalence relation ∼ on X1 ∪ X2 as:
x ∼ x0 ⇔ x ∈ Xi and x0 ∈ Xi for some i ∈ {1, 2}.
We show that ∼ is a finite uniform bisimulation on X1 ∪X2 . For any x, x0 ∈ X1 ∪X2 , and any u ∈ U, if x ∼ x0 ,
we consider two cases: If u ∈ U1ĩ , recall (27), (29), and (31), we see that Ax + Bu ∈ X1 and Ax0 + Bu ∈ X1 ,
therefore Ax + Bu ∼ Ax0 + Bu. Similarly, if u ∈ U2ĩ , then Ax + Bu ∈ X2 and Ax0 + Bu ∈ X2 , therefore
Ax + Bu ∼ Ax0 + Bu. Since (U1ĩ , U2ĩ ) is a binary partition of U, we see that (2) is satisfied.
Since {[x]|x ∈ X1 ∪ X2 } = {X1 , X2 }, we have |{[x]|x ∈ X1 ∪ X2 }| = 2, and (3) is satisfied. Therefore ∼ is
a finite uniform bisimulation on X1 ∪ X2 . This completes the proof.
6.3
Derivation of Theorem 3 & Corollary 1
Proof. To show Theorem 3 and Corollary 1, it suffices to show that Algorithm 2 terminates, and that the
sets X1 , . . . , Xqη+1 :
Xk = Bu1 + ABu2 + · · · + Aη−1 Buη + Aη X̄i , 1 ≤ k ≤ q η+1 ,
(33)
η+1
q
returned by Algorithm 2 afford a finite uniform bisimulation ∼ on ∪k=1
Xk of system (6). By Algorithm 2,
η+1
the number of equivalence classes q
is guaranteed to be greater than z.
By assumption, {Sj1 }qj=1 (9) are disjoint. By Lemma 1, Sj1 is also compact for all j ∈ {1, . . . , q}. Since
the distance between two disjoint compact sets is positive, we have
1
min{d(Sw
, Sv1 ) : w 6= v, 1 ≤ w, v ≤ q} > 0.
Recall
(k)
= {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 = u(1) , u2 , . . . , uk ∈ U},
(k)
= {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 = u(2) , u2 , . . . , uk ∈ U},
C1
C2
..
.
(34)
Cq(k) = {Bu1 + ABu2 + · · · + Ak−1 Buk : u1 = u(q) , u2 , . . . , uk ∈ U},
(k)
we observe that Cj
α∈
(i)
Cv , β
∈
(i)
Cw , w
(34) is a subset of Sj1 for any j ∈ {1, . . . , q} and any k ∈ Z+ , therefore dk = min{kα−βk1 :
6= v, 1 ≤ w, v ≤ q} is uniformly bounded away from zero:
1
dk ≥ min{d(Sw
, Sv1 ) : w 6= v, 1 ≤ w, v ≤ q} > 0, ∀ k ∈ Z+ .
Since lk tends to zero as k tends to infinity, we see that Algorithm 2 terminates.
11
(35)
Recall
(k̃)
X¯1 = C1 + S,
(k̃)
X¯2 = C2 + S,
..
.
(36)
X¯q = Cq(k̃) + S,
we observe that X¯1 , . . . , X¯q afford a finite uniform bisimulation on ∪qj=1 X̄j of system (6) by the derivation
of Theorem 2. We will use this observation to show that sets X1 , . . . , Xqη+1 (33) also afford a finite uniform
bisimulation.
η+1
We first show that ∪qk=1 Xk is an invariant set of system (6). For any x ∈ Xk , by (33), we can write
x = Bu1 + ABu2 + · · · + Aη−1 Buη + Aη x̄
for some (u1 , . . . , uη ) ∈ U η and some x̄ ∈ X̄i with 1 ≤ i ≤ q. Then for any u ∈ U,
Ax + Bu = Bu + ABu1 + A2 Bu2 + · · · + Aη−1 Buη−1 + Aη (Ax̄ + Buη ).
Since ∪qj=1 X̄j is an invariant set of system (6), we have (Ax̄ + Buη ) ∈ X̄j for some 1 ≤ j ≤ q. Recall (33),
η+1
we see that (Ax + Bu) ∈ Xk for some 1 ≤ k ≤ q η+1 , and therefore ∪qk=1 Xk is an invariant set of system (6).
Next, we use an inductive approach to show that the sets Xk , k = 1, . . . , q η+1 (36) are disjoint. Write
U = {u(1) , . . . , u(q) }, we observe that the q 2 sets Bu(i) + AX̄j , i = 1, . . . , q, j = 1, . . . , q are disjoint.
Indeed, consider any Bu(i1 ) + AX̄j 1 and Bu(i2 ) + AX̄j 2 with (i1 , j 1 ) 6= (i2 , j 2 ). If i1 = i2 , then j 1 6=
j 2 . Since X¯1 , . . . , X¯q are disjoint, we have X̄j 1 ∩ X̄j 2 = ∅. Since A is invertible by assumption, we have
AX̄j 1 ∩ AX̄j 2 = ∅, and therefore (Bu(i1 ) + AX̄j 1 ) ∩ (Bu(i2 ) + AX̄j 2 ) = ∅. If i1 6= i2 , from the second part of
the derivation of Theorem 2 (equation (30) through (32)) and the construction of X̄j (34), (36), we see that
(Bu(i1 ) + AX̄j 1 ) ⊂ X̄i1 and (Bu(i2 ) + AX̄j 2 ) ⊂ X̄i2 . Since X¯1 , . . . , X¯q are disjoint, we have X̄i1 ∩ X̄i2 = ∅,
and therefore (Bu(i1 ) + AX̄j 1 ) ∩ (Bu(i2 ) + AX̄j 2 ) = ∅. We conclude that the sets Bu(i) + AX̄j , i = 1, . . . , q,
j = 1, . . . , q are disjoint, where U = {u(1) , . . . , u(q) }.
For the ease of exposition, we use Xj1 , j = 1, . . . , q 2 to denote the q 2 disjoint sets Bu(i) +AX̄j , i = 1, . . . , q,
j = 1, . . . , q. We observe that the q 3 sets Bu(i) + AXj1 , i = 1, . . . , q, j = 1, . . . , q 2 are also disjoint. Indeed,
consider any Bu(i1 ) + AXj11 and Bu(i2 ) + AXj12 with (i1 , j 1 ) 6= (i2 , j 2 ). If i1 = i2 , then j 1 6= j 2 . Since
Xj1 , j = 1, . . . , q 2 are disjoint, we have Xj11 ∩ Xj12 = ∅. Since A is invertible by assumption, we have
AXj11 ∩ AXj12 = ∅, and therefore (Bu(i1 ) + AXj11 ) ∩ (Bu(i2 ) + AXj12 ) = ∅. If i1 6= i2 , by the preceding
paragraph, we see that Xj11 ⊂ X̄l for some 1 ≤ l ≤ q, and therefore
(Bu(i1 ) + AXj11 ) ⊂ (Bu(i1 ) + AX̄l ) ⊂ X̄i1 .
Similarly, we see that (Bu(i2 ) + AXj12 ) ⊂ X̄i2 . Since X¯1 , . . . , X¯q are disjoint, we have X̄i1 ∩ X̄i2 = ∅, and
therefore (Bu(i1 ) + AXj11 ) ∩ (Bu(i2 ) + AXj12 ) = ∅. We conclude that the sets Bu(i) + AXj1 , i = 1, . . . , q,
j = 1, . . . , q 2 are disjoint.
Repeating this argument η times, we conclude that the q η+1 sets Xk , k = 1, . . . , q η+1 (36) are disjoint.
η+1
Next, we define an equivalence relation ∼ on ∪qk=1 Xk as
x ∼ y ⇐⇒ x ∈ Xk and y ∈ Xk for some 1 ≤ k ≤ q η+1 .
We claim that ∼ is a finite uniform bisimulation. Indeed, for any 1 ≤ k ≤ q η+1 , by (33), write Xk as
Xk = Bu1 + ABu2 + · · · + Aη−1 Buη + Aη X̄i . Then for any u ∈ U, AXk + Bu = Bu + ABu1 + A2 Bu2 + · · · +
Aη−1 Buη−1 + Aη (AX̄i + Buη ). Since (AX̄i + Buη ) ⊂ X̄j for some 1 ≤ j ≤ q, we have
(AXk + Bu) ⊂ (Bu + ABu1 + A2 Bu2 + · · · + Aη−1 Buη−1 + Aη X̄j ) = Xk0
for some 1 ≤ k 0 ≤ q η+1 . Therefore (2) is satisfied. Since q η+1 is finite, (3) is also satisfied. This completes
the proof of Theorem 3 and Corollary 1.
12
Lastly, we comment on the fact that the diameter of the equivalence classes Xk can be made arbitrarily
small. For any 1 ≤ k ≤ q η+1 , we have
(k̃)
diam(Xk ) ≤ kAη k1 diam(Ci
+ S) ≤ kAη k1 (diam(A) + diam(S))
Since A is Schur-stable, diam(A) is finite, and kAη k1 can be made arbitrarily small by choosing η large
enough. diam(S) is finite by construction, and we conclude that diam(Xk ) can be made arbitrarily small by
choosing η sufficiently large.
6.4
Derivation of Necessary Conditions
Proof. (of Theorem 4) We will prove by contradiction. Assume ρ(A) > 1, let Av = λv with |λ| > 1, kvk1 = 1,
λ ∈ C, v ∈ Cn . And for any w ∈ Cn , we use Re(w) to denote the real part of w. Define a set O as
O = {α ∈ R+ |Re(γv) ∈ [0], for all |γ| ≤ α, γ ∈ C}.
(37)
We show that O is non-empty and bounded in the following. Write v = [v1 v2 . . . vn ]T , where v1 , . . . , vn ∈ C
and |v1 | + · · · + |vn | = 1. For any γ ∈ C, we have |Re(γvi )| ≤ |γkvi |, therefore
kRe(γv)k1 =
n
X
|Re(γvi )| ≤ |γ|
i=1
n
X
|vi | = |γ|.
i=1
Since Br (0) ⊂ [0] for some r > 0 by assumption, for all γ with |γ| ≤ r/2, Re(γv) ∈ Br (0). Therefore
r/2 ∈ O, and O is nonempty.
Next, we show that O is bounded. Since [0] is bounded by assumption, let [0] ⊂ Bσ (0) for some σ > 0.
Since v = [v1 v2 . . . vn ]T 6= 0n×1 , let |vk | > 0 for some 1 ≤ k ≤ n. Write vk as vk = |vk |eiφ for some
φ ∈ [0, 2π). Assume O is unbounded, then there exist α ∈ O with |α| > 2σ/|vk |. Let γ = (2σ/|vk |)ei(−φ) ,
then |γ| < α. By the definition of O (37), we have Re(γv) ∈ [0]. Observe that
kRe(γv)k1 ≥ |Re(γvk )| = |Re(
2σ i(−φ)
e
|vk |eiφ )| = |Re(2σ)| = 2σ.
|vk |
Therefore Re(γv) ∈
/ Bσ (0), and consequently Re(γv) ∈
/ [0], which draws a contradiction. Therefore O is
bounded.
Next, we define β = sup O. Since O is non-empty and bounded, we have 0 < β < ∞. Then for any
> 0, there is 0 ≤ δ < such that Re(κv) ∈
/ [0] for some κ ∈ C and |κ| = β + δ. Choose = ( |λ|−1
2 )β, and
κ
let κ0 = λ , then
|κ|
β+δ
β+
β + (|λ| − 1)β
|κ0 | =
=
<
<
= β.
|λ|
|λ|
|λ|
|λ|
Therefore |κ0 | < β. Since β = sup O, there exists α ∈ O such that α > |κ0 |. By (37), we see that Re(κ0 v) ∈ [0],
or equivalently Re(κ0 v) ∼ 0. Since ∼ is a finite uniform bisimulation, by (2) and letting the input u be zero,
we have ARe(κ0 v) ∼ 0. We observe that
ARe(κ0 v) = Re(Aκ0 v) = Re(κ0 (Av)) = Re(κ0 λv) = Re(κv),
therefore Re(κv) ∼ 0 ,which draws a contradiction. We conclude that the assumption ρ(A) > 1 is false, and
therefore ρ(A) ≤ 1.
Proof. (of Corollary 2) We will prove by contradiction. Assume |a| > 1, and use [0] to denote the equivalence
class [0] = {x ∈ S|x ∼ 0}. By the assumption Br (0) ⊂ [0] for some r > 0, define β as
β = sup{x ∈ S|[0, x] ⊂ [0]},
(38)
where [0, x] is the closed interval between 0 and x. Since int([0]) is nonempty, there is such that [0, ) ⊂ [0],
therefore the supremum is well defined, and β > 0.
13
First, we consider the case β < ∞. Clearly [0, β) ⊂ [0]. By the definition of β, we have that for any
> 0, there is 0 ≤ δ < such that
β+δ ∈
/ [0].
(39)
Let = (a2 − 1)β > 0, and let δ denote the nonnegative number that satisfy (39). We observe that
z=
β+δ
< β,
a2
therefore z ∼ 0. Since ∼ is a finite uniform bisimulation, when the input is 0 we have az ∼ 0, and a2 z ∼ 0.
This draws a contradiction with (39).
For the case β = ∞, let β 0 = inf{x ∈ S|[x, 0] ⊂ [0]}, then β 0 > −∞, otherwise for any x ∈ R, x ∈ [0],
which implies R = [0] and there is only one equivalence class. Next, for any > 0, there is 0 ≤ δ < such
that β 0 − δ ∈
/ [0]. Choose = (1 − a2 )β 0 and z = (β 0 − δ)/a2 , then the preceding argument follows.
7
Illustrative Examples
In this section, we present a set of illustrative examples: In Example 1, we illustrate the difference between
the notion of finite uniform bisimulation and the notion of finite bisimulation stated in [19]; in Example
2, we show that additional assumptions, besides the conditions in Theorem 1, are needed to guarantee the
existence of arbitrarily many equivalence classes; in Example 3, we show that the condition “[0] is bounded"
in Theorem 4 cannot be dropped; in Example 4, we illustrate the analytical result in Theorem 1, discuss
how to construct a DFM approximation of the original system, and apply Algorithm 2 to construct many
equivalence classes.
Example 1. (Example 2.14, [19]) Consider system
2
0
A = −1 −7
0
4
(6) with parameters
1 2
−1
11 , B = 1 1
1 1
6
According to [19], a finite bisimulation with eight equivalence classes {q1 , . . . , q8 } is constructed. If we
choose x = [1 − 2 − 3]T ∈ q1 , x0 = [8, −18, −24]T ∈ q1 and let input u = [0 60]T , then Ax + Bu =
[125 40 34]T ∈ q2 , and Ax0 + Bu = [160 − 86 − 156]T ∈ q1 . Therefore this finite bisimulation is not a “finite
uniform bisimulation" as defined in Definition 1.
Example 2. Consider system (6) with parameters
0.5 0
A=
,
0 0
and
U=
1
0
B=
1
0
0
1
(40)
0
0
0
,
,
,
1
−1
0
We calculate, and plot cl(A):
cl(A) = {(x, y) ∈ R2 |x = 0, −2 ≤ y ≤ 2} ∪ {(x, y) ∈ R2 |x = 1, −1 ≤ y ≤ 1}
14
(41)
Figure 1: 2 and only 2 equivalence classes.
In the above figure, W and V represents a disconnection of cl(A). We see that both cl(A) ∩ W and
cl(A) ∩ V are connected. Therefore, we cannot apply the analytical result in Theorem 1 to generate more
than two equivalence classes, because such result relies on the disconnectedness of an invariant set.
Example 3. Given system (6) with parameters: A = diag({2, 0.5}) (a diagonal matrix with diagonal entries
2 and 0.5), B is the identity matrix, and U = {[0 0]T }. Let X1 = {(x, y) ∈ R2 : 1 < |y| < 2}, and
X2 = {(x, y) ∈ R2 : |y| < 1}, then we see that X1 , X2 afford a regular finite uniform bisimulation on X1 ∪ X2 ,
which is an invariant set, and Br (0) ⊂ [0] for r = 0.5, and ρ(A) = 2 > 1.
Example 4. Consider system (6) with parameters:
0.25 −0.15
1 0
A=
, B=
(42)
0
0.1
0 1
and
1
−1
0
0
0
U=
,
,
,
,
0
0
1
−1
0
Since A is diagonalizable, we have
(1/4)n (1/10)n − (1/4)n
An =
, n = 0, 1, 2, · · ·
0
(1/10)n
and we can show that cl(A) is a subset of:
[
1 1
4 4
{(±1, ±1), (0, 0)} + {(x, y) : x ∈ [− , ], y ∈ [− , ]}
9 9
9 9
Therefore cl(A) is not connected.
By the derivation of Theorem 1, we find a finite uniform bisimulation ∼ on an invariant set of this system:
Figure 2: 2-d finite uniform bisimulation example.
15
X1 , . . . , X5 shown in Figure 2 afford a finite uniform bisimulation ∼ on an invariant set S = ∪5i=1 Xi of
system (6). The points a, b, c, d are given by:
22 10
10 10
4 1
4 1
, ), b = ( , − ), c = ( , ), d = ( , − )
9 9
9
9
9 9
9 9
and Figure 2 is symmetric with respect to the origin. Particularly, the set S is the convex hull of points:
{a, b, −a, −b}.
Given ∼, we can construct a DFM that is uniformly bisimilar to the original system. Particularly, we
associate each equivalence class Xi to a discrete state qi of the DFM, i = 1, . . . , 5. The state transitions
of the DFM can be determined based on (5): For instance, if the current state of the DFM is q1 , and the
current input is [0 1]T , then the next state of the DFM is q3 .
Since this example also satisfies the conditions in Theorem 3, we can also use Algorithm 2 to generate a
finite uniform bisimulation with an arbitrarily large number of equivalence classes. In particular, we generate
two finite uniform bisimulations with 5 equivalence classes, and 25 equivalence classes respectively.
a=(
(a) 5 equivalence classes.
(c) 25 equivalence classes.
(b) Zoom in on 1 equivalence class.
(d) Zoom in on 1 equivalence class.
Figure 3: Finite uniform bisimulations with many equivalence classes.
In the above, Figure 3a shows the 5 equivalence classes generated by Algorithm 2, and Figure 3b shows
one particular equivalence class (the boxed rectangular area in Figure 3a). Similarly Figure 3c shows the
25 equivalence classes, and Figure 3d shows one particular equivalence class. As shown in Figure 3b and
Figure 3d, an equivalence class computed by Algorithm 2 is the union of all the polytopes (in this case
parallelograms). This is in accordance with the construction of the equivalence classes.
8
Conclusions
In this paper we propose notions of finite uniform bisimulation and regular finite uniform bisimulation. We
then present a sufficient condition for the existence of finite uniform bisimulations: If the forced response of a
Schur stable system is not connected, then the system admits a finite uniform bisimulation. In this case, we
construct an algorithm to compute finite uniform bisimulations. Furthermore, we discuss the existence and
construction of an arbitrarily large number of equivalence classes. We also present a necessary condition for
the existence of regular finite uniform bisimulation. Future works include closing the gap between necessary
conditions and sufficient conditions, and extending the current result to systems with more general dynamics.
16
9
Appendix
For the sake of completeness, we review here relevant mathematical concepts and notation, beginning with
the concept of equivalence relations [14]. Given a set A, a subset ∼ of A × A is called a relation on A. With
some slight abuse of notation, we write a ∼ b, read a is equivalent to b, to mean that (a, b) is an element of
the relation ∼. A relation ∼ on A is an equivalence relation if for any a, b, c ∈ A, we have:
(i) a ∼ a (reflexive),
(ii) If a ∼ b, then b ∼ a (symmetric),
(iii) If a ∼ b and b ∼ c, then a ∼ c (transitive).
An equivalence relation ∼ on a set A can be used to partition A into equivalence classes. We use [x] to
denote the equivalence class of x, defined as [x] = {y ∈ A|y ∼ x}. Note that this indeed defines a partition
as the following properties are satisfied:
(i) [x] 6= ∅, ∀x ∈ A,
(ii) [x] =
6 [y] ⇒ [x] ∩ [y] = ∅,
[
(iii)
[x] = A.
x∈A
Next, we review relevant concepts in analysis. A point x ∈ Rn consists of an n−tuple of real numbers
x = (x1 , x2 , . . . , xn ). For the purpose of illustration, we use the 1-norm to review relevant concepts. The 1norm of x is denoted by kxk1 and is defined as kxk1 = (|x1 |+· · ·+|xn |). The distance between two points x and
y is then simply kx − yk1 . Given a set A in Rn , the diameter of A is defined as diam(A) = sup{ky − xk1 : x ∈
A, y ∈ A}. The open ball in Rn centered at x and of radius r is defined by Br (x) = {y ∈ Rn : ky − xk1 < r}.
Given a set A in Rn , a point x is a closure point of A if for every r > 0, the ball Br (x) contains a point of
A. Similarly, a point x is a limit point of A if for every r > 0, the ball Br (x) contains a point of A that is
distinct from x. The closure of A, cl(A), consists of all closure points of A. A point x ∈ A is an interior
point of A if there exists r > 0 such that Br (x) ⊂ A. The interior of A, int(A), consists of all interior points
of A. A boundary point of A is a point which is in cl(A) but not in int(A). The boundary of A, ∂A, consists
of all boundary points of A.
Lastly, we review the notion of spectral radius of a square matrix. Given a square matrix A, the spectral
radius of A is the nonnegative real number ρ(A) = max{|λ| : λ is an eigenvalue of A}. Given a square
matrix A, the 1-induced norm of A is defined as kAk1 = max kAxk1 . Recall that induced norms satisfy
kxk1 =1
the sub-multiplicative property: kABk1 ≤ kAk1 kBk1 .
10
Acknowledgments
This research was supported by NSF CAREER award 0954601 and AFOSR Young Investigator award
FA9550-11-1-0118.
References
[1] R. Alur, T. A. Henzinger, G. Lafferriere, and G. J. Pappas. Discrete abstractions of hybrid systems.
Proceedings of the IEEE, 88(7):971–984, July 2000.
[2] A. Chutinan and B. H. Krogh. Computing approximating automata for a class of hybrid systems.
Mathematical and Computer Modelling of Dynamical Systems, 6(1):30–50, 2000.
[3] A. Chutinan and B. H. Krogh. Verification of infinite-state dynamic systems using approximate quotient
transition systems. IEEE Transactions on Automatic Control, 46(9):1401–1410, 2001.
[4] J. E. R. Cury, B. H. Krogh, and T. Niinomi. Synthesis of supervisory controllers for hybrid systems
based on approximating automata. IEEE Transactions on Automatic Control, 43(4):564–568, 1998.
17
[5] D. Fan and D. C. Tarraf. On existence of finite uniform bisimulations for linear systems with finite input
alphabets. Proceedings of the 54th IEEE International Conference on Control and Decision, December
2015. To appear.
[6] E. A. Gol, X. Ding, M. Lazar, and C. Belta. Finite bisimulations for switched linear systems. IEEE
Transactions on Automatic Control, 59(12):3122–3134, 2014.
[7] R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, 1990.
[8] M. Kloetzer and C. Belta. A fully automated framework for control of linear systems from temporal
logic specifications. IEEE Transactions on Automatic Control, 53(1):287–297, February 2008.
[9] G. Lafferriere, G. J. Pappas, and S. Sastry. O-minimal hybrid systems. Mathematics of Control, Signals
and Systems, 13(1):1–21, 2000.
[10] J. Liu, N. Ozay, U. Topcu, and R. M. Murray. Synthesis of reactive switching protocols from temporal
logic specifications. IEEE Transactions on Automatic Control, 58(7):1771–1785, 2013.
[11] J. Lunze. Qualitative modeling of linear dynamical systems with quantized state measurements. Automatica, 30:417–431, 1994.
[12] T. Moor and J. Raisch. Supervisory control of hybrid systems whithin a behavioral framework. Systems
& Control Letters, Special Issue on Hybrid Control Systems, 38:157–166, 1999.
[13] T. Moor, J. Raisch, and S. D. O’Young. Discrete supervisory control of hybrid systems by l-complete
approximations. Discrete Event Dynamic Systems: Theory and Applications, 12:83–107, 2002.
[14] W. K. Nicholson. Introduction to Abstract Algebra. Wiley, 2012.
[15] D. Park. Concurrency and automata on infinite sequences. In Proceedings of the Fifth GI Conference
on Theoretical Computer Science, volume 104 of Lecture Notes in Computer Science, pages 167–183.
Springer-Verlag, 1981.
[16] G. Reißig. Computing abstractions of nonlinear systems. IEEE Transactions on Automatic Control,
56(11):2583–2598, November 2011.
[17] E. M. Stein and R. Shakarchi. Real analysis: measure theory, integration, and Hilbert spaces. Princeton
University Press, 2009.
[18] P. Tabuada. Verification and Control of Hybrid Systems: A Symbolic Approach. Springer, 2009.
[19] P. Tabuada and G. J. Pappas. Linear time logic control of discrete-time linear systems. IEEE Transactions on Automatic Control, 51(12):1862–1877, 2006.
[20] D. C. Tarraf. A control-oriented notion of finite state approximation. IEEE Transactions on Automatic
Control, 57(12):3197–3202, 2012.
[21] D. C. Tarraf. An input-output construction of finite state ρ/µ approximations for control design. IEEE
Transactions on Automatic Control, Special Issue on Control of Cyber-Physical Systems, 59(12):3164–
3177, Dec 2014.
[22] D. C. Tarraf, A. Megretski, and M. A. Dahleh. Finite approximations of switched homogeneous systems
for controller synthesis. IEEE Transactions on Automatic Control, 56(5):1140–1145, May 2011.
[23] K. Tsumura. Approximation of discrete time linear systems via bit length of memory. In Proceedings of
the 17th International Symposium on Mathematical Theory of Networks and Systems, pages 1066–1072,
Kyoto, Japan, July 2006.
[24] K. Tsumura. A lower bound of the necessary bit length of memory for approximation of discrete time
systems. In Proceedings of the 46th IEEE Conference on Decision and Control, pages 2241–2246, New
Orleans, LA, U.S.A., December 2007.
18
[25] A. J. van der Schaft. Equivalence of dynamical systems by bisimulation. IEEE Transactions on Automatic Control, 49(12):2160–2172, 2004.
[26] B. Yordanov, J. Tumová, I. Cerná, J. Barnat, and C. Belta. Temporal logic control of discrete-time
piecewise affine systems. IEEE Transactions on Automatic Control, 57(6):1491–1504, 2012.
19
| 3cs.SY
|
C ∗-algebras generated by projective representations
of free nilpotent groups
arXiv:1301.2942v2 [math.OA] 7 Jul 2016
Tron Ånen Omland
Updated version, July 2016
Abstract
We compute the two-cocycles (or multipliers) of the free nilpotent groups of class
2 and rank n and give conditions for simplicity of the corresponding twisted group
C ∗ -algebras. These groups are representation groups for Zn and can be considered as a
family of generalized Heisenberg groups with higher-dimensional center. Their group
1
C ∗ -algebras are in a natural way isomorphic to continuous fields over T 2 n(n−1) with the
noncommutative n-tori as fibers. In this way, the twisted group C ∗ -algebras associated
with the free nilpotent groups of class 2 and rank n may be thought of as “second order”
noncommutative n-tori.
Introduction
The discrete Heisenberg group may be described as the group generated by three elements
u1 , u2 , and v12 satisfying the commutation relations
[u1 , v12 ] = [u2 , v12 ] = 1
and
[u1 , u2 ] = v12 .
The group has received much attention in the literature, partly because it is one of the easiest
examples of a nonabelian torsion-free group. Moreover, the continuous Heisenberg group (see
below) is a connected nilpotent Lie group that arises in certain quantum mechanical systems.
As a natural consequence of this attention, several classes of generalized Heisenberg groups
have been investigated. For example, in [14, 15] Milnes and Walters describe the four and
five-dimensional nilpotent groups, and in [11, 12] Lee and Packer study the finitely generated
torsion-free two-step nilpotent groups with one-dimensional center.
In this paper, on the other hand, we will consider a family of generalized Heisenberg
groups, denoted by G(n) for n ≥ 2, with larger center. The groups G(n) are the so-called
free nilpotent groups of class 2 and rank n and will be defined properly in Section 1. Here we
also provide further motivation for our investigation of these groups. Inspired by the work of
Packer [20] we compute the second cohomology group H 2 (G(n), T) of G(n) and study the
structure of the twisted group C ∗ -algebras C ∗ (G(n), σ) associated with two-cocycles σ of
G(n).
Section 2 is devoted to two-cocycle calculations, where we decompose G(n) into a semidirect
product and apply techniques introduced by Mackey [13]. In particular, we will see that
1
H 2 (G(n), T) ∼
= T 3 (n+1)n(n−1) ,
Published in J. Operator Theory, 73(1):3-25, 2015.
2010 Mathematics Subject Classification: Primary 46L05; Secondary 20C25, 20F18, 20J06, 22D25.
Key words and phrases: free nilpotent group, projective unitary representation, twisted group C ∗ -algebra,
simplicity, two-cocycle, group cohomology, Heisenberg group, noncommutative n-torus.
1
and in Theorem 2.6 we give explicit formulas for the two-cocycles of G(n) up to similarity.
Next, in Section 3 we describe C ∗ (G(n), σ) as a universal C ∗ -algebra of a set of generators
and relations. Then we construct the algebra that in a natural way appear as a continuous
field over the compact space H 2 (G(n), T) with C ∗ (G(n), σ) as fibers. We also explain that
for n = 2, this algebra is the group C ∗ -algebra of the free nilpotent group of class 3 and rank
2.
In Section 4 we investigate the center of C ∗ (G(n), σ) and give conditions for simplicity of
these twisted group C ∗ -algebras in Theorem 4.4 and Corollary 4.6.
Finally, in Section 5 we study the automorphism group of G(n) and discuss isomorphism
invariants of C ∗ (G(n), σ) coming from Aut G(n).
Acknowledgements
This version is from July 2016 and includes several computations and details that were left
out in the published version of the article, in addition to a reformulation of Proposition 2.2
in terms of general semidirect products, and a new Remark 5.6 that makes a connection to
more recent work.
A comprehensive summary of this paper was part of the author’s PhD dissertation [18] at
the Norwegian University of Science and Technology (NTNU) submitted in May 2013.
The author would like to thank Erik Bédos, first for suggesting the problem of computing
the two-cocycles of the group G(3), and then for providing valuable comments throughout
this work. The author would also like to thank the referee of Journal of Operator Theory for
several valuable comments and suggestions.
This research was partially supported by the Research Council of Norway (NFR).
1
The free nilpotent groups G(n) of class 2 and rank n
For each natural number n ≥ 2, let G(n) be the group generated by elements {ui }1≤i≤n and
{vjk }1≤j<k≤n subject to the relations
[vjk , vlm ] = [ui , vjk ] = 1
and
[uj , uk ] = vjk
(1)
for 1 ≤ i ≤ n, 1 ≤ j < k ≤ n, and 1 ≤ l < m ≤ n. Clearly, G(2) is the usual discrete
Heisenberg group. For some purposes, it can be useful to set G(1) = hu1 i ∼
= Z. Note that
G(n) is generated by n + 21 n(n − 1) = 12 n(n + 1) elements.
The group G(n) is called the free nilpotent group of class 2 and rank n. Indeed, G(n) is a
free object on n generators in the category of nilpotent groups of step at most two. To see
this, note first that G(n) is the group generated by {ui }ni=1 subject to the relations that all
commutators of order greater than two involving the generators are trivial. Let G0 (n) be any
other nilpotent group of step at most two and let {u0i }ni=1 be any set of n elements in G0 (n).
Then there is a unique homomorphism from G(n) to G0 (n) that maps ui to u0i for 1 ≤ i ≤ n.
Of course, every free object on n generators in this category is isomorphic to G(n). For a
more extensive treatment of free nilpotent groups, see the article on Terence Tao’s website
[26] (see also 2. in the list below).
e
Furthermore, we will need the following concrete realization, say G(n),
of G(n). For
e
n ≥ 2, we denote the elements of G(n)
by
r = (r1 , . . . , rn , r12 , r13 , . . . , rn−1,n )1 ,
1 To be absolutely precise, the entries with double index are colexicographically ordered, that is, (i, j) < (k, l)
if j < l or if j = l and i < k.
2
where all entries are integers, and define multiplication by
r · s = (r1 + s1 , . . . , rn + sn ,
r12 + s12 + r1 s2 , r13 + s13 + r1 s3 , . . . , rn−1,n + sn−1,n + rn−1 sn ).
By letting ui have 1 in the i’th spot and 0 else and vjk have 1 in the jk’th spot and 0
else, the relations (1) are satisfied for these elements. Next, we define the map
e
G(n)
−→ G(n),
r
r12
n−1,n
r 7−→ v12
· · · vn−1,n
· urnn · · · ur11 ,
e
and then it is not difficult to see that G(n)
is isomorphic to G(n). Henceforth, we will not
e
distinguish between G(n) and the realization G(n)
just described, but this should cause no
confusion.
Denote by V (n) the subgroup of G(n) generated by the vjk ’s. Then V (n) coincides with
the center Z(G(n)) of G(n) and
1
V (n) = Z(G(n)) ∼
= Z 2 n(n−1) .
Indeed, both this and the next observations follow after noticing that
r · s · r−1 = (s1 , . . . , sn , s12 + r1 s2 − s1 r2 , . . . , sn−1,n + rn−1 sn − sn−1 rn ).
Moreover, consider the subgroups G(n − 1) and H(n) of G(n) defined by
G(n − 1) = hui , vjk : 1 ≤ i ≤ n − 1, 1 ≤ j < k ≤ n − 1i,
H(n) = hun , vjn : 1 ≤ j < ni.
Note that G(n − 1) sits inside G(n) as a subgroup and that H(n) ∼
= Zn is a normal subgroup
n
∼
∼
of G(n). Clearly, we have G(n)/V (n) = Z and G(n)/H(n) = G(n − 1). Therefore, there
are short exact sequences
1
V (n)
G(n)
Zn
1
and
1
H(n)
G(n)
G(n − 1)
1
where the second one splits and the first does not. In particular, G(n) is a central extension
1
of Zn by Z 2 n(n−1) and consequently, G(n) is a two-step nilpotent group.
To motivate our investigation of G(n), we present a few aspects about these groups and
some appearances in the literature.
1. Consider in the first place the continuous Heisenberg group. We will represent this group
in two different ways, Gmatrix and Gwedge , both with elements (x, x0 ) = (x1 , x2 , x0 ) ∈ R3 ,
i.e. x = (x1 , x2 ) ∈ R2 , and with multiplication as follows. For Gmatrix we define
(x1 , x2 , x0 )(y1 , y2 , y 0 ) = (x1 + y1 , x2 + y2 , x0 + y 0 + x1 y2 ) ,
and for Gwedge we set
(x1 , x2 , x0 )(y1 , y2 , y 0 ) = x1 + y1 , x2 + y2 , x0 + y 0 + 12 (x1 y2 − x2 y1 ) .
3
∼ Gwedge . To motivate the notation, note that Gmatrix
One can deduce that Gmatrix =
can be represented as matrix multiplication in M3 (R) if one identifies
1 x1 x0
(x1 , x2 , x0 ) ←→ 0 1 x2 ,
0 0
1
and that the multiplication in Gwedge may be written as
(x, x0 )(y, y 0 ) = x + y, x0 + y 0 + 21 (x ∧ y) .
In general, the wedge product on Rn is defined as a certain bilinear map (see e.g. [24,
p. 79])
V2 n
Rn × Rn →
(R ),
V2 n
V2 n
where
(R ) is a 12 n(n − 1)-dimensional real vector space. The elements of
(R )
are called bivectors and if {ei }ni=1 is a basis for Rn , then {ei ∧ ej }i<j is a basis for
V2 n
b R) with elements
(R ). For every n ≥ 2, define the group G(n,
V2 n
(x, x0 ) ∈ Rn ⊕
(R ), where x = (x1 , . . . , xn ), x0 = (x012 , x013 , . . . , x0n−1,n ),
and where multiplication is given by
(x, x0 )(y, y 0 ) = x + y, x0 + y 0 + 12 (x ∧ y) .
This group is of dimension n + 12 n(n − 1) = 12 n(n + 1). Remark especially that if n = 3,
the wedge product can be identified with the vector cross product on R3 . That is, the
b R) is given by
product in G(3,
(x, x0 )(y, y 0 ) = x + y, x0 + y 0 + 12 (x × y) .
b R) is isomorphic to the group consisting of the same
It is not hard to see that G(n,
elements, but with multiplication given by
(x, x0 )(y, y 0 ) = (x + y, x0 + y 0 + (x1 y2 , x1 y3 , . . . , xn−1 yn )) .
(2)
Let G(n, R) denote the group defined by (2). Then G(n) is the integer version of
G(n, R).
We also mention that Nielsen [17] has classified all the six-dimensional connected, simply
connected, nilpotent Lie groups. In this setting, G(3, R) is the group denoted by G6,15 .
2. One may define the free nilpotent group G(m, n) of class m and rank n for every
m ≥ 1. Indeed, G(m, n) is the group generated by {ui }ni=1 subject to the relations that
all commutators of order greater than m involving the generators are trivial. More
precisely, for m = 1, 2, 3 and n ≥ 2, we have that G(m, n) can be described as the
groups with presentations
G(1, n) = h{ui }ni=1 : [ui , uj ] = 1i ∼
= Zn ,
G(2, n) = h{ui }ni=1 : [[ui , uj ], uk ] = 1i = G(n),
G(3, n) =
h{ui }ni=1
(3)
: [[[ui , uj ], uk ], ul ] = 1i,
and it should now be clear how to define G(m, n) for all m ≥ 1 and n ≥ 2. Finally, we
set G(m, 1) = hu1 i ∼
= Z for each m ≥ 1. Moreover, for all m, n ≥ 1, the group G(m, n)
4
is the free object on n generators in the category of nilpotent groups of step at most m.
In particular, notice that G(m, n) is m-step nilpotent and that
∼ G(m + 1, n)/Z(G(m + 1, n)).
G(m, n) =
(4)
Again, we refer to [26] for additional details.
In [15, Section 4] Milnes and Walters describe the simple quotients of the C ∗ -algebra
associated with a five-dimensional group denoted by H5,4 . One can check that H5,4 is
isomorphic to the group G(3, 2). See Remark 3.2 for more about this group.
3. The group G(3) is briefly discussed by Baggett and Packer [4, Example 4.3]. The
purpose of that paper is to describe the primitive ideal space of group C ∗ -algebras of
some two-step nilpotent groups. However, G(3) only serves as an example of a group
the authors could not handle.
4. Let n ≥ 2. It is well-known that the group C ∗ -algebra A = C ∗ (G(n)) may be described as the universal C ∗ -algebra generated by unitaries {Ui }1≤i≤n and {Vjk }1≤j<k≤n
satisfying the relations
[Vjk , Vlm ] = [Ui , Vjk ] = I
and
[Uj , Uk ] = Vjk
for all 1 ≤ i ≤ n, 1 ≤ j < k ≤ n, and 1 ≤ l < m ≤ n.
1
For λ = (λ12 , λ13 , . . . , λn−1,n ) ∈ T 2 n(n−1) , let Aλ be the noncommutative n-torus. It is
the universal C ∗ -algebra generated by unitaries {Wi }ni=1 and relations [Wi , Wj ] = λij I
1
for 1 ≤ i < j ≤ n. The universal property of A gives that for each λ in T 2 n(n−1) there
is a surjective ∗ -homomorphism
πλ : A → Aλ
satisfying πλ (Ui ) = Wi for 1 ≤ i ≤ n and πλ (Vjk ) = λjk I for 1 ≤ j < k ≤ n.
Furthermore, A has center Z(A) = C ∗ ({Vjk }1≤j<k≤n ) ∼
= C ∗ (V (n)). Indeed, this is the
case since G(n) is amenable and its finite conjugacy classes are precisely the one-point
sets of central elements (see Lemma 4.1 below). Therefore, we set
[ = T 21 n(n−1) .
T = Prim Z(A) ∼
= Z(A)
1
Let λ be a primitive ideal of Z(A) identified with an element of T 2 n(n−1) . Let Iλ be the
ideal of A generated by λ, that is, the ideal generated by {Vjk − λjk I : 1 ≤ j < k ≤ n}.
It is clear that Iλ ⊆ ker πλ . By the universal property of Aλ , there is a ∗ -homomorphism
ρ : Aλ → A/Iλ
such that ρ(Wi ) = Ui + Iλ for 1 ≤ i ≤ n. Hence, ρ ◦ πλ coincides with the quotient
∼ A/Iλ and πλ may be
map A → A/Iλ and consequently, ker πλ ⊆ Iλ . Therefore, Aλ =
regarded as the quotient map A → A/Iλ .
F
For an element a of A, let ã be the section T → T Aλ given by ã(λ) = πλ (a) and let
e = {ã | a ∈ A} be the set of all such sections. Then the following can be deduced
A
from the Dauns-Hofmann Theorem [5].
e consisting of the base space T , C ∗ -algebras Aλ
Theorem 1.1. The triple (T, {Aλ }, A)
e is a full continuous field of C ∗ -algebras.
for each λ in T , and the set of sections A,
∗
Moreover, the C -algebra associated with this continuous field is naturally isomorphic
to A.
5
This result may be obtained as a corollary to [23, Theorem 1.2] which employs tools
of Williams [28] related to Fell bundle theory, by taking G = G(n) and σ = 1 in that
theorem. It is also a special case of [3, Corollary 2.3]. Our proof is more direct and
partly inspired by [1, Theorem 1.1] which covers the case where n = 2.
From the above discussion it now follows that G(n) is a representation group for Zn in
the sense of Moore [16]. In this case, that means G(n) is (up to isomorphism) the unique
central extension of Zn by H 2\
(Zn , T) such that the ordinary irreducible representation
theory of G(n) coincides with the projective irreducible representation theory of Zn .
This fact plays an important role in [6], where the noncommutative principal torus
bundles over locally compact spaces are classified up to equivariant Morita equivalence.
As explained in [6, Section 2], the group C ∗ -algebra of G(n) serves as a “universal”
bundle in this classification.
We refer to [7, Section 4] for more information on representation groups, where the
groups G(n, R) and G(n) are treated particularly in [7, Example 4.7].
2
The two-cocycles of the free nilpotent groups G(n)
Let G be any discrete group with identity e. A function σ : G × G → T satisfying
σ(r, s)σ(rs, t) = σ(r, st)σ(s, t)
σ(r, e) = σ(e, r) = 1
for all elements r, s, t ∈ G is called a two-cocycle of G with values in T (or a multiplier of G).
Moreover, two two-cocycles σ and τ are said to be similar, written σ ∼ τ , if
τ (r, s) = β(r)β(s)β(rs)σ(r, s)
for all r, s ∈ G and some function β : G → T. The set of similarity classes of two-cocycles of
G is an abelian group under pointwise multiplication. This group is the second cohomology
group H 2 (G, T).
Let G be a semidirect product of a normal subgroup H and a subgroup K. By properties
of the semidirect product, the elements of G can be uniquely written as products ab, where a
belongs to H and b belongs to K. Define the action α of K on H by αb (a) = bab−1 . One
often writes G = H oα K, but to simplify the notation, we will still denote the elements of G
by ab instead (a, b) and write the group product in G as (ab)(a0 b0 ) = aαb (a0 )bb0 for a, a0 ∈ H
and b, b0 ∈ K. Hopefully, the reader is familiar with semidirect products so that this does not
cause any confusion.
Next, we apply Mackey’s theorem [13, Theorem 9.4] and obtain the following result.
Theorem 2.1. Every two-cocycle of G is similar to a two-cocycle σ of G of the form
σ(a0 b, ab0 ) = σH (a0 , αb (a))g(a, b)σK (b, b0 ),
where σH and σK are two-cocycles of H and K, respectively,
g: H × K → T
6
(5)
is a function such that g(a, e) = g(e, b) = 1 for all a ∈ H, b ∈ K, and σH and g satisfy
g(aa0 , b) = σH (αb (a), αb (a0 ))σH (a, a0 ) · g(a, b)g(a0 , b),
g(a, bb0 ) = g(αb0 (a), b)g(a, b0 ).
(6)
Moreover, for every choice of σH , g, and σK satisfying the conditions above, σ is a two-cocycle
of G.
0
0
Proposition 2.2. Let (σH , g, σK ) and (σH
, g 0 , σK
) be triples satisfying the conditions of
0
Theorem 2.1 and let σ and σ be the corresponding two-cocycles of G. Then σ ∼ σ 0 if and
only if the following conditions hold:
0
(i) σK ∼ σK
,
(ii) There exists a function β : H → T such that
0
σH
(a, a0 ) = β(a)β(a0 )β(aa0 )σH (a, a0 ),
g 0 (a, b) = β(αb (a))β(a)g(a, b).
0
0
Remark 2.3. If (ii) holds, then σH ∼ σH
. If σH ∼ σH
and β and β 0 are two functions
implementing the similarity, then β 0 = f · β for some homomorphism f : H → T.
Proof of Proposition 2.2. Suppose σ ∼ σ 0 . Then there exists some γ : G → T such that
σ(a0 b, ab0 ) = γ(a0 b)γ(ab0 )γ(a0 bab0 )σ 0 (a0 b, ab0 )
(7)
for all a, a0 ∈ H and b, b0 ∈ K. In particular, if a = a0 = e, then
0
σK (b, b0 ) = γ(b)γ(b0 )γ(bb0 )σK
(b, b0 )
0
for all b, b0 ∈ K, so σK ∼ σK
. Moreover, the formula (5) from Theorem 2.1 with a = e and
b = e gives that
σ(a0 , b0 ) = 1 = σ 0 (a0 , b0 )
for all a0 ∈ H and b0 ∈ K. Applying this fact to (7) shows that γ(a0 b0 ) = γ(a0 )γ(b0 ) for all
a0 ∈ H and b0 ∈ K. Define β on H by β(a) = γ(a). Then, by letting b = b0 = e in (5) and
(7), we get
0
σH
(a0 , a) = β(a0 )β(a)β(a0 a)σH (a0 , a)
for all a0 , a ∈ H. Furthermore, by letting a0 = e and b0 = e in (5) and (7), we compute
g(a, b) = γ(b)γ(a)γ(ba)g 0 (a, b)
= γ(b)γ(a)γ(αb (a)b)g 0 (a, b)
= γ(b)γ(a)γ(αb (a))γ(b)g 0 (a, b)
= γ(a)γ(αb (a))g 0 (a, b)
= β(a)β(αb (a))g 0 (a, b)
for all a ∈ H and b ∈ K.
Assume next that β is such that (ii) holds, and that (i) holds through δ, that is,
0
σK (b, b0 ) = δ(b)δ(b0 )δ(bb0 )σK
(b, b0 ).
7
Define γ on G by γ(ab) = β(a)δ(b). Then
σ(a0 b, ab0 ) = σH (a0 , αb (a))g(a, b)σK (b, b0 )
0
(a0 , αb (a))
= β(a0 )β(αb (a))β(a0 αb (a))σH
0
· β(a)β(αb (a))g 0 (a, b) · δ(b)δ(b0 )δ(bb0 )σK
(b, b0 )
= β(a0 )δ(b) · β(a)δ(b0 ) · β(a0 αb (a))δ(bb0 )σ 0 (a0 b, ab0 )
= γ(a0 b)γ(ab0 )γ(a0 bab0 )σ 0 (a0 b, ab0 ).
Fix n ≥ 2. To compute the two-cocycles of G(n) up to similarity, we will proceed in the
following way. Consider G(n) as the split extension of G(n − 1) by H(n) as described in
Section 1. We will identify the elements
a = (0, . . . , 0, an , 0, . . . , 0, a1n , . . . , an−1,n ),
b = (b1 , . . . , bn−1 , 0, b12 , . . . , bn−2,n−1 , 0, . . . , 0),
of H(n) and G(n − 1), respectively, with ones of the form
a ←→ (an , a1n , . . . , an−1,n ),
b ←→ (b1 , . . . , bn−1 , b12 , . . . , bn−2,n−1 ).
The elements of G(n) will be written as products ab, where a belongs to H(n) and b belongs
to G(n − 1), and the action α of G(n − 1) on H(n) is then given by
αb (a) = bab−1 = (an , a1n + b1 an , . . . , an−1,n + bn−1 an ).
Remark 2.4. In the published version, Theorem 2.1 and Proposition 2.2 were only shown
to hold for G(n), not for any semidirect product.
Proposition 2.2 can be deduced from [23, Appendix 2], but in any case it may be useful
to give a proof by a direct computation.
Let τn be a two-cocycle of G(n) coming from a pair (σH(n) , gn ), that is,
τn (a0 b, ab0 ) = σH(n) (a0 , αb (a))gn (a, b),
(8)
where (σH(n) , gn ) satisfies (6). By Theorem 2.1 and Proposition 2.2, every two-cocycle of
G(n) that is trivial on G(n − 1) is similar to one of this form. Denote the abelian group of
e 2 (G(n), T).
similarity classes of two-cocycles of this type by H
Corollary 2.5. The second cohomology group of G(n) may be decomposed as
e 2 (G(n), T) ⊕ H 2 (G(n − 1), T) =
H 2 (G(n), T) = H
n
M
e 2 (G(k), T).
H
k=2
Proof. It follows from Theorem 2.1 and Proposition 2.2 (see our comment above) that
e 2 (G(n), T) ⊕ H 2 (G(n − 1), T).
H 2 (G(n), T) = H
Thus, the second inequality is proven by induction after noticing that
e 2 (G(1), T).
{1} = H 2 (Z, T) = H 2 (G(1), T) = H
8
Theorem 2.6. We have
H 2 (G(n), T) ∼
= T 3 (n+1)n(n−1) ,
1
and for each set of 31 (n + 1)n(n − 1) parameters
{λi,jk : 1 ≤ i ≤ k, 1 ≤ j < k ≤ n} ⊆ T,
the associated [σ] in H 2 (G(n), T) may be represented by
Y s r +s r s r +s (r r −r )
jk i
k ij
ij
ik j
k i j
σ(r, s) =
λi,jk
λj,ik
i<j<k
·
Y
r + 21 sk rj (rj −1) rk (sjk +rj sk )+ 21 rj sk (sk −1)
λk,jk
.
s
(9)
jk j
λj,jk
j<k
The proof of this theorem will be given in Section 2.1.
See the paragraph following Theorem 3.1 for an explanation of why λi,jk for i > k is not
involved in the above.
Example 2.7. For G(1) ∼
= Z there are no nontrivial two-cocycles. The two-cocycles of the
usual Heisenberg group G(2) are, up to similarity, given by two parameters (as computed in
[20, Proposition 1.1]):
r + 12 s2 r1 (r1 −1) r2 (s12 +r1 s2 )+ 12 r1 s2 (s2 −1)
λ2,12
s
12 1
σ(r, s) = λ1,12
(10)
The two-cocycles of G(3) are, up to similarity, given by eight parameters:
s
r +s3 (r1 r2 −r12 )
s23 r1 +s3 r12
13 2
σ(r, s) = λ1,23
λ2,13
s
r + 12 s2 r1 (r1 −1) r2 (s12 +r1 s2 )+ 12 r1 s2 (s2 −1)
λ2,12
s
r + 21 s3 r1 (r1 −1) r3 (s13 +r1 s3 )+ 12 r1 s3 (s3 −1)
λ3,13
s
r + 21 s3 r2 (r2 −1) r3 (s23 +r2 s3 )+ 12 r2 s3 (s3 −1)
λ3,23
12 1
· λ1,12
13 1
· λ1,13
23 2
· λ2,23
Remark 2.8. One may associate a Lyndon-Hochschild-Serre spectral sequence with the
extension (see e.g. [27, 6.8.2]):
1
V (n)
G(n)
Zn
1
By applying [10, Theorem 4] to this sequence, one can compute the second homology group
of G(n) (which is recently also done more generally for G(m, n) in [25, Proposition 2.1]), and
deduce that
∼ Z 13 (n+1)n(n−1) ,
H2 (G(n), Z) =
1
which gives that H 2 (G(n), T) ∼
= T 3 (n+1)n(n−1) after dualizing, using the universal coefficient
theorem for cohomology. However, this does not give an explicit description of H 2 (G(n), T).
2.1
Proof of Theorem 2.6
e 2 (G(n), T) through several lemmas and then use
We will in this proof first compute H
Corollary 2.5 to conclude the argument.
9
e 2 (G(n), T) may be represented by a pair (σH(n) , gn ),
Lemma 2.1.1. Every element of H
where σH(n) is a two-cocycle of H(n) given by
σH(n) (a0 , a) =
n−1
Y
a0 ain
λi n
(11)
(12)
i=1
for some λ1 , . . . , λn−1 ∈ T, and gn satisfies
gn (a + a0 , b) =
n−1
Y
b an a0n
λi i
gn (a, b)gn (a0 , b)
i=1
for all a, a0 ∈ H(n) and b ∈ G(n − 1).
e 2 (G(n), T) may be represented by a two-cocycle of the form (8),
Proof. Every element of H
that is, by a pair (σH(n) , gn ) satisfying (6).
Moreover, it is well-known (see e.g. [2]) that every two-cocycle of H(n) ∼
= Zn is similar to
one of the form
Y
Y
a0 akn
a0 a
σH(n) (a0 , a) =
λi n in ·
µjkjn
1≤i≤n−1
1≤j<k≤n−1
for some sets of scalars {λi }1≤i≤n−1 , {µjk }1≤j<k≤n−1 ⊆ T. Since H(n) is abelian, (6) gives
that
σH(n) (αb (a), αb (a0 ))σH(n) (a, a0 ) = gn (a + a0 , b)gn (a, b)gn (a0 , b)
= gn (a0 + a, b)gn (a0 , b)gn (a, b)
= σH(n) (αb (a0 ), αb (a))σH(n) (a0 , a)
for all a, a0 ∈ H(n) and b ∈ G(n − 1). Furthermore, we have
σH (αb (a), αb (a0 ))σH (a, a0 )
Y
a (a0 +b a0 )−a a0
=
λi n in i n n in ·
1≤i≤n−1
=
Y
1≤i≤n−1
Y
(a
µjkjn
+bj an )(a0kn +bk a0n )−ajn a0kn
1≤j<k≤n−1
b a a0
λi i n n
·
b a0kn an +bk ajn a0n +bj bk an a0n
Y
µjkj
.
1≤j<k≤n−1
This is equal to σH (αb (a0 ), αb (a))σH (a0 , a) for all a, a0 ∈ H(n) and b ∈ G(n − 1) if and only
if the expression remains unchanged under the substitution a ←→ a0 , that is, if and only if
all the µjk ’s are 1.
e 2 (G(n), T) there is a unique associated pair (σH(n) , gn )
Lemma 2.1.2. For every element of H
satisfying the conditions of Lemma 2.1.1 such that
gn (un , ui ) = 1
for all
1 ≤ i ≤ n − 1.
(13)
Proof. Suppose that (σH(n) , gn ) satisfies (11) and (12). Let f : H(n) → T be the homomorphism determined by f (un ) = 1 and f (vin ) = gn (un , ui ) for all 1 ≤ i ≤ n − 1 and define
gn0 by gn0 (a, b) = f (αb (a))f (a)gn (a, b). Then, gn0 (un , ui ) = 1 for all 1 ≤ i ≤ n − 1 and by
Proposition 2.2, (σH(n) , gn0 ) determines a two-cocycle of H(n) in the same similarity class as
the one coming from (σH(n) , gn ).
10
0
Suppose now that there are two pairs (σH(n) , gn ) and (σH(n)
, gn0 ) both satisfying the
0
conditions of Lemma 2.1.1. Then σH(n)
= σH(n) , so by Proposition 2.2 and the succeeding
remark, there is a homomorphism f : H(n) → T such that
gn0 (a, b) = f (αb (a))f (a)gn (a, b) =
n−1
Y
f (vin )an bi gn (a, b)
i=1
for all a ∈ H(n) and b ∈ G(n − 1). In particular,
gn0 (un , ui ) = f (vin )gn (un , ui ) for all
1 ≤ i ≤ n − 1,
so that gn0 = gn if gn0 (un , ui ) = gn (un , ui ) for all 1 ≤ i ≤ n − 1.
e 2 (G(n), T), and let (σH(n) , g) be the
In the forthcoming lemmas we fix an element of H
unique associated pair satisfying (11), (12), and (13) for some set of scalars {λi }n−1
i=1 ⊆ T.
For computational reasons, we introduce the following notation. For a = (an , a1n , . . . , an−1,n )
in H(n), we write a = w(a) + z(a), where w(a) = (an , 0, . . . , 0), and z(a) is the “central
part”, i.e. z(a) = (0, a1n , . . . , an−1,n ). Similarly, for b = (b1 , . . . , bn−1 , b12 , . . . , bn−2,n−1 )
in G(n − 1), we write b = w(b)z(b), where w(b) = (b1 , . . . , bn−1 , 0, . . . , 0) and z(b) =
(0, . . . , 0, b12 , . . . , bn−2,n−1 ). Note that αb (a) = a if either w(a) or w(b) is trivial, i.e. if
either a or b is central.
Lemma 2.1.3. For all a ∈ H(n) and b ∈ G(n − 1) we have
g(a, b) = g(w(a), w(b))g(w(a), z(b))g(z(a), w(b)).
Proof. It follows immediately from Lemma 2.1.1 that if a, a0 ∈ H(n) and w(a) or w(a0 ) is 0,
then
g(a + a0 , b) = g(a, b)g(a0 , b),
(14)
hence,
g(a, b) = g(w(a) + z(a), b) = g(w(a), b)g(z(a), b)
for all a ∈ H(n) and b ∈ G(n − 1). If b0 ∈ G(n − 1) and w(b0 ) = e, then b0 is central and
αb0 (a) = a for all a ∈ H(n). Therefore,
g(a, b)g(a, b0 ) = g(a, bb0 ) = g(a, b0 b) = g(αb (a), b0 )g(a, b)
(15)
for all a ∈ H(n), b ∈ G(n − 1). By (14), we then get
1 = g(αb (a), b0 )g(a, b0 )
= g(a + (0, b1 an , . . . , bn−1 an ), b0 )g(a, b0 )
= g(a, b0 )g((0, b1 an , . . . , bn−1 an ), b0 )g(a, b0 )
= g((0, b1 an , . . . , bn−1 an ), b0 )
for all a ∈ H(n) and b ∈ G(n − 1). Consequently, since this holds for all a ∈ H(n) and
b ∈ G(n − 1), and central b0 ∈ G(n − 1), we get that if ã and b̃ are any elements in H(n) and
G(n − 1), respectively, then g(z(ã), z(b̃)) = 1. Moreover, (15) also imply that if b, b0 ∈ G(n − 1)
and either w(b) or w(b0 ) is equal to e, that is, either b or b0 is central, then
g(a, bb0 ) = g(a, b)g(a, b0 ).
11
(16)
Hence, by (16) and (14),
g(a, b) = g(a, w(b)z(b)) = g(a, w(b))g(a, z(b))
= g(w(a), w(b))g(z(a), w(b))g(w(a), z(b)) · 1
for all a ∈ H(n) and b ∈ G(n − 1).
Lemma 2.1.4. For all a ∈ H(n) and b, b0 ∈ G(n − 1) we have
g(z(a), w(b)) =
n−1
Y
g(vin , uj )ain bj ,
i,j=1
g(w(a), z(b)) =
Y
g(un , vij )an bij =
g(a, bb0 ) =
an bij
,
g(vin , uj )g(vjn , ui )
1≤i<j≤n
1≤i<j≤n
and
Y
n−1
Y
0
g(vin , uj )bi bj an g(a, b)g(a, b0 ).
(17)
i,j=1
Proof. Let z(H(n)) = {z(a) | a ∈ H(n)} and z(G(n − 1)) = {z(b) | b ∈ G(n − 1)}. Then g is
a bihomomorphism when restricted to z(H(n)) × G(n − 1) or H(n) × z(G(n − 1)). Therefore,
the first two identities hold. Indeed, this follows directly from (6) after noticing that since
z(a) and z(b) are central,
αw(b) (z(a)) = z(a) and αz(b) (w(a)) = w(a).
Moreover, for i < j we have ui uj = vij uj ui . By (6) and the previous lemma, one calculates
g(un , ui uj ) = g(αuj (un ), ui )g(un , uj )
= g(un vjn , ui )g(un , uj )
= g(un , ui )g(vjn , ui )g(un , uj )
and
g(un , vij uj ui ) = g(un , vij )g(un , uj ui )
= g(un , vij )g(αui (un ), uj )g(un , ui )
= g(un , vij )g(un vin , uj )g(un , ui )
= g(un , vij )g(un , uj )g(vin , uj )g(un , ui ),
so that
g(vjn , ui ) = g(un , vij )g(vin , uj ),
which gives the last identity in the second line of the statement. Finally, we compute
g(a, bb0 ) = g(αb0 (a), b)g(a, b0 ) = g(a + (0, b01 an , . . . , b0n−1 an ), b)g(a, b0 )
= g((0, b01 an , . . . , b0n−1 an ), w(b))g(a, b)g(a, b0 )
n−1
Y
0
=
g(vin , w(b))bi an g(a, b)g(a, b0 )
i=1
=
n−1
Y n−1
Y
i=1
g(vin , uj )bj
b0i an
j=1
12
g(a, b)g(a, b0 ).
(18)
Lemma 2.1.5. For all a ∈ H(n) and b ∈ G(n − 1) we have
n−1
Y
g(w(a), w(b)) =
1
λi2
bi an (an −1)
1
g(vin , ui ) 2 an bi (bi −1)
i=1
Y
·
g(vin , uj )bi bj an .
1≤i<j≤n−1
Proof. First we see from (17) that if bj ≥ 1, then
b −1
b
g(un , ujj ) = g(un , ujj
uj )
b −1
= g(vjn , uj )bj −1 g(un , ujj
= · · · = g(vjn , uj )
1
2 bj (bj −1)
)g(un , uj )
g(un , uj )bj
and then it is not hard to see that
b
1
g(un , ujj ) = g(vjn , uj ) 2 bj (bj −1) g(un , uj )bj
for negative bj as well, for example by applying (17) again.
bn−1
Moreover, note that w(b) = un−1
· · · ub11 , so that by (17),
b
n−1
g(un , w(b)) = g(un , un−1
· · · ub11 )
n−1
Y
=
bn−1
g(v1n , uj )b1 bj g(un , un−1
· · · ub22 )g(un , ub11 )
j=2
Y
= ··· =
bi bj
g(vin , uj )
Y
n−1
b
g(un , ujj ) .
j=1
1≤i<j≤n−1
Then by (12) for an ≥ 1,
b
n−1
g(w(a), w(b)) = g(an un , un−1
· · · ub11 )
n−1
Y b (a −1)
bn−1
bn−1
λi i n
· g((an − 1)un , un−1
· · · ub11 )g(un , un−1
· · · ub11 )
=
i=1
= ··· =
n−1
Y
b · 12 an (an −1)
λi i
b
n−1
· g(un , un−1
· · · ub11 )an
i=1
=
n−1
Y
i=1
·
n−1
Y
1
2 bi an (an −1)
λi
·
Y
g(vin , uj )bi bj an
1≤i<j≤n−1
1
g(vjn , uj ) 2 an bj (bj −1) g(un , uj )an bj .
j=1
Again, it is not hard to see that a similar argument also works for negative an . Finally, recall
that we have chosen g so that g(un , uj ) = 1 by (13).
Lemma 2.1.6. We have
e 2 (G(n), T) ∼
H
= Tn(n−1) ,
13
and for each set of n(n − 1) parameters
{λi,jn : 1 ≤ i ≤ n, 1 ≤ j ≤ n − 1} ⊆ T,
e 2 (G(n), T) may be represented by
the associated [τ ] in H
Y
τ (a0 b, ab0 ) =
a
b +an bij
jn i
λi,jn
a
b +an (bi bj −bij )
in j
λj,in
·
a0 (a
jn
n
λn,jn
a
b + 21 an bj (bj −1)
jn j
λj,jn
j=1
1≤i<j≤n−1
n−1
Y
n−1
Y
+bj an )+ 21 bj an (an −1)
.
j=1
Proof. If one puts λi,jn = g(vjn , ui ) for i, j < n and λn,jn = λj for j < n, then this is a
consequence of the preceding lemmas. Indeed, by (8) we can represent τ as a pair (σH(n) , g).
Here σH(n) is of the form (11) and g can decomposed as in Lemma 2.1.3 with factors computed
in Lemma 2.1.4 and Lemma 2.1.5.
To complete the proof of Theorem 2.6, we set r =Q
a0 b and s = ab0 and recall that by
n
Corollary 2.5 we can compute σn inductively
as [σn ] = k=2 [τn ].
Pn
Finally, we can also check that k=2 k(k − 1) = 31 (n + 1)n(n − 1).
3
The twisted group C ∗ -algebras C ∗ (G(n), σ) of G(n)
Again, let G be any discrete group, σ a two-cocycle of G and H a nontrivial Hilbert space. A
map U from G into the unitary group of H satisfying
U (r)U (s) = σ(r, s)U (rs)
for all r, s ∈ G is called a σ-projective unitary representation of G on H.
We recall the following facts about twisted group C ∗ -algebras and refer to Zeller-Meier
[29] for further details of the construction.
To each pair (G, σ), we may associate the full twisted group C ∗ -algebra C ∗ (G, σ). Denote
the canonical injection of G into C ∗ (G, σ) by iσ . Then C ∗ (G, σ) satisfies the following
universal property. Every σ-projective unitary representation of G on some Hilbert space H
(or in some unital C ∗ -algebra A) factors uniquely through iσ .
The reduced twisted group C ∗ -algebra Cr∗ (G, σ) is generated by the left regular σ-projective
unitary representation λσ of G on B(`2 (G)). Consequently, λσ extends to a ∗ -homomorphism
of C ∗ (G, σ) onto Cr∗ (G, σ). If G is amenable, then λσ is faithful. Note especially that every
∼ C ∗ (G(n), σ) through λσ for every n ≥ 1
nilpotent group is amenable, so that C ∗ (G(n), σ) =
r
and all two-cocycles σ of G(n).
Finally, we remark that if τ ∼ σ through some function β : G → T, then the assignment
iτ (r) 7→ β(r)iσ (r) induces an isomorphism C ∗ (G, τ ) → C ∗ (G, σ).
Theorem 3.1 (Remark 3.1 in the published version). Fix n ≥ 2, let σ be a two-cocycle of
G(n) of the form (9), that is, determined by the 13 (n + 1)n(n − 1) parameters
{λi,jk : 1 ≤ i ≤ k, 1 ≤ j < k ≤ n} ⊆ T,
and set
λk,ij = λi,jk λj,ik
when 1 ≤ i < j < k ≤ n.
14
(19)
Then the twisted group C ∗ -algebra C ∗ (G(n), σ) is the universal C ∗ -algebra generated by
unitaries {Ui }1≤i≤n and {Vjk }1≤j<k≤n satisfying the relations
[Vjk , Vlm ] = I,
[Ui , Vjk ] = λi,jk I,
[Uj , Uk ] = Vjk
(20)
for 1 ≤ i ≤ n, 1 ≤ j < k ≤ n, and 1 ≤ l < m ≤ n.
Proof. Set Ui = iσ (ui ) and Vjk = iσ (vjk ) and note that (9) gives that σ(ui , vjk ) = λi,jk and
σ(vjk , ui ) = 1 for all 1 ≤ i ≤ n and 1 ≤ j < k ≤ n. Thus,
[Ui , Vjk ] = σ(ui , vjk )σ(vjk , ui )I = λi,jk I for all 1 ≤ i ≤ n, 1 ≤ j < k ≤ n.
Moreover, note that σ(ui , uj ) = 1 for all 1 ≤ i, j ≤ n and σ(vjk , vlm ) = 1 for all 1 ≤ j < k ≤ n
and 1 ≤ l < m ≤ n. Hence, it is clear that C ∗ (G(n), σ) is generated as a C ∗ -algebra by
unitaries satisfying (20).
Next, suppose that A is any C ∗ -algebra generated by a set of unitaries satisfying the
relations (20). For each r in G(n) we define the unitary Wr in A by
r
r12
n−1,n
Wr = V12
· · · Vn−1,n
· Unrn · · · U1r1 .
Then a computation using (20) repeatedly gives that2 Wr Ws = τ (r, s)Wrs , where τ (r, s) is
a scalar in T for all r, s ∈ G(n). Now, the associativity of A immediately implies that τ
is a two-cocycle of G(n), so that W is a τ -projective unitary representation of G(n) in A.
Furthermore, note that τ satisfies
τ (ui , vjk )τ (vjk , ui ) = λi,jk for 1 ≤ i ≤ n, 1 ≤ j < k ≤ n.
By the universal property of the full twisted group C ∗ -algebra, there exists a unique ∗ homomorphism ϕ of C ∗ (G(n), τ ) onto A such that ϕ(iτ (r)) = W (r) for all r ∈ G(n).
Therefore, it is sufficient to show that τ ∼ σ, because then C ∗ (G(n), τ ) is canonically
isomorphic to C ∗ (G(n), σ). By Theorem 2.6, there is some β : G(n) → T such that σ 0 , given
by
σ 0 (r, s) = β(r)β(s)β(rs)τ (r, s),
is of the form (9). We calculate that
σ 0 (ui , vjk )σ 0 (vjk , ui )
= β(ui )β(vjk )β(ui vjk )τ (ui , vjk )β(vjk )β(ui )β(vjk ui )τ (vjk , ui )
= τ (ui , vjk )τ (vjk , ui ) = λi,jk
for all 1 ≤ i ≤ n and 1 ≤ j < k ≤ n. Hence, σ 0 = σ, so τ ∼ σ.
The above relation (19) is a consequence of (18) in the proof of Theorem 2.6 and is
the reason why λi,jk for i > k is not involved in the expression (9). To illustrate this
further, consider the three-dimensional case. Let U1 , U2 , U3 and V12 , V13 , V23 be unitaries in
a C ∗ -algebra B satisfying
[Vjk , Vlm ] = I,
[Ui , Vjk ] = µi,jk I,
[Uj , Uk ] = Vjk
2 In general, it will require much work to compute the formula for τ and it is not needed for this argument.
However, for n = 2, the expression for τ is precisely of the form (10).
15
for 1 ≤ i ≤ 3, 1 ≤ j < k ≤ 3, and 1 ≤ l < m ≤ 3 where {µi,jk } is any set of nine scalars in T.
Then we can compute that
U1 U2 U3 = V12 U2 U1 U3 = · · · = µ2,13 V12 V13 V23 U3 U2 U1 ,
U1 U2 U3 = U1 V23 U3 U2 = · · · = µ1,23 µ3,12 V12 V13 V23 U3 U2 U1 ,
that is, we must have µ2,13 = µ1,23 µ3,12 .
n
For dimensions n > 3, any choice of a triple of unitaries from the family
{U }i=1 gives a
n
1
similar dependence. In the n · 2 n(n − 1) commutation relations, these 3 dependencies are
the only possible ones since
n · 21 n(n − 1) − n3 = 12 n(n − 1) n − 31 (n − 2) = 13 (n + 1)n(n − 1).
Remark 3.2. For n ≥ 2, let ω be the dual two-cocycle of G(n), that is,
1
\ T) ∼
ω : G(n) × G(n) → H 2 (G(n),
= Z 3 (n+1)n(n−1)
is determined by ω(r, s)(σ) = σ(r, s) for a two-cocycle σ of G(n). Let the group R(G(n)) be
1
defined as the set Z 3 (n+1)n(n−1) × G(n) with product
(j, r)(k, s) = (j + k + ω(r, s), rs).
It is not entirely obvious that ω and R(G(n)) are well-defined and we refer to [23, p. 689–690]
and [7, Section 4] for details on this and the fact that R(G(n)) is a representation group for
G(n). Moreover, according to [23, Corollary 1.3] we may construct a continuous field A over
H 2 (G(n), T) with fibers Aλ ∼
= C ∗ (G(n), σλ ) for each λ ∈ H 2 (G(n), T). Then the C ∗ -algebra
associated with this continuous field will be naturally isomorphic to the group C ∗ -algebra of
the group R(G(n)).
Next, we briefly consider the group G(3, 2) generated by u1 , u2 , v12 , w1 , w2 satisfying
[u1 , u2 ] = v12 ,
[u1 , v12 ] = w1 ,
[u2 , v12 ] = w2 ,
w1 , w2 central.
Then we have Z(G(3, 2)) ∼
= Z2 and Z(C ∗ (G(3, 2))) ∼
= C(T2 ).
The following statement can also be deduced from [23, Theorem 1.2 and Examples 1.4 (3)],
but we include the analysis that follows, because it is similar to that used in Theorem 1.1.
Let i denote the canonical injection of G(3, 2) into C ∗ (G(3, 2)). For each λ = (λ1 , λ2 ) ∈ T2 ,
let C ∗ (G(2), σλ ) be generated by unitaries satisfying (20). By a similar argument as in
Theorem 1.1, there is a surjective ∗ -homomorphism
πλ : C ∗ (G(3, 2)) → C ∗ (G(2), σλ )
such that i(ui ) = Ui , i(v12 ) = V12 , and i(wi ) = λi I for i = 1, 2. Moreover, the kernel of πλ
coincides with the ideal of C ∗ (G(3, 2)) generated by
λ ∈ Prim Z(C ∗ (G(3, 2))) ∼
(G(3, 2))) = T2 ∼
= Z(C ∗\
= H 2 (G(2), T).
Again, similarly as in Theorem 1.1, we define a set of sections and apply the DaunsHofmann Theorem. In this way, the triple
^ 2))
H 2 (G(2), T), {C ∗ (G(2), σλ )}λ , C ∗ (G(3,
is a full continuous field of C ∗ -algebras, and the C ∗ -algebra associated with this continuous
field is naturally isomorphic to C ∗ (G(3, 2)).
16
It is not difficult to see that R(G(2)) is isomorphic to G(3, 2). We conjecture that
R(G(n)) ∼
= G(3, n) also for n ≥ 3, where G(3, n) is the free nilpotent group of class 3
and rank n as described in (3), so that A is isomorphic to C ∗ (G(3, n)). For n ≥ 3, the
complicated part is to construct a homomorphism R(G(n)) → G(3, n), find an isomorphism
1
∼ Z(G(3, n)), and then use (4) to produce a commuting diagram:
Z 3 (n+1)n(n−1) =
1
1
Z 3 (n+1)n(n−1)
R(G(n))
∼
=
1
Z(G(3, n))
G(n)
1
=
G(3, n)
G(n)
1
In fact, [25, Proposition 2.2 and Remark 2.3] indicate that the representation group R(G(m, n))
for G(m, n) defined similarly as above may be isomorphic to G(m + 1, n) for all m, n ≥ 1.
4
Simplicity of the twisted group C ∗ -algebras C ∗ (G(n), σ)
Let σ be a two-cocycle of any group G. An element r of G is called σ-regular if σ(r, s) = σ(s, r)
whenever s in G commutes with r. If r is σ-regular, then every conjugate of r is also σ-regular.
Therefore, we say that a conjugacy class of G is σ-regular if it contains a σ-regular element.
Let n ≥ 2. The conjugacy class Cr of r ∈ G(n) is infinite if r ∈
/ V (n) = Z(G(n)). Indeed,
for any s ∈ G(n) we have
(srs−1 )i = ri and (srs−1 )jk = rjk + sj rk − rj sk .
Hence, |Cr | = ∞ if ri 6= 0 for some i. Of course, Cr = {r} if r ∈ V (n).
Now, we fix a two-cocycle σ of G(n) of the form (9).
Lemma 4.1. Let S(G(n)) be the set of σ-regular central elements of G(n), that is,
S(G(n)) = {r ∈ V (n) | σ(r, s) = σ(s, r) for all s ∈ G(n)}.
\
Then S(G(n)) is a subgroup of G(n) and Z(C ∗ (G(n)), σ) ∼
= C(S(G(n))).
Proof. It is not hard to check that S(G(n)) is a subgroup of V (n).
We identify C ∗ (G(n), σ) with Cr∗ (G(n), σ) ⊆ B(`2 (G)). Let δe in `2 (G) be the characteristic function on {e} and for an operator T in B(`2 (G)), set fT = T δe ∈ `2 (G). If T belongs
to the center of C ∗ (G(n), σ), then fT can be nonzero only on the finite σ-regular conjugacy
classes of G(n), that is, on S(G(n)) (see e.g. [19, Lemmas 2.3 and 2.4]).
Next, let C ∗ (S(G(n)), σ) be identified with {λσ (s) | s ∈ S(G(n))} ⊆ B(`2 (G)). This
means that Z(C ∗ (G(n), σ) ⊆ C ∗ (S(G(n)), σ). As the reverse inclusion obviously holds, we
have Z(C ∗ (G(n)), σ) = C ∗ (S(G(n)), σ).
Now, it is not difficult to see that C ∗ (S(G(n))) ∼
= C ∗ (S(G(n)), σ). Indeed, as s 7→ λσ (s)
∗
is a unitary representation of S(G(n)) into C (S(G(n)), σ) and the canonical tracial state τ
on C ∗ (S(G(n)), σ) is faithful and satisfies τ (λσ (s)) = 0 for each nonzero s ∈ S(G(n)), this is
just a consequence of [29, Théorème 4.22]. Altogether, we get
\
Z(C ∗ (G(n)), σ) = C ∗ (S(G(n)), σ) ∼
= C ∗ (S(G(n))) ∼
= C(S(G(n))).
17
Remark 4.2. If S(G(n)) is nontrivial, we can describe C ∗ (G(n), σ) as a continuous field of C ∗ \ The fibers will be isomorphic to C ∗ (G(n)/S(G(n)), ω)
algebras over the base space S(G(n)).
for some two-cocycle ω of G(n)/S(G(n)) (see [11, Theorem 1.1] and [23, Theorem 1.2] for
further details).
Example 4.3 ([11, Lemma 3.8 and Theorem 3.9]). Fix a two-cocycle σ of G(2) of the
form (10) such that both λ1,12 and λ2,12 are torsion elements. Let p and q be the smallest
natural numbers such that λp1,12 = λq2,12 = 1 and set k = lcm(p, q). Clearly, V (2) = Z and
S(G(2)) = kZ. Moreover, G(2)/S(G(2)) can be identified with the group with product
(r1 , r2 , r12 )(s1 , s2 , s12 ) = (r1 + s1 , r2 + s2 , r12 + s12 + r1 s2 mod kZ)
for r1 , r2 , s1 , s2 ∈ Z and r12 , s12 ∈ {0, 1, . . . , k − 1}.
\ ∼
Then C ∗ (G(2), σ) is a continuous field of C ∗ -algebras over the base space S(G(2))
= T.
The fibers will be isomorphic to C ∗ (G(n)/S(G(n)), ωλ ), where λ ∈ T and
ωλ (r, s) = σ(r, s)µr1 s2
for some µ ∈ T with µk = λ.
Characterizations for simplicity of of twisted group C ∗ -algebras of two-step nilpotent
groups have been given in [11, Corollary 1.4] and [23, Corollary 1.6]. For the groups G(n),
the necessary and sufficient conditions for simplicity are somewhat easier to provide.
Theorem 4.4. The following are equivalent:
(i) C ∗ (G(n), σ) is simple.
(ii) C ∗ (G(n), σ) has trivial center.
(iii) There are no nontrivial central σ-regular elements in G(n).
Proof. By [21, Theorem 1.7] C ∗ (G(n), σ) is simple if and only if every nontrivial σ-regular
conjugacy class of G(n) is infinite. Since every finite conjugacy class of G(n) is a one-point
set of a central element, then (i) is equivalent with (iii).
Moreover, (iii) is the same as saying that S(G(n)) is trivial, so therefore, (ii) is equivalent
with (iii) by Lemma 4.1. This also follows from [19, Theorem 2.7].
Lemma 4.5. A central element s = (0, . . . , 0, s12 , s13 , . . . , sn−1,n ) of G(n) is σ-regular if and
only if
Y
sjk
λi;jk
=1
1≤j<k≤n
for all 1 ≤ i ≤ n.
Proof. Clearly, an element s = (0, . . . , 0, s12 , s13 , . . . , sn−1,n ) ∈ V (n) is σ-regular if and only
if σ(s, r) = σ(r, s) for all r ∈ G(n). By a direct calculation from the cocycle formula (9), we
get that
Y
Y
Y
sjk ri sik rj
sjk rj rk sjk
−r s
rk sij
σ(r, s)σ(s, r) =
λi,jk
λj,ik
λj,jk
λk,jk
λi,jkk ij λj,ik
i<j<k
=
n
Y
i=1
j<k
Y
s
jk
λi,jk
i<j<k
ri
1≤j<k≤n
is equal to 1 for all r ∈ G(n) if and only if the inner parenthesis is 1 for each 1 ≤ i ≤ n.
18
Corollary 4.6. We have that C ∗ (G(n), σ) is simple if and only if for each nontrivial central
element s = (0, . . . , 0, s12 , s13 , . . . , sn−1,n ) there is some 1 ≤ i ≤ n such that
Y
sjk
λi,jk
6= 1.
1≤j<k≤n
∗
Example 4.7. In particular, C (G(3), σ) is simple if and only if for each nontrivial central
element s = (0, 0, 0, s12 , s13 , s23 ) at least one of the following hold:
12
13
23
λs1,12
λs1,13
λs1,23
6= 1,
12
13
23
λs2,12
λs2,13
λs2,23
6= 1,
12
13
23
λs3,12
λs3,13
λs3,23
6= 1.
Next, set λi,jk = e2πiti,jk for ti,jk ∈ [0, 1) and consider the n × 12 n(n − 1)-matrix T with
entries ti,jk in the corresponding spots. Then T induces a linear map
1
R 2 n(n−1) → Rn .
Corollary 4.8. Let T be the matrix described above. Then following are equivalent:
(i) C ∗ (G(n), σ) is simple
1
(ii) T −1 (Zn ) ∩ Z 2 n(n−1) = {0}
1
(iii) T (Z 2 n(n−1) \ {0}) ∩ Zn = ∅
Remark 4.9. Clearly, condition (ii) above is equivalent to that T restricts to an injective
map
1
Z 2 n(n−1) → Rn /Zn ∼
= Tn .
Furthermore, for 1 ≤ j < k ≤ n, define
Λjk = {ti,jk ∈ [0, 1), 1 ≤ i ≤ n | e2πiti,jk = λi,jk }
and for 1 ≤ i ≤ n, define
Λi = {ti,jk ∈ [0, 1), 1 ≤ j < k ≤ n | e2πiti,jk = λi,jk }.
Proposition 4.10. If there exists i such that all the elements of Λi are irrational and linearly
independent over Q, then C ∗ (G(n), σ) is simple.
Proof. It follows immediately from Lemma 4.5, that “equation i” cannot be satisfied unless
s = 0. Hence, no nontrivial σ-regular central elements exists.
Proposition 4.11. If there exists j < k such that Λjk consists of only rational elements,
then C ∗ (G(n), σ) is not simple.
Proof. Let q be the least common two-cocycle of the denominators of the elements of Λjk .
Then qvjk is central and σ-regular. Indeed,
σ(r, qvjk )σ(qvjk , r) =
n−1
Y
i
λqr
i,jk = 1
i=1
for all r ∈ G(n).
Remark 4.12. In the case where C ∗ (G(n), σ) is not simple, some more information about
the primitive ideal space can be deduced from [11, Proposition 1.3].
19
5
On isomorphisms invariants of C ∗ (G(n), σ)
Fix n ≥ 2 and let σ be a two-cocycle of G(n). If ϕ is an automorphism of G(n), define the
two-cocycle σϕ of G(n) by
σϕ (r, s) = σ(ϕ(r), ϕ(s)).
(21)
Then it is well-known that the associated twisted group C ∗ -algebras C ∗ (G(n), σ) and
C ∗ (G(n), σϕ ) are isomorphic. Indeed, the map
i(G,σ) (r) 7→ i(G,σϕ ) (ϕ−1 (r))
extends to an isomorphism C ∗ (G(n), σ) → C ∗ (G(n), σϕ ). Moreover, for any automorphism ϕ
of G(n), it is easily seen that two two-cocycles σ and τ of G(n) are similar if and only if σϕ
and τϕ are similar. Hence, there is a well-defined group action of the automorphism group
Aut G(n) on H 2 (G(n), T) defined by ϕ · [σ] = [σϕ ].
Therefore, we will now briefly investigate Aut G(n). Let V (n)n be the subgroup of
Aut G(n) consisting of the automorphisms G(n) → G(n) of the form ui 7→ zi ui for 1 ≤ i ≤ n
and elements zi ∈ V (n) = Z(G(n)). In particular, these automorphisms leave all the vjk ’s
fixed, i.e. V (n)n is the subgroup of Aut G(n) leaving V (n) fixed. Clearly, V (n)n contains
Inn G(n). In fact, in the case n = 2, we have V (2)2 = Inn G(2).
Proposition 5.1. There is a split short exact sequence:
1
V (n)n
Aut G(n)
GL(n, Z)
1
Proof. Assume that ϕ is any endomorphism G(n) → G(n). The image of a central element
under ϕ must be central, so ϕ restricts to an endomorphism ϕ1 : V (n) → V (n). Therefore,
ϕ also induces an endomorphism ϕ2 : G(n)/V (n) → G(n)/V (n) determined by ϕ2 (q(r)) =
q(ϕ(r)). Consider now the following commutative diagram:
1
V (n)
i
ϕ1
1
V (n)
G(n)
q
ϕ
i
G(n)
Zn
1
ϕ2
q
Zn
1
Assume that ϕ2 is an automorphism. Since ϕ2 is surjective, then for all ui there is some
si ∈ G(n) such that ϕ(si ) = zi ui for some zi ∈ V (n). Hence, for all j < k, we have
−1
n
ϕ1 (sj sk s−1
j sk ) = vjk and therefore, ϕ1 is surjective. Every surjective endomorphism of Z
is also injective, so ϕ1 is an automorphism as well. Thus, by the “short five lemma”, ϕ is an
automorphism.
The converse obviously holds and hence, ϕ is an automorphism if and only if ϕ2 is an
automorphism.
Furthermore, the construction of G(n) in terms of generators and relations means that
every endomorphism G(n) → G(n) is uniquely determined by its values at {ui }ni=1 . In
particular, we let ϕ : G(n) → G(n) be determined by the pair of matrices given by its entries
(ϕ(ui )j ), (ϕ(ui )jk ) ∈ Mn (Z) × Mn, 21 n(n−1) (Z)
so that the induced endomorphism ϕ2 is coming from a matrix in Mn (Z).
By the above argument, the map between endomorphism groups defined by
End G(n) → End Zn ,
(ϕ(ui )j ), (ϕ(ui )jk ) 7→ (ϕ(ui )j )
20
(22)
restricts to a surjective map Aut G(n) → Aut Zn = GL(n, Z).
Before concluding the argument, we need the following.
Lemma 5.2 (nested inside the proof). If ϕ and ϕ0 are two endomorphisms of G(n), then
(ϕ ◦ ϕ0 )(ui )j =
n
X
ϕ0 (ui )k ϕ(uk )j .
k=1
If ϕ and ϕ0 are two endomorphisms of G(n) that both induce the trivial map on G(n)/V (n),
then
(ϕ ◦ ϕ0 )(ui )jk = ϕ0 (ui )jk + ϕ(ui )jk .
Proof. For the moment, set ϕ(ui )j = rij and ϕ0 (ui )j = sij . Then
(ϕ ◦ ϕ0 )(ui ) = ϕ(usnin · · · us1i1 z) = (urnnn · · · ur1n1 )sin · · · (urn1n · · · ur111 )si1 z 0
for some elements z, z 0 ∈ V (n). Moreover, we can change the order of the ui ’s in the expression
just by replacing z 0 by another central element z 00 and thus,
(ϕ ◦ ϕ0 )(ui )j = rnj sin + rn−1,j si,n−1 + · · · + r1j si1 =
n
X
sik rkj .
k=1
If both ϕ2 and ϕ02 are trivial, then ϕ(ui ) = zi ui and ϕ0 (ui ) = zi0 ui for all 1 ≤ i ≤ n and some
elements zi , zi0 ∈ V (n). Hence, ϕ(vjk ) = ϕ0 (vjk ) = vjk for all j < k and thus,
(ϕ ◦ ϕ0 )(ui ) = ϕ(zi0 ui ) = zi0 zi ui .
Therefore, (22) restricts to a surjective homomorphism Aut G(n) → GL(n, Z) with kernel
isomorphic to the group Mn, 12 n(n−1) (Z) under addition, that is, to V (n)n .
Moreover, if A is a matrix in GL(n, Z) with entries aij , then one can define an automorphism ϕA of G(n) by ϕA (ui )j = aij . Thus, it should be clear that GL(n, Z) sits inside
Aut G(n) as a subgroup so that the sequence splits.
Proposition 5.3. If ϕ belongs to V (n)n , then σϕ is similar to σ. Thus, the action of V (n)n
on H 2 (G(n), T) given by (21) is trivial.
Proof. It is not hard to see that
σ(ui , vjk )σ(vjk , ui ) = σϕ (ui , vjk )σϕ (vjk , ui ),
that is,
[i(G,σ) (ui ), i(G,σ) (vjk )] = [i(G,σϕ ) (ui ), i(G,σϕ ) (vjk )]
for all 1 ≤ i ≤ n and 1 ≤ j < k ≤ n. It then follows from the universal property of
C ∗ (G(n), σ) described in Theorem 3.1 that σϕ ∼ σ.
To describe the GL(n, Z)-action on H 2 (G(n), T) requires more work. To any A in GL(n, Z)
we may associate a square matrix à of dimension 12 n(n − 1), with entries coming from the
determinant of all 2 × 2-matrices inside A. More precisely, if A = (aij ), Ã is given by entries
ãij,kl for i < j, k < l such that ãij,kl = aik ajl − ail ajk . Then A acts on the matrix T defined
prior to Corollary 4.8 by A · T = AT Ã. Tedious computations of commutation relations and
use of the universal property of C ∗ (G(n), σ) from Theorem 3.1 now lead to the following
result.
21
Proposition 5.4. Let σ and σ 0 be two-cocycles of G(n) of the form (9) and let T and T 0 be
the associated matrices of Corollary 4.8. If there exists a matrix A in GL(n, Z) such that
A · T = T 0 , then C ∗ (G(n), σ) and C ∗ (G(n), σ 0 ) are isomorphic.
For n = 2, it is shown by Packer [21, Theorem 2.9] that C ∗ (G(2), σ) and C ∗ (G(2), σ 0 ),
where σ and σ 0 are of the form (9), are isomorphic if and only if there is a GL(2, Z)-matrix
A taking σ to σ 0 . Note in this case that à = det A = ±1.
For n ≥ 3, it is at the moment not clear whether the GL(n, Z)-action on H 2 (G(n), T)
described above is such that the orbits represent different isomorphism classes of twisted group
C ∗ -algebras. Therefore, the problem of determining the isomorphism classes of C ∗ (G(n), σ)
remains open for future investigation.
σλ
Remark 5.5. Every two-cocycle σλ of G(n) is of the form e2πie
for some two-cocycle σ
eλ
on G(n, R). Moreover, any pair of two-cocycles σλ and σµ of the form described above are
homotopic in the sense of Packer and Raeburn [22, Section 4]. Hence, one may use [22,
Theorem 4.2 and Corollary 4.5] to deduce that
1
i+ n(n−1)
Ki (C ∗ (G(n)), σλ ) ∼
(G(n, R)/G(n)).
= Ki (C ∗ (G(n)), σµ ) ∼
= Ktop2
Remark 5.6 (new). The groups G(n), or more generally all the free nilpotent groups G(m, n)
of class m and rank n are finitely generated (torsion-free) nilpotent groups. Let σ be a
two-cocycle of G(m, n) and suppose that there are no nontrivial central σ-regular elements
in G(m, n), that is, C ∗ (G(m, n), σ) is simple with a unique trace by [20] (see Theorem 4.4
and the subsequent results above for the case of G(n)). Then there exists an irreducible
representation πσ of G(m + 1, n) such that C ∗ (G(m, n), σ) ∼
= C ∗ (πσ (G(m + 1, n))) by using
Theorem 1.1, Remark 3.2, and (4). Thus it follows from [9] and [8] that C ∗ (G(m, n), σ) is
classified by its ordered K-theory within the class of simple, unital, nuclear C ∗ -algebras with
finite nuclear dimension that satisfy the universal coefficient theorem.
References
[1] Joel Anderson and William L. Paschke. The rotation algebra. Houston J. Math.,
15(1):1–26, 1989.
[2] Nigel B. Backhouse and Christopher J. Bradley. Projective representations of space
groups. I. Translation groups. Q. J. Math., 21:203–222, 1970.
[3] Lawrence W. Baggett and Judith A. Packer. C ∗ -algebras associated to two-step nilpotent
groups. In Selfadjoint and nonselfadjoint operator algebras and operator theory, Contemp.
Math., vol. 120, pp. 1–6, Amer. Math. Soc., Providence, RI 1991.
[4] Lawrence W. Baggett and Judith A. Packer. The primitive ideal space of two-step
nilpotent group C ∗ -algebras. J. Funct. Anal., 124(2):389–426, 1994.
[5] John Dauns and Karl Heinrich Hofmann. Representation of rings by sections. Mem.
Amer. Math. Soc., vol. 83, Amer. Math. Soc., Providence, RI 1968.
[6] Siegfried Echterhoff, Ryszard Nest, and Herve Oyono-Oyono. Principal non-commutative
torus bundles. Proc. Lond. Math. Soc., 99(1):1–31, 2009.
[7] Siegfried Echterhoff and Dana P. Williams. Locally inner actions on C0 (X)-algebras. J.
Operator Theory, 45(1):131–160, 2001.
22
[8] Caleb Eckhardt and Elizabeth Gillaspy. Irreducible representations of nilpotent groups
generate classifiable C ∗ -algebras. Münster J. Math., to appaer.
[9] Caleb Eckhardt and Paul McKenney. Finitely generated nilpotent group C ∗ -algebras
have finite nuclear dimension. J. Reine Angew. Math., to appear.
[10] Yuri V. Kuz’min and Yuri S. Semenov. On the homology of a free nilpotent group of
class 2. Mat. Sb., 189(4):49–82, 1998.
[11] Soo Teck Lee and Judith A. Packer. Twisted group C ∗ -algebras for two-step nilpotent
and generalized discrete Heisenberg groups. J. Operator Theory, 34(1):91–124, 1995.
[12] Soo Teck Lee and Judith A. Packer. The cohomology of the integer Heisenberg groups.
J. Algebra, 184(1):230–250, 1996.
[13] George W. Mackey. Unitary representations of group extensions. I. Acta Math., 99:265–
311, 1958.
[14] Paul Milnes and Samuel G. Walters. Simple quotients of the group C ∗ -algebra of a
discrete 4-dimensional nilpotent group. Houston J. Math., 19(4):615–636, 1993.
[15] Paul Milnes and Samuel G. Walters. Simple infinite-dimensional quotients of C ∗ (G) for
discrete 5-dimensional nilpotent groups G. Illinois J. Math., 41(2):315–340, 1997.
[16] Calvin C. Moore. Extensions and low dimensional cohomology theory of locally compact
groups. II. Trans. Amer. Math. Soc., 113:64–86, 1964.
[17] Ole A. Nielsen. Unitary representations and coadjoint orbits of low-dimensional nilpotent
Lie groups. Queen’s Papers Pure Appl. Math., vol. 63, Queen’s Univ., Kingston, ON
1983.
[18] Tron Ånen Omland. On the structure of certain C ∗ -algebras arising from groups.
Doctoral thesis, Norwegian University of Science and Technology (NTNU), 2013.
[19] Tron Ånen Omland. Primeness and primitivity conditions for twisted group C ∗ -algebras.
Math. Scand., 114(2):299–319, 2014.
[20] Judith A. Packer. C ∗ -algebras generated by projective representations of the discrete
Heisenberg group. J. Operator Theory, 18(1):41–66, 1987.
[21] Judith A. Packer. Twisted group C ∗ -algebras corresponding to nilpotent discrete groups.
Math. Scand., 64(1):109–122, 1989.
[22] Judith A. Packer and Iain Raeburn. Twisted crossed products of C ∗ -algebras. II. Math.
Ann., 287(4):595–612, 1990.
[23] Judith A. Packer and Iain Raeburn. On the structure of twisted group C ∗ -algebras.
Trans. Amer. Math. Soc., 334(2):685–718, 1992.
[24] Michael D. Spivak. Calculus on manifolds. A modern approach to classical theorems of
advanced calculus. W. A. Benjamin, Inc., New York-Amsterdam 1965.
[25] Markus Szymik. Twisted homological stability for extensions and automorphism groups
of free nilpotent groups. J. K-Theory, 14(1):185–201, 2014.
[26] Terence Tao. The free nilpotent group. http://terrytao.wordpress.com/2009/12/
21/the-free-nilpotent-group/
23
[27] Charles A. Weibel. An introduction to homological algebra. Cambridge Stud. Adv. Math.,
vol. 38, Cambridge Univ. Press, Cambridge 1994.
[28] Dana P. Williams. The structure of crossed products by smooth actions. J. Aust. Math.
Soc., 47(2):226-235, 1989.
[29] Georges Zeller-Meier. Produits croisés d’une C ∗ -algèbre par un groupe d’automorphismes.
J. Math. Pures Appl., 47:101–239, 1968.
Department of Mathematical Sciences, Norwegian University of Science and Technology
(NTNU), NO-7491 Trondheim, Norway.
24
| 4math.GR
|
arXiv:1608.02085v1 [math.GR] 6 Aug 2016
AN AXIOMATIZABLE PROFINITE GROUP WITH
INFINITELY MANY OPEN SUBGROUPS OF INDEX 2
OR BEN PORATH AND MARK SHUSTERMAN
Abstract. We show that a profinite group with the same first-order
theory as the direct product over all odd primes p of the dihedral group
of order 2p, is necessarily isomorphic to this direct product.
1. Introduction
We say that a profinite group Γ is axiomatizable if for every profinite
group Λ with the same first-order theory as that of Γ, we have Λ ∼
= Γ.
The study of axiomatizable profinite groups began in [3], where it is shown
that finitely generated profinite groups are axiomatizable, and an example
of a profinite group that is not axiomatizable is given (for instance, Zℵ2 0 ).
More generally, it is shown in [2] that a strongly complete profinite group
(that is, a profinite group all of whose finite index subgroups are open) is
axiomatizable. Are there more axiomatizable profinite groups?
By [2, Corollary 3.8] a strongly complete profinite group is small (that
is, it has only finitely many open subgroups of index n, for every n ∈ N).
Thus, a possible precise formulation given by [2, Question 3.15 (i)] for the
question we have just raised is whether every axiomatizable profinite group
is small. Here, a negative answer to this question is given.
Theorem 1.1. The profinite group G given by the direct product of the
dihedral groups Dp over all odd primes p, is axiomatizable.
From our proof of Theorem 1.1 one can extract an explicit (infinite) set
of axioms characterizing G up to an isomorphism.
2. The group G and its properties
Definition 2.1. For a profinite group Γ we set
(2.1)
Inv(Γ) ··= {τ ∈ Aut(Γ) | τ 2 = IdΓ }
and note that this defines a group if Aut(Γ) is abelian.
Definition 2.2. For a prime number p we denote by Cp the cyclic group of
order p. If p is odd, then Aut(Cp ) ∼
= Cp−1 so Inv(Cp ) is a group isomorphic
to C2 . The semidirect product Cp ⋊ Inv(Cp ) is the dihedral group Dp . Let
ρp be a generator of Cp , and let ǫp be a generator of Inv(Cp ) so that they
generate Dp and we have ǫp ρp ǫp = ρ−1
p .
1
2
OR BEN PORATH AND MARK SHUSTERMAN
Proposition 2.3. For an odd prime p we have Cp = {[a, b] | a, b ∈ Dp }.
Proof. For one inclusion note that Dp /Cp is abelian, and for the other one
take some ρnp ∈ Cp . As p is odd, there exists a k ∈ Z such that 2k ≡ n (p).
k
−k
k k
2k
n
We find that [ǫp , ǫp ρkp ] = ǫp ǫp ρkp ǫp ρ−k
p ǫp = ρp ǫp ρp ǫp = ρp ρp = ρp = ρp .
Remark 2.4. A similar argument shows that Cp ≤ Dp is its own centralizer.
Q
Q
Q
Definition 2.5. We set G ··=
Dp , C ··=
Cp , E ··= hǫp i where the
products (here and in the sequel) are always taken over all odd primes p.
Since Cp ⊳ Dp and Dp /Cp ∼
= C2 for all odd primes p, we conclude that C
is a closed normal procyclic subgroup of G with G/C ∼
= (Z/2Z)ℵ0 . Hence,
∀g ∈ G g 2 ∈ C
(2.2)
and G/C is not small. Therefore, G is not small as well. Furthermore, we
have G = C ⋊ E. Since the Sylow subgroups of C are normal, we see that
Y
Y
Cp−1
(2.3)
Aut(C) =
Aut(Cp ) ∼
=
p
p
is an abelian group, so
(2.4)
Inv(C) =
Y
Inv(Cp ) = E.
p
Thus,
(2.5)
G∼
= C ⋊ Inv(C).
Definition 2.6. For a profinite group Γ we denote by Γ′ its profinite commutator, which is the closed subgroup of Γ generated by {[a, b] | a, b ∈ Γ}.
It follows from Proposition 2.3 that
(2.6)
G′ = C = {[a, b] | a, b ∈ G}
so the following first-order sentence holds in G
(2.7)
∀a, b, c, d ∃r, s [a, b][c, d] = [r, s].
Remark 2.7. By (2.2), (2.6) the following first-order sentence is valid in G
(2.8)
∀g ∃h, k g 2 = [h, k].
Remark 2.8. It follows from Remark 2.4 that C is its own centralizer in G.
In view of (2.6), this is tantamount to the following first-order sentence
!
∀y, z x[y, z]x−1 = [y, z] −→ ∃a, b x = [a, b] .
(2.9)
∀x
Remark 2.9. Fix an odd prime p. We can think of ǫp , ρp as elements of E, C
respectively, and thus also as elements of G. The first-order sentence
(2.10)
∃x ∀y, z x[y, z]x−1 = [y, z]−1 ←→ [y, z]p = 1
holds in G since we can take x = ǫp .
AN AXIOMATIZABLE PROFINITE GROUP WITH INFINITELY MANY OPEN SUBGROUPS OF INDEX 2
3
Definition 2.10. We say that profinite groups Γ, Λ are elementarily equivalent if they have the same first-order theory, and denote this by Γ ≡ Λ.
3. The proof of Theorem 1.1
e be a profinite group for which G ≡ G.
e Set
Let G
(3.1)
e ··= {[g, h] | g, h ∈ G}
e
C
e of the compact space G
e2 under the
and note that it is the image in G
2
e
e
e is
continuous map sending (g, h) ∈ G to [g, h] ∈ G. It follows that C
e By (2.7) C
e is a subgroup of G.
e
compact, and thus closed in G.
e
Let us now show that C ≡ C. For that take a first-order sentence ϕ that
holds in C. For every variable x that appears in ϕ, replace each appearance
of Qx by Qx1 , x2 where Q ∈ {∀, ∃} and x1 , x2 are new variables. Furthermore, replace any instance of x in any atomic formula in ϕ by [x1 , x2 ] and
denote the resulting first-order sentence by ψ. It follows from (2.6) that ψ
e By (3.1), ϕ holds in C
e as required.
holds in G, and thus also in G.
e and by (2.8) every element of G/
e C
e is of order
By [3, Theorem A], C ∼
=C
∼
e
e
e
e
dividing 2, so |G/C| is prime to |C| = |C|. Since C = C is abelian and
e the action by conjugation of G
e on C
e gives rise to a continuous
normal in G,
e
e
e
homomorphism τ : G/C → Inv(C). By the Schur-Zassenhaus theorem (see
[1, Lemma 22.10.1]) we get that
(3.2)
e∼
e ⋊τ G/
e C.
e
G
=C
By Remark 2.8 τ is injective, and in order to see that it is also surjective,
e with Inv(C). A generating set for Inv(C)
e is thus given
first identify Inv(C)
by {ǫp }p . By Remark 2.9 the image of τ contains ǫp for each odd prime p,
so τ is a surjection, and thus an isomorphism. We conclude that
(3.3)
2.5
3.2
e ⋊ Inv(C)
e ∼
e ⋊τ G/
e C
e ∼
e
G ∼
= C ⋊ Inv(C) ∼
=C
=C
= G.
Acknowledgments
We would like to sincerely thank Arno Fehm for telling us about the
question that motivated this work, and for many helpful discussions.
References
[1] M. D. Fried, M. Jarden, Field Arithmetic, Third Edition, revised by M. Jarden,
Ergebnisse der Mathematik, 3, 11, Springer, Heidelberg, 2008.
[2] P. Helbig, On small profinite groups, preprint, 2015.
[3] M. Jarden, A. Lubotzky, Elementary equivalence of profinite groups, Bull. London
Math. Soc. 40 (2008), 887-896.
4
OR BEN PORATH AND MARK SHUSTERMAN
Raymond and Beverly Sackler School of Mathematical Sciences, Tel-Aviv
University, Tel-Aviv, Israel
E-mail address: [email protected]
Raymond and Beverly Sackler School of Mathematical Sciences, Tel-Aviv
University, Tel-Aviv, Israel
E-mail address: [email protected]
| 4math.GR
|
Computing All Distinct Squares in Linear Time for Integer
Alphabets
Hideo Bannai1 , Shunsuke Inenaga1 , and Dominik Köppl2
arXiv:1610.03421v2 [cs.DS] 20 Feb 2017
1
Department of Informatics, Kyushu University, Japan,
[email protected], [email protected]
2
Department of Computer Science, TU Dortmund, Germany,
[email protected]
February 21, 2017
Abstract
Given a string on an integer alphabet, we present an algorithm that computes the set of all distinct
squares belonging to this string in time linear to the string length. As an application, we show how
to compute the tree topology of the minimal augmented suffix tree in linear time. Asides from that,
we elaborate an algorithm computing the longest previous table in a succinct representation using
compressed working space.
1
Introduction
A square is a string of the form SS, where S is some non-empty string. It is well-known that a string of
length n contains at most n2 /4 squares. This bound is the number of all squares, i.e., we count multiple
occurrences of the same square, too. If we consider the number of all distinct squares, i.e., we count
exactly one occurrence of each square, then it becomes linear in n: The first linear upper bound was
given by Fraenkel and Simpson [17] who proved that a string of length n contains at most 2n distinct
squares. Later, Ilie [25] showed the slightly improved bound of 2n − Θ(lg n). Recently, Deza et al. [9]
refined this bound to ⌊11n/6⌋. In the light of these results one may wonder whether future results will
“converge” to the upper bound of n: The distinct square conjecture [17, 26] is that a string of length n
contains at most n distinct squares; this number is known to be independent of the alphabet size [33].
However, there still is a big gap between the best known bound and the conjecture. While studying a
combinatorial problem like this, it is natural to think about ways to actually compute the exact number.
This article focuses on a computational problem on distinct squares, namely, we wish to compute
(a compact representation of) the set of all distinct squares in a given string. Gusfield and Stoye [22]
tackled this problem with an algorithm running in O(nσT ) time, where σT denotes the number of different
characters contained in the input text T of length n. Although its running time is optimal O(n) for a
constant alphabet, it becomes O n2 for a large alphabet since σT can be as large as O(n).
We present an algorithm (Sec. 4) that computes this set in O(n) time for a given string of length n
over an integer alphabet of size nO(1) . Like Gusfield and Stoye, we can use the computed set to decorate
1
the suffix tree with all squares (Sec. 6). As an application, we provide an algorithm that computes
the tree topology of the minimal augmented suffix tree [1] in linear time (Sec. 7). The fastest known
algorithm computing this tree topology takes O(n lg n) time [4].
For our approach, we additionally need the longest previous factor table [18, 7]. As a side result of
independent interest, we show in Sec. 3 how to store this table in 2n + o(n) bits, and give an algorithm
that computes it using compressed working space.
2
Definitions
Our computational model is the word RAM model with word size Ω(lg n) for some natural number n.
Let Σ denote an integer alphabet of size σ = |Σ| = nO(1) . An element w in Σ∗ is called a string, and |w|
denotes its length. We denote the i-th character of w with w[i], for 1 ≤ i ≤ |w|. When w is represented
by the concatenation of x, y, z ∈ Σ∗ , i.e., w = xyz, then x, y and z are called a prefix, substring and
suffix of w, respectively. For any 1 ≤ i ≤ j ≤ |w|, let w[i..j] denote the substring of w that begins at
position i and ends at position j in w.
The longest common prefix (LCP) of two strings is the longest prefix shared by both strings.
The longest common extension (LCE) query asks for the longest common prefix of two suffixes of
the same string. The time for an LCE query is denoted by tLCE .
A factorization of a string T is a sequence of non-empty substrings of T such that the concatenations
of the substrings is T . Each substring is called a factor.
In the rest of this paper, we take a string T of length n > 0, and call it the text. We assume that
T [n] = $ is a special character that appears nowhere else in T , so that no suffix of T is a prefix of
another suffix of T . We further assume that T is read-only; accessing a character costs constant time.
We sometimes need the reverse of T , which is given by the concatenation T [n − 1] · · · T [1] · T [n] =
T [n − 1] · · · T [1]$.
The suffix tree of T is the tree obtained by compacting the trie of all suffixes of T ; it has n leaves and
at most n internal nodes. The leaf corresponding to the i-th suffix is labeled with i. Each edge e stores
a string that is called the edge label of e. The string label of a node v is defined as the concatenation
of all edge labels on the path from the root to v; the string depth of a node is the length of its string
label.
SA and ISA denote the suffix array and the inverse suffix array of T , respectively [32]. The access
time to an element of SA is denoted by tSA . LCP is an array such that LCP[i] is the length of the
longest common prefix of T [SA[i]..n] and T [SA[i − 1]..n] for i = 2, . . . , n. For our convenience, we define
LCP[1] := 0.
A range minimum query (RMQ) asks for the smallest value in an integer array for a given range.
There are data structures that can answer an RMQ on an integer array of length n in constant time while
taking 2n+o(n) bits of space [15]. An LCE query for the suffixes T [s..n] and T [t..n] can be answered with
an RMQ data structure on LCP with the range [min(ISA[s], ISA[t]) + 1.. max(ISA[s], ISA[t])] in constant
time.
A bit vector is a string on a binary alphabet. A select query on a bit vector asks the position of
the i-th ‘0’ or ‘1’ in the bit vector. There is a data structure that can be built in O(n) time with O(n)
bits of working space such that it takes o(n) bits on top of the bit vector, and can answer a select query
in constant time [5].
We identify occurrences of substrings with their position and length in the text, i.e., if x is a substring
2
i
1
2
3
4
5
6
7
8
9
10
11
12
T
a
b
a
b
a
a
a
b
a
b
a
$
SA
12
11
5
6
9
3
7
1
10
4
8
2
LCP
0
0
1
2
1
3
3
5
0
2
2
4
PLCP
5
4
3
2
1
2
3
2
1
0
0
0
LPF
0
0
3
2
1
2
5
4
3
2
1
0
Figure 1:
The
arrays SA, LCP,
PLCP and LPF
of the running
example.
of T , then there is a 1 ≤ i ≤ n and a 0 ≤ ℓ ≤ n − i + 1 such that T [i..i + ℓ − 1] = x. In the
following, we will represent the occurrences of substrings by tuples of position and length. When storing
these tuples in a set, we call the set distinct, if there are no two tuples (i, ℓ) and (i′ , ℓ) such that
T [i..i + ℓ − 1] = T [i′ ..i′ + ℓ − 1]. A special kind of substring is a square: A square is a string of the form
SS for S ∈ Σ+ ; we call S and |S| the root and the period of the square SS, respectively. Like with
substrings, we can generate a set containing some occurrences of squares. A set of all distinct squares
is a distinct set of occurrences of squares that is maximal under inclusion.
To compute a set of all distinct squares is the main focus of this paper. We will tackle this problem
theoretically in Sec. 4, and practically in Sec. 5. Finally, we give two applications of this problem in
Sec. 6 and Sec. 7. But before all that, we start with the study of the LPF array needed for our approach
computing all distinct squares:
3
A Compact Representation of the LPF Array
The longest previous factor table LPF of T is formally defined as
LPF[j] := max {ℓ | there exists an i ∈ [1..j − 1] such that T [i, i + ℓ − 1] = T [j, j + ℓ − 1]} .
It is useful for computing the Lempel-Ziv factorization of T = f1 · · · fz , which is defined as fi =
Pi−1
T [k..k + max(1, LPF[k])] with k := j=1 |fj | + 1 for 1 ≤ i ≤ z.
0 0 3 2 1 2 5 4 3 2 1 0
In the following, we will use the text T = ababaaababa$ as our running example whose LPF array is
represented by the small numbers above the characters. The Lempel-Ziv factorization of T is given by
1
2
3
4
5
6
a|b|aba|aa|baba|$, where the small numbers denote the factor indices, and the vertical bars denote the
factor borders. Fig. 1 shows SA, LPF and other used array data structures of our running example.
Corollary 3.1. Given LPF, we can compute the Lempel-Ziv factorization in O(n) time. If the factorization consists of z factors, the factorization can be represented by an array of z lg n bits, where the x-th
entry stores the beginning of the x-th factor.
Alternatively, it can be represented by a bit vector of length n in which we mark the factor beginnings.
A select data structure on top of the bit vector can return the length and the position of a factor in
constant time.
Since we will need LPF in Sec. 4, we are interested in the time and space bounds for computing LPF.
We start with the (to the best of our knowledge) state of the art algorithm with respect to time and
space requirements.
Lemma 3.2 ([8, Theorem 1]). Given SA and LCP, we can compute LPF in O(ntSA ) time. Besides the
output space of n lg n bits, we only need constant working space.
3
Apart from this algorithm, we are only aware of some practical improvements [36, 28].
Let us consider the size of LCP needed in Lem. 3.2. Sadakane [37] showed a 2n + o(n)-bits representation of LCP. Thereto he stores the array PLCP defined as PLCP[SA[i]] = LCP[i] in a bit vector in the
following way (also described in [13]): Since PLCP[1]+ 1, PLCP[2]+ 2, . . . , PLCP[n]+ n is a non-decreasing
sequence with 1 ≤ PLCP[1] + 1 ≤ PLCP[n] + n = n (PLCP[i] ≤ n − i since the terminal $ is a unique
character in T ) the values I[1] := PLCP[1] and I[i] := PLCP[i] − PLCP[i − 1] + 1 (2 ≤ i ≤ n) are nonnegative. By writing I[i] in the unary code 0I[i] 1 to a bit vector S subsequently for each 2 ≤ i ≤ n, we
Pn
can compute PLCP[i] = select1 (S, i) − 2i and LCP[i] = select1 (S, SA[i]) − 2SA[i]. Moreover, i=1 I[i] ≤ n
and therefore S is of length at most 2n.
By using Sadakane’s LCP-representation, we get LPF with the algorithm of Crochemore et al. [8] in
the following time and space bounds:
Corollary 3.3. Having SA and LCP stored in n lg n bits (this allows tSA = O(1)) and 2n + o(n) bits,
respectively, we can compute LPF with O(lg n) additional bits of working space (not counting the space
for LPF) in O(n) time.
By plugging in a suffix array construction algorithm like the in-place construction algorithm by Li
et al. [31], we get the bounds shown in Fig. 2 (since we can build LCP in-place after having SA [23]).
Although this result seems compelling, this approach stores SA and LPF in plain arrays (the former
for getting constant time access). In the following, we will show that the LPF array can be stored more
compactly. We start with a new representation of LPF, for which we use the same trick as for PLCP due
to the following property (which is crucial for squeezing PLCP into 2n + o(n) bits).
Lemma 3.4. n − j ≥ LPF[j] ≥ LPF[j − 1] − 1 for 2 ≤ j ≤ n.
Proof. There is an i with 1 ≤ i < j − 1 such that T [i..i + LPF[j − 1] − 1] = T [j − 1..j − 1 + LPF[j − 1] − 1].
Hence T [i + 1..i + LPF[j − 1] − 1] = T [j..j − 1 + LPF[j − 1] − 1].
We conclude that the sequence LPF[1] + 1, LPF[2] + 2, . . . , LPF[n] + n is non-decreasing with 1 ≤
LPF[1] + 1 ≤ LPF[n] + n ≤ n. We immediately get:
Corollary 3.5. LPF can be represented by a bit vector with a select data structure such that accessing
an LPF value can be performed in constant time. The data structures use 2n + o(n) bits.
To get a better working space bound, we have to come up with a new algorithm since the algorithm
of Lem. 3.2 creates a plain array to get constant time random write-access for computing the entries of
LPF. To this end, we present two algorithms that compute LPF in this representation with the aid of the
suffix tree. The two algorithms are derivatives of the algorithms [30, 16] that compute the Lempel-Ziv
factorization, either in O(n lg lg σ) time using O(n lg σ) bits, or in O n/ǫ2 time using (1+ǫ)n lg n+O(n)
bits, for a constant 0 < ǫ ≤ 1. The current bottleneck of both algorithms is the suffix tree implementation
with respect to space and time. Due to current achievements [35, 31], the algorithms now run in O(n)
time using O(n lg σ) bits, or in O(n/ǫ) time using (1 + ǫ)n lg n + O(n) bits, respectively.
We aim at building the LPF-representation of Cor. 3.5 directly such that we do not need to allocate
the plain LPF array using n lg n bits in the first place. To this end we create a bit vector of length 2n
and store the LPF values in it successively. In more detail, we follow the description of the Lempel-Ziv
factorization algorithms presented in [30, 16]. There, the algorithms are divided into several passes. In
each pass we successively visit leaves in text order (determined by the labels of the leaves). To compute
LPF, we only have to do a single pass. Similarly to the first passes of the two Lempel-Ziv algorithms,
4
algorithm
time
working space
|LPF|
Lem. 3.2,[8]
O(ntSA )
|SA| + |LCP| + O(lg n)
n lg n
Cor. 3.3,[31, 23]
O(n)
n lg n + 2n + O(lg n)
n lg n
Lem. 3.6,[30]
O(n/ǫ)
(1 + ǫ)n lg n + O(n)
2n + o(n)
Lem. 3.6,[16]
O(ntSA )
O(n lg σ)
2n + o(n)
Figure 2: Algorithms
computing
LPF;
space is counted in
bits.
The output
space |LPF| is not
considered as working
space. 0 < ǫ ≤ 1 is a
constant.
we use a bit vector BV to mark already visited internal nodes. On visiting a leaf we climb up the tree
until reaching the root or an already marked node. In the former case (we climbed up to the root) we
output zero. In the latter case, we output the string depth of the marked node. By doing so, we have
computed LPF[1..j] after having processed the leaf with label j.
Lemma 3.6. We can compute LPF in O(ntSA ) time with O(n lg σ) bits of working space, or in O(n/ǫ)
time using (1 + ǫ)n lg n + O(n) bits of working space, for a constant 0 < ǫ ≤ 1. Both variants include
the space of the output in their working spaces.
Proof. Computing the string depth of a node needs access to an RMQ data structure of LCP, and an
access to SA. Both accesses can be emulated by the compressed suffix array in tSA time, given that we
have computed PLCP in the above representation.
4
Computing the Set of All Distinct Squares
Given a string T , our goal is to compute all distinct squares of T . Thereto we return a set of pairs, where
each pair (s, ℓ) consists of a starting position s and a length ℓ such that T [s..s + ℓ − 1] is the leftmost
occurrence of a square. The size of this set is linear due to
Lemma 4.1 (Fraenkel and Simpson [17]). A string of length n can contain at most 2n distinct squares.
We follow the approach of Gusfield and Stoye [22]. Their idea is to compute a set of squares (the set
stores pairs of position and length like described in Sec. 2)1 with which they can generate all distinct
squares. They call this set of squares a leftmost covering set. A leftmost covering set obeys the
property that every square of the text can be constructed by right-rotating a square of this set. A
square (k, ℓ) is constructed by right-rotating a square (i, ℓ) with i ≤ k iff each tuple (i + j, ℓ) with
1 ≤ j ≤ k − i represents a square T [i + j..i + ℓ + j − 1] = T [i + j..i + ℓ − 1]T [i..i + j − 1].
The set of the leftmost occurrences of all squares is a set of all distinct squares. Unfortunately, the
leftmost covering set computed in [22] is not necessarily a set of all distinct squares since (a) it does not
have to be distinct, and (b) a square might be missing that can be constructed by right-rotating a square
of the computed leftmost covering set.
For illustration, the squares of our running example T = ababaaababa$ are highlighted with bars.
The set of all squares is {(1, 4), (2, 4), (5, 2), (6, 2), (7, 4), (8, 4)}. If we take the leftmost occurrences of
all squares, we get {(1, 4), (2, 4), (5, 2)}; this set comprises all squares marked by the solid bars, i.e., the
dotted bars correspond to occurrences of squares that are not leftmost. In this example, the dotted
bars form the set {(6, 2), (7, 4), (8, 4)}, which is a set of all distinct squares. A leftmost covering set is
{(1, 4), (5, 2)}.
1 It differs to the set we want to compute by the fact that they allow, among others, occurrences of the same square in
their set.
5
q
ℓL ℓR ℓ ℓR
L
fx
p
q
ℓL
fx+1
fx−1
ℓR ℓ
p
L
ℓR
fx
Figure 3: Search for squares on Lempel-Ziv borders. The left image corresponds to squares of type
Lem. 4.2(a), the right image to the type Lem. 4.2(b). Given two adjacent factors, we determine a
position q that is p positions away from the border (the direction is determined by the type of square
we want to search for). By two LCE queries we can determine the lengths ℓL and ℓR that indicate the
presence of a square if ℓL + ℓR ≥ p.
Our goal is to compute the set of all leftmost occurrences directly by modifying the algorithm of [22].
To this end, we briefly review how their approach works: They compute their leftmost covering set by
examining the borders between all Lempel-Ziv factors f1 · · · fz = T . That is because of
Lemma 4.2 ([22, Theorem 5]). The leftmost occurrence of a square T [i..i + 2p − 1] touches at least two
Lempel-Ziv factors. Let fx be the factor that contains the center of the square i + p − 1. Then either
(a) the square has its left end (position i) inside fx and its right end (position i + 2p − 1) inside fx+1 ,
or
(b) the left end of the square extends into fx−1 (or even further left). The right end can be contained
inside fx or fx+1 .
Having a data structure for computing LCE queries on the text and on its inverse, they can probe
at the borders of two consecutive factors whether there is a square. Roughly speaking, they have to
check at most |fx | + |fx+1 | many periods at the borders of every two consecutive factors fx and fx+1
Pz
due to the above lemma. This gives x=1 tLCE (|fx | + |fx+1 |) = O(ntLCE ) time, during which they can
compute a leftmost covering set L. Fig. 3 visualizes how the checks are done. Applying the algorithm
on our running example will yield the set L = {(1, 4), (5, 2), (7, 4)}. To transform this set into a set of
all distinct squares, their algorithm runs the so-called Phase II that uses the suffix tree. It begins with
computing the locations of the squares belonging to a subset L′ ⊆ L in the suffix tree in O(n) time. This
subset L′ is still guaranteed to be a leftmost covering set. Finally, their algorithm computes all distinct
squares of the text by right-rotating the squares in L′ . In their algorithm, the right-rotations are done
by suffix link walks over the suffix tree. Their running time analysis is based on the fact that each node
has at most σT incoming suffix links, where σT denotes the number of different characters occurring in
the text T . Given that the number of distinct squares is linear, Phase II runs in O(nσT ) time.
In the following, we will present our modification of the above sketched algorithm. To speed up
the computation, we discard the idea of using the suffix links for right-rotating squares (i.e., we skip
Phase II completely). Instead, we compute a list of all distinct squares directly. To this end, we show
a modification of the sketched algorithm such that it outputs this list sorted first by the lengths (of the
squares), and second by the starting position.
First, we want to show that we can change the original algorithm to output its leftmost covering
set in the above described order. To this end, we iterate over all possible periods, and search not
yet reported squares at all Lempel-Ziv borders, for each period. To achieve linear running time, we
want to skip a factor fx when the period becomes longer than |fx | + |fx+1 |. We can do this with an
array Z of z lg z bits that is zero initialized. When the period p becomes longer than |fx | + |fx+1 |, we
write Z[x] ← min {y > x : |fy | + |fy+1 | ≥ p} such that Z[x] refers to the next factor whose length is
6
sufficiently large. By doing so, if Z[x] 6= 0, we can skip all factors fy with y ∈ [x..Z[x] − 1] in constant
time. This allows us running the modified algorithm still in linear time.
We have to show that the modified algorithm still computes the same set. To this end, let us fix
the period p (over which we iterate in the outer loop). By [22, Lemma 7], processing squares satisfying
Lem. 4.2(a) before processing squares satisfying Lem. 4.2(b) (all squares have the same period p) produces
the desired output for period p.
Finally, we show the modification that computes all distinct squares (instead of the original leftmost
covering set). On a high level, we use an RMQ data structure on LPF to filter already found squares. The
filtered squares are used to determine the leftmost occurrences of all squares by right-rotation. In more
detail, we modify Algorithm 1 of [22] by filtering the squares in the following way (see Algorithm 1): For
each period p, we use a bit vector B marking the beginning positions of all found squares with period p.
On reporting a square, we additionally mark its starting position in B. By doing so, an invariant of the
algorithm below is that all right-rotated squares of a marked square are already reported.
Let us assume that we are searching for the leftmost occurrences of all squares whose periods are
equal to p. Given the starting position s of a square returned by [22, Algorithm 1], we consider the
square (s, 2p) and its right-rotations as candidates of our list: If B[s] = 1, then this square and its
right-rotations have already been reported. Otherwise, we report (s, 2p) if LPF[s] < 2p. In order to
find the leftmost occurrences of all not yet reported right-rotated squares efficiently, we first compute
the rightmost position e of the repetition of period p containing the square (s, 2p) by an LCE query.
Second, we check the interval I := [s + 1.. min(s + p − 1, e − 2p + 1)] for the starting positions of the
squares whose LPF values are less than 2p. To this end, we perform an RMQ query on LPF to find the
position j whose LPF value is minimal in I. If j > 2p, then all squares with period p in the considered
range have already been found, i.e., there is no leftmost occurrence of a square with the period p in this
range. Otherwise, we report (j, 2p) and recursively search for the text position with the minimal LPF
value within the intervals [s + 1..j − 1] and [j + 1.. min(s + p − 1, e − 2p + 1)]. In overall, the time of
the recursion is bounded by twice of the number of distinct squares starting in the interval I, since a
recursion step terminates if it could not report any square.
Theorem 4.3. Given an LCE data structure with tLCE access time and LPF, we can compute all distinct
squares in O(ntLCE + occ) = O(ntLCE ) time, where occ is the number of distinct squares.
Proof. We show that the returned list is the list of all distinct squares. No square occurs in the list
twice since we only report the occurrence of a square (i, ℓ) if LPF[i] < ℓ. Assume that there is a square
missing in the list; let (i, ℓ) be its leftmost occurrence. There is a square (j, ℓ) reported by the (original)
algorithm [22] such that i − ℓ/2 < j ≤ i and right-rotating (j, ℓ) yields (i, ℓ). Since we right-rotate all
found squares, we obviously have reported (j, ℓ).
The occ term in the running time is dominated by the ntLCE term due to Lem. 4.1.
5
Practical Evaluation
We have implemented the algorithm computing the leftmost occurrences of all squares in C++11 [29]. The
primary focus was on the execution time, rather than on a small memory footprint: We have deliberately
chosen plain 32-bit integer arrays for storing all array data structures like SA, LCP and LPF. These data
structures are constructed as follows: First, we generate SA with divsufsort [34]. Subsequently, we
generate LCP with the Φ-algorithm [27], and LPF with the simple algorithm of [8, Proposition 1]. Finally,
we use the bit vector class and the RMQ data structure provided by the sdsl-lite library [20]. In practice,
7
collection
pc-dblp.xml
pc-dna
pc-english
pc-proteins
pc-sources
pcr-cere
pcr-einstein.en
pcr-kernel
pcr-para
σ
maxi LCP[i]
avgLCP
z
maxx |fx |
maxx |fx fx+1 |
|occ|
time
97
17
226
26
231
6
125
161
6
1084
97,979
987,770
45,704
307,871
175,655
935,920
2,755,550
72,544
44
60
9390
278
373
3541
45,983
149,872
2268
7,035,342
13,970,040
13,971,134
20,875,097
11,542,200
1,446,793
49,575
774,532
1,926,563
1060
97,966
987,766
45,703
307,871
175,643
906,995
2,755,550
70,680
1265
97,982
1,094,108
67,809
307,884
185,362
1,634,034
2,755,556
73,735
7412
132,594
13,408
3,108,339
339,818
47,081
18,192,737
9258
37,391
70
310
2639
245
792
535
3953
6608
265
Table 1: Practical evaluation of the algorithm computing all distinct squares on the datasets described in
Sec. 5. Execution time is in seconds. It is the median of several conducted experiments, whose variance
in time was small. We needed approx. 5.73 GB of RAM for each instance. The expression avgLCP is the
average of all LCP values, and z is the number of Lempel-Ziv factors.
it makes sense to use an RMQ only for very large LCP values and periods (i.e., RMQs on LPF) due to
its long execution time. For small values, we naively compared characters, or scanned LPF linearly.
We ran the algorithm on all 200MiB collections of the Pizza&Chili Corpus [12]. The Pizza&Chili
Corpus is divided in a real text corpus with the prefix pc, and in a repetitive corpus with the prefix pcr. The experiments were conducted on a machine with 32 GB of RAM and an Intel Xeon
R
R
CPU E3-1271 v3. The operating system was a 64-bit version of Ubuntu Linux 14.04 with the kernel
version 3.13. We used a single execution thread for the experiments. The source code was compiled
using the GNU compiler g++ 6.2.0 with the compile flags -O3 -march=native -DNDEBUG.
Table 1 shows the running times of the algorithm on the described datasets. It looks like that
large factors tend to slow down the computation, since the algorithm has to check all periods up to
maxx (|fx | + |fx+1 |). This seems to have more impact on the running time than the number of LempelZiv factors z.
6
Decorating the Suffix Tree with All Squares
Gusfield and Stoye described a representation of the set of all distinct squares by a decoration of the
suffix tree, like the highlighted nodes (additionally annotated with its respective square) shown in the
suffix tree of our running example below. This representation asks for a set of tuples of the form (node,
length) such that each square T [i..i + ℓ − 1] is represented by a tuple (v, ℓ), where v is the highest node
whose string label has T [i..i + ℓ − 1] as a (not necessarily proper) prefix.
We show that we can compute this set of tuples
1
in linear time by applying the Phase II algorithm dea
$
b
scribed in Sec. 4 to our computed set of all distinct
12
3
a
a
$
b
5,aa
squares. The Phase II algorithm takes a list Li storing
11 5
14
a
a b
$ a
b
squares starting at text position i, for each 1 ≤ i ≤ n.
5 6
8
10 4
a
$ a
b
Each of these lists has to be sorted in descending order
17,baba
9 3
17
a
$ a
with respect to the squares’ lengths. It is easy to adapt
11,abab
11
8 2
$ a
our algorithm to produce these lists: On reporting a
7 1
square (i, ℓ), we insert it at the front of Li . By doing so, we can fill the lists without sorting, since we
iterate over the period length in the outer loop, while we iterate over all Lempel-Ziv factors in the inner
loop.
Finally, we can conduct Phase II. In the original version, the goal of Phase II was to decorate the suffix
tree with the endpoints of a subset of the original leftmost covering set. We will show that performing
8
exactly the same operations with the set of the leftmost occurrences of all squares will decorate the suffix
tree with all squares directly. In more detail, we first augment the suffix tree leaf having label i with the
list Li , for each 1 ≤ i ≤ n. Subsequently, we follow Gusfield and Stoye [22] by processing every node of
the suffix tree with a bottom-up traversal. During this traversal we propagate the lists of squares from
the leaves up to the root: An internal node u inherits the list of the child whose subtree contains the
leaf with the smallest label among all leaves in the subtree rooted at u. If the edge to the parent node
contains the ending position of one or more squares in the list (these candidates are stored at the front
of the list), we decorate the edge with these squares, and pop them off from the list. By [22, Theorem
8], there is no square of the set L′ (defined in Sec. 4) neglected during the bottom-top traversal. The
same holds if we exchange L′ with our computed set of all distinct squares:
Lemma 6.1. By feeding the algorithm of Phase II with the above constructed lists Li containing the
leftmost occurrences of the squares starting at the text position i, it will decorate the suffix tree with all
distinct squares.
Proof. We adapt the algorithm of Sec. 4 to build the lists Li . These lists contain the leftmost occurrences
of all squares. In the following we show that no square is left out during the bottom-up traversal. Let us
take a suffix tree node u with its children v and w. Without loss of generality, assume that the smallest
label among all leaves contained in the subtree of v is smaller than the label of every leaf contained
in w’s subtree. For the sake of contradiction, assume that the list of w contains the occurrence of a
square (i, ℓ) at the time when we pass the list of v to its parent u. The length ℓ is smaller than v’s string
depth, otherwise it would already have been popped off from the list. But since v’s subtree contains a
leaf whose label j is the smallest among all labels contained in the subtree of w, the square occurs before
at T [j..j + ℓ − 1] = T [i..i + ℓ − 1], a contradiction to the distinctness.
This concludes the correctness of the modified algorithm. We immediately get:
Theorem 6.2. Given LPF, an LCE data structure on the reversed text, and the suffix tree of T , we can
decorate the suffix tree with all squares of the text in O(ntLCE ) time. Asides from these data structures,
we use (occ + n) lg n + z lg z + min(n + o(n) , z lg n) + O(lg n) bits of working space.
Proof. We need (occ + n) lg n bits for storing the lists Li (occ lg n bits for storing the lengths of all
squares in an integer array, and n lg n bits for the pointers to the first element of each list). An LCE
query on the text can be answered by the string depth of a lowest common ancestor in the suffix tree;
most representations can answer string depth and lowest ancestor queries in constant time. The array Z
uses z lg z bits. The Lempel-Ziv factors are represented as in Cor. 3.1.
Corollary 6.3. We can compute the suffix tree and decorate it with all squares of the text in O(n/ǫ)
time using (3n + occ + 2nǫ) lg n + z lg z + O(n) bits, for a constant 0 < ǫ ≤ 1.
Proof. We use Lem. 3.6 to store SA, ISA, LCP, and LPF in (1 + ǫ)n lg n + O(n) bits. Subsequently, we
build an RMQ data structure on LCP such that LCE queries can be answered in constant time. We
additionally need the suffix array, its inverse, and the LCP array (with an RMQ data structure) of the
reversed text to answer LCE queries on the reversed text. Finally, we equip LPF with an RMQ data
structure for the right-rotations.
The values in the lists (i.e., the lengths of the squares starting at a specific position) can be stored in
Pn
Elias-Fano coding [11, 10]. If the list Li stores mi elements, then 2occ + i=1 (mi ⌈lg(n/mi )⌉) + o(occ)
bits are needed to represent the content of all lists. It is easy to implement the popping of the first
9
value from a list with this representation, given that we store an offset value and a pointer to the current
beginning of each list.
As an application, we consider the common squares problem: Given a set of non-empty strings with
a total length n, we want to find all squares that occur in every string in O(n) time. We solve this
problem by first decorating the generalized suffix tree built on all strings with the distinct squares of all
strings. Subsequently, we apply the O(n) time solution of Hui [24] that annotates each internal suffix
tree node v with the number of strings that contain v’s string label. This solves our problem since we
can simply report all squares corresponding to nodes whose string labels are found in all strings. This
also solves the problem asking for the longest common square of all strings in O(n) time, analogously to
the longest common substring problem [21].
Finally, the last section is dedicated to another application of our suffix tree decoration:
7
Computing the Tree Topology of the MAST in Linear Time
A modification of the suffix tree is the minimal aug-
$
12
a
$
11 15
a b
5 6
a
73
71
b
a
14
4
mented suffix tree (MAST) [1]. This tree can answer the number of the non-overlapping occurrences of
b
a substring in T . To this end, it adds some nodes on
4
a
$ a
b
the unary paths of the suffix tree, and augments each
28
10 4
a
$ a
b
internal node with the number of the non-overlapping
9 3
17
2
a
$ a
occurrences of its string label, like in the left tree (the
11
2
8 2
a
$
leaves are shown with their suffix number, each leaf rep7 1
resents a substring that occurs exactly once). The newly created nodes obey the property that the stored
numbers of the MAST nodes on the path from a leaf to the root are strictly increasing. Given a pattern
of length m, the MAST can answer the number of the non-overlapping occurrences of the pattern in
O(m) time. To this end, we traverse the MAST from the root downwards while reading the pattern from
the edge labels. If there is a mismatch, the pattern cannot be found in the text. Otherwise, we end at
reading the label of an edge (u, v); let u be v’s parent. Then the node v is the highest node whose string
label has the pattern as a (not necessarily strict) prefix. By returning the number stored in v we are
done, since this number is the number of non-overlapping occurrences of the pattern in the text. The
MAST can be built in O(n lg n) time [4].
In this section, we show how to compute the tree topology of the MAST in linear time. The topology
of the MAST differs to the suffix tree topology by the fact that the root of each square is the string
label of a MAST node. Our goal is to compute a list storing the information about where to insert the
missing nodes. The list stores tuples consisting of a node v and a length ℓ; we use this information later
to create a new node w splitting the edge (u, v) into (u, w) and (w, v), where u is the (former) parent
of v. We will label (w, v) with the last ℓ characters and (u, v) with the rest of the characters of the edge
label of (u, v).
To this end, we explore the suffix tree with a top-down traversal while locating the roots of the squares
in the order of their lengths. To locate the roots of the squares in linear time we use two data structures.
The first one is a semi-dynamic lowest marked ancestor data structure [19]. It allows marking a node
and querying for the lowest marked ancestor of a node in constant amortized time. We will use it to
mark the area in the suffix tree that has already been processed for finding the roots of the squares.
The second data structure is the list of tuples of the form (node, length) computed in Sec. 6, where
10
each tuple (v, ℓ) consists of the length ℓ of a square T [i..i + ℓ − 1] and the highest suffix tree node v whose
string label has T [i..i + ℓ − 1] as a (not necessarily proper) prefix. We sort this list, which we now call L,
with respect to the square lengths with a linear time integer sorting algorithm.
Finally, we explain the algorithm locating the roots of all squares. We successively process all tuples
of L, starting with the shortest square length. Given a tuple of L containing the node v and the length ℓ,
we want to split an edge on the path from the root to v and insert a new node whose string depth is
ℓ/2. To this end, we compute the lowest marked ancestor u of v. If u’s string depth is smaller than ℓ/2,
we mark all descendants of u whose string depths are smaller than ℓ/2, and additionally the children
of those nodes (this can be done by a DFS or a BFS). If we query for the lowest marked ancestor of u
again, we get an ancestor w whose string depth is at least ℓ/2, and whose parent has a string depth less
than ℓ/2. We report w and the subtraction of ℓ/2 from w’s string depth (if ℓ/2 is equal to the string
depth of w, then w’s string label is equal to the root of v’s string label, i.e., we do not have to report it).
If the suffix tree has a pointer-based representation, it is easy to add the new nodes by splitting each
edge (u, v), where v is a node contained in the output list.
Theorem 7.1. We can compute the tree topology of the MAST in linear time using linear number of
words.
Proof. By using the semi-dynamic lowest marked ancestor data structure, we visit a node as many times
as we have to insert nodes on the edge to its parent, plus one. This gives O(n + 2occ) = O(n) time.
Open Problems
It is left open to compute the number of the non-overlapping occurrences of the string labels of the
MAST nodes in linear time. Since RMQ data structures are practically slow, we wonder whether we can
avoid the use of any RMQ without loosing linear running time.
subsection*Acknowledgements This work was mainly done during a visit at the Kyushu University
in Japan, supported by the Japan Society for the Promotion of Science (JSPS). We thank Thomas
Schwentick for the question whether we can run our algorithm online, for which we provided a solution
in Appendix B.
References
[1] A. Apostolico and F. P. Preparata. Data structures and algorithms for the string statistics problem.
Algorithmica, 15(5):481–494, 1996.
[2] P. Beame and F. E. Fich. Optimal bounds for the predecessor problem and related problems. J.
Comput. Syst. Sci., 65(1):38–72, 2002.
[3] A. Blumer, J. Blumer, D. Haussler, A. Ehrenfeucht, M. T. Chen, and J. I. Seiferas. The smallest
automaton recognizing the subwords of a text. Theor. Comput. Sci., 40:31–55, 1985.
[4] G. Brodal, R. Lyngsø, A. Östlin, and C. Pedersen. Solving the string statistics problem in time
O(n log n). In Automata, Languages and Programming, volume 2380 of LNCS, pages 728–739.
Springer, 2002.
[5] D. R. Clark. Compact Pat Trees. PhD thesis, University of Waterloo, Canada, 1996.
[6] R. Cole and R. Hariharan. Dynamic LCA queries on trees. SIAM J. Comput., 34(4):894–923, 2005.
11
[7] M. Crochemore and L. Ilie. Computing longest previous factor in linear time and applications. Inf.
Process. Lett., 106(2):75–80, 2008.
[8] M. Crochemore, L. Ilie, C. S. Iliopoulos, M. Kubica, W. Rytter, and T. Walen. LPF computation
revisited. In Proc. IWOCA, pages 158–169, 2009.
[9] A. Deza, F. Franek, and A. Thierry. How many double squares can a string contain? Discrete
Applied Mathematics, 180:52–69, 2015.
[10] P. Elias. Efficient storage and retrieval by content and address of static files. Journal of the ACM,
21:246–260, 1974.
[11] R. M. Fano. On the number of bits required to implement an associative memory. Memorandum
61, Computer Structures Group, Project MAC, 1971.
[12] P.
Ferragina
and
G.
Navarro.
The
Pizza
&
Chili
Corpus.
Available
at
http://pizzachili.di.unipi.it and http://pizzachili.dcc.uchile.cl, 2005.
[13] J. Fischer. Wee LCP. Inform. Process. Lett., 110(8–9):317–320, 2010.
[14] J. Fischer. Inducing the LCP-array. In Proc. WADS, volume 6844 of LNCS, pages 374–385. Springer,
2011.
[15] J. Fischer and V. Heun. Space efficient preprocessing schemes for range minimum queries on static
arrays. SIAM J. Comput., 40(2):465–492, 2011.
[16] J. Fischer, T. I, and D. Köppl. Lempel-Ziv computation in small space (LZ-CISS). In Proc. CPM,
volume 9133 of LNCS, pages 172–184. Springer, 2015.
[17] A. S. Fraenkel and J. Simpson. How many squares can a string contain? J. Comb. Theory, Ser. A,
82(1):112–120, 1998.
[18] F. Franek, J. Holub, W. F. Smyth, and X. Xiao. Computing quasi suffix arrays. Journal of Automata,
Languages and Combinatorics, 8(4):593–606, 2003.
[19] H. N. Gabow and R. E. Tarjan. A linear-time algorithm for a special case of disjoint set union. J.
Comput. Syst. Sci., 30(2):209–221, 1985.
[20] S. Gog, T. Beller, A. Moffat, and M. Petri. From theory to practice: Plug and play with succinct
data structures. In Proc. SEA, volume 8504 of LNCS, pages 326–337, 2014.
[21] D. Gusfield. Algorithms on Strings, Trees, and Sequences. Cambridge University Press, 1997.
[22] D. Gusfield and J. Stoye. Linear time algorithms for finding and representing all the tandem repeats
in a string. J. Comput. Syst. Sci., 69(4):525–546, 2004.
[23] W.-K. Hon and K. Sadakane. Space-economical algorithms for finding maximal unique matches. In
Proc. CPM, volume 2373 of LNCS, pages 144–152. Springer, 2002.
[24] L. C. K. Hui. Color set size problem with application to string matching. In Proc. CPM, volume
644 of LNCS, pages 230–243. Springer, 1992.
[25] L. Ilie. A note on the number of squares in a word. Theor. Comput. Sci., 380(3):373–376, 2007.
12
[26] N. Jonoska, F. Manea, and S. Seki. A stronger square conjecture on binary words. In Proc. SOFSEM
2014, volume 8327 of LNCS, pages 339–350. Springer, 2014.
[27] J. Kärkkäinen, G. Manzini, and S. J. Puglisi. Permuted longest-common-prefix array. In Proc.
CPM, volume 5577 of LNCS, pages 181–192. Springer, 2009.
[28] J. Kärkkäinen, D. Kempa, and S. J. Puglisi. Linear time Lempel-Ziv factorization: Simple, fast,
small. In Proc. CPM, volume 7922 of LNCS, pages 189–200. Springer, 2013.
[29] D.
Köppl.
Computing
all
distinct
squares
efficiently.
Available
at
https://github.com/koeppl/distinct_squares, 2017.
[30] D. Köppl and K. Sadakane. Lempel-Ziv computation in compressed space (LZ-CICS). In Proc.
DCC, pages 3–12. IEEE Computer Society, 2016.
[31] Z. Li, J. Li, and H. Huo. Optimal in-place suffix sorting. ArXiv CoRR, abs/1610.08305, 2016.
[32] U. Manber and E. W. Myers. Suffix arrays: A new method for on-line string searches. SIAM J.
Comput., 22(5):935–948, 1993.
[33] F. Manea and S. Seki. Square-density increasing mappings. In Proc. WORDS 2015, volume 9304
of LNCS, pages 160–169. Springer, 2015.
[34] Y. Mori. divsufsort. Available at https://github.com/y-256/libdivsufsort, 2015.
[35] J. Munro, G. Navarro, and Y. Nekrich. Space-efficient construction of compressed indexes in deterministic linear time. In Proc. SODA, pages 408–424. SIAM, 2017.
[36] E. Ohlebusch and S. Gog. Lempel-Ziv factorization revisited. In Proc. CPM, volume 6661 of LNCS,
pages 15–26. Springer, 2011.
[37] K. Sadakane. Compressed suffix trees with full functionality. Theory Comput. Syst, 41(4):589–607,
2007.
[38] Y. Ueki, Diptarama, M. Kurihara, Y. Matsuoka, K. Narisawa, R. Yoshinaka, H. Bannai, S. Inenaga,
and A. Shinohara. Longest common subsequence in at least k length order-isomorphic substrings.
In Proc. SOFSEM, volume 10139 of LNCS, pages 363–374. Springer, 2017.
[39] E. Ukkonen. On-line construction of suffix trees. Algorithmica, 14(3):249–260, 1995.
[40] P. Weiner. Linear pattern matching algorithms. In Proc. Annual Symp. on Switching and Automata
Theory, pages 1–11. IEEE Computer Society, 1973.
13
A
Observations
In [22, Line 6 of Algorithm 1b], the condition start + k < h1 has to be changed to start + k ≤ h1 .
Otherwise, the algorithm would find in T = abaabab$ only the square aa, but not abaaba.
B
Online Variant
In this section, we consider the online setting, where new characters are appended to the end of the text T .
Given the text T [1..i] up to position i with the Lempel-Ziv factorization f1 · · · fy = T , we consider computing the set of all distinct squares of f1 · · · fy−2 , i.e., up to the last two
factors. For this set-
Lempel-Ziv
p
2
ting, we show that we can compute the set of all distinct squares in O n min lg lg n/ lg lg lg n, lg n/ lg lg n
time using O(n) words of space. To this end, we adapt the algorithm of Theorem 4.3 to the online setting.
We need an algorithm computing LPF online, and a semi-dynamic LCE data structure (answering LCE
queries on the text and on the reversed text while supporting appending characters to the text).
The main idea of our solution is to build suffix trees with two online suffix tree construction algorithms.
The first is Ukkonen’s algorithm that computes the suffix tree online in O(ntnav ) time [39], where tnav is
the time for inserting a node and navigating (in particular, selecting the child on the edge starting with a
specific character). We can adapt this algorithm to compute LPF online: Assume that we have computed
the suffix tree of T [1..i − 1]. The algorithm processes the new character T [i] by (1) taking the suffix links
of the current suffix tree, and (2) adding new leaves where a branching occurs. On adding a new leaf with
suffix number i, we additionally set LPF[i] to the string depth of its parent. By doing so, we can update
the LPF values in time linear to the update time of the suffix tree. We build the semi-dynamic RMQ
data structure of Fischer [14] (or of [38] if n is known beforehand) on top of LPF. This data structure
takes O(n) words and can perform query and appending operations in constant amortized time.
The second suffix tree construction algorithm is a modified version [3] of Weiner’s algorithm [40] that
builds the suffix tree in the reversed order of Ukkonen’s algorithm in O(ntnav ) time. Since Weiner’s
algorithm incrementally constructs the suffix tree of a given text from right to left, we can adapt this
algorithm to compute the suffix tree of the reversed
text
nav ) time.
online in O(ntp
2
To get a suffix tree construction time of O n min lg lg n/ lg lg lg n, lg n/ lg lg n , we use the pre-
decessor data structure of Beame and Fich [2]. We create a predecessor
children
datastructure to store the
p
2
of each suffix tree node, such that we get the navigation time tnav = O min lg lg n/ lg lg lg n, lg n/ lg lg n
for both suffix trees. We also create a predecessor data structure to store the out-going suffix links of
each node of the suffix tree constructed by Weiner’s algorithm. Overall, these take a total of O(n) words
of space.
Finally, our last ingredient is a dynamic lowest common ancestor data structure with O(n) words
that performs querying and modification operations in constant time [6]. The lowest common ancestor
of two suffix tree leaves with the labels i and j is the node whose string depth is equal to the longest
common extension of T [i..] and T [j..], where T [i..] denotes the i-th suffix (up to the last position that is
available in the online setting). Building this data structure on the suffix tree of the text T and on the
suffix tree of the reversed text allows us to compute LCE queries in both directions in constant time.
We adapt the algorithm of Sec. 4 by switching the order of the loops (again). The algorithm first
fixes a Lempel-Ziv factor fx and then searches for squares with a period between one and |fx | + |fx+1 |.
Unfortunately, we would need an extra bit vector for each period so that we can track all found leftmost occurrences. Instead, we use the predecessor data structure of [2] storing the found occurrences
of squares as pairs of starting positions and lengths. These pairs can be stored in lexicographic or14
der (first sorted by starting position, then by length). The predecessor data structure will contain at
most
takes O(occ)
= O(n) words of space. An insertion and or a search costs us
occ
elements, hence p
O min lg2 lg n/ lg lg lg n, lg n/ lg lg n
time.
Let us assume that we have computed the set for T [1..i − 1], and that the Lempel-Ziv factorization of T [1..i − 1] is f1 · · · fy . If appending a new character T [i] will result in a new factor fy+1 , we
check for squares of type Lem. 4.2(a) and Lem. 4.2(b) at the borders of fy−1 . Duplicates are filtered
by the predecessor data structure storing all already reported leftmost occurrences. The algorithm
outputs only the leftmost occurrences with the aid of LPF, whose entries are fixed up to the last
two factors (this is sufficient since we search for the starting position of the leftmost occurrence of a
square
with type Lem.
|f1 · · · fy−1
|], including right-rotations). In overall, we need
4.2(a) only in T [1..
p
time.
O (|fy−1 | + |fy |) min lg2 lg n/ lg lg lg n, lg n/ lg lg n
The current bottleneck of the online algorithm is the predecessor data structure in terms of the
running time. Future integer dictionary data structures can improve the overall performance of this
algorithm.
C
Algorithm Execution with one Step at a Time
In this section, we process the running example T = ababaaababa$ with the algorithm devised in Sec. 4
step by step. SA, LCP, PLCP, and LPF are given in the table below:
1
2
i
1
2
3
4
5
6
7
8
9
10 11 12
T
a
b
a
b
a
a
a
b
a
b
a
$
SA
12 11
5
6
9
3
7
1
10
4
8
2
LCP
0
0
1
2
1
3
3
5
0
2
2
4
PLCP 5
4
3
2
1
2
3
2
1
0
0
0
LPF
0
0
3
2
1
2
5
4
3
2
1
0
LZ
f1
f2
3
4
f3
5
f4
f5
f6
6
The text T = a—b—aba—aa—baba—$ = f1 · · · f6 is factorized in six Lempel-Ziv factors. We call
T [1 + |f1 · · · fi−1 |] (first position of the i-th factor) and T [1 + |f1 · · · fi |] (position after the i-th factor) the
left border and the right border of fi , respectively. The idea of the algorithm is to check the presence
of a square at a factor border and at an offset value q of the border with LCE queries. q is either the
addition of p to the left border, or the subtraction of p from the right border.
The algorithm finds the leftmost occurrences of all squares in the order (first) of their lengths and (second)
of their starting positions. We start with the period p = 1 and try to detect squares at each Lempel-Ziv
factor border. To this end, we create a bit vector B marking all found squares with period p. A square
of this period is found at the right border of f3 . It is of type Lem. 4.2(a), since its starting position is
in f3 . To find it, we take the right border b = 6 of f3 , and the position q := b − p = 5. We perform
an LCE query at b and q in the forward and backward direction. Only the forward query returns the
non-zero value of one. But this is sufficient to find the square aa of period one. Its LPF value is smaller
than 2p = 2, so it is the leftmost occurrence. It is not yet marked in B, thus we have not yet reported
it. Right-rotations are not necessary for period 1. Having found all squares with period 1, we clear B.
Next, we search for squares with period 2. We find a square of type Lem. 4.2(b) at the left border b = 2
of f2 . To this end, we perform an LCE query starting from b and q := b + p = 4 in both directions. Both
15
LCE queries show that T [1..5] is a repetition with period p = 2. Thus we know that T [1..4] is a square.
It is not yet marked in B, and has an LPF value smaller than 2p = 4, i.e., it is a not yet reported leftmost
occurrence. On finding a leftmost occurrence of a square, we right-rotate it, and report all right-rotations
whose LPF values are below 2p. This is the case for T [2..5], which is the leftmost occurrence of the square
baba.
After some unsuccessful checks at the next factor borders, we come to factor f5 and search for a
square of type Lem. 4.2(b). Two LCE queries in both directions at the left border b = 8 of f5 and
q := b + p = 10 reveal that T [7..11] is a repetition of period 2. The substring T [7..10] is a square, but its
LPF value is 5(≥ 2p), i.e., we have already reported this square. Although we have already reported it,
some right-rotation of it might not have been reported yet (see Appendix D for an example). This time,
all right-rotations (i.e., T [8..12]) have an LPF value ≥ 2p, i.e., there is no leftmost occurrence of a square
of period 2 found by right-rotations. In overall, we have found and reported the leftmost occurrences of
all squares once.
D
Need for RMQ on LPF
In Sec. 4, we perform the right-rotations of a square (s, 2p) with an RMQ on the interval I := [s +
1.. min(s + p − 1, e − 2p + 1)], where e is the last position of the maximal repetition of period p that
contains the square. Instead of an RMQ, we can linearly scan all LPF values in I, giving O(p) = O(n)
time. We cannot do better since the LPF values can be arbitrary. For instance, consider the text
T = abaaabaababaaabaaa$. The text aligned with LPF is shown in the table below.
i
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19
T
a
b
a
a
a
b
a
a
b
a
b
a
a
a
b
a
a
a
$
LPF
0
0
1
2
4
3
4
3
2
8
7
6
5
5
4
3
2
1
0
The square abaaabaa has two occurrences starting at the positions 1 and 10. The square baaabaaa at
position 11 is found by right-rotating the occurrence of abaaabaa at position 10. It is found by a linear
scan over LPF or an RMQ on LPF. A slight modification of this example can change the LPF values
around this occurrence. This shows that we cannot perform a shortcut in general (like stopping the
search when the LPF value is at least twice as large as p).
16
E
More Evaluation
collection
1MiB
10MiB
50MiB
100MiB
200MiB
pc-dblp.xml
pc-dna
0.2
0.3
3
3
16
23
33
56
70
310
pc-english
pc-proteins
0.2
0.3
5
4
42
25
500
74
2639
245
pc-sources
pcr-cere
0.2
0.6
3
6
31
30
286
79
792
535
pcr-einstein.en
pcr-kernel
0.4
0.2
12
8
83
233
1419
1274
3953
6608
pcr-para
0.4
4
26
98
265
Table 2: Running times in seconds, evaluated on different input sizes. We took prefixes of 1MiB, 10MiB,
and 100MiB of all collections.
F
Pseudo Code
17
Algorithm 1: Modified Algorithm 1 of [22]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
b(f ) denotes the left end of a factor f = T [b(f )..b(f ) + |f | − 1], lcp and lcs compute the LCE in T
and the LCE in the reverse of T (mirroring the input indices by i 7→ n − i for 1 ≤ i ≤ n − 1),
respectively.
Let f1 , . . . , fz be the factors of the Lempel-Ziv factorization
fz+1 ← T [n]
// dummy factor
Function recursive-rotate(s : starting position, e: ending position)
m ← LPF.RMQ[s..e]
if m > 2p then return
report(m, 2p) and B[m] ← 1
recursive-rotate(s,m − 1) and recursive-rotate(m + 1,e)
Function right-rotate(s : starting position of square, p: period of square)
if B[s] = 1 then return
if LPF[s] < 2p then report(s, 2p) and B[s] ← 1
ℓ ← lcp(s, s + p)
recursive-rotate(s + 1, s + p − 1, s + ℓ − p)
Z ← array of size z lg z bits, zero initialized
m ← max(|f1 | + |f2 | , . . . , |fz−1 | + |fz |)
for p = 1, . . . , m do
B ← bit vector of length n, zero initialized
for x = 1, . . . , z do
if |fx | + |fx+1 | < p then
y←x
while |fy | + |fy+1 | < p do
if Z[y] 6= 0 then y ← Z[y]
else incr y
Z[x] ← y and x ← y
if |fx | ≥ p then // probe for squares satisfying Lem. 4.2(a)
q ← b(fx+1 ) − p
ℓR ← lcp(b(fx+1 ), q) and ℓL ← lcs(b(fx+1 ) − 1, q − 1)
if ℓR + ℓL ≥ p and ℓR > 0 then // found a square of length 2p with its right
end in fx+1
s ← max(q − ℓL , q − p + 1) // square starts at s
right-rotate(s, p)
q ← b(fx ) + p // probe for squares satisfying Lem. 4.2(b)
ℓR ← lcp(b(fx ), q) and ℓL ← lcs(b(fx ) − 1, q − 1)
s ← max(b(fx ) − ℓL , b(fx ) − p + 1) // square starts in a factor preceding fx
if ℓR + ℓL ≥ p and ℓR > 0 and s + p ≤ b(fx+1 ) and ℓL > 0 then // found a square of
length 2p whose center is in fx
right-rotate(s, p)
18
| 8cs.DS
|
A Novel Weighted Distance Measure for Multi-Attributed Graph
Muhammad Abulaish, SMIEEE
Jahiruddin∗
Department of Computer Science
South Asian University
New Delhi-21, India
[email protected]
Department of Computer Science
Jamia Millia Islamia
New Delhi-25, India
[email protected]
arXiv:1801.07150v1 [cs.SI] 22 Jan 2018
ABSTRACT
Due to exponential growth of complex data, graph structure has
become increasingly important to model various entities and their
interactions, with many interesting applications including, bioinformatics, social network analysis, etc. Depending on the complexity of the data, the underlying graph model can be a simple
directed/undirected and/or weighted/un-weighted graph to a complex graph (aka multi-attributed graph) where vertices and edges are
labelled with multi-dimensional vectors. In this paper, we present
a novel weighted distance measure based on weighted Euclidean
norm which is defined as a function of both vertex and edge attributes, and it can be used for various graph analysis tasks including classification and cluster analysis. The proposed distance
measure has flexibility to increase/decrease the weightage of edge
labels while calculating the distance between vertex-pairs. We have
also proposed a MAGDist algorithm, which reads multi-attributed
graph stored in CSV files containing the list of vertex vectors and
edge vectors, and calculates the distance between each vertex-pair
using the proposed weighted distance measure. Finally, we have
proposed a multi-attributed similarity graph generation algorithm,
MAGSim, which reads the output of MAGDist algorithm and generates a similarity graph that can be analysed using classification
and clustering algorithms. The significance and accuracy of the
proposed distance measure and algorithms is evaluated on Iris and
Twitter data sets, and it is found that the similarity graph generated
by our proposed method yields better clustering results than the
existing similarity graph generation methods.
CCS CONCEPTS
• Information systems → Similarity measures; Data analytics; Clustering; • Human-centered computing → Social network
analysis;
KEYWORDS
Data mining, Clustering, Multi-attributed graph, Weighted distance
measure, Similarity measure
∗ To
whom correspondence should be made
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
Compute ’17, Bhopal, India
© 2017 ACM. 978-1-4503-5323-6/17/11. . . $15.00
DOI: 10.1145/3140107.3140114
ACM Reference format:
Muhammad Abulaish, SMIEEE and Jahiruddin. 2017. A Novel Weighted
Distance Measure for Multi-Attributed Graph. In Proceedings of 10th Annual ACM India Compute Conference, Bhopal, India, November 16–18, 2017
(Compute ’17), 9 pages.
DOI: 10.1145/3140107.3140114
1
INTRODUCTION
Due to increasing popularity of Internet and Web2.0, the UserGenerated Contents (UGCs) are growing exponentially. Since most
of the UGCs are not independent, rather linked, the graph data structure is being considered as a suitable mathematical tool to model
the inherent relationships among data. Simple linked data are generally modelled using simple graph G = (V , E), where V is the set
of vertices representing key concepts or entities and E ⊆ V × V is
the set of links between the vertices representing the relationships
between the concepts or entities. Depending on the nature of data
to be modelled, the graph G could be weighted/un-weighted or
directed/undirected, and it may have self-loops. However, there are
many complex data like online social networks, where an entity is
characterized by a set of features and multiple relationships exist
between an entity-pair. To model such data, the concept of multiattributed graph can be used wherein each vertex is represented by
an n-dimensional vector and there may be multiple weighted edges
between each pair of vertices. In other words, in a multi-attributed
graph, both vertices and edges are assigned multiple labels. One of
the important tasks related to graph data analysis is to decompose
a given graph into multiple cohesive sub-graphs, called clusters,
based on some common properties. The clustering is an unsupervised learning process to identify the underlying structure (in the
form of clusters) of data, which is generally based on some similarity/distance measures between data elements. Graph clustering
is special case of clustering process which divides an input graph
into a number of connected components (sub-graphs) such that
intra-component edges are maximum and inter-components edges
are minimum. Each connected component is called a cluster (aka
community) [13]. Though graph clustering has received attention
of many researchers and a number of methods for graph clustering
has been proposed by various researchers [19], to the best of our
knowledge, the field of clustering multi-attributed graph is still
unexplored.
Since similarity/distance measure is the key requirement for
any clustering algorithm, in this paper, we have proposed a novel
weighted distance measure based on weighted Euclidean norm to
calculate the distance between the vertices of a multi-attributed
graph. The proposed distance measure considers both vertex and
edge weight-vectors, and it is flexible enough to assign different
weightage to different edges and scale the overall edge-weight
Compute ’17, November 16–18, 2017, Bhopal, India
while computing the weighted distance between a vertex-pair. We
have also proposed a MAGDist algorithm that reads the lists of
vertex and edge vectors as two separate CSV files and calculates the
distance between each vertex-pairs using the proposed weighted
distance measure. Finally, we have proposed a multi-attributed similarity graph generation algorithm, MAGSim, which reads the output
of MAGDist algorithm and generates a similarity graph. In other
words, MAGDist and MAGSim algorithms can be used to transform
a multi-attributed graph into a simple weighted similarity graph,
wherein a single weighted edge exists between each vertex-pair.
Thereafter, the weighted similarity graph can be analysed using
existing classification and clustering algorithms for varied purposes.
The efficacy of the proposed distance measure and algorithms is
tested over the well-known Iris data set and a Twitter data set
related to three different events. The proposed similarity graph generation approach is compared with other existing similarity graph
generation methods like Gaussian kernel and k-nearest neighbours
methods, and it is found that our proposed approach yields better
clustering results in comparison to the existing methods. Moreover,
the proposed distance measure can be applied to any graph data
where both vertices are edges are multi-dimensional real vectors.
In case, the data is not linked, i.e., edges do not exist between the
edges, the proposed distance measure simple works as an Euclidean
distance between the node vectors.
The rest of the paper is organized as follows. Section 2 presents
a brief review on various distance measures and graph clustering methods. Section 3 presents the formulation of our proposed
weighted distance measure for multi-attributed graph. It also presents
the founding mathematical concepts and formal descriptions of the
MAGDist and MAGSim algorithms. Section 4 presents the experimental setup and evaluation results. Finally, section 5 concludes
the paper with future directions of work.
2
RELATED WORK
Multi-attributed graph is used to model many complex problems,
mainly those in which objects or entities are characterized using a
set of features and linked together in different ways. For example,
an online social network user can be characterized using a set of
features like ID, display name, demographic information, interests,
etc. Similarly, two person in an online social network may be associated through different relationships like friendship, kinship,
common interests, common friends, etc [4]. In [15], the authors
proposed a new principled framework for estimating graphs from
multi-attribute data. Their method estimates the partial canonical
correlations that naturally accommodate complex vertex features
instead of estimating the partial correlations. However, when the
vertices of a multi-attributed graph have different dimensions, it
is unclear how to combine the estimated graphs to obtain a single
Markov graph reflecting the structure of the underlying complex
multi-attributed data. In [14], Katenka and Kolaczyk proposed a
method for estimating association networks from multi-attribute
data using canonical correlation as a dependence measure between
two groups of attributes.
Clustering is an unsupervised learning technique which aims to
partition a data set into smaller groups in which elements of a group
are similar and that elements from different groups are dissimilar.
Abulaish et al.
As a result, clustering has broad applications, including the analysis
of business data, spatial data, biological data, social network data,
and time series data [24], and a large number of researchers have
targeted clustering problem [2, 12, 17]. Since graph is a popular data
structure to model structural as well as contextual relationships of
data objects, recently graph data clustering is considered as one of
the interesting and challenging research problems [16, 22], and it
aims to decompose a large graph into different densely connected
components. Some of the applications of the graph clustering is
community detection in online social networks [5, 7, 8], social bot
detection [10, 11], spammer detection [1, 3, 6, 9], functional relation
identification in large protein-protein interaction networks, and so
on. Generally graph clustering methods are based on the concept of
normalized cut, structural density, or modularity[16, 22]. However,
like [20], graphs can be partitioned on the basis of attribute similarity in such a way that vertices having similar attribute vectors
are grouped together to form a cluster. Mainly graph clustering
and simple data clustering differs in the way associations between
objects are calculated. In graph clustering, the degree of association between a pair of objects is calculated as closeness between
the respective nodes of the graph, generally in terms of number
of direct links or paths between them. Whereas, in simple data
clustering, the degree of association between a pair of objects is
calculated in terms of similarity/distance measure between the vector representations of the objects. Both topological structure and
vertex properties of a graph play an important role in many real
applications. For example, in a social graph, vertex properties can
be used to characterize network users, whereas edges can be used
to model different types of relationships between the users.
It can be seen from the discussions mentioned above that most
of the graph analysis approaches have considered only one aspect
of the graph structure and ignored the other aspects, due to which
the generated clusters either have loose intra-cluster structures
or reflect a random vertex properties distribution within clusters.
However, as stated in [24], “an ideal graph clustering should generate clusters which have a cohesive intra-cluster structure with
homogeneous vertex properties, by balancing the structural and
attribute similarities". Therefore, considering both vertex attributes
and links for calculating similarity/distance between the pairs of
vertices in a graph is one of basic requirements for clustering complex multi-attributed graphs. In this paper, we have proposed a
weighted distance function that can transform a multi-attributed
graph into a simple similarity graph on which existing graph clustering algorithms can be applied to identify densely connected
components.
3
PRELIMINARIES AND DISTANCE
MEASURES
In this section, we present the mathematical basis and formulation
of the proposed weighted distance measures for multi-attributed
graph. Starting with the basic mathematical concepts in subsection
3.1, the formulation of the proposed distance function along with
an example is presented in the subsequent subsections.
A Novel Weighted Distance Measure for Multi-Attributed Graph
3.1
Founding Mathematical Concepts
Since the Linear Algebra concepts inner product and norm generally
form the basis for any distance function, we present a brief overview
of these mathematical concepts in this section.
Inner Product: An inner product is a generalized form of the
vector dot product to multiply vectors together in such a way that
the end result is a scalar quantity. If a® = (a 1 , a 2 , ..., an )T and b® =
(b1 , b2 , ..., bn )T are two vectors in the vector space <n , then the
® and defined using
inner product of a® and b® is denoted by h®
a , bi
equation 1 [18].
n
Õ
® = a® · b® = a 1b1 + a 2b2 + · · · + an bn =
h®
a , bi
a i bi
(1)
Compute ’17, November 16–18, 2017, Bhopal, India
(1) kak
® > 0 with kak
® = 0 iff a® = 0
(positivity)
(2) knak
® = nkak
®
(homogeneity)
(3) ka® + b®k ≤ kak
® + kb®k
(triangle inequality)
Some of the well-known norms defined in [21] are as follows:
(1) L1 -norm (aka Manhattan norm): The L 1 -norm of a vector
a® ∈ <n is simply obtained by adding the absolute values
of its components. Formally, the L1 -norm of the vector a® is
represented as kak
® 1 and it can be defined using equation 5.
n
Õ
kak
®1=
|ai |
(5)
i=1
(2) L2 -norm (aka Euclidean norm): The L 2 -norm of a vector
a® ∈ <n is obtained by taking the positive square root of
the sum of the square of its components. Formally, the
L2 -norm of a vector a® is represented as kak
® 2 and it can be
defined using equation 6.
! 1/2
n
Õ
2
kak
®2=
ai
(6)
i=1
b 1
b 2
T
®
h®
a , bi = a b = a 1 a 2 . . . an .
..
b
n
= a 1b 1 + a 2b 2 + · · · + a n b n
(2)
i=1
(3) Infinity-norm (aka max-norm): The Infinity-norm of a vector a® ∈ <n is obtained as maximum of the absolute values
of the components of the vector. Formally, the infinitynorm of the vector a® is represented as kak
® ∞ and it can be
defined using equation 7.
Alternatively, inner product of a® and b® can be defined as a matrix
product of row vector a®T and column vector b® using equation
2. However, as stated in [18], an inner product must satisfy the
® and c® belong to <n ,
following four basic properties, where a,
® b,
and n is a scalar quantity.
(1) h®
a + b® , c®i = h®
a , c®i + hb® , c®i
®
®
(2) hna® , bi = nh®
a , bi
®
®
(3) h®
a , bi = hb , ai
®
(4) h®
a , ai
® > 0 and h®
a , ai
® = 0 iff a® = 0
n
kak
® ∞ = max |ai |
i=1
(4) Lp -norm: The Lp norm of a vector a® ∈ <n is the positive
p th root of the sum of p th power of the absolute value of
the components of the vector. Formally, the Lp -norm of the
vector a® is represented as kak
® p and it can be defined using
equation 8.
! 1/p
n
Õ
p
kak
®p=
|ai |
(8)
Weighted Inner Product: The weighted inner product is generally
used to emphasize certain features (dimensions) through assigning
weight to each dimension of the vector space <n . If d 1 , d 2 , . . . , dn ∈
< are n positive real numbers between 0 and 1 and D is a corresponding diagonal matrix of order n × n, then the weighted inner
product of vectors a® and b® is defined by equation 3 [18]. It can be easily verified that weighted inner product also satisfies the four basic
properties of the inner product mentioned earlier in this section.
d 1
0
T
®
h®
a , biD = a Db =
di ai bi , where D= .
..
i=1
0
n
Õ
0
d2
..
.
0
...
...
..
.
...
0
0
..
.
dn
i=1
(5) Weighted Euclidean norm: The weighted Euclidean norm of
a vector a® ∈ <n is a special case of the Euclidean norm in
which different dimensions of a® can have different weights.
If AT = a®T is a row matrix of order 1 × n, and D is a
diagonal matrix of order n × n whose i th diagonal element
is the weight of the i th dimension of the vector a,
® then the
weighted Euclidean norm of a® is the positive square root of
the product matrix AT DA. In case D is an identity matrix,
this gives simply the Euclidean norm. Formally, for a given
vector a® ∈ <n and a weight (diagonal) matrix D of order
n × n, the weighted Euclidean norm of a® is represented as
kak
® D and defined using equation 9.
(3)
Norm: In linear algebra and related area of mathematics, the
norm on a vector space <n is a function used to assign a non
negative real number to each vector a® ∈ <n . Every inner product
gives a norm that can be used to calculate the length of a vector.
However, not every norm is derived from an inner product [18].
The norm of a vector a® ∈ <n is denoted by kak
® and can be defined
using equation 4.
p
kak
® = h®
a , ai
®
(4)
T
T
®
As given in [18], if a® = (a 1 , a 2 , ..., an ) and b = (b1 , b2 , ..., bn ) are
two vectors of the vector space <n and n > 0 is a scalar quantity,
then every norm satisfies the following three properties.
(7)
! 1/2
n
1/2
Õ
T
2
kak
® D = A DA
=
di ai
(9)
i=1
3.2
Multi-Attributed Graph
In this section, we present a formal definition of multi-attributed
graph, in which both vertices and edges can be represented as multidimensional vectors. Mathematically, a multi-attributed graph can
Compute ’17, November 16–18, 2017, Bhopal, India
Abulaish et al.
be defined as a quadruple Gm = (V , E, Lv , Le ), where V , ϕ
represents the set of vertices, E ⊆ V × V represents the set of
edges, Lv : V → <n is a vertex-label function that maps each
vertex to an n-dimensional real vector, and Le : E → <m is an
edge-label function that maps each edge to an m-dimensional real
vector. Accordingly, a vertex v ∈ V in a multi-attributed graph can
be represented as an n-dimensional vector v® = (v 1 , v 2 , . . . , vn )T
and an edge between a vertex-pair (u, v) can be represented as an
m-dimensional vector e®(u, v) = (e 1 (u, v), e 2 (u, v), . . . , em (u, v))T .
3.3
Proposed Weighted Distance Measure
In this section, we present our proposed weighted distance measure
for multi-attributed graph that is based on weighted Euclidean norm.
If u = u® = (u 1 , u 2 , . . . , un )T and v = v® = (v 1 , v 2 , . . . , vn )T are
two vertices of a multi-attributed graph and (u, v) = e®(u, v) =
(e 1 (u, v), e 2 (u, v), . . . , em (u, v))T is a multi-labelled edge between
u and v, then the distance between u and v is denoted by ∆(u, v) and
calculated using equation 10, where λ is a scalar value (equation 11),
and I is an identity matrix of order n × n. The value of λ depends on
the aggregate weight, ω(u, v), of the edges between the vertex-pair
(u, v), which is calculated using equation 12. The value of γ > 0 is
a user-supplied threshold, providing flexibility to tune the value of
λ for calculating distance between a vertex-pair. In equation 12, α i
Í
is a constant such that α i ≥ 0 and ni=1 α i = 1.
⇒ (1 + ω 1 ) ≤ (1 + ω 2 )
⇒ (1 + ω 1 )γ ≤ (1 + ω 2 )γ , where γ ≥ 1
⇒ (1+ω1 )γ ≥ (1+ω1 )γ
1
2
⇒ f (ω 1 ) ≥ f (ω 2 )
Hence, λ is a monotonically decreasing function of ω for γ ≥ 1.
Similarly, it can be shown that λ is a monotonically decreasing
function of γ for ω > 0.
Let λ = f (γ ) be a function of γ
Let γ 1 , γ 2 ≥ 1 such that γ 1 ≤ γ 2
⇒ (1 + ω)γ1 ≤ (1 + ω)γ2 , where ω > 0
1
1
⇒ (1+ω)
γ 1 ≥ (1+ω)γ 2
⇒ f (γ 1 ) ≥ f (γ 2 )
Hence, λ is a monotonically decreasing function of γ for ω > 0.
ω=1
ω=0.75
ω=0.50
ω=0.25
1.1
1
0.9
0.8
0.7
0.6
λ
0.5
0.4
∆(u, v) = ∆(®
u, v)
® = ku® − v®kλI = (u − v)T .λI .(u − v)
u 1 − v 1 1/2
...
©
. ª®
..
= u 1 − v 1 . . .
.. ®
.
u − v
...
n¬
«
n
1/2
= λ(u 1 − v 1 )2 + λ(u 2 − v 2 )2 + · · · + λ(un − vn )2
! 1/2
n
Õ
√
2
(ui − vi )
= λ×
λ
.
un − vn ..
0
0.3
1/2
0
..
.
λ
0.2
0.1
0
0
(10)
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
γ
Figure 1: Visualization of the monotonically decreasing nature of λ for varying γ and ω values
i=1
1
λ=
(1 + ω(u, v))γ
ω(u, v) = α 1e 1 (u, v) + α 2e 2 (u, v) + · · · + αm em (u, v)
n
Õ
α i ei (u, v)
=
1
(11)
It should be noted that if there is no direct link (edge) between a
vertex pair (u, v), then the value of ω(u, v) in equation 11 becomes
zero, leading the value of λ to 1. In this case, the value of ∆(u, v)
is simply an Euclidean distance between the vertex-pair (u, v), as
proved in theorem 3.2.
(12)
i=1
It may be noted that the λ function defined using equation 11 is
a monotonic decreasing function, as proved in theorem 3.1, so that
the distance between a vertex-pair could decrease with increasing
ties between them, and vice-versa. The novelty of the proposed
distance function lies in providing flexibility (i) to assign different
weights to individual edges using α, and (ii) to control the degree
of monotonicity using γ , as shown in figure 1. It may also be noted
that the value of λ will always be 1, if either if ω(u, v) = 0 or γ = 0.
Theorem 3.1. The λ function defined in equation 11 is a monotonically decreasing function.
Proof. Let λ = f (ω) be a function of ω, where ω is the aggregate
weight of edges between a vertex-pair (u, v).
Let ω 1 , ω2 ≥ 0 such that ω 1 ≤ ω 2
Theorem 3.2. In a multi-attributed graph, if there is no edge
between a vertex-pair then the distance between them is simply an
Euclidean distance.
Proof. Let u = u® = (u 1 , u 2 , . . . , un )T and v = v® = (v 1 , v 2 , . . . , vn )T
be two vertices not connected by any edge. Since there is no edge between the vertex-pair (u, v), the edge vector e®(u, v) = (e 1 (u, v), e 2 (u, v), . . . ,
em (u, v))T = 0®.
⇒ e 1 (u, v) = e 2 (u, v) = · · · = em (u, v) = 0
Hence, ω(u, v) = α 1 .e 1 (u, v) + α 2 .e 2 (u, v) + · · · + αm .em (u, v)
= α 1 .0 + α 2 .0 + · · · + αm .0 = 0.
[using equation 12]
1
1
Hence, λ = (1+ω(u,v))
[using equation 11]
γ = (1+0)γ = 1
√
1/2
Ín
2
Finally, ∆(u, v) = λ × i=1 (ui − vi )
[using equation 10]
√
1/2
Ín
Ín
2
2 1/2 , which is an
= 1×
=
i=1 (ui − vi )
i=1 (ui − vi )
Euclidean distance between the vertex-pair (u, v).
A Novel Weighted Distance Measure for Multi-Attributed Graph
Algorithm 1 presents a formal way to calculate the distance
between all pairs of vertices of a given multi-attributed graph
using MAGDist. The proposed MAGDist algorithm reads a multiattributed graph using two separate CSV files – one containing
the list of vertex vectors and the other containing the list of edge
vectors, and produces the distance between each vertex-pairs as a
CSV file, wherein each tuple contains a vertex-pair and distance
value. We have also proposed MAGSim algorithm to generate similarity graph using the distance values calculated by the MAGDist
algorithm. The proposed algorithm reads each vertex-pair and its
distance value < i, j, ∆(i, j) > and calculates similarity between the
vertex-pair (i, j) using equation 13.
sim(i, j) = 1 −
∆(i, j)
max {∆(x, y)}
Algorithm 1: MAGDist(Lv , Le , α, γ )
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Algorithm 2: MAGSim(D)
1
2
3
4
// generating similarity graph corresponding to a
multi-attributed graph
Input : A CSV file D containing vertex-pairs along with
distance output by the MAGDist.
Output : A CSV file G s containing edge-listed similarity graph.
dmax ← getMaxDistance(D);
for each tuple < i, j, ∆(i, j > in D do
∆(i, j)
sim(i, j) = 1 − d
write tuple < i, j, sim(i, j) > into G s ;
max
end
(13)
x,y ∈V
1
Compute ’17, November 16–18, 2017, Bhopal, India
// computing distance between all pairs of
vertices of a multi-attributed graph
Input : CSV files Lv and Le containing list of vertex vectors,
and edge vectors, respectively. An 1D array α[1..m],
Í
wherein α i ≥ 0 and m
i=1 α i = 1 for calculating the
linear combination of edge weights between a pair of
vertices. A positive integer threshold γ representing
the weightage of edges in distance calculation.
Output : A CSV file D containing distance between each pairs
of vertices.
nv ← vertexCount[Lv ] ;
// number of vertices
n ← vertexDimCount[Lv ] ; // vertex vector dimension
m ← edдeDimCount[Le ] ;
// edge vector dimension
V [nv ][n] ← Lv ;
// reading Lv into array V .
for each vertex-pair (i, j) ∈ Le do
// calculating aggregate weight ω(i, j) of the
edges between vertex-pair (i, j).
ω(i, j) ← 0;
for k ← 1 to m do
ω(i, j) = ω(i, j) + α[k] × ek (i, j) ;
// [eqn. 12]
end
// calculating the value of scalar quantity λ
1
;
// [eqn. 11]
λ = (1+ω(i,
j))γ
// calculating distance ∆(i, j) between the
vertex-pair (i, j).
d ← 0;
for k ← 1 to n do
d = d + (V [i][k] − V [j][k])2 ;
end
√
√
∆(i, j) = λ × d ;
// [eqn. 10]
write tuple < i, j, ∆(i, j) > into D ;
end
Example: Figure 2 presents a simpler case of multi-attributed
graph having four vertices in which each vertex is represented as
a two dimensional feature vector and each edge is represented as
an one dimensional vector. In case, there is no direct edge between
a pair of vertices (e.g., v 1 and v 3 ), the corresponding edge vector
is a zero vector. If we simply calculate the Euclidean distance between the vertex-pairs, then the distance between the vertex-pairs
(v 1 , v 2 ), (v 2 , v 3 ), (v 3 , v 4 ) and (v 4 , v 1 ) is same (10 unit), whereas the
distance calculated by the proposed MAGDist differs, based on
the weight of the edges between them. Similarly, the Euclidean
distance
between the vertex-pairs (v 1 , v 3 ) and (v 2 , v 4 ) are same
√
(10 2), but the distance values calculated using MAGDist are different. The distance values between each vertex-pairs of the multiattributed graph calculated using MAGDist for γ = 1 and γ = 2
are is shown in D 1 and D 2 matrices, respectively of figure 2. It can
be observed that giving more weightage to edges by increasing
the value of γ reduces the distance between the respective vertexpairs. Figure 3 presents a multi-attributed graph in which both
vertex and edge labels are multi-dimensional vectors. For example, the edge vector corresponding to the edge connecting v 2 and
v 3 is e®(2, 3) = (0.36, 0.64)T . If we simply calculate the Euclidean
distance between the vertex-pairs, then the distance between the
vertex-pairs (v 1 , v 2 ), (v 2 , v 3 ), (v 3 , v 4 ) and (v 4 , v 1 ) is same (10 unit),
whereas the distance calculated by the proposed MAGDist differs,
based on the weight of the edges between them. The distance values
between each vertex-pairs of the multi-attributed graph calculated
using MAGDist for γ = 1 and γ = 2 are is shown in D 1 and D 2
matrices, respectively of figure 3. It can be observed from these two
distance matrices too that giving more weightage to edges by increasing the value of γ reduces the distance between the respective
vertex-pairs.
4
EXPERIMENTAL SETUP AND RESULTS
To establish the efficacy of the proposed MAGDist distance measure,
we have considered the well-known Iris data set 1 , which contains
total 150 instances of three different categories of Iris flower; 50
instances of each category. The Iris data set can be represented as
a 150 × 5 data matrix, wherein each row represents an Iris flower,
first four columns represent four different attribute values in centimetre, and the 5th column represents class labels (categories) of
the Iris flower as Setosa, Virginica, or Versicolour. Table 1 shows
a partial snapshot of the Iris data set. We model Iris data set as a
multi-attributed similarity graph in which each vertex v ∈ <4 is a
4-dimensional vector representing a particular instance of the Iris
1 http://archive.ics.uci.edu/ml/datasets/Iris
Compute ’17, November 16–18, 2017, Bhopal, India
0 v04
10 10
v3 v1010
3
v4 1.01.0
1010
0 0 9.81
.14
9.81 1414
.14 8.894
.94
9.81
0
7
.
81
1010
9
.
81
0
7
.
81
1.01.0 0.64
0.64 D1D1
1414
.14.14 7.81
7.81 0 0 7.707
.07
8
.
94
10
7
.
07
0
8.94 10
7.07
0
1010
0.250.25
0 0
v
0 01
Abulaish et al.
0.04
v1 0.04
DD2 2
00
99.62
.62
14
14.14
.14
88.00
.00
99..62
62 14
14..14
14
00
66..10
10
66..10
00
10
77..07
5
.
00
07
00
88..00
7
.
07
7.07
00
55..00
0
v2 v20 0
Figure 2: A multi-attributed graph with vertices as multi-dimensional vectors, and distance matrices D 1 , D 2 calculated using
MAGDist algorithm for γ = 1 and γ = 2, respectively
0 0v4
10 10
v4
1.01.0
9.81 1414
.14 8.816
.16
0 0 9.81
.14
9
.
81
0
8
.
16
14
.14
0
8.16 14.14
D1 9.81
D
DD2 2
1
14
.
14
8
.
16
0
7
.
0.36 0.64
0.64
14.14 8.16
0
7.0707
0.36
8.16 1414
.07 0 0
8.16
.14.14 7.707
1010
1.01.0
0.250.25 0.750.75
0 0
v01
0
v3 v31010
v1
0.04
0.04
0.04
0.04
v2
1010
0
0
00
9
.62
9.62
14
.14
14.14
.67
66.67
62 14
14..14
14 66..67
99..62
67
0
6
.
67
14
.
0
6.67 14.14
14
6
.
67
55..00
6.67
00
00
14..14
14 55..00
00
00
14
v2
Figure 3: A multi-attributed graph with both vertices and edges as multi-dimensional vectors, and distance matrices D 1 , D 2
calculated using MAGDist algorithm for γ = 1 and γ = 2, respectively
Table 1: A partial view of the Iris data set
Sepal length
5.1
4.9
4.7
4.6
5.0
...
6.4
6.9
5.5
6.5
5.7
...
6.3
5.8
7.1
6.3
6.5
Sepal width
3.5
3.0
3.2
3.1
3.6
...
3.2
3.1
2.3
2.8
2.8
...
3.3
2.7
3.0
2.9
3.0
Petal length
1.4
1.4
1.3
1.5
1.4
...
4.5
4.9
4.0
4.6
4.5
...
6.0
5.1
5.9
5.6
5.8
Petal width
0.2
0.2
0.2
0.2
0.2
...
1.5
1.5
1.3
1.5
1.3
...
2.5
1.9
2.1
1.8
2.2
Species
Setosa
Setosa
Setosa
Setosa
Setosa
...
versicolor
versicolor
versicolor
versicolor
versicolor
...
virginica
virginica
virginica
virginica
virginica
flower. The edge (similarity) between a vertex-pair (u, v) is determined using equation 14 on the basis of the Gaussian kernel value
defined in equation 15, where σ = 1 is a constant value [23].
(
κ G (u, v) if κ G (u, v) ≥ 0.55
e(u, v) =
(14)
0
otherwise
κ G (u, v) = e
− ku−v2k
2σ
2
(15)
The resulting multi-attributed Iris similarity graph is shown in
figure 4 (termed hereafter as G 1 for rest of the paper), which contains 150 vertices and 2957 edges. In this graph, the instances of
Setosa, Virginica, or Versicolour are shown using triangles, squares,
and circles, respectively. Although κ G (u, u) = 1.0 for all vertices,
we haven’t shown self-loops in this graph. The proposed MAGDist
algorithm is applied over G 1 to calculate distance between all vertexpairs, and finally MAGSim algorithm is applied to generate Iris similarity graph (termed hereafter as G 2 for rest of the paper), which is
shown in figure 5. In [23], the authors have also used Gaussian kernel followed by the concept of nearest-neighbours to generate Iris
similarity graph (termed hereafter as G 3 for rest of the paper) and
applied Markov Clustering (MCL) to classify the instances of the
Iris data into three different categories. Therefore, in order to verify
the significance of our MAGDist and MAGSim algorithms, we have
also applied the MCL over the Iris similarity graphs G 1 and G 2 , and
present a comparative analysis of all clustering results. Figures 6
and 7 present the clustering results after applying MCL over the Iris
similarity graphs G 1 and G 2 , respectively. Table 2 presents the contingency table for the discovered clusters versus true Iris types from
all three different Iris similarity graphs. It can be observed from
this table that, in case of G 2 , only six instances of iris-versicolor
are wrongly grouped with iris-virginica in C 2 and three instances
of iris-virginica are wrongly grouped with iris-versicolor in C 3 ;
whereas in G 1 forty iris-versicolor instances are wrongly grouped
with iris-virginica in C 2 , and in G 3 , one instance of iris-versicolor
is wrongly grouped with iris-setosa in C 1 and fourteen instances of
iris-virginica are wrongly grouped with iris-versicolor in C 3 .
A Novel Weighted Distance Measure for Multi-Attributed Graph
Figure 4: G 1 : Iris data graph modelled as a multi-attributed
graph in which vertices are 4-dimensional real vectors and
edges are the Gaussian similarity (≥ 0.55) between the respective vertices
Compute ’17, November 16–18, 2017, Bhopal, India
Figure 6: Clustering results after applying MCL over the Iris
data graph G 1
Figure 7: Clustering results after applying MCL over the Iris
data graph G 2
Figure 5: G 2 : Iris data graph modelled as a multi-attributed
graph in which vertices are 4-dimensional real vectors and
edges are MAGSim similarity (≥ 0.80) calculated using equation 13
We have also analyzed the significance of different similarity
graph generation methods in terms of True Positive Rate (TPR) and
False Positive Rates (FPR) that are defined using equations 16 and
17, respectively. In these equations, TP is the True Positives, representing the number of correctly classified positive instances, FP is
the False Positives, representing the number of wrongly classified
negative instances, and P and N represent the total number of positive and negative instances, respectively in the data set. Table 3
presents TPR and FPR values for all three different similarity graphs,
showing best results for G 2 , which has been generated using our
proposed MAGDist and MAGSim algorithms.
Table 2: Contingency table for MCL clusters from similarity
graphs generated by three different methods
Clusters
C 1 (triangle)
C 2 (square)
C 3 (circle)
Set
50
0
0
G1
Vir
0
50
0
Ver
0
40
10
Set
50
0
0
G2
Vir
0
47
3
Ver
0
6
44
Set
50
0
0
G 3 [23]
Vir Ver
0
1
36
0
14
49
TPR =
TP
P
(16)
FPR =
FP
N
(17)
Compute ’17, November 16–18, 2017, Bhopal, India
Abulaish et al.
Table 3: Performance comparison of three different methods on Iris data set
Clusters
C 1 (triangle)
C 2 (square)
C 3 (circle)
Average
4.1
G1
TPR FPR
1.00
0.00
1.00
0.40
0.20
0.00
0.73 0.13
G2
TPR FPR
1.00 0.00
0.94 0.06
0.88 0.03
0.94 0.03
G 3 [23]
TPR FPR
1.00 0.01
0.72 0.00
0.98 0.14
0.90 0.05
Evaluation Results on Twitter Data Set
In order to illustrate the application of the proposed distance measure over a real multi-attributed graph, we have considered a Twitter data set of 300 tweets, comprising equal number of tweets related to three different events – NoteBandi (NTB), RyanInternationalSchool (RIS), and Rohingya (ROH). The tweets are modelled as
a multi-attributed graph, wherein each vertex represents a tweet
as an 110-dimensional binary vector based on the top-110 keyterms identified using tf-idf, and two different edges exist between
a vertex-pair – one representing the degree of Hashtags overlap
and the other representing the tweet-time overlap. The similarity
graph is generated using MAGSim algorithm, and shown in figure 8 wherein the instances of NoteBandi, RyanInternationalSchool,
and Rohingya are represented using squares, triangles, and circles,
respectively. Finally, MCL is applied over the similarity graph to
group the tweets into different clusters shown in figure 9. The evaluation of the obtained clusters is given in table 4. It can be seen
from this table that only five instances of RyanInternationalSchool
are wrongly clustered with Rohingya in C 3 .
Figure 8: Similarity graph generated from Twitter data set
(only edges having similarity value > 0.5 are shown)
5
CONCLUSION AND FUTURE WORKS
In this paper, we have proposed a novel weighted distance measure
based on weighted Euclidean norm that can be used to calculate
the distance between vertex-pairs of a multi-attributed graph containing multi-labelled vertices and multiple edges between a single
Figure 9: Clustering results obtained by applying MCL over
the similarity graph of the Twitter data set
Table 4: Evaluation results of the proposed method on Twitter data set
Clusters
C 1 (triangle)
C 2 (square)
C 3 (circle)
Contingency Table
NTB RIS ROH
0
95
0
100
0
0
0
5
100
Average
Evaluation Results
TPR
FPR
1.00
0.00
1.00
0.00
1.00
0.025
1.00
0.008
vertex-pair. The proposed distance measure considers both vertex
and edge weight-vectors, and it is flexible enough to assign different
weightage to different edges and scale the overall edge-weight while
computing the weighted distance between a vertex-pair. We have
also proposed a MAGDist algorithm that reads the lists of vertex and
edge vectors as two separate CSV files and calculates the distance
between each vertex-pairs using the proposed weighted distance
measure. Finally, we have proposed a multi-attributed similarity
graph generation algorithm, MAGSim, which reads the output produced by the MAGDist algorithm and generates a similarity graph,
which can be used by the existing classification and clustering algorithms for various analysis tasks. Since the proposed MAGDist
algorithm reads multi-attributed graph as CSV files containing
vertex and edge vectors, it can be scaled to handle large complex
graphs (aka big graphs). Applying proposed distance measure and
algorithms on (research articles) citation networks and online social
networks to analyze them at different levels of granularity is one
of the future directions of work.
REFERENCES
[1] Muhammad Abulaish and Sajid Yousuf Bhat. 2015. Classifier Ensembles using
Structural Features for Spammer Detection in Online Social Networks. Foundations of Computing and Decision Sciences 40, 2 (2015), 89–105.
A Novel Weighted Distance Measure for Multi-Attributed Graph
[2] Rakesh Agrawal, Johannes Gehrke, Dimitrios Gunopulos, and Prabhakar Raghavan. 1998. Automatic Subspace Clustering of High Dimensional Data for Data
Mining Applications. In Proceedings of the ACM SIGMOD international conference
on Management of data. Seattle, Washington, USA, 94–105.
[3] Faraz Ahmed and Muhammad Abulaish. 2013. A Generic Statistical Approach
for Spam Detection in Online Social Networks. Computer Communications 36,
10-11 (2013), 1120–1129.
[4] Sajid Yousuf Bhat and Muhammad Abulaish. 2013. Analysis and Mining of
Online Social Networks - Emerging Trends and Challenges. WIREs Data Mining
and Knowledge Discovery 3, 6 (2013), 408–444.
[5] Sajid Yousuf Bhat and Muhammad Abulaish. 2014. HOCTracker: Tracking the
Evolution of Hierarchical and Overlapping Communities in Dynamic Social
Networks. IEEE Transactions on Knowledge and Data Engineering 27, 4 (2014),
1019–1032. DOI:http://dx.doi.org/10.1109/TKDE.2014.2349918
[6] Sajid Yousuf Bhat and Muhammad Abulaish. 2014. Using communities against
deception in online social networks. Computer Fraud & Security 2014, 2 (2014),
8–16.
[7] Sajid Yousuf Bhat and Muhammad Abulaish. 2015. OCMiner: A Density-Based
Overlapping Community Detection Method for Social Networks. Intelligent Data
Analysis 19, 4 (2015), 1–31.
[8] Sajid Yousuf Bhat and Muhammad Abulaish. 2017. A Unified Framework for
Community Structure Analysis in Dynamic Social Networks. In Hybrid Intelligence for Social Networks, Hema Banati, Siddhartha Bhattacharyya, Ashish Mani,
and Mario Koppen (Eds.). Springer, 1–21.
[9] Sajid Y. Bhat, Muhammad Abulaish, and Abdulrahman A. Mirza. 2014. Spammer Classification using Ensemble Methods over Structural Social Network
Features. In Proceedings of the 14th IEEE/WIC/ACM International Conference on
Web Intelligence (WIâĂŹ14). Warsaw, Poland, 11–14.
[10] Mohd Fazil and Muhammad Abulaish. 2017. Identifying Active, Reactive, and
Inactive Targets of Socialbots in Twitter. In Proceedings of the IEEE/WIC/ACM
International Conference on Web Intelligence (WIâĂŹ17), Leipzig, Germany. ACM,
573–580.
[11] Mohd Fazil and Muhammad Abulaish. 2017. Why a Socialbot is Effective in
Twitter? A Statistical Insight. In Proceedings of the 9th International Conference on
Communication Systems and Networks (COMSNETS), Social Networking Workshop,
Bengaluru, India. 562–567.
Compute ’17, November 16–18, 2017, Bhopal, India
[12] David Gibson, Jon Kleinberg, and Prabhakar Raghavan. 1998. Inferring Web
Communities from Link Topology. In Proceedings of the Ninth ACM Conference
on Hypertext and Hypermedia. ACM, New York, USA, 225–234.
[13] M. Girvan and M. E. J. Newman. 2002. Community structure in social and
biological networks. In Proceedings of the National Academy of Sciences. USA,
8271–8276.
[14] Natallia Katenka and Eric D. Kolaczyk. 2011. Multi-Attribute Networks and the
Impact of Partial Information on Inference and Characterization. The Annals of
Applied Statistics 6, 3 (2011), 1068–1094.
[15] Mladen Kolar, Han Liu, and Eric P. Xing. 2014. Graph Estimation From MultiAttribute Data. Journal of Machine Learning Research 15 (2014), 1713–1750.
http://jmlr.org/papers/v15/kolar14a.html
[16] M. E. Newman and M. Girvan. 2004. Finding and evaluating community structure
in networks. Physical Review E69, 026113 (Feb 2004).
[17] Raymond T. Ng and Jiawei Han. 1994. Efficient and Effective Clustering Methods
for Spatial Data Mining. In Proceedings of the 20th International Conference on
Very Large Data Bases. Santiago, Chile, 144–155.
[18] Peter J. Olver. 2008. Numerical Analysis Lecture Note. Retrieved on 18.03.2017
from http://www-users.math.umn.edu/~olver/num_/lnn.pdf.
[19] S. E. Schaeffer. 2007. Graph clustering. Computer Science Review 1, 1 (2007),
27–64. DOI:http://dx.doi.org/10.1016/j.cosrev.2007.05.001
[20] Yuanyuan Tian, Richard A. Hankins, and Jignesh M. Patel. 2008. Efficient Aggregation for Graph Summarization. In Proceedings of the ACM SIGMOD International
Conference on Management of Data. Vancouver, Canada, 567–580.
[21] E.W. Weisstein. 2002. Vector Norm. Wolfram MathWorld. Retrieved on 18.03.2017
from http://mathworld.wolfram.com/VectorNorm.html.
[22] Xiaowei Xu, Nurcan Yuruk, Zhidan Feng, and Thomas A. J. Schweiger. 2007.
SCAN: A Structural Clustering Algorithm for Networks. In Proceedings of the
13th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining. San Jose, California, USA, 824–833.
[23] Mohammed J. Zaki and Wagner Meira Jr. 2014. Data Mining and Analysis:
Fundamental Concepts and Algorithms. Cambridge University Press, New York,
USA.
[24] Yang Zhou, Hong Cheng, and Jeffrey Xu Yu. 2009. Graph Clustering Based on
Structural/Attribute Similarities. VLDB Endowment 2, 1 (Aug 2009), 718–729.
DOI:http://dx.doi.org/10.14778/1687627.1687709
| 2cs.AI
|
arXiv:1603.07051v1 [cs.AI] 23 Mar 2016
Cosolver2B: An Efficient Local Search Heuristic for
the Travelling Thief Problem
Mohamed El Yafrani
Belaı̈d Ahiod
LRIT, associated unit to CNRST (URAC 29)
Faculty of Science, Mohammed V University of Rabat
B.P. 1014 Rabat, Morocco
Email: [email protected]
LRIT, associated unit to CNRST (URAC 29)
Faculty of Science, Mohammed V University of Rabat
B.P. 1014 Rabat, Morocco
Email: [email protected]
Abstract—Real-world problems are very difficult to optimize.
However, many researchers have been solving benchmark problems that have been extensively investigated for the last decades
even if they have very few direct applications. The Traveling Thief
Problem (TTP) is a NP-hard optimization problem that aims to
provide a more realistic model. TTP targets particularly routing
problem under packing/loading constraints which can be found in
supply chain management and transportation. In this paper, TTP
is presented and formulated mathematically. A combined local
search algorithm is proposed and compared with Random Local
Search (RLS) and Evolutionary Algorithm (EA). The obtained
results are quite promising since new better solutions were found.
I. I NTRODUCTION
The travelling thief problem (TTP) is a novel NP-hard
problem introduced in [1] to provide a model that better
represents real-world problems. The particularity of TTP is
that it is composed of two other NP-hard problems, namely
the travelling salesman problem and the knapsack problem,
which are interdependent.
The problem can be introduced with the following simplified statement:
”Given n cities and m items scattered among these cities, a
thief with his rented knapsack should visit all n cities, once
and only once each, and pick up some items. The more the
knapsack gets heavier, the more the thief becomes slower. What
is the best path and picking plan to adopt to achieve the best
benefit ?”.
Two particular properties can be extracted from the statement above. First, the overall problem is composed of two
sub-problems: choosing the best path and picking up the most
profitable items. Second, the sub-problems are interdependent:
when the knapsack gets heavier, the speed of the thief decreases. The composition increases the number of possible
solutions, and the interdependence makes it impossible to
isolate sub-problems.
This kind of multiple interdependent components is widely
found in logistics and supply chain management. Nevertheless, few attempts have been made to study and solve these
problems as a whole, and a lot of effort have been made to
solve the components independently. In [3], the authors explain
why there is a gap between the work done by the Evolutionary Computation researchers and real-world applications. As
far as we know, these observations can be extended to all
metaheuristics community.
Examples of other realistic problems with multiple interdependent sub-problems include most supply chain optimization
problems [7], [8], routing problems with loading constraints
such as 2L-CVRP and 3L-CVRP [9], [10], [3], and water tank
delivery [14].
Since TTP was introduced, some algorithms were proposed
to solve it. An Evolutionary Algorithm and a Random Local
Search were proposed in [13] to provide a starting point to
other researchers. In [2], an approach named Cosolver was
introduced to solve TTP by separating the sub-problems and
managing a communication between them. Lastly, an interesting work on TTP introduces many complexity reduction
techniques in order to solve very large instances in a time
budget of 10 minutes [11].
In this paper, we follow the ideas proposed in [2] and [11]
to propose further improvements and introduce an algorithm
based on the Cosolver framework that can solve TTP instances
efficiently.
This paper is organized as follows. In section II TTP is
mathematically defined and investigated. Section III is dedicated to introduce our approach for solving TTP. The tests and
results are presented in section IV. Section V concludes the
paper and outlines areas for future research.
II. BACKGROUND
In this section, we present some background information
about TTP. The problem is formulated and an abstract algorithm is presented.
A. The Travelling Thief Problem
Herein we formulate the Travelling Thief Problem (TTP)
which combines two other well known benchmark problems,
namely the Travelling Salesman Problem (TSP) and the Knapsack Problem (KP).
In TTP, we consider n cities and the associated distance
matrix {dij }. There are m items scattered in these cities, each
item k have a profit pk and a weight wk . A thief with his
knapsack is going to visit all these cities (once and only once
978-1-5090-0478-2/15/$31.00 c 2015 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any
current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
each), and pick up some items to fill his knapsack. We note
W the maximum capacity of the knapsack, vmin and vmax are
the minimum and maximum possible velocity respectively.
Also, we consider also the following constraints and parameters:
•
•
•
Each item is available in only one city. We note {Ai } the
availability vector, Ai ∈ {1, ..., n} contains the reference
to the city that contains the item i.
We suppose that the knapsack is rented. We note R the
renting price per time unit.
The velocity of the thief changes accordingly to the
knapsack weight. We note vxi the velocity of the thief
at city xi (see equation 1).
vxi = vmax − C ∗ wxi
(1)
−vmin
where C = vmaxW
is a constant value, and wxi
represents the weight of the knapsack at city xi .
The goal of the problem is to find the tour x and the picking
plan z that optimize the total travel gain defined in equation
2.
G(x, z) = g(z) − R ∗ f (x, z)
(2)
P
Where g(z) = m pm ∗ zm is the total value of the items
Pn−1
P
subject to m wm ∗ zm ≤ W , and f (x, z) = i=1 txi ,xi+1 +
txn ,x1 is the total travel time.
The solutions could be naturally coded as follows. The tour
x = (x1 , ..., xn ) is a vector containing the ordered list of cities,
and the picking plan z = (z1 , ..., zn ) is a binary vector such
as zi is equal to 1 if the item i is picked, 0 otherwise.
The interdependence between KP and TSP have been investigated. As shown in [1], [11], optimizing the sub-problems
in isolation (even to optimality) does not guarantee finding
good solutions for the overall problem. Therefore, finding good
global solutions requires an algorithm that takes interdependence of components in consideration, which makes the design
of such an algorithm quite difficult.
B. Cosolver
Cosolver is an abstract algorithm proposed in [2] to solve
TTP. The idea behind the algorithm is simple. Decompose
the overall problem and solve components separately using a
fitness function that takes in consideration the interdependence
between the two components.
Thus, given an initial solution (x0 , z0 ), TTP is decomposed
into the following problems.
1) The Travelling Salesman with Knapsack Problem
(TSKP): consists of finding the best tour x′ that, combined
with the last found picking plan z ′ , optimizes the TTP objective function G. Also, because the total profit function g
does not depend on the tour, instead of maximizing G, we
can consider minimizing the total travel cost T (see equation
3).
T (x, z) = R ∗ f (x, z)
(3)
2) The Knapsack on the Route Problem (KRP): consists of
finding the best picking plan z ′ that maximizes the total gain
G when combined with the last found tour x′ .
Figure 1 represents a flowchart of the Cosolver framework.
III. P ROPOSED
APPROACH
In this section, we propose an algorithm that implements
the Cosolver framework and uses some techniques proposed
in [11].
A. Solution initialization
In TTP, the initial solution has a big impact on the search
algorithm. Thus, the initialization strategy should be chosen
carefully. In our implementation, we use a heuristic approach
to initialize both the tour and the picking plan.
Firstly, the tour is generated using a good TSP algorithm
such as the Lin-Kernighan heuristic [6] or the Ant Colony
Optimization algorithm [5].
Then, the picking plan is initialized using the following
approach.
1) The insertion heuristic proposed in [11] is used to
fill the knapsack with items according to three fitness
approximations.
2) A simple bit-flip search on inserted items is performed
to eliminate some useless items from the picking plan
using the objective function.
We will refer to this heuristic as the insertion & elimination
heuristic.
B. Complexity reduction techniques
Our algorithm uses the following complexity reduction
techniques.
1) TSKP neighborhood reduction: The Delaunay triangulation [4] is used as a candidate generator for the 2-OPT
heuristic. This strategy was also proposed in [11].
2) Objective value recovery: Instead of using the objective
function to calculate the objective value of neighbors, this
value can be recovered by keeping track of time and weight
information at each tour’s city. The following vectors are used
to perform such operation.
acc
• Time accumulator (t
): a vector that contains the current
time at each city of the tour.
acc
• Weight accumulator (w
): a vector that contains the
current weight at each city of the tour.
reg
• Time register (t
): a vector that contains the added time
at each city of the tour.
reg
• Weight register (w
): a vector that contains the added
weight at each city of the tour.
Figure 2 shows the effect (in gray) of mutating a TTP
solution on the total travel time. Two mutations are shown
(in black): the one-bit-flip on the picking plan, and a 2-OPT
exchange on the tour.
A similar technique named incremental evaluation was
proposed in [11]. In our implementation, we go further and
use these techniques also to recover the vectors.
Fig. 1.
bit- ip
x(1)
A simplified Cosolver flowchart
x(i)
x(n)
t(1)
t(i-1)
x(1)
2-opt
t(n)
x(j)
x(i)
x(n)
t(1)
t(n)
t(i)
t(j)
Fig. 2. Example of changes that can be applied to a TTP solution (in black)
and its effect on the total traveling time function (in gray)
3) Objective function calculation: The objective function
has a complexity of O(m ∗ n) as proposed in [12], where
m is the number of items and n is the number of cities. The
complexity could be reduced to O(k∗n) where k is the number
of items per city. In the CEC ’2015 competition website 1 ,
the Java implementations already has a reduced complexity
but it uses a technique highly dependent on the TTP instance
generator. We believe that a better way is to class items per
city.
Figure 3 explains this technique using a simple example. In
the example, 4 cities are considered: city 1 contains 4 items (1,
2, 3 and 4), city 2 contains 2 items (5 and 6), city 3 contains
no items, city 4 contains 3 items (7, 8 and 9).
Fig. 4.
Fig. 3.
Data structure for classing items by cities
C. Proposed algorithm
Our heuristic, namely Cosolver2B, is based on local search
algorithms, and it implements the Cosolver framework. In
addition to the techniques presented above, we use the 2-OPT
heuristic to solve TSKP and a bit-flip search to solve the KRP.
The algorithm is summarized in the flowchart in figure 4.
1 http://cs.adelaide.edu.au/
optlog/CEC2015Comp/
Cosolver2B: Cosolver with 2-opt and bit-flip
The algorithm starts with a tour generated using the LinKernighan heuristic. The insertion & elimination heuristic is
then used to initialize the picking plan. The obtained tour
and picking plan are then given to the TSKP optimizer that
tries to improve the current tour using a 2-OPT heuristic.
The resulting tour is combined with the current picking plan
and given to the KRP optimizer that uses a simple bit-flip to
find a decent picking plan. If there is improvement, the new
picking plan is combined with the current tour and given to
the TSKP optimizer. This process is repeated until there is no
improvement.
IV. E XPERIMENTAL RESULTS
We tested our algorithm on 15 TTP instances of various
sizes [13]. The obtained results are compared with RLS and
TABLE I
R ESULTS OF C OSOLVER 2B COMPARED TO EA AND RLS
instance
berlin52 n255 uncorr-similar-weights 05
kroA100 n495 uncorr-similar-weights 05
ch150 n745 uncorr-similar-weights 05
ts225 n1120 uncorr-similar-weights 05
a280 n1395 uncorr-similar-weights 05.ttp
lin318 n1585 uncorr-similar-weights 05
u574 n2865 uncorr-similar-weights 05
dsj1000 n4995 uncorr-similar-weights 05
rl1304 n6515 uncorr-similar-weights 05
fl1577 n7880 uncorr-similar-weights 05
d2103 n10510 uncorr-similar-weights 05
pcb3038 n15185 uncorr-similar-weights 05
fnl4461 n22300 uncorr-similar-weights 05
rl11849 n59240 uncorr-similar-weights 05
usa13509 n67540 uncorr-similar-weights 05
pla33810 n169045 uncorr-similar-weights 05
EA
Cosolver2B
firstfit
Cosolver2B
bestfit
objective
time
objective
RLS
time
objective
time
objective
time
20591
38864
40922
84907
104365
75387
223165
320165
573667
589756
801110
1119168
1477673
2152954
2226676
-9929644
1.28
1.92
2.69
3.82
5.64
6.87
21
47
95
127
214
427
600
600
600
600
20591
38864
40922
84907
104365
75387
223165
320165
573667
589756
801110
1119169
1478963
4328939
5077892
-3158915
0.89
1.06
1.38
1.69
1.84
3.81
5.42
12
21
28
44
101
265
600
600
600
25440
39599
52225
88925
107233
114915
243571
332916
571634
665866
869584
1117659
1220469
1785671
4420300
14887314
0.13
0.21
0.38
0.61
0.88
1.25
5.36
5.95
68
139
357
600
600
600
600
600
26658
39597
53075
89079
112003
114915
253770
332917
586576
684731
869584
1086092
696650
478884
4082169
15085752
0.11
0.27
0.43
0.92
1.09
1.54
6.51
8.44
83
148
389
600
600
600
600
600
EA proposed in the same paper 2 .
We have implemented two versions of Cosolver2B. The first
is best fit (also known as best improvement) which searches
the entire neighborhood on a given iteration and selects the
best neighbor. The second is first fit (also known as first
improvement) which stops the neighborhood search once a
better solution is found.
Since our algorithms are deterministic, one run per instance
is sufficient. The algorithms have a maximum runtime limit
of 600 seconds. Note that our implementation uses the Java
platform, and the tests are performed on a core i3-2370M CPU
2.40GHz machine with 4 GB of RAM, running Linux.
On the other hand, due to their random behavior, 30 runs
per instance were performed using RLS and EA. The tests on
EA and RLS were performed on a different machine (Matlab
2014, core i7 2600 CPU 3.4 GHz, with 4GB of RAM, running
Windows 7). Thus, The runtimes for these algorithms are given
for guidance. Furthermore, in our comparison, we only select
the best found solution for these two algorithms.
The results are reported in table I. Note that the runtimes
are measured with seconds and the best objective values are
made bolder.
The results show that our algorithm was able to find new
better solutions for most tested instances only by combining
local search algorithms. Furthermore, we have made the following observations:
• Cosolver2B surpasses EA and RLS on various TTP sizes.
• The runtime of Cosolver2B is very decent for small and
mid-size instances. However, for very large instances the
runtime is quite high compared to RLS and EA.
2 for
instance files and raw results of RLS and EA, the reader is referred to
https://sites.google.com/site/mohammadrezabonyadi/standarddatabases/travellingthief-problem-data-bases-and-raw-results
•
•
•
Both EA and RLS have an unpredictable behavior, while
Cosolver2B is deterministic and garantees decent solutions for most instances.
For many instances, most optimization runtime and gain
improvement is done on the KRP sub-problem.
The best fit strategy performed better than first fit for
most small and mid-size instances.
V. C ONCLUSION
In this paper, a combined local search heuristic has been
proposed. Our approach is based on the Cosolver framework
and complexity reduction techniques. The experimental results
have shown promising results both in terms of solution quality
and runtime. The good performance of Cosolver2B and the
memetic algorithm MALTS on very large scale instances
[11] proves the importance of making strong assumptions
about the problem domain when dealing with multi-component
problems.
Furthermore, we engage to improve our algorithm in terms
of space exploration and runtime by using more sophisticated
search heuristics.
ACKNOWLEDGMENT
The authors express their gratefulness to the following
persons who have contributed to this work with their useful
comments: Asmae El Ghazi, Aziz Ouaarab, and Mehdi El
Krari.
R EFERENCES
[1] Mohammad Reza Bonyadi, Zbigniew Michalewicz, and Luigi Barone.
The travelling thief problem: the first step in the transition from
theoretical problems to realistic problems. In Evolutionary Computation
(CEC), 2013 IEEE Congress on, pages 1037–1044. IEEE, 2013.
[2] Mohammad Reza Bonyadi, Zbigniew Michalewicz, Michaŏ Roman Przybyŏek, and Adam Wierzbicki. Socially inspired algorithms
for the travelling thief problem. In Proceedings of the 2014 conference
on Genetic and evolutionary computation, pages 421–428. ACM, 2014.
[3] Andreas Bortfeldt. A hybrid algorithm for the capacitated vehicle
routing problem with three-dimensional loading constraints. Computers
& Operations Research, 39(9):2248–2257, 2012.
[4] Boris Delaunay. Sur la sphere vide. Izv. Akad. Nauk SSSR, Otdelenie
Matematicheskii i Estestvennyka Nauk, 7(793-800):1–2, 1934.
[5] Marco Dorigo and Luca Maria Gambardella. Ant colonies for the
travelling salesman problem. BioSystems, 43(2):73–81, 1997.
[6] Keld Helsgaun. An effective implementation of the lin–kernighan
traveling salesman heuristic. European Journal of Operational Research,
126(1):106–130, 2000.
[7] Maksud Ibrahimov, Arvind Mohais, Sven Schellenberg, and Zbigniew
Michalewicz. Evolutionary approaches for supply chain optimisation:
part i: single and two-component supply chains. International Journal
of Intelligent Computing and Cybernetics, 5(4):444–472, 2012.
[8] Maksud Ibrahimov, Arvind Mohais, Sven Schellenberg, and Zbigniew
Michalewicz. Evolutionary approaches for supply chain optimisation.
part ii: multi-silo supply chains. International Journal of Intelligent
Computing and Cybernetics, 5(4):473–499, 2012.
[9] Manuel Iori and Silvano Martello. Routing problems with loading
constraints. Top, 18(1):4–27, 2010.
[10] Stephen CH Leung, Zhenzhen Zhang, Defu Zhang, Xian Hua, and
Ming K Lim. A meta-heuristic algorithm for heterogeneous fleet vehicle
routing problems with two-dimensional loading constraints. European
Journal of Operational Research, 225(2):199–210, 2013.
[11] Yi Mei, Xiaodong Li, and Xin Yao. Improving efficiency of heuristics
for the large scale traveling thief problem. In Simulated Evolution and
Learning, pages 631–643. Springer, 2014.
[12] Yi Mei, Xiaodong Li, and Xin Yao. On investigation of interdependence
between sub-problems of the travelling thief problem. Soft Computing,
pages 1–16, 2014.
[13] Sergey Polyakovskiy, Mohammad Reza, Markus Wagner, Zbigniew
Michalewicz, and Frank Neumann. A comprehensive benchmark set
and heuristics for the traveling thief problem. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO), Vancouver,
Canada, 2014.
[14] Jacob Stolk, Isaac Mann, Arvind Mohais, and Zbigniew Michalewicz.
Combining vehicle routing and packing for optimal delivery schedules
of water tanks. OR Insight, 26(3):167–190, 2013.
| 8cs.DS
|
arXiv:cs/0204047v2 [cs.CE] 22 Apr 2002
Sampling Strategies for Mining in Data-Scarce Domains
Chris Bailey-Kellogg
Department of Computer Sciences
Purdue University, IN 47907
Tel: (765) 494-9025
Email: [email protected]
Naren Ramakrishnan
Department of Computer Science
Virginia Tech, VA 24061
Tel: (540) 231-8451
Email: [email protected]
Abstract
Data mining has traditionally focused on the task of drawing inferences from large datasets. However, many
scientific and engineering domains, such as fluid dynamics and aircraft design, are characterized by scarce data,
due to the expense and complexity of associated experiments and simulations. In such data-scarce domains, it
is advantageous to focus the data collection effort on only those regions deemed most important to support a
particular data mining objective. This paper describes a mechanism that interleaves bottom-up data mining, to
uncover multi-level structures in spatial data, with top-down sampling, to clarify difficult decisions in the mining
process. The mechanism exploits relevant physical properties, such as continuity, correspondence, and locality,
in a unified framework. This leads to effective mining and sampling decisions that are explainable in terms of
domain knowledge and data characteristics. This approach is demonstrated in two diverse applications — mining
pockets in spatial data, and qualitative determination of Jordan forms of matrices.
1 Introduction
A number of important scientific and engineering applications, such as fluid dynamics simulation and aircraft design,
require analysis of spatially-distributed data from expensive experiments and/or complex simulations demanding
days, weeks, or even years on petaflops-class computing systems. For example, consider the conceptual design of
a high-speed civil transport (HSCT), which involves the disciplines of aerodynamics, structures, controls (missionrelated), and propulsion. 80% of the aircraft lifecycle cost is determined at this stage. Fig. 1 shows a cross-section
of the design space for such a problem involving 29 design variables with 68 constraints [10]. Frequently, the
engineer will change some aspect of a nominal design point, and run a simulation to see how the change affects
the objective function and various constraints dealing with aircraft geometry and performance/aerodynamics. Or
the design process is made configurable, so the engineer can concentrate on accurately modeling some aspect (e.g.,
the interaction between the wing root and the fuselage) while replacing the remainder of the design with fixed
boundary conditions surrounding the focal area. Both these approaches are inadequate for exploring such large
high-dimensional design spaces, even at low fidelity. Ideally, the design engineer would like a high-level mining
system to identify the pockets that contain good designs and which merit further consideration; traditional tools
from optimization and approximation theory can then be applied to fine-tune such preliminary analyses.
Three important characteristics distinguish such applications. First, they are characterized not by an abundance
of data, but rather by a scarcity of data (owing to the cost and time involved in conducting simulations). Second, the
computational scientist has complete control over the data acquisition process (e.g. regions of the design space where
data can be collected), especially via computer simulations. And finally, there exists significant domain knowledge
in the form of physical properties such as continuity, correspondence, and locality. It is natural therefore to use
such information to focus data collection for data mining. In this paper, we are interested in the question: ‘Given a
simulation code, knowledge of physical properties, and a data mining goal, at what points should data be collected?’
1
no constraint violations
constraints active
constraints violated
range constraint
landing CLmax and tip scrape
tip spike constraint
Optimum 2
TOGW
658000
657000
656000
655000
654000
653000
652000
651000
650000
649000
648000
647000
646000
645000
644000
Optimum 1
Figure 1: A pocket in an aircraft design space viewed as a slice through three design points [10] (courtesy Layne T.
Watson).
By suitably formulating an objective function and constraints around this question, we can pose it as a problem
of minimizing the number of samples needed for data mining. Such a combination of {data-scarcity + control
over data collection + need to exploit domain knowledge} characterizes many important computational science
applications. Data mining is now recognized as a key solution approach for such applications, supporting analysis,
visualization, and design tasks [17]. It serves a primary role in many domains (e.g., microarray bioinformatics) and a
complementary role in others, by augmenting traditional techniques from numerical analysis, statistics, and machine
learning.
The goal of this paper is to describe focused sampling strategies for mining scientific data. Our approach is
based on the spatial aggregation language (SAL) [3], which supports construction of data interpretation and control
design applications for spatially-distributed physical systems. Used as a basis for describing data mining algorithms,
SAL programs also help exploit knowledge of physical properties such as continuity and locality in data fields. They
work in a bottom-up manner to uncover regions of uniformity in spatially distributed data. In conjunction with
this process, we introduce a top-down sampling strategy that focuses data collection in only those regions that are
deemed most important to support a data mining objective. Together, they help define a methodology for mining in
data-scarce domains. We describe this methodology at a high-level and devote the major part of the paper to two
applications that employ it.
2
2 A Methodology for Mining in Data-Scarce Domains
It is possible to study the problem of sampling for targeted data mining activities, such as clustering, finding association rules, and decision tree construction [9]. This is the approach taken by work such as [12]. In this paper,
however, we are interested in a general framework or language to express data mining operations on datasets and
which can be used to study the design of data collection and sampling strategies. The spatial aggregation language
(SAL) [3, 19] is such a framework.
2.1 SAL: The Spatial Aggregation Language
As a data mining framework, SAL is based on successive manipulations of data fields by a uniform vocabulary
of aggregation, classification, and abstraction operators. Programming in SAL follows a philosophy of building a
multi-layer hierarchy of aggregations of data. These increasingly abstract descriptions of data are built using explicit
representations of physical knowledge, expressed as metrics, adjacency relations, and equivalence predicates. This
allows a SAL program to uncover and exploit structures in physical data.
SAL programs employ what has been called an imagistic reasoning style [20]. They employ vision-like routines
to manipulate multi-layer geometric and topological structures in spatially distributed data. SAL adopts a field
ontology, in which the input is a field mapping from one continuum to another (e.g. 2-D temperature field: R2 → R1 ;
3-D fluid flow field: R3 → R3 ). Multi-layer structures arise from continuities in fields at multiple scales. Due to
continuity, fields exhibit regions of uniformity, and these regions of uniformity can be abstracted as higher-level
structures which in turn exhibit their own continuities. Task-specific domain knowledge specifies how to uncover
such regions of uniformity, defining metrics for closeness of both field objects and their features. For example,
isothermal contours are connected curves of nearby points with equal (or similar enough) temperature.
The identification of structures in a field is a form of data reduction: a relatively information-rich field representation is abstracted into a more concise structural representation (e.g. pressure data points into isobar curves or pressure
cells; isobar curve segments into troughs). Navigating the mapping from field to abstract description through multiple layers rather than in one giant step allows the construction of more modular programs with more manageable
pieces that can use similar processing techniques at different levels of abstraction. The multi-level mapping also
allows higher-level layers to use global properties of lower-level objects as local properties of the higher-level objects. For example, the average temperature in a region is a global property when considered with respect to the
temperature data points, but a local property when considered with respect to a more abstract region description. As
this paper demonstrates, analysis of higher-level structures in such a hierarchy can guide interpretation of lower-level
data.
SAL supports structure discovery through a small set of generic operators, parameterized with domain-specific
knowledge, on uniform data types. These operators and data types mediate increasingly abstract descriptions of the
input data (see Fig. 2) to form higher-level abstractions and mine patterns. The primitives in SAL are contiguous
regions of space called spatial objects; the compounds are (possibly structured) collections of spatial objects; the
abstraction mechanisms connect collections at one level of abstraction with single objects at a higher level.
SAL is currently available as a C++ library1 providing access to a large set of data type implementations and
operations. In addition, an interpreted, interaction environment layered over the library supports rapid prototyping
of data mining applications. It allows users to inspect data and structures, test the effects of different predicates, and
graphically interact with representations of the structures.
To illustrate SAL programming style, consider the task of bundling vectors in a given vector field (e.g. wind
velocity or temperature gradient) into a set of streamlines (paths through the field following the vector directions).
This process can be depicted as shown in Fig. 3 and the corresponding SAL data mining program is shown in
1
The SAL implementation can be downloaded from http://www.cis.ohio-state.edu/insight/sal-code.html.
3
Abstract Description
Higher-Level Objects
Disaggregate
Redescribe
Structural description
Equivalence
classes
Map, Filter,
geometric ops
Update
Classify
Spatial
objects
Redescribe
Behavioral description
Aggregate
Disaggregate
N-graph
Search
Consistent?
Lower-Level Objects
Input Field
Figure 2: SAL multi-layer spatial aggregates, uncovered by a uniform vocabulary of operators utilizing domain
knowledge. A variety of scientific data mining tasks, such as vector field bundling, contour aggregation, correspondence abstraction, clustering, and uncovering regions of uniformity can be expressed as multi-level computations
with SAL aggregates.
4
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 3: Example steps in SAL implementation of vector field analysis application. (a) Input vector field. (b) 8adjacency neighborhood graph. (c) Forward neighbors. (d) Best forward neighbors. (e) Ngraph transposed from best
forward neighbors. (f) Best backward neighbors. (g) Resulting adjacencies redescribed as curves. (h) Higher-level
aggregation and classification of curves whose flows converge.
5
// (a) Read vector field.
vect_field = read_point_point_field(infile);
points = domain_space(vect_field);
// (b) Aggregate with 8-adjacency (i.e. within 1.5 units).
point_ngraph = aggregate(points, make_ngraph_near(1.5));
// (c) Compare vector directions with node-neighbor direction.
angle = function (p1, p2) {
dot(normalize(mean(feature(vect_field, p1), feature(vect_field, p2))),
normalize(subtract(p2, p1)))
}
forward_ngraph = filter_ngraph(adj in point_ngraph, {
angle(from(adj), to(adj)) > angle similarity
})
// (d) Find best forward neighbor, comparing vector direction
// with ngraph edge direction and penalizing for distance.
forward_metric = function (adj) {
angle(from(adj), to(adj)) - distance penalty * distance(from(adj),to(adj))
}
best_forward_ngraph = best_neighbors_ngraph(forward_ngraph, forward_metric);
// (e) Find backward neighbors by transposing best forward neighbors.
backward_ngraph = transpose_ngraph(best_forward_ngraph);
// (f) At junctions, keep best backward neighbor using metric
// similar to that for best forward neighbors.
backward_metric = function (adj) {
angle(to(adj), from(adj)) - distance penalty*distance(from(adj),to(adj))
}
best_backward_ngraph = best_neighbors_ngraph(backward_ngraph, backward_metric);
// (g) Move to a higher abstraction level by forming equivalence classes
// from remaining groups and redescribing them as curves.
final_ngraph = symmetric_ngraph(best_backward_ngraph, extend=true);
point_classes = classify(points, make_classifier_transitive(final_ngraph));
points_to_curves = redescribe(classes(point_classes),
make_redescribe_op_path_nline(final_ngraph));
trajs = high_level_objects(points_to_curves);
Figure 4: SAL data mining program for the vector field analysis application of Fig. 3.
6
Fig. 4. The steps in this program are as follows: (a) Establish a field mapping points (locations) to points (vector
directions, assumed here to be normalized). (b) Localize computation with a neighborhood graph, so that only
spatially proximate points are compared. (c)–(f) Use a series of local computations on this representation to find
equivalence classes of neighboring vectors with respect to vector direction (systematically eliminate all edges but
those whose directions best match the vector direction at both endpoints). (g) Redescribe equivalence classes of
vectors into more abstract streamline curves. (h) Aggregate and classify these curves into groups with similar flow
behavior, using the exact same operators but with different metrics (code not shown). As this example illustrates,
SAL provides a vocabulary for expressing the knowledge required (e.g., distance metrics and similarity metrics) for
uncovering multi-level structures in spatial datasets. It has been applied to applications ranging from decentralized
control design [2] to analysis of diffusion-reaction morphogenesis [16].
2.2 Data Collection and Sampling
The above example illustrated the use of SAL in a data-rich domain. The exploitation of physical properties is
a central tenet of SAL since it drives the computation of multi-level spatial aggregates. Many important physical
properties can be expressed as SAL computations by suitably defining adjacency relations and aggregation metrics.
To extend the use of SAL to data-scarce settings, we present the sampling methodology outlined in Fig. 5.
Once again, it is easy to understand the methodology in the context of the vector-field bundling application
(Fig. 3). Assume that we apply the SAL data mining program of Fig. 4 with a small dataset and have navigated upto
the highest level of the hierarchy (streamlines bundled with convergent flows). The SAL program computes different
streamline aggregations from a neighborhood graph and chooses one based on how well its curvature matches the
direction of the vectors it aggregates. If data is scarce, it is likely that some of these classification decisions will be
ambiguous, i.e., there may exist multiple streamline aggregations. In such a case, we would like to choose a new
data sample that reduces the ambiguity and clarifies what the correct classification should be.
This is the essence of our sampling methodology: using SAL aggregates, we identify an information-theoretic
measure (here, ambiguity) that can be used to drive stages of future data collection. For instance, the ambiguous
streamline classifications can be summarized as a 2D ambiguity distribution that has a spike for every location
where an ambiguity was detected. Reduction of ambiguity can be posed as the problem of minimization of (or
maximization, as the case may be) a functional involving the (computed) ambiguity. The functional could be the
entropy in the underlying data field, as revealed by the ambiguity distribution. Such a minimization will lead us to
selecting a data point(s) that clarifies the distribution of streamlines, and hence makes more effective use of data
for data mining purposes. The net effect of this methodology is that we are able to capture the desirability of a
particular design (data layout) in terms of computations involving SAL aggregates. Thus, sampling is conducted
for the express purpose of improving the quality and efficacy of data mining. The dataset is updated with the newly
collected value and the process is repeated till a desired stopping criteria is met. For instance, we could terminate
if the functional is within accepted bounds, or when there is no improvement in confidence of data mining results
between successive rounds of data collection. In our case, when there is no further ambiguity.
This idea of sampling to satisfy particular design criteria has been studied in various contexts, especially spatial
statistics [6, 11, 18]. Many of these approaches (including ours) rely on capturing properties of a desirable design
in terms of a novel objective function. The distinguishing feature of our work is that it uses spatial information
gleaned from a higher level of abstraction to focus data collection at the field/simulation code layer. The applications
presented here are also novel in that they span and connect arbitrary levels of abstraction, thus suggesting new ways
to integrate qualitative and quantitative simulation [4].
We present concrete realizations of the above methodology in the next section. But before we proceed, it is
pertinent to note an optional step in our methodology. The newly collected data value can be used to improve a
surrogate model which then generates a dense data field for mining. A surrogate function is something that is used
in lieu of the real data source, so as to generate sufficient data for mining purposes. This is often more advantageous
7
Step 2: Calculate information
SAL Objects
theoretic measure M from SAL
......
Step 3: Capture design merits
as a functional involving M
Step 1: Data mining using
successive spatial aggregation
SAL Objects
Step 4: Sample data point(s)
to minimize the functional
SAL Objects
Data field
Optional: Improve surrogate model
Figure 5: The sampling methodology for SAL mining in data-scarce domains.
than working directly with sparse data. Surrogate models are widely used in engineering design, optimization, and
in response surface approximations [13, 15].
Together, SAL and our focused sampling methodology address the main issues raised in the beginning of the
paper: SAL’s uniform use of fields and abstraction operators allows us to exploit prior knowledge in a bottom-up
manner. Discrepancies as suggested by our knowledge of physical properties (e.g., ambiguities) are used in a topdown manner by the sampling methodology. Continuing these two stages alternatively leads to a closed-loop data
mining solution for data-scarce domains.
3 Example Applications
3.1 Mining Pockets in Spatial Data
Our first application is motivated by the aircraft design problem and is meant to illustrate the basic idea of our
methodology. Here, we are given a spatial vector field and we wish to identify pockets underlying the gradient. In
a weather map, this might mean identifying pressure troughs, for instance. The question is: ‘where should data be
collected so that we are able to mine the pockets with high confidence?’ We begin by presenting a mathematical
function that gives rise to pockets in spatial fields. This function will be used to validate and test our data mining
and sampling methodology.
de Boor’s function
Carl de Boor invented a pocket function that exploits containment properties of the n-sphere of radius 1 centered
at the origin (Σxi 2 ≤ 1) with respect to the n-dimensional hypercube defined by xi ∈ [−1, 1], i = 1 · · · n. Even
though the sphere is embedded inside the cube, notice that the ratio of the volume of the cube (2n ) to that of the sphere
(π n/2 /(n/2)!) grows unboundedly with n. This means that the volume of a high-dimensional cube is concentrated
in its corners (a counterintuitive notion at first). de Boor exploited this property to design a difficult-to-optimize
8
1
0.5
0
−0.5
−1
−1.5
−2
1
0.5
1
0.5
0
0
−0.5
−0.5
−1
−1
Figure 6: A 2D pocket function.
function which assumes a pocket in each corner of the cube (Fig. 6), that is just outside the sphere. Formally, it can
be defined as:
!
n
X
xi
i
2 1+
α(X) = cos
−2
(1)
| xi |
i=1
δ(X) = kX − 0.5Ik
(2)
2
p(X) = α(X)(1 − δ (X)(3 − 2δ(X))) + 1
(3)
where X is the n-dimensional point (x1 , x2 , · · · , xn ) at which the pocket function p is evaluated, I is the identity
n-vector, and k · k is the L2 norm.
It is easily seen that p has 2n pockets (local minima); if n is large (say, 30, which means it will take more than
half a million points to just represent the corners of the n-cube!), naive global optimization algorithms will require
an unreasonable number of function evaluations to find the pockets. Our goal for data mining here is to obtain a
qualitative indication of the existence, number, and locations of pockets, using low-fidelity models and/or as few
data points as possible. The results can then be used to seed higher-fidelity calculations. This is also fundamentally
different from DACE [18], polynomial response surface approximations [13], and other approaches in geo-statistics
where the goal is accuracy of functional prediction at untested data points. Here, accuracy of estimation is traded for
the ability to mine pockets.
Surrogate Function
In this study, we use the SAL vector-field bundling code presented earlier along with a surrogate model as the basis
for generating a dense field of data. Surrogate theory is an established area in engineering optimization and there are
several ways in which we can build a surrogate. However, the local nature of SAL computations means that we can
be selective about our choice of surrogate representation. For example, global, least-squares type approximations
are inappropriate since measurements at all locations are equally considered to uncover trends and patterns in a
particular region. We advocate the use of kriging-type interpolators [18], which are local modeling methods with
roots in Bayesian statistics. Kriging can handle situations with multiple local extrema (for example, in weather data,
remote sensing data, etc.) and can easily exploit anisotropies and trends. Given k observations, the interpolated
model gives exact responses at these k sites and estimates values at other sites by minimizing the mean squared error
(MSE), assuming a random data process with zero mean and a known covariance function.
Formally (for two dimensions), the true function p is assumed to be the realization of a random process such as:
p(x, y) = β + Z(x, y)
9
(4)
where β is typically a uniform random variate, estimated based on the known k values of p, and Z is a correlation
function. Kriging then estimates a model p′ of the same form, based on the k observations:
p′ (xi , yi ) = E(p(xi , yi ) | p(x1 , y1 ), · · · , p(xk , yk ))
(5)
and minimizing mean squared error between p′ and p:
M SE = E(p′ (x, y) − p(x, y))2
(6)
A typical choice for Z in p′ is σ 2 R, where scalar σ 2 is the estimated variance, and correlation matrix R encodes
domain-specific constraints and reflects the current fidelity of data. We use an exponential function for entries in R,
with two parameters C1 and C2 :
2
2
(7)
Rij = e−C1 |xi −xj | −C2 |yi −yj |
Intuitively, values at closer points should be more highly correlated.
The estimator minimizing mean squared error is then obtained by multi-dimensional optimization (the derivation
from Eqs. 6 and 7 is beyond the scope of this paper):
max
C
−k
(ln σ 2 + ln |R|)
2
(8)
This expression satisfies the conditions that there is no error between the model and the true values at the chosen k
sites, and that all variability in the model arises from the design of Z. The multi-dimensional optimization is often
performed by gradient descent or pattern search methods. More details are available in [18], which demonstrates
this methodology in the context of the design and analysis of computer experiments.
Data Mining and Sampling Methodology
The bottom-up computation of SAL aggregates from the surrogate model’s outputs will possibly lead to some ambiguous streamline classifications, as discussed earlier. Ambiguity can reflect the desirability of acquiring data at
or near a specified point, to clarify the correct classification and to serve as a mathematical criterion of information
content. There are several ways in which we can use information about ambiguity to drive data collection. In this
study, we express the ambiguities as a distribution describing the number of possible good neighbors (for a streamline). This ambiguity distribution provides a novel mechanism to include qualitative information — streamlines that
agree will generally contribute less to data mining, for information purposes. The information-theoretic measure M
(ref. Fig. 5) was thus defined to be the ambiguity distribution ℘.
The functional was defined as the posterior entropy E(− log d), where d is the conditional density of ℘ over the
design space not covered by the current data values. By a reduction argument, minimizing this posterior entropy
can be shown to be maximizing the prior entropy over the unsampled design space [18]. In turn, this means that the
amount of information obtained from an experiment (additional data collection) is maximized. In addition, we also
incorporated ℘ as an indicator covariance term in our surrogate model (this is a conventional method for including
qualitative information in an interpolatory model [11]).
Experimental Results
The initial experimental configuration used a face-centered design (4 points in the 2D case). A surrogate model by
kriging interpolation then generated data on a 41n -point grid. de Boor’s function was used as the source for data
values; we also employed pseudorandom perturbations of it that shift the pockets from the corners in a somewhat
unpredictable way (see [1] for details). In total, we experimented with 100 perturbed variations (each) of the 2D
10
35
30
Krig
Ambig
Freq (%)
25
20
15
10
5
0
5
10
15
# Samples
20
25
Figure 7: Pocket-finding results (2D) show that focused sampling using a measure of ambiguity always requires
fewer total samples (7-15) than conventional kriging (17-23).
and 3D pocket functions. For each of these cases, data collection was organized in rounds of one extra sample each
(that minimizes the above functional). The number of samples needed to mine all the pockets by SAL was recorded.
We also compared our results with those obtained from a pure DACE/kriging approach (i.e., where sampling was
directed at improving accuracy of function estimation). In other words, we used the DACE methodology to suggest
new locations for data collection and determined how these choices fared with respect to mining the pockets.
Fig. 7 shows the distributions of total number of data samples required to mine the four pockets for the 2D
case. We were thus able to mine the 2D pockets using 3 to 11 additional samples, whereas the conventional kriging
approach required 13 to 19 additional samples. The results were were more striking in the 3D case: at most 42
additional samples for focused sampling and upto 151 points for conventional kriging. This shows that our focused
sampling methodology performs 40-75% better than sampling by conventional kriging.
Fig. 8 (left) describes a 2D design involving only 7 total data points that is able to mine the four pockets.
Counterintuitively, no additional sample is required in the lower left quadrant! While this will lead to a highly
sub-optimal design (from the traditional viewpoint of minimizing variance in predicted values), it is nevertheless an
appropriate design for data mining purposes. In particular, this means that neighborhood calculations involving the
other three quadrants are enough to uncover the pocket in the fourth quadrant. Since the kriging interpolator uses
local modeling and since pockets in 2D effectively occupy the quadrants, obtaining measurements at ambiguous
locations serves to capture the relatively narrow regime of each dip, which in turn helps to distinguish the pocket in
the neighboring quadrant. This effect is hard to achieve without exploiting knowledge of physical properties, in this
case, locality of the dips.
3.2 Qualitative Jordan Form Determination
In our second application, we use our methodology to identify the most probable Jordan form of a given matrix.
This is a good application for data mining since the direct computation of the Jordan form leads to a numerically
unstable algorithm.
11
1
0.5
0
−0.5
−1
−1
−0.5
0
0.5
1
Figure 8: Mining pockets in 2D from only 7 sample points. (left) The chosen sample locations: 4 initial facecentered samples (marked as blue circles) plus 3 samples selected by our methodology (marked as red diamonds).
Note that no additional sample is required in the lower-left quadrant. (right) SAL structures in surrogate model data,
confirming the existence of four pockets.
Jordan forms
A matrix A (real or complex) that has r independent eigenvectors has a Jordan form that consists of r blocks. Each
of these blocks is an upper triangular matrix that is associated with one of the eigenvectors of A and whose size
describes the multiplicity of the corresponding eigenvalue. For the given matrix A, the diagonalization thus posits a
nonsingular matrix B such that:
J1
J2
B −1 AB =
(9)
·
Jr
where
λi 1
· 1
Ji =
·
1
λi
(10)
and λi is the eigenvalue revealed by the ith Jordan block (Ji ). The Jordan form is most easily explained by looking
at how eigenvectors are distributed for a given eigenvalue. Consider, for example, the matrix
1
1 −1
0
0
2
0 −1
3
that has eigenvalues at 1, 1, and 2. This matrix has only two eigenvectors, as revealed by the two-block structure of
its Jordan form:
1 1 0
0 1 0
0 0 2
12
−3
1.5
−4
x 10
1.5
1
1
0.5
0.5
0
0
−0.5
−0.5
−1
−1
−1.5
6.998
6.9985
6.999
6.9995
7
7.0005
7.001
7.0015
7.002
x 10
6.9998
6.9999
6.9999
7
7
7.0001
7.0001
7.0002
7.0002
Figure 9: Superimposed spectra for assessing the Jordan form of the Brunet matrix. Two Jordan blocks of multiplicity 3 are observed for eigenvalue 7, at different (left, right) perturbation levels.
The Jordan form is unique modulo shufflings of the blocks and, in this case, shows that there is one eigenvalue (1)
of multiplicity 2 and one eigenvalue (2) of multiplicty 1. We say that the matrix has the Jordan structure given by
(1)2 (2)1 . In contrast, the matrix
1 0 0
0 2 0
0 0 1
has the same eigenvalues but a three-block Jordan structure:
1 0 0
0 1 0
0 0 2
This is because there are three independent eigenvectors (the unit vectors, actually). The diagonalizing matrix is
thus the identity matrix and the Jordan form has three permutations. The Jordan structure is therefore given by
(1)1 (1)1 (2)1 . These two examples show that a given eigenvalue’s multiplicity could be distributed across one,
many, or all Jordan blocks. Correlating the eigenvalue with the block structure is an important problem in numerical
analysis.
The typical approach to computing the Jordan form is to ‘follow the staircase’ pattern of the structure and
perform rank determinations in conjunction with ascertaining the eigenvalues. One of the more serious caveats with
such an approach involves mistaking an eigenvalue of multiplicity > 1 for multiple eigenvalues [8]. In the first
example matrix above, this might lead to inferring that the Jordan form has three blocks. The extra care needed to
safeguard staircase algorithms usually involves more complexity than the original computation to be performed! The
ill-conditioned nature of this computation has thus traditionally prompted numerical analysts to favor other, more
stable, decompositions.
Qualitative assessment of Jordan forms
A recent development has been the acceptance of a qualitative approach to Jordan structure determination, proposed
by Chaitin-Chatelin and Frayssé [5]. This approach does not employ the staircase idea and, instead, exploits a
semantics of eigenvalue perturbations to infer multiplicity. This leads to a geometrically intuitive algorithm that can
be implemented using SAL.
Consider a matrix that has eigenvalues λ1 , λ2 , · · · , λn with multiplicities ρ1 , ρ2 , · · · , ρn (resp). Any attempt at
finding the eigenvalues (e.g., determining the roots of the characteristic polynomial) is intrinsically subject to the
13
0.1
0.1
0
0
−0.1
6.9
7
−0.1
6.9
7.1
0.1
0
−0.1
6.9
7
7.1
0.1
0
7
−0.1
6.9
7.1
7
7.1
Figure 10: Mining Jordan forms from (left) a small sample set, and (right) large sample set. (top) Approximately
congruent triangles. (bottom) Evaluation of correspondence of rotated triangles in terms of match between original
(red dots) and rotated (green circles) samples.
numerical analysis dogma: the problem being solved will actually be a perturbed version of the original problem.
This allows the expression of the computed eigenvalues in terms of perturbations on the actual eigenvalues. It can
be easily seen that the computed eigenvalue corresponding to any λk will be distributed on the complex plane as:
1
iφ
λk + |∆| ρk e ρk
where the phase φ of the perturbation ∆ ranges over {2π, 4π, . . . , 2ρk π} if ∆ is positive and over {3π, 5π, . . . , 2(ρk +
1)π} if ∆ is negative. The insight in [5] is to superimpose numerous such perturbed calculations graphically so that
the aggregate picture reveals the ρk of the eigenvalue λk . Notice that the phase variations imply that the computed
eigenvalues will be lying on the vertices of a regular polygon centered on the actual eigenvalue and where the number of sides is two times the multiplicity of the considered eigenvalue (this takes into account both positive and
negative ∆). Since the diameter of the polygon is influenced by ∆, iterating this process over many ∆ will lead to a
‘sticks’ depiction of the Jordan form.
To illustrate, we choose a matrix whose computations will be more prone to finite precision errors. Perturbations
on the 8-by-8 Brunet matrix [5] with Jordan structure (−1)1 (−2)1 (7)3 (7)3 induce the superimposed structures
shown in Fig. 9. The left part of Fig. 9 depicts normwise relative perturbations in the scale of [2−50 , 2−40 ]. The
six sticks around the eigenvalue at 7 clearly reveal that its Jordan block is of size 3. The other Jordan block, also
centered at 7, is revealed if we conduct our exploration at a finer perturbation level. Fig. 9 reveals the second Jordan
block using perturbations in the range [2−53 , 2−50 ]. The noise in both pictures is a consequence of (i) having two
Jordan blocks with the same size, and (ii) a ‘ring’ phenomenon studied in [7]; we do not attempt to capture these
effects in this paper.
14
Data Mining and Sampling Methodology
For this study, we collect data by random normwise perturbations in a given region and a SAL program determines
multiplicity by detecting symmetry correspondence in the samples. The first aggregation level collects the samples for a given perturbation into triangles. The second aggregation level finds congruent triangles via geometric
hashing [14], and uses congruence to establish an analogy relation among triangle vertices. This relation is then
abstracted into a rotation about a point (the eigenvalue), and evaluated for whether each point rotates onto another
and whether matches define regular polygons. A third level then compares rotations across different perturbations,
re-visiting perturbations or choosing new perturbations in order to disambiguate (see Fig. 10). The end result of this
analysis is a confidence measure on models of possible Jordan forms. Each model is defined by its estimate of λ
and ρ (notice that we are working only within one region at a time). The measure M was defined to be the joint
probability distribution over the space of λ and ρ.
Experimental Results
Since our Jordan form computation treats multiple perturbations (irresp. of level) as independent estimates of eigenstructure, the idea of sampling here is not ‘where to collect,’ but ‘how much to collect.’ The goal of data mining is
hence to improve our confidence in model evaluation. We organized data collection into rounds of 6-8 samples each,
varied a tolerance parameter for triangle congruence from 0.1 to 0.5 (effectively increasing the number of models
posited), and determined the number of rounds needed to determine the Jordan form. As test cases, we used the
set of matrices studied in [5]. On average, our focused sampling approach required 1 round of data collection at a
tolerance of 0.1 and up to 2.7 rounds at 0.5. Even with a large number of models posited, additional data quickly
weeded out bad models. Fig. 10 demonstrates this mechanism on the Brunet matrix discussed above for two sets of
sample points. To the best of our knowledge, this is the only known known focused sampling methodology for this
domain; we hence are unable to present any comparisons. However, it is clear that by harnessing domain knowledge
about correspondences, we have arrived at an intelligent sampling methodology that resembles what a human would
obtain by visual inspection.
4 Discussion
The presented methodology for mining in data-scarce domains has several intrinsic benefits. First, it is based on a
uniform vocabulary of operators that can be exploited for a rich diversity of applications. Second, it demonstrates a
novel factorization to the problem of mining when data is scarce, namely, formulating an experiment design methodology to clarify, disambiguate, and improve confidences in higher-level aggregates of data. This allows us to bridge
qualitative and quantitative information in a unified framework. SAL programs thus uncover bottom-up structures in
data systematically and use difficulties encountered in this process (ambiguities, lack of correspondences) to guide
top-down selection of additional data samples. By using knowledge of physical properties explicitly, our approach
can provide more holistic and explainable results than off-the-shelf data mining algorithms. Third, our methodology
can co-exist with more traditional approaches to problem solving (numerical analysis, optimization) and is not meant
to be a replacement or a contrasting approach. This is amply demonstrated in each of the two applications above,
where connections with various traditional methodologies have been carefully established.
The methodology makes several intrinsic assumptions which we only briefly mention here. All of our applications have been such that the cause, formation, and effect of the relevant physical properties are well understood.
This is precisely what allows us to act decisively based on higher-level information from SAL aggregates, through
the measure M . It also assumes that the problems that will be encountered by the mining algorithm are the same
as the problems for which it was designed. This is an inheritance from Bayesian inductive inference and leads to
fundamental limitations on what can be done in such a setting. For instance, if new data does not help clarify an
15
ambiguity, does the fault lie with the model (SAL higher-level aggregate) or with the data? We can summarize this
problem by saying that the approach requires strong a priori information about what is possible and what is not.
Nevertheless, by advocating targeted use of domain specific knowledge and aiding qualitative model selection, our methodology is more efficient at determining high level models from empirical data. Together, SAL and
our information-theoretic measure M encapsulate knowledge about physical properties and this is what makes our
methodology a viable one for data mining purposes. In future we aim to characterize more formally the particular
forms of domain knowledge that help overcome sparsity and noise in scientific datasets.
It should be mentioned that while the two studies formulate their sampling objectives differently, they are naturally supported by the SAL framework:
• (pockets) Where should I collect data in order to mine the pockets with high confidence?
• (Jordan forms) How much data should I collect in order to determine the right Jordan form with high confidence?
One could imagine extending our framework to also take into account the expense of data samples. If the cost of
data collection is non-uniform across the domain, then including this in the design of our functional will allow us
to tradeoff the cost of gathering information with the expected improvement in problem solving performance. This
area of data mining is referred to as active learning.
Data mining can sometimes be a controversial term in a discipline that is used to mathematical rigor; this is
because it often used synonymously with ‘lack of a hypothesis or theory.’ We hope to have convinced the reader that
this need not be the case and that data mining can indeed be sensitive to knowledge about the domain, especially
physical properties of the kind we have harnessed here. As data mining applications become more prevalent in
science, the need to incorporate a priori domain knowledge will only become more important.
References
[1] C. Bailey-Kellogg and N. Ramakrishnan. Ambiguity-Directed Sampling for Qualitative Analysis of Sparse
Data from Spatially Distributed Physical Systems. In Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence (IJCAI-01), pages 43–50, 2001.
[2] C. Bailey-Kellogg and F. Zhao. Influence-Based Model Decomposition for Reasoning about Spatially Distributed Physical Systems. Artificial Intelligence, Vol. 130(2):pages 125–166, 2001.
[3] C. Bailey-Kellogg, F. Zhao, and K. Yip. Spatial Aggregation: Language and Applications. In Proceedings of
the Thirteenth National Conference on Artificial Intelligence (AAAI’96), pages 517–522, 1996.
[4] D. Berleant and B. Kuipers. Qualitative and Quantitative Simulation: Bridging the Gap. Artificial Intelligence,
Vol. 95(2):pages 215–255, 1998.
[5] F. Chaitin-Chatelin and V. Frayssé. Lectures on Finite Precision Computations. SIAM Monographs, 1996.
[6] R.G. Easterling. Comment on ‘Design and Analysis of Computer Experiments’. Statistical Science, Vol.
4(4):pages 425–427, 1989.
[7] A. Edelman and Y. Ma. Non-Generic Eigenvalue Perturbations of Jordan Blocks. Linear Algebra & Applications, Vol. 273(1-3):pages 45–63, 1998.
[8] A. Edelman and Y. Ma. Staircase Failures Explained by Orthogonal Versal Forms. SIAM Journal on Matrix
Analysis and Applications, Vol. 21(3):pages 1004–1025, 2000.
16
[9] V. Ganti, J. Gehrke, and R. Ramakrishnan. Mining Very Large Databases. IEEE Computer, Vol. 32(8):pages
38–45, August 1999.
[10] A. Goel, C.A. Baker, C.A. Shaffer, B. Grossman, W.H. Mason, L.T. Watson, and R.T. Haftka. VizCraft:
A Problem-Solving Environment for Aircraft Configuration Design. IEEE/AIP Computing in Science and
Engineering, Vol. 3(1):pages 56–66, 2001.
[11] A. Journel. Constrainted Interpolation and Qualitative Information - The Soft Kriging Approach. Mathematical
Geology, Vol. 18(2):pages 269–286, November 1986.
[12] J. Kivinen and H. Mannila. The Use of Sampling in Knowledge Discovery. In Proceedings of the Thirteenth
ACM Symposium on Principles of Database Systems, pages 77–85, 1994.
[13] D.L. Knill, A.A. Giunta, C.A. Baker, B. Grossman, W.H. Mason, R.T. Haftka, and L.T. Watson. Response
Surface Models Combining Linear and Euler Aerodynamics for Supersonic Transport Design. Journal of
Aircraft, 36(1):pages 75–86, 1999.
[14] Y. Lamdan and H. Wolfson. Geometric Hashing: A General and Efficient Model-Based Recognition Scheme.
In Proceedings of the Second International Conference on Computer Vision (ICCV), pages 238–249, 1988.
[15] R.H. Myers and D.C. Montgomery. Response Surface Methodology: Process and Product Optimization using
Designed Experiments. Wiley, Jan 2002.
[16] I. Ordóñez and F. Zhao. STA: Spatio-Temporal Aggregation with Applications to Analysis of DiffusionReaction Phenomena. In Proceedings of the Seventeenth National Conference on Artificial Intelligence
(AAAI’00), pages 517–523, 2000.
[17] N. Ramakrishnan and A.Y. Grama. Mining Scientific Data. Advances in Computers, Vol. 55:pages 119–169,
Sep 2001.
[18] J. Sacks, W.J. Welch, T.J. Mitchell, and H.P. Wynn. Design and Analysis of Computer Experiments. Statistical
Science, Vol. 4(4):pages 409–435, 1989.
[19] K.M. Yip and F. Zhao. Spatial Aggregation: Theory and Applications. Journal of Artificial Intelligence
Research, Vol. 5:pages 1–26, 1996.
[20] K.M. Yip, F. Zhao, and E. Sacks. Imagistic Reasoning. ACM Computing Surveys, Vol. 27(3):pages 363–365,
1995.
17
| 5cs.CE
|
Concentrated Differential Privacy:
Simplifications, Extensions, and Lower Bounds
arXiv:1605.02065v1 [cs.CR] 6 May 2016
Mark Bun∗
Thomas Steinke†
{mbun,tsteinke}@seas.harvard.edu
Abstract
“Concentrated differential privacy” was recently introduced by Dwork and Rothblum as a relaxation of differential privacy, which permits sharper analyses of many
privacy-preserving computations. We present an alternative formulation of the concept of concentrated differential privacy in terms of the Rényi divergence between the
distributions obtained by running an algorithm on neighboring inputs. With this reformulation in hand, we prove sharper quantitative results, establish lower bounds, and
raise a few new questions. We also unify this approach with approximate differential
privacy by giving an appropriate definition of “approximate concentrated differential
privacy.”
∗
†
Supported by an NDSEG Fellowship and NSF grant CNS-1237235.
Supported by NSF grants CCF-1116616, CCF-1420938, and CNS-1237235.
1
Contents
1 Introduction
1.1 Our Reformulation: Zero-Concentrated Differential Privacy .
1.1.1 Comparison to the Definition of Dwork and Rothblum
1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Relationship between zCDP and Differential Privacy .
1.2.2 Gaussian Mechanism . . . . . . . . . . . . . . . . . . .
1.2.3 Basic Properties of zCDP . . . . . . . . . . . . . . . .
1.2.4 Group Privacy . . . . . . . . . . . . . . . . . . . . . .
1.2.5 Lower Bounds . . . . . . . . . . . . . . . . . . . . . .
1.2.6 Approximate zCDP . . . . . . . . . . . . . . . . . . .
1.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
. 4
. 5
. 6
. 6
. 6
. 7
. 7
. 8
. 9
. 9
. 10
2 Rényi Divergence
11
2.1 Composition and Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Gaussian Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Relation to Differential Privacy
13
3.1 Pure DP versus zCDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Approximate DP versus zCDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4 Zero- versus Mean-Concentrated Differential Privacy
17
5 Group Privacy
18
6 Lower Bounds
20
6.1 Example Applications of the Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . 22
7 Obtaining Pure DP Mechanisms from zCDP
25
8 Approximate zCDP
30
8.1 Approximate DP Implies Approximate zCDP . . . . . . . . . . . . . . . . . . . . . . 31
8.2 Approximate zCDP Implies Approximate DP . . . . . . . . . . . . . . . . . . . . . . 32
8.3 Application of Approximate zCDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Acknowledgements
34
References
34
A Postprocessing and mCDP
37
B Miscellaneous Proofs and Lemmata
39
B.1 Proof of Lemma 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
C Privacy versus Sampling
49
2
1
Introduction
Differential privacy [DMNS06] is a formal mathematical standard for protecting individuallevel privacy in statistical data analysis. In its simplest form, (pure) differential privacy is
parameterized by a real number ε > 0, which controls how much “privacy loss”1 an individual
can suffer when a computation (i.e., a statistical data analysis task) is performed involving
his or her data.
One particular hallmark of differential privacy is that it degrades smoothly and predictably under the composition of multiple computations. In particular, if one performs k
computational tasks that are each ε-differentially private and combines the results of those
tasks, then the computation as a whole is kε-differentially private. This property makes differential privacy amenable to the type of modular reasoning used in the design and analysis
of algorithms: When a sophisticated algorithm is comprised of a sequence of differentially
private steps, one can establish that the algorithm as a whole remains differentially private.
A widely-used relaxation of pure differential privacy is approximate or (ε, δ)-differential
privacy [DKM+ 06], which essentially guarantees that the probability that any individual
suffers privacy loss exceeding ε is bounded by δ. For sufficiently small δ, approximate (ε, δ)differential privacy provides a comparable standard of privacy protection as pure ε-differential
privacy, while often permitting substantially more useful analyses to be performed.
Unfortunately, there are situations where, unlike pure differential privacy, approximate
differential privacy is not a very elegant abstraction for mathematical analysis, particularly
the analysis of composition. The “advanced composition theorem” of Dwork, Rothblum,
and Vadhan [DRV10] (subsequently improved by [KOV15, MV16])√shows that the composition of k tasks which are each (ε, δ)-differentially private is (≈ kε, ≈ kδ)-differentially
private. However, these bounds can be unwieldy; computing the tightest possible privacy
guarantee for the composition of k arbitrary mechanisms with differing (εi , δi )-differential
privacy guarantees is #P-hard [MV16]! Furthermore, these bounds are not tight even for
simple and natural privacy-preserving computations. For instance, consider the mechanism
which approximately answers k statistical queries on a given database by adding independent
Gaussian noise to each answer. Even for this basic computation, the advanced composition
theorem does not yield a tight analysis.2
Dwork and Rothblum [DR16] recently put forth a different relaxation of differential privacy called concentrated differential privacy. Roughly, a randomized mechanism satisfies
concentrated differentially privacy if the privacy loss has small mean and is subgaussian.
Concentrated differential privacy behaves in a qualitatively similar way as approximate
1
The privacy loss is a random variable which quantifies how much information is revealed about an
individual by a computation involving their data; it depends on the outcome of the computation, the way
the computation was performed, and the information that the individual wants to hide. We discuss it
informally in this introduction and define it precisely in Definition 1.2 on page 4.
2
In particular, consider answering k statistical queries on a dataset of n individualspby adding noise
drawn from N (0, (σ/n)2 ) independently for each query. Each individual query satisfies (O( log(1/δ)/σ), δ)differential privacy for any δ >
√ 0. Applying the advanced composition theorem shows that the composition
of all k queries satisfies (O( k log(1/δ)/σ), (k + 1)δ)-differential
privacy for any δ > 0. However, it is
p
well-known that this bound can be improved to (O( k log(1/δ)/σ), δ)-differential privacy.
3
(ε, δ)-differential privacy under composition. However, it permits sharper analyses of basic
computational tasks, including a tight analysis of the aforementioned Gaussian mechanism.
Using the work of Dwork and Rothblum [DR16] as a starting point, we introduce an
alternative formulation of the concept of concentrated differential privacy that we call “zeroconcentrated differential privacy” (zCDP for short). To distinguish our definition from that
of Dwork and Rothblum, we refer to their definition as “mean-concentrated differential privacy” (mCDP for short). Our definition uses the Rényi divergence between probability distributions as a different method of capturing the requirement that the privacy loss random
variable is subgaussian.
1.1
Our Reformulation: Zero-Concentrated Differential Privacy
As is typical in the literature, we model a dataset as a multiset or tuple of n elements (or
“rows”) in X n , for some “data universe” X , where each element represents one individual’s
information. A (privacy-preserving) computation is a randomized algorithm M : X n → Y,
where Y represents the space of all possible outcomes of the computation.
Definition 1.1 (Zero-Concentrated Differential Privacy (zCDP)). A randomised mechanism
M : X n → Y is (ξ, ρ)-zero-concentrated differentially private (henceforth (ξ, ρ)-zCDP) if, for
all x, x0 ∈ X n differing on a single entry and all α ∈ (1, ∞),
Dα (M (x)kM (x0 )) ≤ ξ + ρα,
(1)
where Dα (M (x)kM (x0 )) is the α-Rényi divergence3 between the distribution of M (x) and
the distribution of M (x0 ).
We define ρ-zCDP to be (0, ρ)-zCDP.4
Equivalently, we can replace (1) with
E e(α−1)Z ≤ e(α−1)(ξ+ρα) ,
(2)
where Z = PrivLoss (M (x)kM (x0 )) is the privacy loss random variable:
Definition 1.2 (Privacy Loss Random Variable). Let Y and Y 0 be random variables on Ω.
We define the privacy loss random variable between Y and Y 0 – denoted Z = PrivLoss (Y kY 0 )
3
Rényi divergence has a parameter α ∈ (1, ∞) which allows it to interpolate between KL-divergence
(α → 1) and max-divergence (α → ∞). It should be thought of as a measure of dissimilarity between
distributions. We define it formally in Section 2. Throughout, we assume that all logarithms are natural
unless specified otherwise — that is, base e ≈ 2.718. This includes logarithms in information theoretic
quantities like entropy, divergence, and mutual information, whence these quantities are measured in nats
rather than in bits.
4
For clarity of exposition, we consider only ρ-zCDP in the introduction and give more general statements
for (ξ, ρ)-zCDP later. We also believe that having a one-parameter definition is desirable.
4
– as follows. Define a function f : Ω → R by f (y) = log(P [Y = y] /P [Y 0 = y]).5 Then Z is
distributed according to f (Y ).
Intuitively, the value of the privacy loss Z = PrivLoss (M (x)kM (x0 )) represents how well
we can distinguish x from x0 given only the output M (x) or M (x0 ). If Z > 0, then the
observed output of M is more likely to have occurred if the input was x than if x0 was the
input. Moreover, the larger Z is, the bigger this likelihood ratio is. Likewise, Z < 0 indicates
that the output is more likely if x0 is the input. If Z = 0, both x and x0 “explain” the output
of M equally well.
A mechanism M : X n → Y is ε-differentially private if and only if P [Z > ε] = 0, where
Z = PrivLoss (M (x)kM (x0 )) is the privacy loss of M on arbitrary inputs x, x0 ∈ X n differing
in one entry. On the other hand, M being (ε, δ)-differentially private is equivalent, up to a
small loss in parameters, to the requirement that P [Z > ε] ≤ δ.
In contrast, zCDP
entails a bound on the moment generating function of the privacy
(α−1)Z
loss Z — that is, E e
as a function of α − 1. The bound (2) implies that Z is a
subgaussian random variable6 with small mean. Intuitively, this means that Z resembles a
Gaussian distribution with mean ξ + ρ and variance 2ρ. In particular, we obtain strong tail
bounds on Z. Namely (2) implies that
P [Z > λ + ξ + ρ] ≤ e−λ
2 /4ρ
for all λ > 0.7
Thus zCDP requires that the privacy loss random variable is concentrated around zero
(hence the name). That is, Z is “small” with high probability, with larger deviations from
zero becoming increasingly unlikely. Hence we are unlikely to be able to distinguish x from
x0 given the output of M (x) or M (x0 ). Note that the randomness of the privacy loss random
variable is taken only over the randomnesss of the mechanism M .
1.1.1
Comparison to the Definition of Dwork and Rothblum
For comparison, Dwork and Rothblum [DR16] define (µ, τ )-concentrated differential privacy
for a randomized mechanism M : X n → Y as the requirement that, if Z = PrivLoss (M (x)kM (x0 ))
5
Throughout we abuse notation by letting P [Y = y] represent either the probability mass function or
the probability density function of Y evaluated at y. Formally, P [Y = y] /P [Y 0 = y] denotes the RadonNikodym derivative of the measure Y with respect to the measure Y 0 evaluated at y, where we require Y to
be absolutely continuous with respect to Y 0 , i.e. Y Y 0 .
6
A randomhvariable X being
is characterised by hthe following
i four 2equivalent conditions
i subgaussian
t(X−E[X])
−Ω(λ2 )
[Riv12]. (i) P |X − E [X] | > λ ≤ e
for all λ > 0. (ii) E e
≤ eO(t ) for all t ∈ R. (iii)
h
i
c(X−E[X])2
E (X − E [X])2k ≤ O(k)k for all k ∈ N. (iv) E e
≤ 2 for some c > 0.
7
We only discuss bounds on the upper tail of Z. We can obtain similar bounds on the lower tail of
Z = PrivLoss (M (x)kM (x0 )) by considering Z 0 = PrivLoss (M (x0 )kM (x)).
5
is the privacy loss for x, x0 ∈ X n differing on one entry, then
h
i
21 2
(α−1)(Z−E[Z])
E [Z] ≤ µ
and
E e
≤ e(α−1) 2 τ
for all α ∈ R. That is, they require both a bound on the mean of the privacy loss and that
the privacy loss is tightly concentrated around its mean. To distinguish our definitions, we
refer to their definition as mean-concentrated differential privacy (or mCDP).
Our definition, zCDP, is a relaxation of mCDP. In particular, a (µ, τ )-mCDP mechanism
is also (µ − τ 2 /2, τ 2 /2)-zCDP (which is tight for the Gaussian mechanism example), whereas
the converse is not true. (However, a partial converse holds; see Lemma 4.3.)
1.2
1.2.1
Results
Relationship between zCDP and Differential Privacy
Like Dwork and Rothblum’s formulation of concentrated differential privacy, zCDP can be
thought of as providing guarantees of (ε, δ)-differential privacy for all values of δ > 0:
p
Proposition 1.3. If M provides ρ-zCDP, then M is (ρ + 2 ρ log(1/δ), δ)-differentially
private for any δ > 0.
We also prove a slight strengthening of this result (Lemma 3.6). Moreover, there is
a partial converse, which shows that, up to a loss in parameters, zCDP is equivalent to
differential privacy with this ∀δ > 0 quantification (see Lemma 3.7).
There is also a direct link from pure differential privacy to zCDP:
Proposition 1.4. If M satisfies ε-differential privacy, then M satisfies ( 21 ε2 )-zCDP.
Dwork and Rothblum [DR16, Theorem 3.5] give a slightly weaker version of Proposition
1.4, which implies that ε-differential privacy yields ( 21 ε(eε − 1))-zCDP; this improves on an
earlier bound [DRV10] by the factor 12 .
We give proofs of these and other properties using properties of Rényi divergence in
Sections 2 and 3.
Propositions 1.3 and 1.4 show that zCDP is an intermediate notion between pure differential privacy and approximate differential privacy. Indeed, many algorithms satisfying
approximate differential privacy do in fact also satisfy zCDP.
1.2.2
Gaussian Mechanism
Just as with mCDP, the prototypical example of a mechanism satisfying zCDP is the Gaussian mechanism, which answers a real-valued query on a database by perturbing the true
answer with Gaussian noise.
Definition 1.5 (Sensitivity). A function q : X n → R has sensitivity ∆ if for all x, x0 ∈ X n
differing in a single entry, we have |q(x) − q(x0 )| ≤ ∆.
6
Proposition 1.6 (Gaussian Mechanism). Let q : X n → R be a sensitivity-∆ query. Consider
the mechanism M : X n → R that on input x, releases a sample from N (q(x), σ 2 ). Then M
satisfies (∆2 /2σ 2 )-zCDP.
We remark that either inequality defining zCDP — (1) or (2) — is exactly tight for the
Gaussian mechanism for all values of α. Thus the definition of zCDP seems tailored to the
Gaussian mechanism.
1.2.3
Basic Properties of zCDP
Our definition of zCDP satisfies the key basic properties of differential privacy. Foremost,
these properties include smooth degradation under composition, and invariance under postprocessing:
Lemma 1.7 (Composition). Let M : X n → Y and M 0 : X n → Z be randomized algorithms.
Suppose M satisfies ρ-zCDP and M 0 satisfies ρ0 -zCDP. Define M 00 : X n → Y ×Z by M 00 (x) =
(M (x), M 0 (x)). Then M 00 satisfies (ρ + ρ0 )-zCDP.
Lemma 1.8 (Postprocessing). Let M : X n → Y and f : Y → Z be randomized algorithms.
Suppose M satisfies ρ-zCDP. Define M 0 : X n → Z by M 0 (x) = f (M (x)). Then M 0 satisfies
ρ-zCDP.
These properties follow immediately from corresponding properties of the Rényi divergence outlined in Lemma 2.2.
We remark that Dwork and Rothblum’s definition of mCDP is not closed under postprocessing; we provide a counterexample in Appendix A. (However, an arbitrary amount of
postprocessing can worsen the guarantees of mCDP by at most constant factors.)
1.2.4
Group Privacy
A mechanism M guarantees group privacy if no small group of individuals has a significant
effect on the outcome of a computation (whereas the definition of zCDP only refers to
individuals, which are groups of size 1). That is, group privacy for groups of size k guarantees
that, if x and x0 are inputs differing on k entries (rather than a single entry), then the outputs
M (x) and M (x0 ) are close.
Dwork and Rothblum [DR16, Theorem 4.1] gave nearly tight bounds on the group privacy
guarantees of concentrated differential privacy, showing that a (µ = τ 2 /2, τ )-concentrated
differentially private mechanism affords (k 2 µ · (1 + o(1)), kτ · (1 + o(1)))-concentrated differential privacy for groups of size k = o(1/τ ). We are able to show a group privacy guarantee
for zCDP that is exactly tight and works for a wider range of parameters:
Proposition 1.9. Let M : X n → Y satisfy ρ-zCDP. Then M guarantees (k 2 ρ)-zCDP for
groups of size k — i.e. for every x, x0 ∈ X n differing in up to k entries and every α ∈ (1, ∞),
we have
Dα (M (x)kM (x0 )) ≤ (k 2 ρ) · α.
In particular, this bound is achieved (simultaneously for all values α) by the Gaussian
mechanism. Our proof is also simpler than that of Dwork and Rothblum; see Section 5.
7
1.2.5
Lower Bounds
The strong group privacy guarantees of zCDP yield, as an unfortunate consequence, strong
lower bounds as well. We show that, as with pure differential privacy, zCDP is susceptible to
information-based lower bounds, as well as to so-called packing arguments [HT10, MMP+ 10,
De12]:
Theorem 1.10. Let M : X n → Y satisfy ρ-zCDP. Let X be a random variable on X n .
Then
I (X; M (X)) ≤ ρ · n2 ,
where I(·; ·) denotes the mutual information between the random variables (in nats, rather
than bits). Furthermore, if the entries of X are independent, then I(X; M (X)) ≤ ρ · n.
Theorem 1.10 yields strong lower bounds for zCDP mechanisms, as we can construct
distributions X such that, for any accurate mechanism M , M (X) reveals a lot of information
about X (i.e. I(X; M (X)) is large for any accurate M ).
In particular, we obtain a strong separation between approximate differential privacy
and zCDP. For example, we can show that releasing an accurate approximate histogram (or,
equivalently, accurately answering
all point queries) on a data domain of size k requires an
√
input with at least n = Θ( log k) entries to satisfy zCDP. In contrast, under approximate
differential privacy, n can be independent of the domain size k [BNS13]! In particular,
our lower bounds show that “stability-based” techniques (such as those in the propose-testrelease framework [DL09]) are not compatible with zCDP.
Our lower bound exploits the strong group privacy guarantee afforded by zCDP. Group
privacy has been used to prove tight lower bounds for pure differential privacy [HT10, De12]
and approximate differential privacy [SU15a]. These results highlight the fact that group
privacy is often the limiting factor for private data analysis. For (ε, δ)-differential privacy,
group privacy becomes vacuous for groups of size k = Θ(log(1/δ)/ε). Indeed, stability-based
techniques exploit precisely this breakdown in group privacy.
As a result of this strong lower bound, we show that any mechanism for answering
statistical queries that satisfies zCDP can be converted into a mechanism satisfying pure
differential privacy with only a quadratic blowup in its sample complexity. More precisely,
the following theorem illustrates a more general result we prove in Section 7.
Theorem 1.11. Let n ∈ N and α ≥ 1/n be arbitrary. Set ε = α and ρ = α2 . Let
q : X → [0, 1]k be an arbitrary family of statistical queries. Suppose M : X n → [0, 1]k
satisfies ρ-zCDP and
E [kM (x) − q(x)k∞ ] ≤ α
M
0
for all x ∈ X . Then there exists M 0 : X n → [0, 1]k for n0 = 5n2 satisfying ε-differential
privacy and
E0 [kM 0 (x) − q(x)k∞ ] ≤ 10α
n
M
n0
for all x ∈ X .
8
For some classes of queries, this reduction is essentially tight. For example,
√ for k oneway marginals, the Gaussian mechanism achieves sample complexity n = Θ( k) subject
to zCDP, whereas the Laplace mechanism achieves sample complexity n = Θ(k) subject to
pure differential privacy, which is known to be optimal.
For more details, see Sections 6 and 7.
1.2.6
Approximate zCDP
To circumvent these strong lower bounds for zCDP, we consider a relaxation of zCDP in the
spirit of approximate differential privacy that permits a small probability δ of (catastrophic)
failure:
Definition 1.12 (Approximate Zero-Concentrated Differential Privacy (Approximate zCDP)).
A randomized mechanism M : X n → Y is δ-approximately (ξ, ρ)-zCDP if, for all x, x0 ∈ X n
differing on a single entry, there exist events E (depending on M (x)) and E 0 (depending on
M (x0 )) such that P [E] ≥ 1 − δ, P [E 0 ] ≥ 1 − δ, and
∀α ∈ (1, ∞)
Dα (M (x)|E kM (x0 )|E 0 ) ≤ ξ+ρ·α
∧
Dα (M (x0 )|E 0 kM (x)|E ) ≤ ξ+ρ·α,
where M (x)|E denotes the distribution of M (x) conditioned on the event E. We further
define δ-approximate ρ-zCDP to be δ-approximate (0, ρ)-zCDP.
In particular, setting δ = 0 gives the original definition of zCDP. However, this definition
unifies zCDP with approximate differential privacy:
Proposition 1.13. If M satisfies (ε, δ)-differential privacy, then M satisfies δ-approximate
1 2
ε -zCDP.
2
Approximate zCDP retains most of the desirable properties of zCDP, but allows us to
incorporate stability-based techniques and bypass the above lower bounds. This also presents
a unified tool to analyse a composition of zCDP with approximate differential privacy; see
Section 8.
1.3
Related Work
Our work builds on the aforementioned prior work of Dwork and Rothblum [DR16].8 We
view our definition of concentrated differential privacy as being “morally equivalent” to their
definition of concentrated differential privacy, in the sense that both definitions formalize the
same concept.9 (The formal relationship between the two definitions is discussed in Section
8
Although Dwork and Rothblum’s work only appeared publicly in March 2016, they shared a preliminary
draft of their paper with us before we commenced this work. As such, our ideas are heavily inspired by
theirs.
9
We refer to our definition as “zero-concentrated differential privacy” (zCDP) and their definition as
“mean-concentrated differential privacy” (mCDP). We use “concentrated differential privacy” (CDP) to
refer to the underlying concept formalized by both definitions.
9
4.) However, the definition of zCDP generally seems to be easier to work with than that of
mCDP. In particular, our formulation in terms of Rényi divergence simplifies many analyses.
Dwork and Rothblum prove several results about concentrated differential privacy that
are similar to ours. Namely, they prove analogous properties of mCDP as we prove for
zCDP (cf. Sections 1.2.1, 1.2.2, 1.2.3, and 1.2.4). However, as noted, some of their bounds
are weaker than ours; also, they do not explore lower bounds.
Several of the ideas underlying concentrated differential privacy are implicit in earlier
works. In particular, the proof of the advanced composition theorem of Dwork, Rothblum,
and Vadhan [DRV10] essentially uses the ideas of concentrated differential privacy. Their
proof contains analogs of Propositions 1.7, 1.3, and 1.4,.
We also remark that Tardos [Tar08] used Rényi divergence to prove lower bounds for cryptographic objects called fingerprinting codes. Fingerprinting codes turn out to be closely related to differential privacy [Ull13, BUV14, SU15b], and Tardos’ lower bound can be (loosely)
viewed as a kind of privacy-preserving algorithm.
1.4
Further Work
We believe that concentrated differential privacy is a useful tool for analysing private computations, as it provides both simpler and tighter bounds. We hope that CDP will be prove
useful in both the theory and practice of differential privacy.
Furthermore, our lower bounds show that CDP can really be a much more stringent
condition than approximate differential privacy. Thus CDP defines a “subclass” of all (ε, δ)differentially private algorithms. This subclass includes most differentially private algorithms
in the literature, but not all — the most notable exceptions being algorithms that use the
propose-test-release approach [DL09] to exploit low local sensitivity.
This “CDP subclass” warrants further exploration. In particular, is there a “complete”
mechanism for this class of algorithms, in the same sense that the exponential mechanism
[MT07, BLR13] is complete for pure differential privacy? Can we obtain a simple characterization of the sample complexity needed to satisfy CDP? The ability to prove stronger
and simpler lower bounds for CDP than for approximate DP may be useful for showing
the limitations of certain algorithmic paradigms. For example, any differentially private algorithm that only uses the Laplace mechanism, the exponential mechanism, the Gaussian
mechanism, and the “sparse vector” technique, along with composition and postprocessing
will be subject to the lower bounds for CDP.
There is also room to examine how to interpret the zCDP privacy guarantee. In particular, we leave it as an open question to understand the extent to which ρ-zCDP provides a
stronger privacy guarantee than the implied (ε, δ)-DP guarantees (cf. Proposition 1.3).
In general, much of the literature on differential privacy can be re-examined through the
lens of CDP, which may yield new insights and results.
10
2
Rényi Divergence
Recall the definition of Rényi divergence:
Definition 2.1 (Rényi Divergence [Rén61, Equation (3.3)]). Let P and Q be probability
distributions on Ω. For α ∈ (1, ∞), we define the Rényi divergence of order α between P
and Q as
Z
1
α
1−α
log
P (x) Q(x) dx
Dα (P kQ) =
α−1
Ω
α
P (x)
1
log E
=
x∼Q
α−1
Q(x)
"
α−1 #!
1
P (x)
=
log E
,
x∼P
α−1
Q(x)
where P (·) and Q(·) are the probability mass/density functions of P and Q respectively or,
more generally, P (·)/Q(·) is the Radon-Nikodym derivative of P with respect to Q.10 We
also define the KL-divergence
Z
P (x)
P (x) log
dx
D1 (P kQ) = lim Dα (P kQ) =
α→1
Q(x)
Ω
and the max-divergence
D∞ (P kQ) = lim Dα (P kQ) = sup log
α→∞
x∈Ω
P (x)
Q(x)
.
Alternatively, Rényi divergence can be defined in terms of the privacy loss (Definition
1.2) between P and Q:
(α−1)Z
e(α−1)Dα (P kQ) =
E
e
Z∼PrivLoss(P kQ)
for all α ∈ (1, ∞). Moreover, D1 (P kQ) =
E
Z∼PrivLoss(P kQ)
[Z].
We record several useful and well-known properties of Rényi divergence. We refer the
reader to [vEH14] for proofs and discussion of these (and many other) properties. Selfcontained proofs are given in Appendix B.1.
Lemma 2.2. Let P and Q be probability distributions and α ∈ [1, ∞].
• Non-negativity: Dα (P kQ) ≥ 0 with equality if and only if P = Q.
10
If P is not absolutely continuous with respect to Q (i.e. it is not the case that P Q), we define
Dα (P kQ) = ∞ for all α ∈ [1, ∞].
11
• Composition: Suppose P and Q are distributions on Ω × Θ. Let P 0 and Q0 denote the
marginal distributions on Ω induced by P and Q respectively. For x ∈ Ω, let Px0 and
Q0x denote the conditional distributions on Θ induced by P and Q respectively, where
x specifies the first coordinate. Then
Dα (P 0 kQ0 ) + min Dα (Px0 kQ0x ) ≤ Dα (P kQ) ≤ Dα (P 0 kQ0 ) + max Dα (Px0 kQ0x ) .
x∈Ω
x∈Ω
In particular if P and Q are product distributions, then the Rényi divergence between
P and Q is just the sum of the Rényi divergences of the marginals.
• Quasi-Convexity: Let P0 , P1 and Q0 , Q1 be distributions on Ω, and let P = tP0 + (1 −
t)P1 and Q = tQ0 +(1−t)Q1 for t ∈ [0, 1]. Then Dα (P kQ) ≤ max{Dα (P0 kQ0 ) , Dα (P1 kQ1 )}.
Moreover, KL divergence is convex:
D1 (P kQ) ≤ tD1 (P0 kQ0 ) + (1 − t)D1 (P1 kQ1 ) .
• Postprocessing: Let P and Q be distributions on Ω and let f : Ω → Θ be a function.
Let f (P ) and f (Q) denote the distributions on Θ induced by applying f to P or Q
respectively. Then Dα (f (P )kf (Q)) ≤ Dα (P kQ).
Note that quasi-convexity allows us to extend this guarantee to the case where f is a
randomized mapping.
• Monotonicity: For 1 ≤ α ≤ α0 ≤ ∞, Dα (P kQ) ≤ Dα0 (P kQ).
2.1
Composition and Postprocessing
The following lemma gives the postprocessing and (adaptive) composition bounds (extending
Lemmas 1.7 and 1.8).
Lemma 2.3 (Composition & Postprocessing). Let M : X n → Y and M 0 : X n × Y → Z.
Suppose M satisfies (ξ, ρ)-zCDP and M 0 satisfies (ξ 0 , ρ0 )-zCDP (as a function of its first
argument). Define M 00 : X n → Z by M 00 (x) = M 0 (x, M (x)). Then M 00 satisfies (ξ+ξ 0 , ρ+ρ0 )zCDP.
The proof is immediate from Lemma 2.2. Note that, while Lemma 2.3 is only stated for
the composition of two mechanisms, it can be inductively applied to analyse the composition
of arbitrarily many mechanisms.
2.2
Gaussian Mechanism
The following lemma gives the Rényi divergence between two Gaussian distributions with
the same variance.
12
Lemma 2.4. Let µ, ν, σ ∈ R and α ∈ [1, ∞). Then
α(µ − ν)2
Dα N (µ, σ ) N (ν, σ ) =
2σ 2
Consequently, the Gaussian mechanism,
2 which answers a sensitivity-∆ query by adding
∆
2
noise drawn from N (0, σ ), satisfies 2σ
2 -zCDP (Proposition 1.6).
2
2
Proof. We calculate
Z
1
(x − ν)2
(x − µ)2
dx
exp (α − 1)Dα N (µ, σ ) N (ν, σ ) = √
− (1 − α)
exp −α
2σ 2
2σ 2
2πσ 2 R
Z
1
(x − (αµ + (1 − α)ν))2 − (αµ + (1 − α)ν)2 + αµ2 + (1 − α)ν 2
=√
dx
exp −
2σ 2
2πσ 2 R
−(αµ + (1 − α)ν)2 + αµ2 + (1 − α)ν 2
=
E
exp −
x∼N (αµ+(1−α)ν,σ 2 )
2σ 2
α(α − 1)(µ − ν)2
= exp
.
2σ 2
2
2
For the multivariate Gaussian mechanism, Lemma 2.4 generalises to the following.
Lemma 2.5. Let µ, ν ∈ Rd , σ ∈ R, and α ∈ [1, ∞). Then
αkµ − νk22
Dα N (µ, σ 2 Id ) N (ν, σ 2 Id ) =
2σ 2
Thus, if M : X n → Rd is the mechanism that, on input x, releases a sample from
N (q(x), σ 2 Id ) for some function q : X n → Rd , then M satisfies ρ-zCDP for
ρ=
3
1
2σ 2
sup
x,x0 ∈X n
differing in one entry
kq(x) − q(x0 )k22 .
(3)
Relation to Differential Privacy
We now discuss the relationship between zCDP and the traditional definitions of pure and
approximate differential privacy. There is a close relationship between the notions, but not
an exact characterization.
For completeness, we state the definition of differential privacy:
Definition 3.1 (Differential Privacy (DP) [DMNS06, DKM+ 06]). A randomized mechanism
M : X n → Y satisfies (ε, δ)-differential privacy if, for all x, x0 ∈ X differing in a single entry,
we have
P [M (x) ∈ S] ≤ eε P [M (x0 ) ∈ S] + δ
for all (measurable) S ⊂ Y. Further define ε-differential privacy to be (ε, 0)-differential
privacy.
13
3.1
Pure DP versus zCDP
Pure differential privacy is exactly characterized by (ξ, 0)-zCDP:
Lemma 3.2. A mechanism M : X n → Y satisfies ε-DP if and only if it satisfies (ε, 0)-zCDP.
Proof. Let x, x0 ∈ X n be neighbouring. Suppose M satisfies ε-DP. Then D∞ (M (x)kM (x0 )) ≤
ε. By monotonicity,
Dα (M (x)kM (x0 )) ≤ D∞ (M (x)kM (x0 )) ≤ ε = ε + 0 · α
for all α. So M satisfies (ε, 0)-zCDP. Conversely, suppose M satisfies (ε, 0)-zCDP. Then
D∞ (M (x)kM (x0 )) = lim Dα (M (x)kM (x0 )) ≤ lim ε + 0 · α = ε.
α→∞
α→∞
Thus M satisfies ε-DP.
We now show that ε-differential privacy implies ( 21 ε2 )-zCDP (Proposition 1.4).
Proposition 3.3. Let P and Q be probability distributions on Ω satisfying D∞ (P kQ) ≤ ε
and D∞ (QkP ) ≤ ε. Then Dα (P kQ) ≤ 21 ε2 α for all α > 1.
Remark 3.4. In particular, Proposition 3.3 shows that the KL-divergence D1 (P kQ) ≤ 12 ε2 .
A bound on the KL-divergence between random variables in terms of their max-divergence is
an important ingredient in the analysis of the advanced composition theorem [DRV10]. Our
bound sharpens (up to lower order terms) and, in our opinion, simplifies the previous bound
of D1 (P kQ) ≤ 21 ε(eε − 1) proved by Dwork and Rothblum [DR16].
Proof of Proposition 3.3. We may assume 21 εα ≤ 1, as otherwise 12 ε2 α > ε, whence the result
follows from monotonicity. We must show that
α
1
P (x)
2
(α−1)Dα (P kQ)
≤ e 2 α(α−1)ε .
e
= E
x∼Q
Q(x)
We know that e−ε ≤
E [A(x)] =
A
P (x)
Q(x)
≤ eε for all x. Define a random function A : Ω → {e−ε , eε } by
for all x. By Jensen’s inequality,
E
P (x)
Q(x)
x∼Q
P (x)
Q(x)
α
= E
x∼Q
h
α i
h
i
α
E [A(x)]
≤ E E [A(x) ] = E [Aα ] ,
A
x∼Q
A
A
where A denotes A(x) for a random x ∼ Q. We also have E [A] = E
A
x∼Q
h
P (x)
Q(x)
equation, we can conclude that
eε − 1
P A = e−ε = ε
A
e − e−ε
and
14
P [A = eε ] =
A
1 − e−ε
.
eε − e−ε
i
= 1. From this
Thus
e(α−1)Dα (P kQ) ≤E [Aα ]
A
eε − 1
1 − e−ε αε
−αε
·
e
+
·e
eε − e−ε
eε − e−ε
(eαε − e−αε ) − (e(α−1)ε − e−(α−1)ε )
=
eε − e−ε
sinh(αε) − sinh((α − 1)ε)
.
=
sinh(ε)
=
The result now follows from the following inequality, which is proved in Lemma B.1.
0 ≤ y < x ≤ 2 =⇒
3.2
1
sinh(x) − sinh(y)
≤ e 2 xy .
sinh(x − y)
Approximate DP versus zCDP
The statements in this section show that, up to some loss in parameters, zCDP is equivalent
to a family of (ε, δ)-DP guarantees for all δ > 0.
Lemma 3.5. Let M : X n → Y satisfy (ξ, ρ)-zCDP. Then M satisfies (ε, δ)-DP for all δ > 0
and
p
ε = ξ + ρ + 4ρ log(1/δ).
Thus to achieve a given (ε, δ)-DP guarantee it suffices to satisfy (ξ, ρ)-zCDP with
p
2
p
(ε − ξ)2
ρ=
.
ε − ξ + log(1/δ) − log(1/δ) ≈
4 log(1/δ)
Proof. Let x, x0 ∈ X n be neighbouring. Define f (y) = log(P [M (x) = y] /P [M (x0 ) = y]). Let
Y ∼ M (x) and Z = f (Y ). That is, Z = PrivLoss (M (x)kM (x0 )) is the privacy loss random
variable. Fix α ∈ (1, ∞) to be chosen later. Then
!α−1
P
[M
(x)
=
Y
]
= e(α−1)Dα (M (x)kM (x0 )) ≤ e(α−1)(ξ+ρα) .
E e(α−1)Z = E
0
Y ∼M (x)
P [M (x ) = Y ]
By Markov’s inequality
E e(α−1)Z
(α−1)Z
P [Z > ε] = P e
> e(α−1)ε ≤
≤ e(α−1)(ξ+ρα−ε) .
e(α−1)ε
Choosing α = (ε − ξ + ρ)/2ρ > 1 gives
2 /4ρ
P [Z > ε] ≤ e−(ε−ξ−ρ)
15
≤ δ.
Now, for any measurable S ⊂ Y,
P [M (x) ∈ S] =P [Y ∈ S]
≤P [Y ∈ S ∧ Z ≤ ε] + P [Z > ε]
≤P [Y ∈ S ∧ Z ≤ ε] + δ
Z
= P [M (x) = y] · I(y ∈ S) · I(f (y) ≤ ε) dy + δ
ZY
≤ eε P [M (x0 ) = y] · I(y ∈ S) dy + δ
Y
ε
=e P [M (x0 ) ∈ S] + δ.
Lemma 3.5 is not tight. In particular, we have the following refinement of Lemma 3.5,
the proof of which is deferred to the appendix (Lemma B.2).
Lemma 3.6. Let M : X n → Y satisfy (ξ, ρ)-zCDP. Then M satisfies (ε, δ)-DP for all δ > 0
and
q
√
ε = ξ + ρ + 4ρ · log( π · ρ/δ).
Alternatively M satisfies (ε, δ)-DP for all ε ≥ ξ + ρ and
√
π·ρ
1
2
1+(ε−ξ−ρ)/2ρ
δ = e−(ε−ξ−ρ) /4ρ · min
r 2
1+ ε−ξ−ρ + (1+ ε−ξ−ρ )2 + 4
2ρ
2ρ
.
πρ
Note that the last of three options in the minimum dominates the first two options. We
have included the first two options as they are simpler.
Now we show a partial converse to Lemma 3.5, which is proved in the appendix (Lemma
B.3).
Lemma 3.7. Let M : X n → Y satisfy (ε, δ)-DP for all δ > 0 and
p
ε = ξˆ + ρ̂ log(1/δ)
1
4
ˆ ρ̂ ∈ [0, 1]. Then M is ξˆ − 1 ρ̂ + 5√
for some constants ξ,
ρ̂,
ρ̂
-zCDP.
4
4
(4)
Thus zCDP and DP are equivalent up to a (potentially substantial) loss in parameters
and the quantification over all δ.
16
4
Zero- versus Mean-Concentrated Differential Privacy
We begin by stating the definition of mean-concentrated differential privacy:
Definition 4.1 (Mean-Concentrated Differential Privacy (mCDP) [DR16]). A randomized
mechanism M : X n → Y satisfies (µ, τ )-mean-concentrated differential privacy if, for all
x, x0 ∈ X n differing in one entry, and letting Z = PrivLoss (M (x)kM (x0 )), we have
E [Z] ≤ µ
and
λ Z−E[Z]
E e
≤ eλ
2 ·τ 2 /2
for all λ ∈ R.
In contrast (ξ, ρ)-zCDP requires that, for all α ∈ (1, ∞), E e(α−1)Z ≤ e(α−1)(ξ+ρα) , where
Z ∼ PrivLoss (M (x)kM (x0 )) is the privacy loss random variable. We now show that these
definitions are equivalent up to a (potentially significant) loss in parameters.
Lemma 4.2. If M : X n → Y satisfies (µ, τ )-mCDP, then M satisfies (µ−τ 2 /2, τ 2 /2)-zCDP.
Proof. For all α ∈ (1, ∞),
h
i
(α−1)Z
2 2
2
2
(α−1)(Z−E[Z])
(α−1)E[Z]
E e
=E e
·e
≤ e(α−1) τ /2 · e(α−1)µ = e(α−1)(µ−τ /2+τ /2·α) .
√
Lemma 4.3. If M : X n → Y satisfies (ξ, ρ)-zCDP, then M satisfies (ξ + ρ, O( ξ + 2ρ))mCDP.
The proof of Lemma 4.3 is deferred to the appendix.
p
Thus we can convert (µ, τ )-mCDP into (µ−τ 2 /2, τ 2 /2)-zCDP and then back to (µ, O( µ + τ 2 /2))mCDP. This may result in a large loss in parameters, which is why, for example, pure DP
can be characterised in terms of zCDP, but not in terms of mCDP.
We view zCDP as a relaxation of mCDP; mCDP requires the privacy loss to be “tightly”
concentrated about its mean and that the mean is close to the origin. The triangle inequality
then implies that the privacy loss is “weakly” concentrated about the origin. (The difference
between “tightly” and “weakly” accounts for the use of the triangle inequality.) On the other
hand, zCDP direcly requires that the privacy loss is weakly concentrated about the origin.
That is to say, zCDP gives a subgaussian bound on the privacy loss that is centered at
zero, whereas mCDP gives a subgaussian bound that is centered at the mean and separately
bounds the mean.
There may be some advantage to the stronger requirement of mCDP, either in terms of
what kind of privacy guarantee it affords, or how it can be used as an analytic tool. However,
it seems that for most applications, we only need what zCDP provides.
17
5
Group Privacy
In this section we show that zCDP provides privacy protections to small groups of individuals.
Definition 5.1 (zCDP for Groups). We say that a mechanism M : X n → Y provides
(ξ, ρ)-zCDP for groups of size k if, for every x, x0 ∈ X n differing in at most k entries, we
have
∀α ∈ (1, ∞)
Dα (M (x)kM (x0 )) ≤ ξ + ρ · α.
The usual definition of zCDP only applies to groups of size 1. Here we show that it
implies bounds for all group sizes. We begin with a technical lemma.
Lemma 5.2 (Triangle-like Inequality for Rényi Divergence). Let P , Q, and R be probability
distributions. Then
Dα (P kQ) ≤
kα
D kα−1 (P kR) + Dkα (RkQ)
kα − 1 k−1
(5)
for all k, α ∈ (1, ∞).
Proof. Let p =
kα−1
α(k−1)
and q =
(α−1)Dα (P kQ)
e
kα−1
.
α−1
Z
Then
1
p
1
q
+
=
α(k−1)+(α−1)
kα−1
= 1. By Hölder’s inequality,
P (x)α Q(x)1−α dx
=
ZΩ
P (x)α R(x)−α · R(x)α−1 Q(x)1−α · R(x)dx
Ω
"
α
α−1 #
P (x)
R(x)
= E
·
x∼R
R(x)
Q(x)
"
pα 1/p
q(α−1) #1/q
R(x)
P (x)
· E
≤ E
x∼R
x∼R
R(x)
Q(x)
=
=e(pα−1)Dpα (P kR)/p · eq(α−1)Dq(α−1)+1 (RkQ)/q .
Taking logarithms and rearranging gives
Dα (P kQ) ≤
Now pα =
kα−1
,
k−1
pα−1
p(α−1)
kα
kα−1
=
pα − 1
Dpα (P kR) + Dq(α−1)+1 (RkQ) .
p(α − 1)
q(α − 1) + 1 = kα, and
pα − 1 kα − 1
·
=
pα
k(α − 1)
kα−1
−
k−1
kα−1
k−1
1
·
kα − 1
kα − 1 − k + 1 kα − 1
=
·
= 1,
k(α − 1)
kα − 1
k(α − 1)
as required.
18
Proposition 5.3. If M : X n → Y satisfies (ξ, ρ)-zCDP, then M gives (ξ · k
zCDP for groups of size k.
Note that
k
X
1
i=1
i
Z
=1+
1
k
1
dx ≤ 1 +
dxe
Z
k
1
Pk
1
i=1 i , ρ
· k 2 )-
1
dx = 1 + log k.
x
Thus (ξ, ρ)-zCDP implies (ξ · O(k log k), ρ · k 2 )-zCDP for groups of size k.
The Gaussian mechanism shows that k 2 ρ is the optimal dependence on ρ. However,
O(k log k)ξ is not the optimal dependence on ξ: (ξ, 0)-zCDP implies (kξ, 0)-zCDP for groups
of size k.
Proof. We show this by induction on k. The statement is clearly true for groups of size 1.
We now assume the statement holds for groups of size k − 1 and will verify it for groups of
size k.
Let x, x0 ∈ X n differ in k entries. Let x̂ ∈ X n be such that x and x̂ differ in k − 1 entries
and x0 and x̂ differ in one entry.
Then, by the induction hypothesis,
Dα (M (x)kM (x̂)) ≤ ξ · (k − 1)
k−1
X
1
i=1
i
+ ρ · (k − 1)2 · α
and, by zCDP,
Dα (M (x̂)kM (x0 )) ≤ ξ + ρ · α
for all α ∈ (1, ∞).
By (5), for any α ∈ (1, ∞),
kα
D kα−1 (M (x)kM (x̂)) + Dkα (M (x̂)kM (x0 ))
kα − 1 k−1
!
k−1
X
1
kα
−
1
kα
ξ · (k − 1)
+ ρ · (k − 1)2 ·
+ ξ + ρ · kα
≤
kα − 1
i
k
−
1
i=1
!
k−1
X
kα
1
kα
2 kα − 1
=ξ · 1 +
(k − 1)
+ρ·
(k − 1)
+ kα
kα − 1
i
kα − 1
k−1
i=1
!
k−1
X
kα
1
=ξ · 1 +
(k − 1)
+ ρ · k2 · α
kα − 1
i
i=1
!
k−1
X1
k
≤ξ · 1 +
(k − 1)
+ ρ · k2 · α
k−1
i
i=1
Dα (M (x)kM (x0 )) ≤
=ξ · k
k
X
1
i=1
i
+ ρ · k 2 · α,
where the last inequality follows from the fact that
α > 1.
19
kα
kα−1
is a decreasing function of α for
6
Lower Bounds
In this section we develop tools to prove lower bounds for zCDP. We will use group privacy to
bound the mutual information between the input and the output of a mechanism satisfying
zCDP. Thus, if we are able to construct a distribution on inputs such that any accurate
mechanism must reveal a high amount of information about its input, we obtain a lower
bound showing that no accurate mechanism satisfying zCDP can be accurate for this data
distribution.
We begin with the simplest form of our mutual information bound, which is an analogue
of the bound of [MMP+ 10] for pure differential privacy:
Proposition 6.1. Let M : X n → Y satisfy (ξ, ρ)-zCDP. Let X be a random variable in X n .
Then
I(X; M (X)) ≤ ξ · n(1 + log n) + ρ · n2 ,
where I denotes mutual information (measured in nats, rather than bits).
P
Proof. By Proposition 5.3, M provides (ξ · n ni=1 1i , ρ · n2 )-zCDP for groups of size n. Thus
0
D1 (M (x)kM (x )) ≤ ξ · n
n
X
1
i=1
i
+ ρ · n2 ≤ ξ · n(1 + log n) + ρ · n2
for all x, x0 ∈ X n . Since KL-divergence is convex,
I(X; M (X)) = E [D1 (M (x)kM (X))]
x←X
h
i
0
≤ E
E [D1 (M (x)kM (x ))]
x←X x0 ←X
h
i
2
≤ E
E
ξ
·
n(1
+
log
n)
+
ρ
·
n
0
x←X
x ←X
=ξ · n(1 + log n) + ρ · n2 .
The reason this lower bound works is the strong group privacy guarantee — even for
groups of size n, we obtain nontrivial privacy guarantees. While this is good for privacy it is
bad for usefulness, as it implies that even information that is “global” (rather than specific
to a individual or a small group) is protected. These lower bounds reinforce the connection
between group privacy and lower bounds [HT10, De12, SU15a].
In contrast, (ε, δ)-DP is not susceptible to such a lower bound because it gives a vacuous
privacy guarantee for groups of size k = O(log(1/δ)/ε). This helps explain the power of the
propose-test-release paradigm.
Furthermore, we obtain even stronger mutual information bounds when the entries of
the distribution are independent:
20
Lemma 6.2. Let M : X m → Y satisfy (ξ, ρ)-zCDP. Let X be a random variable in X m with
independent entries. Then
I (X; M (X)) ≤ (ξ + ρ) · m,
where I denotes mutual information (measured in nats, rather than bits).
Proof. First, by the chain rule for mutual information,
X
I(X; M (X)) =
I(Xi ; M (X)|X1···i−1 ),
i∈[m]
where
I(Xi ; M (X)|X1···i−1 ) =
E
[I(Xi |X1···i−1 = x; M (X)|X1···i−1 = x)]
=
E
[I(Xi ; M (x, Xi···m ))] ,
x←X1···i−1
x←X1···i−1
by independence of the Xi s.
We can define mutual information in terms of KL-divergence:
I(Xi ; M (x, Xi···m )) = E [D1 (M (x, Xi···m )|Xi = ykM (x, Xi···m ))]
y←Xi
= E [D1 (M (x, y, Xi+1···m )kM (x, Xi···m ))] .
y←Xi
By zCDP, we know that for all x ∈ X i−1 , y, y 0 ∈ X , and z ∈ X m−i , we have
D1 (M (x, y, z)kM (x, y 0 , z)) ≤ ξ + ρ.
Thus, by the convexity of KL-divergence,
D1 (M (x, y, Xi+1···m )kM (x, Xi···m )) ≤ ξ + ρ
for all x and y. The result follows.
More generally, we can combine dependent and independent entries as follows.
Theorem 6.3. Let M : X n → Y satisfy (ξ, ρ)-zCDP. Take n = m · `. Let X 1 , · · · , X m be
independent random variables on X ` . Denote X = (X 1 , · · · , X m ) ∈ X n . Then
I (X; M (X)) ≤ m · ξ · `(1 + log `) + ρ · `2 ,
where I denotes the mutual information (measured in nats, rather than bits).
21
Proof. By Proposition 5.3, M provides (ξ · `
D1 (M (x1 , · · · , xi , · · · , xm )kM (x1 , · · ·
P`
1
i=1 i , ρ
, x0i , · · ·
· `2 )-zCDP for groups of size `. Thus
, xm )) ≤ ξ ·`
`
X
1
i=1
i
+ρ·`2 ≤ ξ ·`(1+log `)+ρ·`2
(6)
for all x1 , · · · , xm , x0i ∈ X ` .
By the chain rule for mutual information,
X
I(X; M (X)) =
I(Xi ; M (X)|X1···i−1 ),
i∈[m]
where
I(Xi ; M (X)|X1···i−1 ) =
E
[I(Xi |X1···i−1 = x; M (X)|X1···i−1 = x)]
=
E
[I(Xi ; M (x, Xi···m ))] ,
x←X1···i−1
x←X1···i−1
by independence of the Xi s.
We can define mutual information in terms of KL-divergence:
I(Xi ; M (x, Xi···m )) = E [D1 (M (x, Xi···m )|Xi = ykM (x, Xi···m ))]
y←Xi
= E [D1 (M (x, y, Xi+1···m )kM (x, Xi···m ))] .
y←Xi
By (6) and the convexity of KL-divergence,
D1 (M (x, y, Xi+1···m )kM (x, Xi···m )) ≤ ξ · `(1 + log `) + ρ · `2
for all x and y. The result follows.
6.1
Example Applications of the Lower Bound
We informally discuss a few applications of our information-based lower bounds to some
simple and well-studied problems in differential privacy.
One-Way Marginals Consider M : X n → Y where X = {0, 1}d and Y = [0, 1]d . The
goal of M is to estimate the attribute means, or one-way marginals, of its input database x:
1X
M (x) ≈ x =
xi .
n
i∈[n]
It is known that this is possible subject to ε-DP if andp
only if n = Θ(d/ε) [HT10, SU15a].
This is possible subject to (ε, δ)-DP if and only if n = Θ̃( d log(1/δ)/ε), assuming δ 1/n
[BUV14, SU15a].
22
We now analyze what can be accomplished with zCDP. Adding independent noise drawn
from N (0, d/2n2 ρ) to each
p of the d coordinates of x satisfies ρ-zCDP. This gives accurate
answers as long as n d/ρ.
For a lower bound, consider sampling X1 ∈ {0, 1}d uniformly at random. Set Xi = X1
for all i ∈ [n]. By Proposition 6.1,
I(X; M (X)) ≤ n2 ρ
for any ρ-zCDP M : ({0, 1}d )n → [0, 1]d . However, if M is accurate, we can recover (most
p of)
X1 from M (X), whence I(X; M (X)) ≥ Ω(d). This yields a lower bound of n ≥ Ω( d/ρ),
which is tight up to constant factors.
Histograms (a.k.a. Point Queries) Consider M : X n → Y, where X = [T ] and Y = RT .
The goal of M is to estimate the histogram of its input:
M (x)t ≈ ht (x) = |{i ∈ [n] : xi = t}|
For ε-DP it is possible to do this if and only if n = Θ(log(T )/ε)); the optimal algorithm
is to independently sample
M (x)t ∼ ht (x) + Laplace(2/ε).
However, for (ε, δ)-DP, it is possible to attain sample complexity n = O(log(1/δ)/ε)
[BNS13,
p
BNS16b, Theorem 3.13]. Interestingly, for zCDP we can show that n = Θ( log(T )/ρ) is
sufficient and necessary:
Sampling
M (x)t ∼ ht (x) + N (0, 1/ρ)
independently for t ∈ [T ] satisfies ρ-zCDP. Moreover,
2
P max |M (x)t − ht (x)| ≥ λ ≤ T · P [|N (0, 1/ρ)| > λ] ≤ T · e−λ ρ/2 .
t∈[T ]
i
h
p
In particular P maxt∈[T ] |M (x)t − ht (x)| ≥ log(T /β)/ρ ≤ β for all β > 0. Thus this
p
algorithm is accurate if n log(T )/ρ.
On the other hand, if we sample X1 ∈ [T ] uniformly at random and set Xi = X1 for all
i ∈ [n], then I(X; M (X)) ≥ Ω(log T ) for any accurate M , p
as we can recover X1 from M (X)
if M is accurate. Proposition 6.1 thus implies that n ≥ Ω( log(T )/ρ) is necessary to obtain
accuracy.
This gives a strong separation between approximate DP and zCDP.
23
Randomized Response and the Exponential Mechanism Consider M : {±1}n →
{±1}n where the goal is to maximize hx, M (x)i subject to ρ-zCDP.
One solution is randomized response [War65]: Each output bit i of M is chosen independently with
eε
.
P [M (x)i = xi ] = ε
e +1
This satisfies ε-DP and, hence, 12 ε2 -zCDP. And E [hx, M (x)i] = n(eε − 1)/(eε + 1) = Θ(nε).
Alternatively, we can independently choose the output bits i according to
√
M (x)i = sign N (xi , 2/ρ) ,
which satisfies ρ-zCDP.
Turning our attention to lower bounds: Let X ∈ {±1}n be uniformly random. By
Lemma 6.2, since the bits of X are independent, we have I(X; M (X)) ≤ ρ · n for any ρzCDP M . However, if M is accurate, we can recover part of X from M (X) [BSU16], whence
I(X; M (X)) ≥ Ω(n).
Randomized response is a special case of the exponential mechanism [MT07, BLR13].
Consequently this can be interpreted as a lower bound for search problems.
Lower Bounds with Accuracy The above examples can be easily discussed in terms of a
more formal and quantitative definition of accuracy. In particular, we consider the histogram
example again:
Proposition 6.4. If M : [T ]n → RT satisfies ρ-zCDP and
n
∀x ∈ [T ]
E max M (x)t − ht (x) ≤ αn,
M
t∈[T ]
p
then n ≥ Ω( log(α2 T )/ρα2 ).
Proof. Let m = 1/10α and ` = n/m. For simplicity, assume that both m and n are integral.
Let X1 , X2 , · · · , Xm ∈ [T ]` be independent, where each Xi is ` copies of a uniformly
random element of [T ]. By Theorem 6.3,
I(X; M (X)) ≤ ρ · m · `2 = 10ραn2 ,
where X = (X1 , · · · , Xm ) ∈ X n . However,
I(X; M (X)) ≥I(f (X); g(M (X)))
=H(f (X)) − H(f (X)|g(M (X)))
=H(X) − H(X|f (X)) − H(f (X)|g(M (X)))
for any functions f and g, where H is the entropy (in nats). In particular, we let
f (x) = {t ∈ T : ∃i ∈ [n] xi = t}
and
24
g(y) = {t ∈ T : yt ≥ 5αn}.
(7)
Clearly H(X) = m log T . Furthermore, H(X|f (X)) ≤ m log m, since X can be specified by
naming m elements of f (X), which is a set of at most m elements.
If
max M (X)t − ht (X) < 5α,
(8)
t∈[T ]
then g(M (X)) contains exactly all the values in X — i.e. f (X) = g(M (X)). By Markov’s
inequality, (8) holds with probability at least 4/5.
Now we can upper bound H(f (X)|g(M (X))) by giving a scheme for specifying f (X)
given g(M (X)). If (8) holds, we simply need one bit to say so. If (8) does not hold, we need
one bit to say this and m log2 T bits to describe f (X). This gives
H(f (X)|g(M (X))) ≤ log 2 + P [f (X) 6= g(M (X))] · m log T.
Combining these inequalities gives
4
1
I(X; M (X)) ≥ m log T −m log m−log 2− m log T ≥ m log(T m−5/4 )−1 ≥ Ω(log(α1.25 T )/α).
5
5
Combining this with (7) completes the proof.
We remark that our lower bounds for zCDP can be converted to lower bounds for mCDP
using Lemma 4.2.
7
Obtaining Pure DP Mechanisms from zCDP
We now establish limits on what more can be achieved with zCDP over pure differential
privacy. In particular, we will prove that any mechanism satisfying zCDP can be converted
into a mechanism satisfying pure DP with at most a quadratic blowup in sample complexity.
Formally, we show the following theorem.
Theorem 7.1. Fix n ∈ N, n0 ∈ N, k ∈ N α > 0, and ε > 0. Let q : X → Rk and let k · k be
a norm on Rk . Assume maxx∈X kq(x)k ≤ 1.
Suppose there exists a (ξ, ρ)-zCDP mechanism M : X n → Rk such that for all x ∈ X n ,
E [kM (x) − q(x)k] ≤ α.
M
Assume ξ ≤ α2 , ρ ≤ α2 , and
4
ρ · n2 + ξ · n · (1 + log n) + 1 .
n0 ≥
εα
0
Then there exists a (ε, 0)-differentially private M 0 : X n → Rk satisfying
E [kM 0 (x) − q(x)k] ≤ 10α
M0
and
1
4
0
≤β
P0 kM (x) − q(x)k > 10α + 0 log
M
εn
β
0
for all x ∈ X n and β > 0.
25
Before discussing the proof of Theorem 7.1, we make some remarks about its statement:
• Unfortunately, the theorem only works for families of statistical queries q : X → Rk .
However, it works equally well for k · k∞ and k · k1 error bounds.
• If ξ = 0, we have n0 = O(n2 ρ/εα). So, if ρ, ε, and α are all constants, we have
n0 = O(n2 ). This justifies our informal statement that we can convert any mechanism satisfying zCDP into one satisfying pure DP with a quadratic blowup in sample
complexity.
• Suppose M : X n → Rk is the Gaussian mechanism scaled to satisfy ρ-zCDP and
k · k = k · k1 /k. Then
s
!
k
α = E [kM (x) − q(x)k] = Θ
.
ρn2
p
0
In particular, n = Θ( k/ρα2 ). The theorem then gives us a ε-DP M 0 : X n → Rk
with E [kM 0 (x) − q(x)k] ≤ O(α) for
0
n =Θ
n2 ρ
εα
=Θ
k
α3 ε
.
However, the Laplace mechanism achieves ε-DP and E [kM 0 (x) − q(x)k] ≤ α with
n = Θ(k/αε).
This example illustrates that the theorem is not tight in terms of α; it loses a 1/α2
factor here. However, the other parameters are tight.
• The requirement that ξ, ρ ≤ α2 is only used to show that
max min kq(x) − q(x̂)k ≤ 2α
n
x∈X n0 x̂∈X
(9)
using Lemma 7.5. However, in many situations (9) holds even when ξ, ρ α2 . For
example, if n ≥ O(log(k)/α2 ) or even n ≥ O(V C(q)/α2 ) then (9) is automatically
satisfied.
The technical condition (9) is needed to relate the part of the proof with inputs of size
n to the part with inputs of size n0 .
Thus we can restate Theorem 7.1 with the condition ξ, ρ ≤ α2 replaced by (9). This
would be more general, but also more mysterious.
Alas, the proof of Theorem 7.1 is not constructive. Rather than directly constructing a mechanism satisfying pure DP from any mechanism satisfying zCDP, we show the contrapositive
statement: any lower bound for pure DP can be converted into a lower bound for zCDP.
Pure DP is characterized by so-called packing lower bounds and the exponential mechanism.
We begin by giving a technical lemma showing that for any output space and any desired
accuracy we have a “packing” and a “net:”
26
Lemma 7.2. Let (Y, d) be a metric space. Fix α > 0. Then there exists a countable T ⊂ Y
such that both of the following hold.
• (Net:) Either T is infinite or for all y 0 ∈ Y there exists y ∈ T with d(y, y 0 ) ≤ α.
• (Packing:) For all y, y 0 ∈ T , if y 6= y 0 , then d(y, y 0 ) > α.
Proof. Consider the following procedure for producing T .
• Initialize A ← Y and T ← ∅.
• Repeat:
– If A = ∅, terminate.
– Pick some y ∈ A.
– Update T ← T ∪ {y}.
– Update A ← {y 0 ∈ A : d(y 0 , y) > α}.
This procedure either terminates giving a finite T or runs forever enumerating a countably
infinite T .
(Net:) If T is infinite, we immediately can dispense the first condition, so suppose the
procedure terminates and T is finite. Fix y 0 ∈ Y. Since the procedure terminates, A = ∅ at
the end, which means y 0 was removed from A at some point. This means some y ∈ T was
added such that d(y 0 , y) ≤ α, as required.
(Packing:) Fix y 6= y 0 ∈ T . We assume, without loss of generality, that y was added to
T before y 0 . This means y 0 was not removed from A when y was added to T . In particular,
this means d(y 0 , y) > α.
It is well-known that a net yields a pure DP algorithm:
Lemma 7.3 (Exponential Mechanism [MT07, BLR13]). Let ` : X n ×T → R satisfy |`(x, y)−
`(x0 , y)| ≤ ∆ for all x, x0 ∈ X n differing in one entry and all y ∈ T . Then, for all ε > 0,
there exists an ε-differentially private M : X n → T such that
|T |
2∆
log
P `(x, M (x)) ≤ min `(x, y) +
≥1−β
M
y∈T
ε
β
and
E [`(x, M (x))] ≤ min `(x, y) +
M
y∈T
2∆
log |T |
ε
n
for all x ∈ X and β > 0.
Proof. The mechanism is defined by
P [M (x) = y] = P
M
e−`(x,y)ε/2∆
.
−`(x,y)ε/2∆
y 0 ∈T e
The analysis can be found in [DR14, Theorems 3.10 and 3.11] and [BNS+ 16a, Lemma 7.1].
27
We also show that a packing yields a lower bound for zCDP:
Lemma 7.4. Let (Y, d) be a metric space and q : X n → Y a function. Let M : X n → Y be
a (ξ, ρ)-zCDP mechanism satisfying
P [d(M (x), q(x)) > α/2] ≤ β
M
for all x ∈ X n . Let T ⊂ Y be such that d(y, y 0 ) > α, for all y, y 0 ∈ T with y 6= y 0 . Assume
that for all y ∈ T there exists x ∈ X n with q(x) = y. Then
(1 − β) log |T | − log 2 ≤ ξ · n(1 + log n) + ρ · n2 .
In particular, if ξ = 0, we have
s
p
(1 − β) log |T | − log 2
n≥
= Ω( log |T |/ρ).
ρ
Proof. Let q −1 : T → X n be a function such that q(q −1 (y)) = y for all y ∈ T . Define
f : Y → T by
f (y) = argmin d(y, y 0 )
y 0 ∈T
(breaking ties arbitrarily). Then
P f (M (q −1 (y))) = y ≥ 1 − β
M
for all y ∈ T , as P [d(M (q −1 (y)), y) > α/2] ≤ β and d(y 0 , y) > α for all y 0 ∈ T \ {y}.
M
Let Y be a uniformly random element of T and let X = q −1 (Y ). By the data processing
inequality and Proposition 6.1,
I(Y ; f (M (q −1 (Y )))) = I(q(X); f (M (X))) ≤ I(X; M (X)) ≤ ξ · n(1 + log n) + ρ · n2 .
However, P [f (M (q −1 (Y ))) = Y ] ≥ 1 − β. Denote Z = f (M (q −1 (Y ))) and let E be the
indicator of the event that Z = Y . We have
I(Y ; Z) = H(Y ) − H(Y |Z) = H(Y ) − H(Y, E|Z) = H(Y ) − H(Y |E, Z) − H(E|Z).
Clearly H(Y ) = log |T | and H(E|Z) ≤ H(E) ≤ log 2. Moreover,
H(Y |E, Z) = E [H(Y |Z, E = e)]
e←E
=P [Y = Z] · 0 + P [Y 6= Z] · H(Y |Z, Y 6= Z)
≤β · H(Y ).
Thus
I(Y ; Z) ≥ log |T | − log 2 − β log |T |.
The result now follows by combining inequalities.
28
We need one final technical lemma:
Lemma 7.5. Let q : X → Rk satisfy maxx∈X kq(x)k ≤ 1, where k · k is some norm. Let
M : X n → Rk satisfy (ξ, ρ)-zCDP and
E [kM (x) − q(x)k] ≤ α
M
for all x ∈ X n . For all n0 ,
max0 minn kq(x̂) − q(x)k ≤ 2α +
x∈X n x̂∈X
p
2(ξ + ρ).
The proof of Lemma 7.5 is deferred to the appendix.
Now we can combine Lemmas 7.2, 7.3, 7.4, and 7.5 to prove Theorem 7.1:
o
n
P
Proof of Theorem 7.1. Apply Lemma 7.2 with Y = q(x) = n1 i∈[n] q(xi ) : x ∈ X n ⊂ Rk
and d being the metric induced by the norm to obtain T ⊂ Y:
• (Net:) Either T is infinite or for all y 0 ∈ {q(x) : x ∈ X n } ⊂ Rk there exists y ∈ T with
ky − y 0 k ≤ 4α.
• (Packing:) For all y, y 0 ∈ T , if y 6= y 0 , then ky − y 0 k > 4α.
By Markov’s inequality
1
P [kM (x) − q(x)k > 2α] ≤ .
M
2
Thus, by Lemma 7.4,
1
log |T | − log 2 ≤ ξ · n(1 + log n) + ρ · n2 .
2
This gives an upper bound on |T |. In particular, T must be finite.
0
Let M 0 : X n → Rk be the exponential mechanism (Lemma 7.3) instantiated with T and
`(x, y) = ky − q(x)k. We have
4
|T |
P kM (x) − q(x)k ≤ min ky − q(x)k + 0 log
≥1−β
M
y∈T
εn
β
and
4
log |T |
M
y∈T
εn0
0
0
for all x ∈ X n . For x ∈ X n , by the Net property and Lemma 7.5,
E [kM (x) − q(x)k] ≤ min ky − q(x)k +
min ky − q(x)k ≤ min min
ky − y 0 k + ky 0 − q(x)k
y∈T
y∈T y 0 ∈Y
0
0
= min
min ky − y k + ky − q(x)k
0
y ∈Y
y∈T
≤ min
(4α + ky 0 − q(x)k)
0
y ∈Y
= minn (4α + kq(x̂) − q(x)k)
x̂∈X
p
≤4α + 2α + 2(ξ + ρ).
29
Furthermore,
4
8
log |T | ≤ 0 ξ · n(1 + log n) + ρ · n2 + log 2 ≤ 2α.
0
εn
εn
The theorem now follows by combining inequalities.
8
Approximate zCDP
In the spirit of approximate DP, we propose a relaxation of zCDP:
Definition 8.1 (Approximate zCDP). A randomised mechanism M : X n → Y is δapproximately (ξ, ρ)-zCDP if, for all x, x0 ∈ X n differing on a single entry, there exist events
E = E(M (x)) and E 0 = E 0 (M (x0 )) such that, for all α ∈ (1, ∞),
Dα (M (x)|E kM (x0 )|E 0 ) ≤ ξ + ρ · α
and P [E] ≥ 1 − δ and
M (x)
and
Dα (M (x0 )|E 0 kM (x)|E ) ≤ ξ + ρ · α
P [E 0 ] ≥ 1 − δ.
M (x0 )
Clearly 0-approximate zCDP is simply zCDP. Hence we have a generalization of zCDP.
As we will show later in this section, δ-approximate (ε, 0)-zCDP is equivalent to (ε, δ)-DP.
Thus we have also generalized approximate DP. Hence, this definition unifies both relaxations
of pure DP.
Approximate zCDP is a three-parameter definition which allows us to capture many
different aspects of differential privacy. However, three parameters is quite overwhelming.
We believe that use of the one-parameter ρ-zCDP (or the two-parameter δ-approximate
ρ-zCDP if necessary) is sufficient for most purposes.
It is easy to verify that the definition of approximate zCDP satisfies the following basic
properties.
Lemma 8.2 (Composition & Postprocessing). Let M : X n → Y and M 0 : X n × Y → Z be
randomized algorithms. Suppose M satisfies δ-approximate (ξ, ρ)-zCDP and, for all y ∈ Y,
M 0 (·, y) : X n → Z satisfies δ 0 -approximate (ξ 0 , ρ0 )-zCDP. Define M 00 : X n → Z by M 00 (x) =
M 0 (x, M (x)). Then M 00 satisfies (δ + δ 0 − δ · δ 0 )-approximate (ξ + ξ 0 , ρ + ρ0 )-zCDP.
Lemma 8.3 (Tradeoff). Suppose M : X n → Y satisfies δ-approximate (ξ, 0)-zCDP. Then
M satisfies δ-approximate ξ-zCDP and δ-approximate 21 ξ 2 -zCDP.
However, the strong group privacy guarantees of Section 5 no longer apply to approximate
zCDP and, hence, the strong lower bounds of Section 6 also no longer hold. Circumventing
these lower bounds is part of the motivation for considering approximate zCDP. However,
approximate zCDP is not necessarily the only way to relax zCDP that circumvents our lower
bounds:
Proving the group privacy bound requires “inflating” the parameter α: Suppose M :
n
X → Y satisfies ρ-zCDP and x, x0 ∈ X n differ on k entries. To prove Dα (M (x)kM (x0 )) ≤
30
k 2 ρα, the proof of Proposition 5.3 requires a bound on Dkα (M (x00 )kM (x000 )) for x00 , x000 ∈ X n
differing on a single entry.
Consider relaxing the definition of zCDP to only require the bound (1) or (2) to hold
when α ≤ m:
Definition 8.4 (Bounded zCDP). We say that M : X n → Y satisfies m-bounded (ξ, ρ)zCDP if, for all x, x0 ∈ X n differing in only one entry and all α ∈ (1, m), Dα (M (x)kM (x0 )) ≤
ξ + ρ · α.
This relaxed definition may also be able to circumvent the group privacy-based lower
bounds, as our group privacy proof would no longer work for groups of size larger than m.
We do not know what group privacy guarantees Definition 8.4 provides for groups of size
k m. This relaxed definition may be worth exploring, but is beyond the scope of our
work.
8.1
Approximate DP Implies Approximate zCDP
We can convert approximate DP to approximate zCDP using the following lemma.
First we define a approximate DP version of the randomized response mechanism:
Definition 8.5. For ε ≥ 0 and δ ∈ [0, 1], define M̃ε,δ : {0, 1} → {0, 1} × {⊥, >} by
h
i
h
i
P M̃ε,δ (b) = (b, >) =δ,
P M̃ε,δ (b) = (1 − b, >) =0,
h
i
h
i
1
eε
,
P
M̃
(b)
=
(1
−
b,
⊥)
=(1
−
δ)
P M̃ε,δ (b) = (b, ⊥) =(1 − δ)
ε,δ
1 + eε
1 + eε
for both b ∈ {0, 1}.
The above mechanism is “complete” for approximate DP:
Lemma 8.6 ([KOV15], [MV16, Lemma 3.2]). For every (ε, δ)-DP M : X n → Y and all
x0 , x1 ∈ X n differing in one entry, there exists a randomized T : {0, 1} × {⊥, >} → Y such
that T (M̃ε,δ (b)) has the same distribution as M (xb ) for both b ∈ {0, 1}.
Corollary 8.7. If M : X n → Y satisfies (ε, δ)-DP, then M satisfies δ-approximate (ε, 0)zCDP, which, in turn, implies δ-approximate (0, 21 ε2 )-zCDP.
Proof. Fix neighbouring x0 , x1 ∈ X n . Let T : {0, 1} × {⊥, >} → Y be as in Lemma 8.6.
Now we can write M (xb ) = T (M̃ε,δ (b)) for b ∈ {0, 1}. Define events E0 and E1 by
h
i
Eb ≡ M̃ε,δ (b) ∈ {0, 1} × {⊥} .
By definition, for both b ∈ {0, 1},
P
[Eb ] = 1 − δ and
M̃ε,δ (b)
M (xb )|Eb = T M̃ε,δ (b)|M̃ε,δ (b)∈{0,1}×{⊥} = T (M̃ε,0 (b)).
31
We have D∞ M̃ε,0 (b) M̃ε,0 (1 − b) ≤ ε for both b ∈ {0, 1}. By postprocessing and
monotonicity, this implies
Dα M (xb )|Eb M (x1−b )|E1−b ≤ ε
for both b ∈ {0, 1} and all α ∈ (1, ∞). Thus we have satisfied the definition of δ-approximate
(ε, 0)-zCDP.
Applying Proposition 3.3 shows that this also implies δ-approximate (0, 12 ε2 )-zCDP.
8.2
Approximate zCDP Implies Approximate DP
Lemma 8.8. Suppose M : X n → Y satisfies δ-approximate (ξ, ρ)-zCDP. If ρ = 0, then M
satisfies (ξ, δ)-DP. In general, M satisfies (ε, δ + (1 − δ)δ 0 )-DP for all ε ≥ ξ + ρ, where
1
√
π·ρ
2
1
.
δ 0 = e−(ε−ξ−ρ) /4ρ · min
1+(ε−ξ−ρ)/2ρ
2
r
1+ ε−ξ−ρ + (1+ ε−ξ−ρ )2 + 4
2ρ
2ρ
πρ
Proof. Fix neighbouring x, x0 ∈ X n and let E and E 0 be the events promised by definition
8.1. We can assume, without loss of generality that P [E] = P [E 0 ] = 1 − δ.
Fix S ⊂ Y. Then
P [M (x) ∈ S] =P [M (x) ∈ S | E] · P [E] + P [M (x) ∈ S | ¬E] · P [¬E]
≤P [M (x) ∈ S | E] · (1 − δ) + δ,
P [M (x0 ) ∈ S] =P [M (x0 ) ∈ S | E 0 ] · P [E 0 ] + P [M (x0 ) ∈ S | ¬E 0 ] · P [¬E 0 ]
≥P [M (x0 ) ∈ S | E 0 ] · (1 − δ).
Firstly, if ρ = 0, then
P [M (x) ∈ S | E] ≤ eξ P [M (x0 ) ∈ S | E 0 ]
and
P [M (x) ∈ S] ≤ P [M (x) ∈ S | E]·(1−δ)+δ ≤ eξ P [M (x0 ) ∈ S | E 0 ]·(1−δ)+δ ≤ eξ P [M (x0 ) ∈ S]+δ,
which proves the first half of the lemma.
Secondly, by Lemma B.2 (cf. Lemma 3.5), for all ε ≥ ξ + ρ,
1
√
π·ρ
2
1
P [M (x) ∈ S | E] ≤ eε P [M (x0 ) ∈ S | E 0 ] + e−(ε−ξ−ρ) /4ρ · min
1+(ε−ξ−ρ)/2ρ
r 2
1+ ε−ξ−ρ + (1+ ε−ξ−ρ )2 + 4
2ρ
32
2ρ
πρ
.
Thus
2 /4ρ
P [M (x) ∈ S] ≤ eε P [M (x0 ) ∈ S] + δ + (1 − δ) · e−(ε−ξ−ρ)
· min
1
√
π·ρ
8.3
1
1+(ε−ξ−ρ)/2ρ
r 2
1+ ε−ξ−ρ
+
2ρ
(1+ ε−ξ−ρ
2ρ )
.
2
4
+ πρ
Application of Approximate zCDP
Approximate zCDP subsumes approximate DP. A result of this is that we can apply our
tightened lemmas to give a tighter version of the so-called advanced composition theorem
[DRV10].
Note that the following results are subsumed by the bounds of Kairouz, Oh, and Viswanath
[KOV15] and Murtagh and Vadhan [MV16]. However, these bounds may be extended to analyse the composition of mechanisms satisfying CDP with mechanisms satisfying approximate
DP. We believe that such a “unified” analysis of composition will be useful.
Applying Corollary 8.7, Lemma 8.2, and Lemma 8.8 yields the following result.
Corollary 8.9. Let M1 , · · · , Mk : X n → Y and
let M : X n → Y k be their composition.
P
k
Suppose each Mi satisfies (εi , δi )-DP. Set ρ = 12 i ε2i . Then M satisfies
!
k
Y
ε, 1 − (1 − δ 0 ) (1 − δi ) -DP
i
for all ε ≥ ρ and
2 /4ρ
δ 0 = e−(ε−ρ)
· min
1
√
π·ρ
1
1+(ε−ρ)/2ρ
r 2
1+ ε−ρ
+
2ρ
.
2
(1+ ε−ρ
2ρ )
4
+ πρ
A slight restatement is the following
n
k
Corollary 8.10. Let M1 , · · · , Mk : X n → Y and
Pklet2 M : X → Y be their composition.
1
2
Suppose each Mi satisfies (εi , δi )-DP. Set ε = 2 i εi . Then M satisfies
!
k
Y
2
0
ε + 2λε, 1 − (1 − δ ) (1 − δi ) -DP
i
for all λ ≥ 0 and
2
δ 0 = e−λ · min
1√
π·ε
.
1
1+λ/ε
1+ λε +
33
q 2
2
(1+ λε ) +
4
πε2
Finally, by picking the second term in the minimum and using 1 −
we have the following simpler form of the lemma.
Q
i (1
− δi ) ≤
P
i δi ,
Corollary 8.11. Let M1 , · · · , Mk : X n → Y and let M : X n → Y k be their composition.
Suppose each Mi satisfies (εi , δi )-DP. Then M satisfies
r
√
π
1
2
−λ2
kεk2 + 2λkεk2 ,
· kεk2 · e
+ kδk1 -DP
2
2
for all λ ≥ 0. Alternatively M satisfies
q
p
1
2
0
0
kεk2 + 2 log( π/2 · kεk2 /δ ) · kεk2 , δ + kδk1 -DP
2
for all δ 0 ≥ 0.
In comparison to the composition theorem
p of [DRV10], we save modestly by a constant
factor in the first term and, in most cases π/2kεk2 < 1, whence the logarithmic term is an
improvement over the usual advanced composition theorem.
Acknowledgements We thank Cynthia Dwork and Guy Rothblum for sharing a preliminary draft of their work with us. We also thank Ilya Mironov, Kobbi Nissim, Adam Smith,
Salil Vadhan, and the Harvard Differential Privacy Research Group for helpful discussions
and suggestions.
References
[AS64]
Milton
Abramowitz
and
mathematical
functions:
ematical
tables,
volume
http://people.math.sfu.ca/
Irene
A
Stegun.
Handbook of
with
formulas,
graphs,
and
math55.
Courier
Corporation,
1964.
cbm/aands/abramowitz and stegun.pdf.
[BLR13]
Avrim Blum, Katrina Ligett, and Aaron Roth. A learning theory approach to
noninteractive database privacy. J. ACM, 60(2):12, 2013.
[BNS13]
Amos Beimel, Kobbi Nissim, and Uri Stemmer. Private learning and sanitization:
Pure vs. approximate differential privacy. In Approximation, Randomization, and
Combinatorial Optimization. Algorithms and Techniques - 16th International
Workshop, APPROX 2013, and 17th International Workshop, RANDOM 2013,
Berkeley, CA, USA, August 21-23, 2013. Proceedings, pages 363–378, 2013.
[BNS+ 16a] Raef Bassily, Kobbi Nissim, Adam Smith, Thomas Steinke, Uri Stemmer, and
Jonathan Ullman. Algorithmic stability for adaptive data analysis. In STOC,
2016.
34
[BNS16b]
Mark Bun, Kobbi Nissim, and Uri Stemmer. Simultaneous private learning of
multiple concepts. In Proceedings of the 2016 ACM Conference on Innovations
in Theoretical Computer Science, ITCS ’16, pages 369–380, New York, NY, USA,
2016. ACM.
[BNSV15] Mark Bun, Kobbi Nissim, Uri Stemmer, and Salil Vadhan. Differentially private
release and learning of threshold functions. In Foundations of Computer Science
(FOCS), 2015 IEEE 56th Annual Symposium on, pages 634–649. IEEE, 2015.
[BSU16]
Mark Bun, Thomas Steinke, and Jonathan Ullman. Make up your mind: The
price of online queries in differential privacy. CoRR, abs/1604.04618, 2016.
[BUV14]
Mark Bun, Jonathan Ullman, and Salil P. Vadhan. Fingerprinting codes and the
price of approximate differential privacy. In Symposium on Theory of Computing,
STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 1–10, 2014.
[Coo09]
John D Cook. Upper and lower bounds for the normal distribution function,
2009.
[De12]
Anindya De. Lower bounds in differential privacy. In Proceedings of the 9th
International Conference on Theory of Cryptography, TCC’12, pages 321–338,
Berlin, Heidelberg, 2012. Springer-Verlag.
[DKM+ 06] Cynthia Dwork, Krishnaram Kenthapadi, Frank McSherry, Ilya Mironov, and
Moni Naor. Our data, ourselves: Privacy via distributed noise generation. In
Advances in Cryptology - EUROCRYPT 2006, 25th Annual International Conference on the Theory and Applications of Cryptographic Techniques, St. Petersburg, Russia, May 28 - June 1, 2006, Proceedings, pages 486–503, 2006.
[DL09]
Cynthia Dwork and Jing Lei. Differential privacy and robust statistics. In Proceedings of the 41st Annual ACM Symposium on Theory of Computing, STOC
2009, Bethesda, MD, USA, May 31 - June 2, 2009, pages 371–380, 2009.
[DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating
noise to sensitivity in private data analysis. In Theory of Cryptography, Third
Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March
4-7, 2006, Proceedings, pages 265–284, 2006.
[DR14]
Cynthia Dwork and Aaron Roth. The Algorithmic Foundations of Differential
Privacy, volume 9. Foundations and Trends R in Theoretical Computer Sc, 2014.
[DR16]
Cynthia Dwork and Guy Rothblum. Concentrated differential privacy. CoRR,
abs/1603.01887, 2016.
[DRV10]
Cynthia Dwork, Guy N. Rothblum, and Salil P. Vadhan. Boosting and differential privacy. In IEEE Symposium on Foundations of Computer Science (FOCS
’10), pages 51–60. IEEE, 23–26 October 2010.
35
[DSS+ 15]
Cynthia Dwork, Adam Smith, Thomas Steinke, Jonathan Ullman, and Salil
Vadhan. Robust traceability from trace amounts. In FOCS, 2015.
[Due10]
L. Duembgen. Bounding Standard Gaussian Tail Probabilities. ArXiv e-prints,
December 2010.
[hh]
hbp
(http://math.stackexchange.com/users/131476/hbp).
Hyperbolic
trig
inequality.
Mathematics
Stack
Exchange.
URL:http://math.stackexchange.com/q/1461426 (version: 2015-10-02).
[HT10]
Moritz Hardt and Kunal Talwar. On the geometry of differential privacy. In Proceedings of the Forty-second ACM Symposium on Theory of Computing, STOC
’10, pages 705–714, New York, NY, USA, 2010. ACM.
[KOV15]
Peter Kairouz, Sewoong Oh, and Pramod Viswanath. The composition theorem
for differential privacy. In Proceedings of the 32nd International Conference on
Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 1376–1385,
2015.
[MMP+ 10] Andrew McGregor, Ilya Mironov, Toniann Pitassi, Omer Reingold, Kunal Talwar, and Salil P. Vadhan. The limits of two-party differential privacy. In 51th
Annual IEEE Symposium on Foundations of Computer Science, FOCS 2010,
October 23-26, 2010, Las Vegas, Nevada, USA, pages 81–90, 2010.
[MT07]
F. McSherry and K. Talwar. Mechanism design via differential privacy. In Foundations of Computer Science, 2007. FOCS ’07. 48th Annual IEEE Symposium
on, pages 94–103, Oct 2007.
[MV16]
Jack Murtagh and Salil P. Vadhan. The complexity of computing the optimal
composition of differential privacy. In Theory of Cryptography - 13th International Conference, TCC 2016-A, Tel Aviv, Israel, January 10-13, 2016, Proceedings, Part I, pages 157–175, 2016.
[Rén61]
Alfréd Rényi. On measures of entropy and information. In Proceedings of the
Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume
1: Contributions to the Theory of Statistics, pages 547–561, Berkeley, Calif.,
1961. University of California Press.
[Riv12]
Omar Rivasplata. Subgaussian random variables: An expository note, 2012.
http://www.stat.cmu.edu/ arinaldo/36788/subgaussians.pdf.
[SU15a]
Thomas Steinke and Jonathan Ullman. Between pure and approximate differential privacy. CoRR, abs/1501.06095, 2015.
[SU15b]
Thomas Steinke and Jonathan Ullman.
Interactive fingerprinting codes
and the hardness of preventing false discovery.
In COLT, 2015.
http://arxiv.org/abs/1410.1228.
36
[Tar08]
Gábor Tardos. Optimal probabilistic fingerprint codes. J. ACM, 55(2), 2008.
[Ull13]
Jonathan Ullman. Answering n {2+ o (1)} counting queries with differential
privacy is hard. In Proceedings of the forty-fifth annual ACM symposium on
Theory of computing, pages 361–370. ACM, 2013.
[vEH14]
T. van Erven and P. Harremos. Rényi divergence and Kullback-Leibler divergence. IEEE Transactions on Information Theory, 60(7):3797–3820, July 2014.
[War65]
Stanley L. Warner. Randomized response: A survey technique for eliminating
evasive answer bias. Journal of the American Statistical Association, 60(309):63–
69, 1965. PMID: 12261830.
A
Postprocessing and mCDP
In this appendix we give a family of counterexamples showing that mCDP is not closed
under postprocessing (unlike zCDP).
Fix a parameter σ > 0, and consider the Gaussian mechanism for a single bit M :
{−1, 1} → R, where M (x) samples from N (x, σ 2 ). The mechanism M satisfies (2/σ 2 , 2/σ)mCDP (and also (2/σ 2 )-zCDP).
Now consider the postprocessing function T : R → {−1, 0, 1} defined as follows:
if y > t
1
T (y) = −1 if y < −t
0
if − t ≤ y ≤ t.
We examine the mCDP guarantees of the postprocessed mechanism M 0 : {−1, 1} →
{−1, 0, 1} defined by M 0 (x) = T (M (x)):
Proposition A.1. Let σ ≥ 1 and let t ≥ 6σ 3 + 1. Then while the mechanism M is
(2/σ 2 , 2/σ)-mCDP, the postprocessed mechanism M 0 is not (2/σ 2 , 2/σ)-mCDP.
Proof. For each x ∈ {−1, 1}, let
p = P [M 0 (x) = x] = P N (0, σ 2 ) > t − 1 ,
q = P [M 0 (x) = −x] = P N (0, σ 2 ) > t + 1 .
Note that p > q. Hence, for each x ∈ {−1, 1},
P [M 0 (x) = 0] = P N (0, σ 2 ) ∈ [−t − 1, t − 1] = 1 − p − q.
Let f (y) = log(P [M (1) = y] /P [M (−1) = y]), and observe that
p
f (1) = log ,
q
q
f (−1) = log ,
p
37
f (0) = 0.
Now consider the privacy loss random variable Z = PrivLoss (M (1)kM (−1)) = f (M (1)).
Then Z is distributed according to
p
log q w.p. p
Z = log pq w.p. q
.
0
w.p. 1 − p − q
This gives E [Z] = (p − q) log(p/q) ≥ 0. For λ ∈ R, we have
!
λ
λ
−λ(p−q)
q
p
p
λ(Z−E[Z])
+q
+1−p−q ·
.
E e
= p
q
p
q
h
i
2
2
λ(Z−E[Z])
If M 0 were to satisfy (2/σ 2 , 2/σ)-mCDP, we would have E e
≤ e2λ /σ for all
λ > 0. We will show that this does not hold for any setting of parameters σ ≥ 1 and
t ≥ 6σ 3 + 1, which shows that mCDP is not closed under postprocessing.11
h
i
Lemma A.2. The values p, q satisfy the following inequalities:
q
q
2
2
1
σ
1
−(t−1)2 /2σ 2
·
·
e
≤
p
≤
· σ · e−(t−1) /2σ
1.
2π t+σ−1
2π t−1
2.
p
q
≥ e2t/σ
2
Proof. We have [Coo09, Equation (5)]
q
q
2 −x2 /2
2 −x2 /2
e
·
x
e
π
π
≤
P
[N
(0,
1)
>
x]
≤
x2 + 1
x
for all x ≥ 0. Thus
q
q
2 −(t−1)2 /2σ 2
2 −(t−1)2 /2σ 2
e
·
(t
−
1)
e
·σ
t−1
π
π
≤
p
=
P
N
(0,
1)
>
≤
.
σ · ((t − 1)2 /σ 2 + 1)
σ
t−1
The first part now follows from the fact that σ ≤ t − 1.
To establish the second inequality, we write
Z ∞
1
2
2
p= √
e−x /2σ dx
2
2πσ Zt−1
∞
1
2
2
e−(u−2) /2σ du
=√
2
2πσ Zt+1
∞
1
2
2
2
=√
e(4u−4)/2σ · e−u /2σ du
2
2πσ t+1 Z
∞
1
2
2
2
≥ e2t/σ √
e−u /2σ du
2
2πσ t+1
2t/σ 2
=e
· q.
11
For specific settings of parameters, this can be verified numerically. (For example, with the values σ = 1,
t = 3, and λ = 2.)
38
By Lemma A.2,
h
λ(Z−E[Z])
E e
i
!
λ
λ
−λ(p−q)
p
q
p
= p
+q
+1−p−q ·
q
p
q
λ(1−(p−q))
p
≥p
q
r
σ
1
2
2
2
·
· e−(t−1) /2σ · (e2t/σ )λ(1−p) .
≥
2π t + σ − 1
Now set λ = 2t . Then this quantity becomes
r
σ
1
2
2
2
2
2
2
·
· e−(t−1) /2σ · et (1−p)/σ = et /2σ ·
2π t + σ − 1
We now examine the term
r
2
2
1 σt2 e−(t−1) /2σ
2
pt ≤
2π
t−1
r
5/2 −(t−1)
1 t e
≤
4π t − 1
1
≤
2
r
σ
1
·
· exp
2π t + σ − 1
t
pt2
1
− 2 − 2 .
σ2
σ
2σ
(10)
by Lemma A.2
for t ≥ 2σ 2 + 1
for t ≥ 3.
We may thus bound (10) from below by
r
r
σ
t−1
t−1
1
1 1
t2 /2σ 2
t2 /2σ 2
e
·
·
· exp
≥e
·
· · exp
2π t + σ − 1
2σ 2
2π 2t
2σ 2
2
The expression
is monotone increasing for t ≥ 2σ 2 . Thus, it is strictly
√ exp((t − 1)/2σ )/2t
larger than 2π as long as t ≥ 6σ 3 + 1.
B
Miscellaneous Proofs and Lemmata
Lemma B.1.
0 ≤ y < x ≤ 2 =⇒
1
sinh(x) − sinh(y)
≤ e 2 xy .
sinh(x − y)
1
This technical lemma may be “verified” numerically by inspecting a plot of z = e 2 xy ·
sinh(x − y) − (sinh(x) − sinh(y)) for (x, y) ∈ [0, 2]2 .
39
For intuition, consider the third-order Taylor approximation to sinh(x) about 0:
1
sinh(x) = x + x3 ± O(x5 ).
6
Then we can approximate
x + 61 x3 − (y + 16 y 3 )
sinh(x) − sinh(y)
≈
sinh(x − y)
(x − y) + 16 (x − y)3
3xy(x − y)
=1+
6(x − y) + (x − y)3
1
≤ 1 + xy
2
1
xy
≤ e2 .
Unfortunately, turning this intuition into an actual proof is quite involved. We instead
provide a proof from [hh]:
Proof. We need the following hyperbolic trigonometric identities.
cosh (w + z) = cosh(w) cosh(z) + sinh(w) sinh(z)
= cosh(w) cosh(z) (1 + tanh(w) tanh(z)) ,
1
sinh(x − y) = ex−y + 1 − 1 − e−(x−y)
2
1 (x−y)/2
= e
− e−(x−y)/2 e(x−y)/2 + e−(x−y)/2
2
x−y
x−y
=2 sinh
cosh
2
2
x
x−y
−y
−y
x
=2 sinh
cosh
tanh
cosh
1 + tanh
2
2
2
2
2
x
y
x
y
x−y
cosh
cosh
1 − tanh
tanh
,
=2 sinh
2
2
2
2
2
1
sinh(x) − sinh(y) = ex − e−x − ey + e−y
2
1 (x−y)/2
= e
− e−(x−y)/2 e(x+y)/2 + e−(x+y)/2
2
x−y
x+y
=2 sinh
cosh
2
2
y
x
y
x
x−y
=2 sinh
cosh
cosh
1 + tanh
tanh
.
2
2
2
2
2
40
We also use the fact that 0 ≤ tanh(z) < min{z, 1} for all z > 0. Thus
1 + tanh x2 tanh y2
sinh(x) − sinh(y)
1+t
=
,
y =
x
sinh(x − y)
1−t
1 − tanh 2 tanh 2
where t = tanh x2 tanh y2 < 1. Now
1+t
≤ exy/2 ⇐⇒ 1 + t ≤ (1 − t)exy/2 ⇐⇒ (exy/2 + 1)t ≤ exy/2 − 1
1−t
xy
exy/4 − e−xy/4
exy/2 − 1
= xy/4
.
⇐⇒ t ≤ xy/2
= tanh
e
+1
e
+ e−xy/4
4
So it only remains to show that tanh x2 tanh y2 ≤ tanh xy
. This is clearly true when
4
y = 0. Now
x
y
x
y 1
y x
∂
tanh
tanh
= tanh
cosh−2
≤ cosh−2
∂y
2
2
2
2 2
2 4
and
xy
xy x
∂
tanh
= cosh−2
.
∂y
4
4 4
Since 0 ≤ y < x ≤ 2, we have 0 ≤ xy/4 ≤ y/2 and, hence, cosh(y/2) ≥ cosh(xy/4). Thus
x
y
xy
∂
∂
tanh
tanh
≤
tanh
.
∂y
2
2
∂y
4
The inequality now follows by integration of the inequality on the derivatives.
The following lemma immediately implies Lemma 3.6 and is also used to prove Lemma
8.8.
Lemma B.2. Let P and Q be probability distributions on Y with Dα (P kQ) ≤ ξ + ρ · α for
all α ∈ (1, ∞). Then, for any ε ≥ ξ + ρ and
1
√
π·ρ
2
1
δ = e−(ε−ξ−ρ) /4ρ · min
,
1+(ε−ξ−ρ)/2ρ
2
r
1+ ε−ξ−ρ + (1+ ε−ξ−ρ )2 + 4
2ρ
2ρ
πρ
we have
P (S) ≤ eε Q(S) + δ
for all (measurable) S.
41
Proof. Define f : Y → R by f (y) = log(P (y)/Q(y)). Let Y ∼ P , Y 0 ∼ Q and let Z = f (Y )
be the privacy loss random variable. That is, Z = PrivLoss (P kQ). For any measurable
S ⊂ Y,
P (S) =P [Y ∈ S]
=P [Y ∈ S ∧ f (Y ) ≤ ε] + P [Y ∈ S ∧ f (Y ) > ε]
Z
= P (y)I[f (y) ≤ ε]dy + P [Y ∈ S ∧ f (Y ) > ε]
ZS
= P (y)I[P (y) ≤ eε Q(y)]dy + P [Y ∈ S ∧ f (Y ) > ε]
ZS
≤ eε Q(y)I[f (y) ≤ ε]dy + P [Y ∈ S ∧ f (Y ) > ε]
S
=eε P [Y 0 ∈ S ∧ f (Y 0 ) ≤ ε] + P [Y ∈ S ∧ f (Y ) > ε]
=eε P [Y 0 ∈ S] − eε P [Y 0 ∈ S ∧ f (Y 0 ) > ε] + P [Y ∈ S ∧ f (Y ) > ε]
≤eε P [Y 0 ∈ S] + P [f (Y ) > ε] − eε P [f (Y 0 ) > ε] .
Thus we want to bound
δ =P [f (Y ) > ε] − eε P [f (Y 0 ) > ε]
=E [I[f (Y ) > ε]] − E0 [eε I[f (Y 0 ) > ε]]
Y
Y
Z
eε I[f (y) > ε]Q(y)dy
=E [I[f (Y ) > ε]] −
Y
ZY
Q(y)
eε I[f (y) > ε]
=E [I[f (Y ) > ε]] −
P (y)dy
Y
P (y)
Y
=E [I[f (Y ) > ε]] − E eε I[f (Y ) > ε]e−f (Y )
Y
Y
ε−Z
=E I[Z > ε] 1 − e
Z
=E max 0, 1 − eε−Z
Z
Z ∞
=
1 − eε−z P [Z = z] dz.
ε
In particular,
δ ≤ E [I[Z > ε]] = P [Z > ε] .
Z
Alternatively, by integration by parts,
Z
δ=
∞
eε−z P [Z > z] dz.
ε
Now it remains to bound P [Z > z].
42
As in Lemma 3.5, by Markov’s inequality, for all α > 1 and λ > ξ + ρ,
0
E e(α−1)Z
e(α−1)Dα (M (x)kM (x ))
P [Z > λ] ≤
≤
≤ e(α−1)(ξ+ρα−λ) .
e(α−1)λ
e(α−1)λ
Choosing α = (λ − ξ + ρ)/2ρ > 1 gives
P [Z > λ] ≤ e−(λ−ξ−ρ)
2 /4ρ
Thus δ ≤ P [Z > ε] ≤ e−(ε−ξ−ρ)
Z
. Furthermore,
eε−z P [Z > z] dz
Zε ∞
≤
Z
.
∞
δ=
ε
2 /4ρ
2 /4ρ
eε−z e−(z−ξ−ρ)
∞
dz
2
e−(z−ξ+ρ) /4ρ−ξ+ε dz
ε
√
Z
eε−ξ · 2π · 2ρ ∞ −(z−ξ+ρ)2 /4ρ
= √
e
dz
2π · 2ρ
ε
p
=eε−ξ · 2π · 2ρ · P [N (ξ − ρ, 2ρ) > ε]
√
ε+ρ−ξ
ε−ξ
=e · 2 π · ρ · P N (0, 1) > √
.
2ρ
√
2
Define h(x) = P [N (0, 1) > x] · 2π · ex /2 . Then
=
δ≤e
ε−ξ
√
·2 π·ρ· √
h
ε+ρ−ξ
√
2ρ
2
2π · e
ε+ρ−ξ
√
2ρ
/2
p
p
ε−ξ−ρ
−(ε−ξ−ρ)2 /4ρ
= 2ρ · e
·h
2ρ + √
.
2ρ
Now we can substitute upper bounds on h to obtain the desired results.
p
2
First we use P [N (0, 1) > x] ≤ 12 e−x /2 , which holds for all x ≥ 0. That is, h(x) ≤ π/2,
whence
r
p
π
−(ε−ξ−ρ)2 /4ρ
δ ≤ 2ρ · e
·
.
2
This rearranges to
q
√
ε ≤ ξ + ρ + 4ρ · log( π · ρ/δ).
√
2
Alternatively, we can use P [N (0, 1) > x] ≤ e−x /2 / 2πx, which holds for all x > 0. That
is, h(x) ≤ 1/x and
p
2
δ ≤ 2ρ · e−(ε−ξ−ρ) /4ρ · h
ε−ξ+ρ
√
2ρ
2
2ρ
e−(ε−ξ−ρ) /4ρ
2
=
· e−(ε−ξ−ρ) /4ρ =
·.
ε+ρ−ξ
1 + (ε − ξ − ρ)/2ρ
43
Finally, we can use P [N (0, 1) > x] ≤ e−x
2 /2
· 2/(
p
√
x2 + 8/π + x) 2π, which holds for
all x ≥ 0 p
[Due10, Equation (4)][Coo09, Equation (7)][AS64, Equation 7.1.13]. That is,
h(x) ≤ 2/( x2 + 8/π + x) and
p
2
δ ≤ 2ρ · e−(ε−ξ−ρ) /4ρ · h
ε−ξ+ρ
√
2ρ
2
2 · e−(ε−ξ−ρ) /4ρ
r
=
2
ε−ξ−ρ
1 + 2ρ +
1 + ε−ξ−ρ
+
2ρ
.
4
πρ
Lemma B.3 (Restating Lemma 3.7). Let M : X n → Y satisfy (ε, δ)-DP for all δ > 0 and
p
(11)
ε = ξˆ + ρ̂ log(1/δ)
1
4
ˆ ρ̂ ∈ [0, 1]. Then M is ξˆ − 1 ρ̂ + 5√
-zCDP.
for some constants ξ,
ρ̂
ρ̂,
4
4
Proof. Let x, x0 ∈ X n be neighbouring. Define f (y) = log(P [M (x) = y] /P [M (x0 ) = y]). Let
Y ∼ M (x) and Z = f (Y ). That is, Z = PrivLoss (M (x)kM (x0 )) is the privacy loss random
variable. Let Y 0 ∼ M (x0 ) and Z 0 = f (Y 0 ). That is, −Z 0 is the privacy loss random variable
if we swap x and x0 .
Let ε, δ > 0 satisfy (11). By postprocessing, for all t ∈ R,
P [Z > t] =P [f (M (x)) > t]
≤eε P [f (M (x0 )) > t] + δ
Z
ε
=e
P [M (x0 ) = y] · I(P [M (x) = y] > et P [M (x0 ) = y])dy + δ
ZY
<eε P [M (x) = y] · e−t · I(P [M (x) = y] > et P [M (x0 ) = y])dy + δ
Y
=eε−t P [Z > t] + δ,
whence P [Z > t] ≤
δ
.
1−eε−t
2
Then we can set ε = ξˆ + λ and δ = e−λ /ρ̂ to obtain
2
e−λ /ρ̂
.
0<λ<t 1 − eλ−t
h
i
ˆ
∀t > 0 P Z > ξ + t ≤ inf
In particular, for all t ≥ 0,
2
h
P Z > ξˆ +
p
4
i
ρ̂ + t ≤
e−t /ρ̂
2 −t2 /ρ̂
√
≤ √
e
.
4
4
−
ρ̂
ρ̂
1−e
44
We use the inequality 1 + xeξ ≤ ex+ξ for all x, ξ ≥ 0. We have
Z ∞
(α−1)Z
P e(α−1)Z > t dt
E e
=
Z0 ∞
log t
P Z>
=
dt
α−1
0
Z ∞
dt
P [Z > z] dz
=
dz
Z−∞
∞
(α − 1)e(α−1)z · P [Z > z] dz
=
−∞
√
ξ̂+ 4 ρ̂
∞
√
2 −t2 /ρ̂
4
e
dt
(α − 1)e(α−1)(t+ξ̂+ ρ̂) · √
4
ρ̂
−∞
0
Z ∞
√
√
2
(α−1)(ξ̂+ 4 ρ̂)
(α−1)(ξ̂+ 4 ρ̂) 2
√
=e
+ (α − 1)e
e(α−1)t · e−t /ρ̂ dt
4
ρ̂ 0
Z ∞
√
√
(α−1)(ξ̂+ 4 ρ̂) 2
(α−1)(ξ̂+ 4 ρ̂)
−(t− 12 (α−1)ρ̂)2 /ρ̂+ 14 (α−1)2 ρ̂
√
+ (α − 1)e
dt
≤e
e
4
ρ̂ −∞
√
√
2 1 (α−1)2 ρ̂ p
4
4
=e(α−1)(ξ̂+ ρ̂) + (α − 1)e(α−1)(ξ̂+ ρ̂) √
e4
π ρ̂
4
ρ̂
√
√ p
1
4
(α−1)2 ρ̂
(α−1)(ξ̂+ 4 ρ̂)
4
1 + (α − 1)2 π ρ̂e
=e
Z
≤
(α − 1)e
√
4
(α−1)z
Z
· 1dz +
√ √
4
1
≤e(α−1)(ξ̂+ ρ̂) e(α−1)2 π ρ̂+ 4 (α−1)
√
√ √
1
4
4
=e(α−1)(ξ̂+ ρ̂+2 π ρ̂+ 4 (α−1)ρ̂)
√
=e(α−1)(ξ̂+(1+2
√
π) 4 ρ̂− ρ̂4 + ρ̂4 α)
2 ρ̂
.
0
Since E e(α−1)Z = e(α−1)Dα (M (x)kM (x )) , this completes the proof.
We make use of the following technical lemma taken from [DSS+ 15].
Lemma B.4. Let X be a random variable. Then
h
i 1 1
X−E[X]
E e
≤ E e2X + E e−2X .
2
2
Proof. Let X 0 be an independent copy of X and let Y ∈ {0, 1} be uniformly random and
independent from X and X 0 . By Jensen’s inequality,
h
i
h E [X−X 0 ] i
h
i
h 2E[Y X−(1−Y )X 0 ] i
0
X−E[X]
E e
=E eX 0
≤ E 0 eX−X = E 0 e Y
X
X,X
X,X
i 1 1
h
i
h
1
1
2Y X−2(1−Y )X 0
2X−0
0−2X 0
= E0 e
+ E0 e
= E e2X + E e−2X .
≤ E0 e
X,X ,Y
2 X,X
2 X,X
2
2
45
n
Lemma B.5
√ (Restating Lemma 4.3). If M : X → Y satisfies (ξ, ρ)-zCDP, then M satisfies
(ξ + ρ, O( ξ + 2ρ))-CDP.
Proof. Let x and x0 be neighbouring databases and Z ∼ PrivLoss (M (x)kM (x0 )) the privacy
loss random variable We have E [Z] = D1 (M (x)kM (x0 )) ∈ [0, ξ + ρ] by non-negativity and
zCDP. By zCDP, for all α ∈ (1, ∞), we have
0
E e(α−1)Z = e(α−1)Dα (M (x)kM (x )) ≤ e(α−1)(ξ+ρα)
and
E e−αZ =
E
Y ∼M (x)
=
P [M (x) = Y ]
P [M (x0 ) = Y ]
"
P [M (x0 ) = Y ]
E
!−α
!α #
P [M (x) = Y ]
Y ∼M (x)
0
=e(α−1)Dα (M (x )kM (x))
≤e(α−1)(ξ+ρα) .
By Lemma B.4, for λ ≥ 1/2,
h
i 1
1
1
1
2
λ(Z−E[Z])
E e
≤ E e2λZ + E e−2λZ ≤ e2λ(ξ+ρ(2λ+1)) + e(2λ−1)(ξ+2ρλ) ≤ e4(ξ+2ρ)λ
2
2
2
2
and, for λ ≤ −1/2,
h
i 1
1
1
1
2
λ(Z−E[Z])
E e
≤ E e2λZ + E e−2λZ ≤ e(−2λ−1)(ξ−2ρλ) + e−2λ(ξ+ρ(1−2λ)) ≤ e4(ξ+2ρ)λ .
2
2
2
2
Now suppose |λ| < 1/2. Then
h
i 1
1
λ(Z−E[Z])
E e
≤ E e2λZ + E e−2λZ
2
2
∞
X
(2λ)2k 2k
=1 +
E Z
(2k)!
k=1
2
≤1 + (2λ)
∞
X
1
E Z 2k
(2k)!
k=1
1 Z 1 −Z
2
=1 + 4λ
E e + E e
−1
2
2
≤1 + 4λ2 eξ+2ρ − 1
≤eλ
2 O(ξ+2ρ)
.
46
B.1
Proof of Lemma 2.2
Proof of Non-Negativity. Let h(t) = tα . Then h00 (t) = α(α − 1)tα−2 > 0 for all t > 0
and α > 1. Thus h is strictly convex. Hence e(α−1)Dα (P kQ)) = E [h(P (x)/Q(x))] ≥
x∼Q
h( E [P (x)/Q(x)]) = h(1) = 1, as required.
x∼Q
Proof of Composition.
(α−1)Dα (P kQ)
e
Z
P (x, y)α Q(x, y)1−α d(x, y)
ZΩ×Θ
Z
0
α 0
1−α
= P (x) Q (x)
Px0 (y)α Q0x (y)1−α dydx
Θ
ZΩ
0
0
= P 0 (x)α Q0 (x)1−α e(α−1)Dα (Px kQx ) dx
ZΩ
0
0
≤ P 0 (x)α Q0 (x)1−α dx · max e(α−1)Dα (Px kQx )
=
Ω
(α−1)Dα (P 0 kQ0 )
=e
x
(α−1) maxx Dα (Px0 kQ0x )
·e
.
The other side of the inequality is symmetric.
Proof of Quasi-Convexity. Unfortunately Rényi divergence is not convex for α > 1. (Although KL-divergence is.) However, the following property implies that Rényi divergence is
quasi-convex.
Lemma B.6. Let P0 , P1 , Q0 , Q1 be distributions on Ω. For t ∈ [0, 1], define Pt = tP1 + (1 −
t)P0 and Qt = tQ1 + (1 − t)Q0 to be the convex combinations specified by t. Then
e(α−1)Dα (Pt kQt ) ≤ te(α−1)Dα (P1 kQ1 ) + (1 − t)e(α−1)Dα (P0 kQ0 ) .
Moreover, the limit as α → 1+ gives
D1 (Pt kQt ) ≤ tD1 (P1 kQ1 ) + (1 − t)D1 (P0 kQ0 ) .
Proof of Lemma B.6. Let f (t) = e(α−1)Dα (Pt kQt ) . Since the equality is clearly true for t = 0
47
and t = 1, it suffices to show that f 00 (t) ≥ 0 for all t ∈ [0, 1]. We have
Z
f (t) = Pt (x)α Qt (x)1−α dx,
ZΩ
d
f 0 (t) =
Pt (x)α Qt (x)1−α dx
dt
ZΩ
d
d
α−1
1−α
α
−α
= αPt (x)
Pt (x) Qt (x)
+ (1 − α)Pt (x) Qt (x)
Qt (x) dx
dt
dt
Ω
Z
= αPt (x)α−1 (P1 (x) − P0 (x)) Qt (x)1−α + (1 − α)Pt (x)α Qt (x)−α (Q1 (x) − Q0 (x)) dx,
ZΩ
f 00 (t) = α(α − 1)Pt (x)α−2 (P1 (x) − P0 (x))2 Qt (x)1−α
Ω
+ α(1 − α)Pt (x)α−1 (P1 (x) − P0 (x)) Qt (x)−α (Q1 (x) − Q0 (x))
+ (1 − α)αPt (x)α−1 (P1 (x) − P0 (x)) Qt (x)−α (Q1 (x) − Q0 (x))
+ (1 − α)(−α)Pt (x)α Qt (x)−α−1 (Q1 (x) − Q0 (x))2 dx
Z p
2
p
α−2
1−α
α
−α−1
=α(α − 1)
Pt (x) Qt (x)
(P1 (x) − P0 (x)) − Pt (x) Qt (x)
(Q1 (x) − Q0 (x)) dx
Ω
≥0.
Proof of Postprocessing. Let h(x) = xα . Note that h is convex. Let f −1 (y) = {x ∈ Ω :
f (x) = y}. Let Qy be the conditional distribution on x ∼ Q conditioned on f (x) = y. By
Jensen’s inequality,
α
P (x)
(α−1)Dα (P kQ)
e
= E
x∼Q
Q(x)
P (x)
= E
E h
y∼f (Q) x∼Qy
Q(x)
P (x)
≥ E
h
E
x∼Qy Q(x)
y∼f (Q)
Z
Q(x) P (x)
= E
h
dx
−1 (y)) Q(x)
y∼f (Q)
f −1 (y) Q(f
P (f −1 (y))
= E
h
y∼f (Q)
Q(f −1 (y))
=e(α−1)Dα (f (P )kf (Q)) .
48
α0 −1
Proof of Monotonicity. Let 1 < α ≤ α0 < ∞. Let h(x) = x α−1 . Then
0
α −1
α0 − 1 α0 − 1
00
h (x) =
− 1 x α−1 −2 ≥ 0,
α−1 α−1
so h is convex on (0, ∞). Thus
(α0 −1)Dα (P kQ)
e
=h e
(α−1)Dα (P kQ)
"
=h
E
x∼P
P (x)
Q(x)
"
α−1 #!
≤ E
x∼P
h
P (x)
Q(x)
α−1 !#
which gives the result.
C
Privacy versus Sampling
In this appendix we prove the technical Lemma 7.5. Essentially we show that if a private
mechanism can accurately answer a set of queries with a given sample complexity, then those
queries can be approximated on an unknown distribution with the same sample complexity.
This is related to lower bounds on private sample complexity using Vapnik-Chervonenkis
dimension e.g. [DR14, Theorem 4.8] and [BNSV15, Theorem 5.5].
First we state Pinsker’s inequality [vEH14, Theorem 31].
Lemma C.1 (Pinsker’s Inequality). Let X and Y be random variables on [−1, 1]. Then
p
E [X] − E [Y ] ≤ 2D1 (XkY )
We can also generalise Pinsker’s inequality using Rényi divergence:
Lemma C.2. Let P and Q be distributions on Ω and f : Ω → R. Then
E [f (x)] − E [f (x)] ≤
x∼P
x∼Q
q
E [f (x)2 ] ·
x∼Q
p
eD2 (P kQ) − 1.
In particular, if M : X n → Y satisfies (ξ, ρ)-zCDP. Then, for any f : Y → R and all
neighbouring x, x0 ∈ X n ,
p
q
E [f (M (x0 ))] − E [f (M (x))] ≤ E [f (M (x))2 ] · eξ+2ρ − 1.
Proof. By Cauchy-Schwartz,
P (x)
E [f (x)] − E [f (x)] = E f (x)
−1
x∼P
x∼Q
x∼Q
Q(x)
v
"
u
2 #
u
q
P
(x)
≤ E [f (x)2 ] · t E
−1
.
x∼Q
x∼Q
Q(x)
49
0
= e(α −1)Dα0 (P kQ) ,
Now
"
E
x∼Q
"
#
2 #
2
P (x)
P (x)
P (x)
−1
= E
−2
+1
x∼Q
Q(x)
Q(x)
Q(x)
"
2 #
P (x)
= E
−2+1
x∼Q
Q(x)
=eD2 (P kQ) − 1.
Proposition C.3. Let q : X → Rk satisfy maxx∈X kq(x)k ≤ 1, where k · k is some norm.
Let M : X n → Rk satisfy (ξ, ρ)-zCDP and
E [kM (x) − q(x)k] ≤ α
M
for all x ∈ X n . Then, for any distribution D on X ,
En
x∼D ,M
[kM (x) − q(D)k] ≤ α +
p
2(ξ + ρ)
and
E n [kq(x) − q(D)k] ≤ 2α +
x∼D
p
2(ξ + ρ),
where q(D) = E [q(z)].
z∼D
Proof. First define the dual norm: For x ∈ Rk ,
kxk∗ :=
max
y∈Rk :kyk=1
hx, yi.
By definition, hx, yi = hy, xi ≤ kxk∗ ·kyk for all x, y ∈ Rk . Moreover, kzk = maxy∈Rk :kyk∗ =1 hz, yi
for all z ∈ Rk .
Fix a distribution D. Define W : X n → Rk × Rk as follows. On input x ∈ X n , compute
a = M (x) and
s = argmax ha − q(D), vi,
v∈Rk :kvk∗ =1
and output (a, s).
By postprocessing, W satisfies (ξ, ρ)-zCDP and
E [ha − q(x), si | (a, s) = W (x)] ≤ E [ka − q(x)k · ksk∗ | (a, s) = W (x)] = E [kM (x) − q(x)k] ≤ α
W
W
M
(12)
for all x ∈ X n .
50
The following is similar to [BNS+ 16a, Lemma 3.1]. Let x ∼ Dn and y ∼ D. Now
1X
E [hq(x), si | (a, s) = W (x)] =
E [hq(xi ), si | (a, s) = W (x)]
x,W
x,W
n
(13)
i∈[n]
=
1X
E [f (xi , W (x))]
x,W
n
i∈[n]
(letting f : X × Rk × Rk → [−1, 1] be f (z, a, s) = hq(z), si/ksk∗ )
≤
1X
E [f (xi , W (x1 , · · · , xi−1 , y, xi+1 , · · · , xn ))]
x,y,W
n
i∈[n]
p
+ 2D1 (f (xi , W (x))kf (xi , W (x1 , · · · , xi−1 , y, xi+1 , · · · , xn )))
(by Pinsker’s inequality)
≤
1X
E [f (y, W (x1 , · · · , xi−1 , xi , xi+1 , · · · , xn ))]
x,y,W
n
i∈[n]
p
+ 2D1 (W (x)kW (x1 , · · · , xi−1 , y, xi+1 , · · · , xn ))
(by postprocessing and convexity and the fact that xi and y are interchangable)
≤
1X
E [hq(y), si | (j, s, a) = W (x)]
x,y,W
n
i∈[n]
p
+ 2(ξ + ρ)
(by zCDP)
= E [hq(D), si | (a, s) = W (x)] +
x,W
p
2(ξ + ρ).
Combining (12) and (13) gives
E [kM (x) − q(D)k] = E [ha − q(D), si | (a, s) = W (x)] ≤ α +
x,M
x,W
p
2(ξ + ρ).
Finally, combining (14) and (12) gives
E [kq(x) − q(D)k] ≤ E [kM (x) − q(x)k] + E [kM (x) − q(D)k] ≤ 2α +
x
x,M
x,M
51
p
2(ξ + ρ).
(14)
0
Proof of Lemma 7.5. Fix x ∈ X n . Let D be the uniform distribution on elements of x so
that q(D) = q(x). By Proposition C.3,
p
E n [kq(x̂) − q(D)k] ≤ 2α + 2(ξ + ρ).
x̂∼D
In particular, there must exist x̂ ∼ Dn such that kq(x̂) − q(D)k ≤ 2α +
required.
52
p
2(ξ + ρ), as
| 8cs.DS
|
Weight Normalization: A Simple Reparameterization
to Accelerate Training of Deep Neural Networks
arXiv:1602.07868v3 [cs.LG] 4 Jun 2016
Tim Salimans
OpenAI
[email protected]
Diederik P. Kingma
OpenAI
[email protected]
Abstract
We present weight normalization: a reparameterization of the weight vectors
in a neural network that decouples the length of those weight vectors from their
direction. By reparameterizing the weights in this way we improve the conditioning
of the optimization problem and we speed up convergence of stochastic gradient
descent. Our reparameterization is inspired by batch normalization but does not
introduce any dependencies between the examples in a minibatch. This means
that our method can also be applied successfully to recurrent models such as
LSTMs and to noise-sensitive applications such as deep reinforcement learning
or generative models, for which batch normalization is less well suited. Although
our method is much simpler, it still provides much of the speed-up of full batch
normalization. In addition, the computational overhead of our method is lower,
permitting more optimization steps to be taken in the same amount of time. We
demonstrate the usefulness of our method on applications in supervised image
recognition, generative modelling, and deep reinforcement learning.
1
Introduction
Recent successes in deep learning have shown that neural networks trained by first-order gradient
based optimization are capable of achieving amazing results in diverse domains like computer vision,
speech recognition, and language modelling [5]. However, it is also well known that the practical
success of first-order gradient based optimization is highly dependent on the curvature of the objective
that is optimized. If the condition number of the Hessian matrix of the objective at the optimum is
low, the problem is said to exhibit pathological curvature, and first-order gradient descent will have
trouble making progress [18, 28]. The amount of curvature, and thus the success of our optimization,
is not invariant to reparameterization [1]: there may be multiple equivalent ways of parameterizing
the same model, some of which are much easier to optimize than others. Finding good ways of
parameterizing neural networks is thus an important problem in deep learning.
While the architectures of neural networks differ widely across applications, they are typically mostly
composed of conceptually simple computational building blocks sometimes called neurons: each such
neuron computes a weighted sum over its inputs and adds a bias term, followed by the application of
an elementwise nonlinear transformation. Improving the general optimizability of deep networks is a
challenging task [4], but since many neural architectures share these basic building blocks, improving
these building blocks improves the performance of a very wide range of model architectures and
could thus be very useful.
Several authors have recently developed methods to improve the conditioning of the cost gradient for
general neural network architectures. One approach is to explicitly left multiply the cost gradient
with an approximate inverse of the Fisher information matrix, thereby obtaining an approximately
whitened natural gradient. Such an approximate inverse can for example be obtained by using a
Kronecker factored approximation to the Fisher matrix and inverting it (KFAC, [19]), by using an
approximate Cholesky factorization of the inverse Fisher matrix (FANG, [8]), or by whitening the
input of each layer in the neural network (PRONG, [3]).
Alternatively, we can use standard first order gradient descent without preconditioning, but change
the parameterization of our model to give gradients that are more like the whitened natural gradients
of these methods. For example, Raiko et al. [23] propose to transform the outputs of each neuron
to have zero output and zero slope on average. They show that this transformation approximately
diagonalizes the Fisher information matrix, thereby whitening the gradient, and that this leads to
improved optimization performance. Another approach in this direction is batch normalization [11],
a method where the output of each neuron (before application of the nonlinearity) is normalized by
the mean and standard deviation of the outputs calculated over the examples in the minibatch. This
reduces covariate shift of the neuron outputs and the authors suggest it also brings the Fisher matrix
closer to the identity matrix.
Following this second approach to approximate natural gradient optimization, we propose a simple
but general method, called weight normalization, for improving the optimizability of the weights
of neural network models. The method is inspired by batch normalization, but it is a deterministic
method that does not share batch normalization’s property of adding noise to the gradients. In
addition, the overhead imposed by our method is lower: no additional memory is required and the
additional computation is negligible. The method show encouraging results on a wide range of deep
learning applications.
2
Weight Normalization
We consider standard artificial neural networks where the computation of each neuron consists in
taking a weighted sum of input features, followed by an elementwise nonlinearity:
y = φ(w · x + b),
(1)
where w is a k-dimensional weight vector, b is a scalar bias term, x is a k-dimensional vector of input
features, φ(.) denotes an elementwise nonlinearity such as the rectifier max(., 0), and y denotes the
scalar output of the neuron.
After associating a loss function to one or more neuron outputs, such a neural network is commonly
trained by stochastic gradient descent in the parameters w, b of each neuron. In an effort to speed up
the convergence of this optimization procedure, we propose to reparameterize each weight vector w
in terms of a parameter vector v and a scalar parameter g and to perform stochastic gradient descent
with respect to those parameters instead. We do so by expressing the weight vectors in terms of the
new parameters using
g
w=
v
(2)
||v||
where v is a k-dimensional vector, g is a scalar, and ||v|| denotes the Euclidean norm of v. This
reparameterization has the effect of fixing the Euclidean norm of the weight vector w: we now
have ||w|| = g, independent of the parameters v. We therefore call this reparameterizaton weight
normalization.
The idea of normalizing the weight vector has been proposed before (e.g. [27]) but earlier work
typically still performed optimization in the w-parameterization, only applying the normalization
after each step of stochastic gradient descent. This is fundamentally different from our approach: we
propose to explicitly reparameterize the model and to perform stochastic gradient descent in the new
parameters v, g directly. Doing so improves the conditioning of the gradient and leads to improved
convergence of the optimization procedure: By decoupling the norm of the weight vector (g) from the
direction of the weight vector (v/||v||), we speed up convergence of our stochastic gradient descent
optimization, as we show experimentally in section 5.
Instead of working with g directly, we may also use an exponential parameterization for the scale, i.e.
g = es , where s is a log-scale parameter to learn by stochastic gradient descent. Parameterizing the g
parameter in the log-scale is more intuitive and more easily allows g to span a wide range of different
magnitudes. Empirically, however, we did not find this to be an advantage. In our experiments,
the eventual test-set performance was not significantly better or worse than the results with directly
learning g in its original parameterization, and optimization was slightly slower.
2
2.1
Gradients
Training a neural network in the new parameterization is done using standard stochastic gradient
descent methods. Here we differentiate through (2) to obtain the gradient of a loss function L with
respect to the new parameters v, g. Doing so gives
∇w L · v
g
g∇g L
∇g L =
v,
(3)
,
∇v L =
∇w L −
||v||
||v||
||v||2
where ∇w L is the gradient with respect to the weights w as used normally.
Backpropagation using weight normalization thus only requires a minor modification to the usual
backpropagation equations, and is easily implemented using standard neural network software. We
provide reference implementations for Theano at https://github.com/TimSalimans/weight_
norm. Unlike with batch normalization, the expressions above are independent of the minibatch size
and thus cause only minimal computational overhead.
An alternative way to write the gradient is
g
ww0
∇v L =
,
(4)
Mw ∇w L,
with
Mw = I −
||v||
||w||2
where Mw is a projection matrix that projects onto the complement of the w vector. This shows
that weight normalization accomplishes two things: it scales the weight gradient by g/||v||, and it
projects the gradient away from the current weight vector. Both effects help to bring the covariance
matrix of the gradient closer to identity and benefit optimization, as we explain below.
Due to projecting away from w, the norm of v grows monotonically with the number of weight
updates when learning a neural network with weight normalization using standard gradient descent
without momentum: Let v0 = v + ∆v denote our parameter update, with ∆v ∝ ∇v L (steepest
ascent/descent), then ∆v is necessarily orthogonal to the current weight vector w since we project
away from it when calculating ∇v L (equation 4). Since v is proportional to w, the update is thus also
orthogonal to v and increases its norm by the Pythagorean
theorem. Specifically,
if ||∆v||/||v|| = c
p
√
the new weight vector will have norm ||v0 || = ||v||2 + c2 ||v||2 = 1 + c2 ||v|| ≥ ||v||. The rate
of increase will depend on the the variance of the weight gradient. If our gradients are noisy, c will be
high and the norm of v will quickly increase,√which in turn will decrease the scaling factor g/||v||.
If the norm of the gradients is small, we get 1 + c2 ≈ 1, and the norm of v will stop increasing.
Using this mechanism, the scaled gradient self-stabilizes its norm. This property does not strictly
hold for optimizers that use separate learning rates for individual parameters, like Adam [12] which
we use in experiments, or when using momentum. However, qualitatively we still find the same effect
to hold.
Empirically, we find that the ability to grow the norm ||v|| makes optimization of neural networks
with weight normalization very robust to the value of the learning rate: If the learning rate is too
large, the norm of the unnormalized weights grows quickly until an appropriate effective learning rate
is reached. Once the norm of the weights has grown large with respect to the norm of the updates, the
effective learning rate stabilizes. Neural networks with weight normalization therefore work well
with a much wider range of learning rates than when using the normal parameterization. It has been
observed that neural networks with batch normalization also have this property [11], which can also
be explained by this analysis.
By projecting the gradient away from the weight vector w, we also eliminate the noise in that
direction. If the covariance matrix of the gradient with respect to w is given by C, the covariance
matrix of the gradient in v is given by D = (g 2 /||v||2 )Mw CMw . Empirically, we find that w is
often (close to) a dominant eigenvector of the covariance matrix C: removing that eigenvector then
gives a new covariance matrix D that is closer to the identity matrix, which may further speed up
learning.
2.2
Relation to batch normalization
An important source of inspiration for this reparameterization is batch normalization [11], which
normalizes the statistics of the pre-activation t for each minibatch as
t − µ[t]
t0 =
,
σ[t]
3
with µ[t], σ[t] the mean and standard deviation of the pre-activations t = v · x. For the special
case where our network only has a single layer, and the input features x for that layer are whitened
(independently distributed with zero mean and unit variance), these statistics are given by µ[t] = 0
and σ[t] = ||v||. In that case, normalizing the pre-activations using batch normalization is equivalent
to normalizing the weights using weight normalization.
Convolutional neural networks usually have much fewer weights than pre-activations, so normalizing
the weights is often much cheaper computationally. In addition, the norm of v is non-stochastic, while
the minibatch mean µ[t] and variance σ 2 [t] can in general have high variance for small minibatch
size. Weight normalization can thus be viewed as a cheaper and less noisy approximation to batch
normalization. Although exact equivalence does not usually hold for deeper architectures, we still
find that our weight normalization method provides much of the speed-up of full batch normalization.
In addition, its deterministic nature and independence on the minibatch input also means that our
method can be applied more easily to models like RNNs and LSTMs, as well as noise-sensitive
applications like reinforcement learning.
3
Data-Dependent Initialization of Parameters
Besides a reparameterization effect, batch normalization also has the benefit of fixing the scale of the
features generated by each layer of the neural network. This makes the optimization robust against
parameter initializations for which these scales vary across layers. Since weight normalization lacks
this property, we find it is important to properly initialize our parameters. We propose to sample the
elements of v from a simple distribution with a fixed scale, which is in our experiments a normal
distribution with mean zero and standard deviation 0.05. Before starting training, we then initialize
the b and g parameters to fix the minibatch statistics of all pre-activations in our network, just like
in batch normalization, but only for a single minibatch of data and only during initialization. This
can be done efficiently by performing an initial feedforward pass through our network for a single
minibatch of data X, using the following computation at each neuron:
v·x
t − µ[t]
t=
,
and y = φ
,
(5)
||v||
σ[t]
where µ[t] and σ[t] are the mean and standard deviation of the pre-activation t over the examples in
the minibatch. We can then initialize the neuron’s biase b and scale g as
g←
1
,
σ[t]
b←
−µ[t]
,
σ[t]
(6)
so that y = φ(w · x + b). Like batch normalization, this method ensures that all features initially have
zero mean and unit variance before application of the nonlinearity. With our method this only holds
for the minibatch we use for initialization, and subsequent minibatches may have slightly different
statistics, but experimentally we find this initialization method to work well. The method can also be
applied to networks without weight normalization, simply by doing stochastic gradient optimization
on the parameters w directly, after initialization in terms of v and g: this is what we compare to in
section 5. Independently from our work, this type of initialization was recently proposed by different
authors [20, 14] who found such data-based initialization to work well for use with the standard
parameterization in terms of w.
The downside of this initialization method is that it can only be applied in similar cases as where
batch normalization is applicable. For models with recursion, such as RNNs and LSTMs, we will
have to resort to standard initialization methods.
4
Mean-only Batch Normalization
Weight normalization, as introduced in section 2, makes the scale of neuron activations approximately
independent of the parameters v. Unlike with batch normalization, however, the means of the neuron
activations still depend on v. We therefore also explore the idea of combining weight normalization
with a special version of batch normalization, which we call mean-only batch normalization: With
this normalization method, we subtract out the minibatch means like with full batch normalization,
4
but we do not divide by the minibatch standard deviations. That is, we compute neuron activations
using
t = w · x,
t̃ = t − µ[t] + b,
y = φ(t̃)
(7)
where w is the weight vector, parameterized using weight normalization, and µ[t] is the minibatch
mean of the pre-activation t. During training, we keep a running average of the minibatch mean
which we substitute in for µ[t] at test time.
The gradient of the loss with respect to the pre-activation t is calculated as
∇t L = ∇t̃ L − µ[∇t̃ L],
(8)
where µ[.] denotes once again the operation of taking the minibatch mean. Mean-only batch normalization thus has the effect of centering the gradients that are backpropagated. This is a comparatively
cheap operation, and the computational overhead of mean-only batch normalization is thus lower
than for full batch normalization. In addition, this method causes less noise during training, and
the noise that is caused is more gentle as the law of large numbers ensures that µ[t] and µ[∇t̃ ] are
approximately normally distributed. Thus, the added noise has much lighter tails than the highly
kurtotic noise caused by the minibatch estimate of the variance used in full batch normalization. As
we show in section 5.1, this leads to improved accuracy at test time.
5
Experiments
We experimentally validate the usefulness of our method using four different models for varied
applications in supervised image recognition, generative modelling, and deep reinforcement learning.
5.1
Supervised Classification: CIFAR-10
To test our reparameterization method for the application of supervised classification, we consider the
CIFAR-10 data set of natural images [15]. The model we are using is based on the ConvPool-CNN-C
architecture of [26], with some small modifications: we replace the first dropout layer by a layer
that adds Gaussian noise, we expand the last hidden layer from 10 units to 192 units, and we use
2 × 2 max-pooling, rather than 3 × 3. The only hyperparameter that we actively optimized (the
standard deviation of the Gaussian noise) was chosen to maximize the performance of the network
on a holdout set of 10000 examples, using the standard parameterization (no weight normalization
or batch normalization). A full description of the resulting architecture is given in table A in the
supplementary material.
We train our network for CIFAR-10 using Adam [12] for 200 epochs, with a fixed learning rate and
momentum of 0.9 for the first 100 epochs. For the last 100 epochs we set the momentum to 0.5 and
linearly decay the learning rate to zero. We use a minibatch size of 100. We evaluate 5 different
parameterizations of the network: 1) the standard parameterization, 2) using batch normalization,
3) using weight normalization, 4) using weight normalization combined with mean-only batch normalization, 5) using mean-only batch normalization with the normal parameterization. The network
parameters are initialized using the scheme of section 3 such that all four cases have identical parameters starting out. For each case we pick the optimal learning rate in {0.0003, 0.001, 0.003, 0.01}.
The resulting error curves during training can be found in figure 1: both weight normalization
and batch normalization provide a significant speed-up over the standard parameterization. Batch
normalization makes slightly more progress per epoch than weight normalization early on, although
this is partly offset by the higher computational cost: with our implementation, training with batch
normalization was about 16% slower compared to the standard parameterization. In contrast, weight
normalization was not noticeably slower. During the later stage of training, weight normalization
and batch normalization seem to optimize at about the same speed, with the normal parameterization
(with or without mean-only batch normalization) still lagging behind.
After optimizing the network for 200 epochs using the different parameterizations, we evaluate their
performance on the CIFAR-10 test set. The results are summarized in table 2: weight normalization,
the normal parameterization, and mean-only batch normalization have similar test accuracy (≈ 8.5%
error). Batch normalization does significantly better at 8.05% error. Mean-only batch normalization
combined with weight normalization has the best performance at 7.31% test error, and interestingly
does much better than mean-only batch normalization combined with the normal parameterization:
This suggests that the noise added by batch normalization can be useful for regularizing the network,
5
Model
Test Error
Maxout [6]
11.68%
0.1
Network in Network [17]
10.41%
Deeply Supervised [16]
9.6%
ConvPool-CNN-C [26]
9.31%
0.05
ALL-CNN-C [26]
9.08%
our CNN, mean-only B.N.
8.52%
0
our CNN, weight norm.
8.46%
0
50
100
150
200
training epochs
our CNN, normal param.
8.43%
our CNN, batch norm.
8.05%
Figure 1: Training error for CIFAR-10 using differ- ours, W.N. + mean-only B.N. 7.31%
ent network parameterizations. For weight normal- Figure 2: Classification results on CIFAR-10
ization, batch normalization, and mean-only batch without data augmentation.
normalization we show results using Adam with a
learning rate of 0.003. For the normal parameterization we instead use 0.0003 which works best in
this case. For the last 100 epochs the learning rate
is linearly decayed to zero.
training error
normal param.
weight norm.
batch norm.
WN + mean−only BN
mean−only BN
but that the reparameterization provided by weight normalization or full batch normalization is also
needed for optimal results. We hypothesize that the substantial improvement by mean-only B.N. with
weight normalization over regular batch normalization is due to the distribution of the noise caused
by the normalization method during training: for mean-only batch normalization the minibatch mean
has a distribution that is approximately Gaussian, while the noise added by full batch normalization
during training has much higher kurtosis. As far as we are aware, the result with mean-only batch
normalization combined with weight normalization represents the state-of-the-art for CIFAR-10
among methods that do not use data augmentation.
5.2
Generative Modelling: Convolutional VAE
Next, we test the effect of weight normalization applied to deep convolutional variational autoencoders (CVAEs) [13, 24, 25], trained on the MNIST data set of images of handwritten digits and
the CIFAR-10 data set of small natural images.
Variational auto-encoders are generative models that explain the data vector x as arising from a set of
latent variables z, through a joint distribution of the form p(z, x) = p(z)p(x|z), where the decoder
p(x|z) is specified using a neural network. A lower bound on the log marginal likelihood log p(x)
can be obtained by approximately inferring the latent variables z from the observed data x using
an encoder distribution q(z|x) that is also specified as a neural network. This lower bound is then
optimized to fit the model to the data.
We follow a similar implementation of the CVAE as in [25] with some modifications, mainly that the
encoder and decoder are parameterized with ResNet [9] blocks, and that the diagonal posterior is
replaced with auto-regressive variational inference1 . For MNIST, the encoder consists of 3 sequences
of two ResNet blocks each, the first sequence acting on 16 feature maps, the others on 32 feature
maps. The first two sequences are followed by a 2-times subsampling operation implemented using
2 × 2 stride, while the third sequence is followed by a fully connected layer with 450 units. The
decoder has a similar architecture, but with reversed direction. For CIFAR-10, we used a neural
architecture with ResNet units and multiple intermediate stochastic layers1 . We used Adamax [12]
with α = 0.002 for optimization, in combination with Polyak averaging [22] in the form of an
exponential moving average that averages parameters over approximately 10 epochs.
In figure 3, we plot the test-set lower bound as a function of number of training epochs, including
error bars based on multiple different random seeds for initializing parameters. As can be seen, the
parameterization with weight normalization has lower variance and converges to a better optimum.
We observe similar results across different hyper-parameter settings.
1
Manuscript in preparation
6
Convolutional VAE on MNIST
84.0
Convolutional VAE on CIFAR-10
10000
85.0
bound on marginal likelihood
bound on marginal likelihood
84.5
85.5
86.0
86.5
87.0
normal parameterization
weight normalization
87.5
88.0
50
100
150
200
training epochs
250
9500
9000
8500
normal parameterization
weight normalization
8000
0
300
50
100
150
200 250 300
training epochs
350
400
450
Figure 3: Marginal log likelihood lower bound on the MNIST (top) and CIFAR-10 (bottom) test
sets for a convolutional VAE during training, for both the standard implementation as well as our
modification with weight normalization. For MNIST, we provide standard error bars to indicate
variance based on different initial random seeds.
5.3
Generative Modelling: DRAW
Next, we consider DRAW, a recurrent generative model by [7]. DRAW is a variational auto-encoder
with generative model p(z)p(x|z) and encoder q(z|x), similar to the model in section 5.2, but with
both the encoder and decoder consisting of a recurrent neural network comprised of Long Short-Term
Memory (LSTM) [10] units. LSTM units consist of a memory cell with additive dynamics, combined
with input, forget, and output gates that determine which information flows in and out of the memory.
The additive dynamics enables learning of long-range dependencies in the data.
At each time step of the model, DRAW uses the same set of weight vectors to update the cell states
of the LSTM units in its encoder and decoder. Because of the recurrent nature of this process it is not
clear how batch normalization could be applied to this model: Normalizing the cell states diminishes
their ability to pass through information. Fortunately, weight normalization can be applied trivially to
the weight vectors of each LSTM unit, and we find this to work well empirically.
We take the Theano implementation of DRAW provided at https://github.com/jbornschein/
draw and use it to model the MNIST data set of handwritten digits. We then make a single modification to the model: we apply weight normalization to all weight vectors. As can be seen in figure 4,
this significantly speeds up convergence of the optimization procedure, even without modifying the
initialization method and learning rate that were tuned for use with the normal parameterization.
bound on marginal log likelihood
−80
−85
−90
−95
−100
−105
−110
normal parameterization
weight normalization
−115
−120
0
10
20
30
40
50
60
70
80
90
100
training epochs
Figure 4: Marginal log likelihood lower bound on the MNIST test set for DRAW during training, for
both the standard implementation as well as our modification with weight normalization. 100 epochs
is not sufficient for convergence for this model, but the implementation using weight normalization
clearly makes progress much more quickly than with the standard parameterization.
7
5.4
Reinforcement Learning: DQN
Next we apply weight normalization to the problem of Reinforcement Learning for playing games on
the Atari Learning Environment [2]. The approach we use is the Deep Q-Network (DQN) proposed
by [21]. This is an application for which batch normalization is not well suited: the noise introduced
by estimating the minibatch statistics destabilizes the learning process. We were not able to get
batch normalization to work for DQN without using an impractically large minibatch size. In
contrast, weight normalization is easy to apply in this context, as is the initialization method of
section 3. Stochastic gradient learning is performed using Adamax [12] with momentum of 0.5. We
search for optimal learning rates in {0.0001, 0.0003, 0.001, 0.003}, generally finding 0.0003 to work
well with weight normalization and 0.0001 to work well for the normal parameterization. We also
use a larger minibatch size (64) which we found to be more efficient on our hardware (Amazon
Elastic Compute Cloud g2.2xlarge GPU instance). Apart from these changes we follow [21]
as closely as possible in terms of parameter settings and evaluation methods. However, we use a
Python/Theano/Lasagne reimplementation of their work, adapted from the implementation available
at https://github.com/spragunr/deep_q_rl, so there may be small additional differences in
implementation.
Figure 5 shows the training curves obtained using DQN with the standard parameterization and
with weight normalization on Space Invaders. Using weight normalization the algorithm progresses
more quickly and reaches a better final result. Table 6 shows the final evaluation scores obtained
by DQN with weight normalization for four games: on average weight normalization improves the
performance of DQN.
2500
Game
Breakout
Enduro
Seaquest
Space Invaders
test reward per episode
2000
1500
1000
500
0
0
normal parameterization
weight normalization
50
100
150
200
training epochs
Figure 5: Evaluation scores for Space Invaders obtained by DQN after each epoch of
training, for both the standard parameterization and using weight normalization. Learning rates for both cases were selected to maximize the highest achieved test score.
6
normal
410
1,250
7,188
1,779
weightnorm
403
1,448
7,375
2,179
Mnih
401
302
5,286
1,975
Figure 6: Maximum evaluation scores obtained by
DQN, using either the normal parameterization or
using weight normalization. The scores indicated by
Mnih et al. are those reported by [21]: Our normal
parameterization is approximately equivalent to their
method. Differences in scores may be caused by small
differences in our implementation. Specifically, the
difference in our score on Enduro and that reported
by [21] might be due to us not using a play-time limit
during evaluation.
Conclusion
We have presented weight normalization, a simple reparameterization of the weight vectors in a
neural network that accelerates the convergence of stochastic gradient descent optimization. Weight
normalization was applied to four different models in supervised image recognition, generative
modelling, and deep reinforcement learning, showing a consistent advantage across applications. The
reparameterization method is easy to apply, has low computational overhead, and does not introduce
dependencies between the examples in a minibatch, making it our default choice in the development
of new deep learning architectures.
Acknowledgments
We thank John Schulman for helpful comments on an earlier draft of this paper.
8
References
[1] S. Amari. Neural learning in structured parameter spaces - natural Riemannian gradient. In Advances in
Neural Information Processing Systems, pages 127–133. MIT Press, 1997.
[2] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An evaluation
platform for general agents. Journal of Artificial Intelligence Research, 47:253–279, 06 2013.
[3] G. Desjardins, K. Simonyan, R. Pascanu, et al. Natural neural networks. In Advances in Neural Information
Processing Systems, pages 2062–2070, 2015.
[4] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In
International conference on artificial intelligence and statistics, pages 249–256, 2010.
[5] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. Book in preparation for MIT Press, 2016.
[6] I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Bengio. Maxout networks. In ICML,
2013.
[7] K. Gregor, I. Danihelka, A. Graves, and D. Wierstra. Draw: A recurrent neural network for image
generation. arXiv preprint arXiv:1502.04623, 2015.
[8] R. Grosse and R. Salakhudinov. Scaling up natural gradient by sparsely factorizing the inverse fisher
matrix. In ICML, pages 2304–2313, 2015.
[9] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. arXiv preprint
arXiv:1512.03385, 2015.
[10] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.
[11] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal
covariate shift. In ICML, 2015.
[12] D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
[13] D. P. Kingma and M. Welling. Auto-Encoding Variational Bayes. Proceedings of the 2nd International
Conference on Learning Representations, 2013.
[14] P. Krähenbühl, C. Doersch, J. Donahue, and T. Darrell. Data-dependent initializations of convolutional
neural networks. arXiv preprint arXiv:1511.06856, 2015.
[15] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images, 2009.
[16] C.-Y. Lee, S. Xie, P. Gallagher, Z. Zhang, and Z. Tu. Deeply-supervised nets. In Deep Learning and
Representation Learning Workshop, NIPS, 2014.
[17] M. Lin, C. Qiang, and S. Yan. Network in network. In ICLR: Conference Track, 2014.
[18] J. Martens. Deep learning via hessian-free optimization. In Proceedings of the 27th International
Conference on Machine Learning (ICML-10), pages 735–742, 2010.
[19] J. Martens and R. Grosse. Optimizing neural networks with kronecker-factored approximate curvature.
arXiv preprint arXiv:1503.05671, 2015.
[20] D. Mishkin and J. Matas. All you need is a good init. arXiv preprint arXiv:1511.06422, 2015.
[21] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller,
A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. Nature,
518(7540):529–533, 2015.
[22] B. T. Polyak and A. B. Juditsky. Acceleration of stochastic approximation by averaging. SIAM Journal on
Control and Optimization, 30(4):838–855, 1992.
[23] T. Raiko, H. Valpola, and Y. LeCun. Deep learning made easier by linear transformations in perceptrons.
In International Conference on Artificial Intelligence and Statistics, pages 924–932, 2012.
[24] D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and approximate inference in
deep generative models. In ICML, pages 1278–1286, 2014.
[25] T. Salimans, D. P. Kingma, and M. Welling. Markov chain Monte Carlo and variational inference: Bridging
the gap. In ICML, 2015.
9
[26] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. Striving for simplicity: The all convolutional net. In ICLR Workshop Track, 2015.
[27] N. Srebro and A. Shraibman. Rank, trace-norm and max-norm. In Proceedings of the 18th Annual
Conference on Learning Theory, pages 545—-560, 2005.
[28] I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the importance of initialization and momentum in
deep learning. In ICML, pages 1139–1147, 2013.
10
A
Neural network architecure for CIFAR-10 experiments
Layer type
# channels x, y dimension
raw RGB input
3
32
ZCA whitening
3
32
Gaussian noise σ = 0.15
3
32
3 × 3 conv leaky ReLU
96
32
3 × 3 conv leaky ReLU
96
32
3 × 3 conv leaky ReLU
96
32
2 × 2 max pool, str. 2
96
16
dropout with p = 0.5
96
16
3 × 3 conv leaky ReLU
192
16
3 × 3 conv leaky ReLU
192
16
3 × 3 conv leaky ReLU
192
16
2 × 2 max pool, str. 2
192
8
dropout with p = 0.5
192
8
3 × 3 conv leaky ReLU
192
6
1 × 1 conv leaky ReLU
192
6
1 × 1 conv leaky ReLU
192
6
global average pool
192
1
softmax output
10
1
Table 1: Neural network architecture for CIFAR-10.
11
| 9cs.NE
|
Proper affine actions in non-swinging
representations
arXiv:1605.03833v1 [math.GR] 12 May 2016
Ilia Smilga
May 13, 2016
For a semisimple real Lie group G with an irreducible representation ρ on
a finite-dimensional real vector space V , we give a sufficient criterion on ρ for
existence of a group of affine transformations of V whose linear part is Zariskidense in ρ(G) and that is free, nonabelian and acts properly discontinuously
on V .
1 Introduction
1.1 Background and motivation
The present paper is part of a larger effort to understand discrete groups Γ of affine
transformations (subgroups of the affine group GLn (R) ⋉ Rn ) acting properly discontinuously on the affine space Rn . The case where Γ consists of isometries (in other words,
Γ ⊂ On (R) ⋉ Rn ) is well-understood: a classical theorem by Bieberbach says that such
a group always has an abelian subgroup of finite index.
We say that a group G acts properly discontinuously on a topological space X if for
every compact K ⊂ X, the set {g ∈ G | gK ∩ K 6= ∅} is finite. We define a crystallographic group to be a discrete group Γ ⊂ GLn (R) ⋉ Rn acting properly discontinuously
and such that the quotient space Rn /Γ is compact. In [Aus64], Auslander conjectured
that any crystallographic group is virtually solvable, that is, contains a solvable subgroup
of finite index. Later, Milnor [Mil77] asked whether this statement is actually true for
any affine group acting properly discontinuously. The answer turned out to be negative:
Margulis [Mar83, Mar87] gave a nonabelian free group of affine transformations with linear part Zariski-dense in SO(2, 1), acting properly discontinuously on R3 . On the other
hand, Fried and Goldman [FG83] proved the Auslander conjecture in dimension 3 (the
cases n = 1 and 2 are easy). Recently, Abels, Margulis and Soifer [AMS13] proved it in
dimension n ≤ 6. See [Abe01] for a survey of already known results.
Margulis’s breakthrough was soon followed by the construction of other counterexamples to Milnor’s conjecture. The first advance was made by Abels et al. [AMS02]: they
1
generalized Margulis’s construction to subgroups of the affine group
SO(2n + 2, 2n + 1) ⋉ R4n+3 ,
for all values of n. The author further generalized this in his previous paper [Smi16], by
finding such subgroups in the affine group G ⋉ g, where G is any noncompact semisimple
real Lie group, acting on its Lie algebra g by the adjoint representation. Recently Danciger et al. [DGK] found examples of affine groups acting properly discontinuously that
were neither virtually solvable nor virtually free.
Proliferation of these counterexamples leads to the following question. Consider a
semisimple real Lie group G; for every representation ρ of G on a finite-dimensional
real vector space V , we may consider the affine group G ⋉ V . Which of those affine
groups contain a nonabelian free subgroup with linear part Zariski-dense in G and acting
properly discontinuously on V ?
In this paper, we give a fairly general sufficient condition on the representation ρ for
existence of such subgroups. Before stating this condition, we need to introduce a few
classical notations.
1.2 Basic notations
For the remainder of the paper, we fix a semisimple real Lie group G; let g be its Lie
algebra. Let us introduce a few classical objects related to g and G (defined for instance
in Knapp’s book [Kna96], though our terminology and notation differ slightly from his).
We choose in g:
• a Cartan involution θ. Then we have the corresponding Cartan decomposition
g = k ⊕ q, where we call k the space of fixed points of θ and q the space of fixed
points of −θ. We call K the maximal compact subgroup with Lie algebra k.
• a Cartan subspace a compatible with θ (that is, a maximal abelian subalgebra of g
among those contained in q). We set A := exp a.
• a system Σ+ of positive restricted roots in a∗ . Recall that a restricted root is a
nonzero element α ∈ a∗ such that the restricted root space
gα := {Y ∈ g | ∀X ∈ a, [X, Y ] = α(X)Y }
is nontrivial. They form a root system Σ; a system of positive roots Σ+ is a subset
of Σ contained in a half-space and such that Σ = Σ+ ⊔ −Σ+ .
We call Π be the set of simple restricted roots in Σ+ . We call
a++ := X ∈ a ∀α ∈ Σ+ , α(X) > 0
the (open) dominant Weyl chamber of a corresponding to Σ+ , and
a+ := X ∈ a ∀α ∈ Σ+ , α(X) ≥ 0 = a++
the closed dominant Weyl chamber.
2
Then we call:
• M the centralizer of a in K, m its Lie algebra.
• L the centralizer of a in G, l its Lie algebra. It is clear that l = a ⊕ m, and well
known (see e.g. [Kna96], Proposition 7.82a) that L = M A.
• n+ (resp. n− ) the sum of the restricted root spaces of Σ+ (resp. of −Σ+ ), and
N + := exp(n+ ) and N − := exp(n− ) the corresponding Lie groups.
• p+ := l ⊕ n+ and p− := l ⊕ n− the corresponding minimal parabolic subalgebras,
P + := LN + and P − := LN − the corresponding minimal parabolic subgroups.
• W := NG (A)/ZG (A) the restricted Weyl group.
• w0 the longest element of the Weyl group, that is, the unique element such that
w0 (Σ+ ) = Σ− .
See Examples 2.3 and 2.4 in the author’s previous paper [Smi16] for working through
those definitions in the cases G = PSLn (R) and G = PSO+ (n, 1).
Finally, if ρ is a representation of G on a finite-dimensional real vector space V , we
call:
• the restricted weight space in V corresponding to a form λ ∈ a∗ the space
V λ := {v ∈ V | ∀X ∈ a, X · v = λ(X)v} ;
• a restricted weight of the representation ρ any form λ ∈ a∗ such that the corresponding weight space is nonzero.
Remark 1.1. The reader who is unfamiliar with the theory of noncompact semisimple
real Lie groups may focus on the case where G is split, i.e. its Cartan subspace a is
actually a Cartan subalgebra (just a maximal abelian subalgebra, without any additional
hypotheses). In that case the restricted roots are just roots, the restricted weights are just
weights, and the restricted Weyl group is just the usual Weyl group. Also the algebra m
vanishes and M is a discrete group.
However, the case where G is split does not actually require the full strength of this
paper, in particular because quasi-translations (see Section 4.5) then reduce to ordinary
translations.
1.3 Statement of main result
Let ρ be an irreducible representation of G on a finite-dimensional real vector space V .
Without loss of generality, we may assume that G is connected and acts faithfully. We
may then identify the abstract group G with the linear group ρ(G) ⊂ GL(V ). Let VAff be
the affine space corresponding to V . The group of affine transformations of VAff whose
linear part lies in G may then be written G ⋉ V (where V stands for the group of
translations). Here is the main result of this paper.
3
Main Theorem. Suppose that ρ satisfies the following conditions:
(i) there exists a vector v ∈ V such that:
(a) ∀l ∈ L, l(v) = v, and
(b) w̃0 (v) 6= v, where w̃0 is any representative in G of w0 ∈ NG (A)/ZG (A);
(ii) there exists an element X0 ∈ a such that −w0 (X0 ) = X0 and for every λ nonzero
restricted weight of ρ, we have λ(X0 ) 6= 0.
Then there exists a subgroup Γ in the affine group G ⋉ V whose linear part is Zariskidense in G and that is free, nonabelian and acts properly discontinuously on the affine
space corresponding to V .
(Note that the choice of the representative w̃0 in (i)(b) does not matter, precisely
because by (i)(a) the vector v is fixed by L = ZG (A).)
A few words about the significance of these hypotheses. We call representations satisfying condition (ii) "non-swinging" representations (see Section 3.3 to understand why).
This seems to be merely a technical assumption: the author conjectures that condition (i)
alone is sufficient, and in fact necessary and sufficient.
We shall start working with an arbitrary representation ρ, and gradually make stronger
and stronger hypotheses on it (except for the "no swinging" assumption which is logically
independent), introducing each one when we need it to make the construction work (so
that it is at least partially motivated). Here is the complete list of places where new
assumptions on ρ are introduced:
• Assumption 3.2, which is a necessary condition for (i)(a);
• Assumption 3.8, which is (ii) (the "no swinging" assumption);
• Assumption 4.23, which is (i)(a);
• Assumption 10.1, which is (i).
For each of these, we postpone the discussion of which groups do and which groups do
not satisfy it until the spot where it is introduced in the text. For the moment, let us
just say that the previously-known examples do fall under the scope of this theorem:
Example 1.2.
1. For G = SO+ (2n + 2, 2n + 1), the standard representation (acting on V = R4n+3 )
satisfies these conditions (see Remark 3.9 and Examples 4.22.1.b and 10.2.1 for
details). So Theorem A from [AMS02] is a particular case of this theorem.
2. If the real semisimple Lie group G is noncompact, the adjoint representation satisfies these conditions (see Remark 3.9 and Examples 4.22.3 and 10.2.2 for details).
So the main theorem of [Smi16] is a particular case of this theorem.
4
Remark 1.3. When G is compact, no representation can satisfy these conditions: indeed
in that case L is the whole group G and condition (i)(a) fails. So for us, only noncompact
groups are interesting. This is not surprising: indeed, any compact group acting on a
vector space preserves a positive-definite quadratic form, and so falls under the scope of
Bieberbach’s theorem.
1.4 Strategy of the proof
The proof has a lot in common with the author’s previous paper [Smi16]. The main idea
(which comes back to Margulis’s seminal paper [Mar83]) is to introduce, for some affine
maps g, an invariant that measures the translation part of g along a particular affine
subspace of V . The key part of the argument (just as in [Mar83] and in [Smi16]) is then
to show that, under some conditions, the invariant of the product of two maps is roughly
equal to the sum of their invariants (Proposition 8.1). Here are the two main difficulties
that were not present in [Smi16].
• The first one is that [Smi16] crucially relies on the following fact: if two maps are
R-regular (i.e. the dimension of their centralizer is the lowest possible), in general
position with respect to each other and strongly contracting (when acting on g),
their product is still R-regular. The natural generalization of the notion of an Rregular map that is adapted to an arbitrary representation is that of a "generic"
map, i.e. a map that has as few eigenvalues of modulus 1 (counted with multiplicity)
as possible. Unfortunately, the corresponding statement is then no longer true in
an arbitrary representation. If the representation is "too large", i.e. if it contains
restricted weights that are not multiples of restricted roots (see Example 3.5.2),
there are several different "types" of generic maps, depending on the region where
their Jordan projection (see Definition 2.3) falls.
In order to ensure that the product of two generic maps g and h (that are in general
position and strongly contracting) is still generic, we need to control the Jordan
projection Jd(gh) of the product based on the Jordan projections Jd(g), Jd(h) of
the factors. To do this, we use ideas developed by Benoist in [Ben96, Ben97]: when
g and h are in general position and sufficiently contracting, he show that Jd(gh) is
approximately equal to Jd(g) + Jd(h). So if we restrict all maps to have the same
"type", our argument works.
• Here is where the second difficulty comes: the argument of [Ben96] only works for
maps that are actually R-regular in addition of being generic. In most representations this is automatically true: if every restricted root occurs as a restricted
weight, then every generic map is in particular R-regular. But when the representation is "too small", this is not the case. (A surprising fact is that a handful
of representations are actually "too large" and "too small" at the same time: see
Example 3.5.4!)
As an example, consider the subgroup G of GL5 (R) consisting of transformations
5
preserving the quadratic form
x1 x3 + x2 x4 + x25 .
This is a form of signature (3, 2), so G ≃ SO(3, 2). Now take any real number
λ > 1 and any x ∈ R; then the element
λ λx
0
0 0
0 λ
0
0 0
−1
λ
0 0
g=
0 0
∈G
0 0 −λ−1 x λ−1 0
0 0
0
0 1
is generic in the standard representation ("pseudohyperbolic" in the terminology
of [AMS02] and [Smi14]), but not R-regular (when x 6= 0 it is not even semisimple!).
Just as there are two different notions of being "generic" (the notion of R-regularity,
which is adapted to the adjoint representation, and the notion of being generic in ρ),
there are also two different notions of being "in general position", two different
notions of being "strongly contracting" and so on. The results of [Ben96] crucially
rely on the stronger version of every property.
If we had used them as such, our Proposition 6.14 about products of maps "of given
type" (and the subsequent propositions that rely on it) would no longer include, as a
particular case, the corresponding result for G = SO+ (n+1, n) (namely Lemma 5.6,
point (1) in [AMS02]). Instead, we would need to duplicate all definitions, and to
always require that the maps we deal with satisfy both versions of the constraints.
(In particular, we would probably lose the benefit of the unified treatment of the
linear part and translation part, as outlined in Remark 5.3).
This weaker version is in theory sufficient for us, because it is known that "almost
all" elements are R-regular. So it is actually possible to construct the group Γ in
such a way that its elements all have this additional property, and thus provide a
working proof of the Main Theorem. But we felt that the simpler, stronger version
of Proposition 6.14 was interesting in its own right.
To prove it, we needed a generalization of the results of [Ben96]. Benoist’s subsequent paper [Ben97] does seem to provide such a generalization, by proving similar theorems with the hypothesis of R-regularity replaced by what he calls "θproximality", for θ some subset of Π. This is quite close to what we are looking
for; but unfortunately, the results of [Ben97] rely on the assumption that the Jordan projections of the maps lie in a vector subspace of a (see Remark 6.13 for
details), which is unacceptably restrictive for us. So in Section 6 of this paper, we
redeveloped this theory from scratch, in a suitably general way.
1.5 Plan of the paper
In Section 2, we give some background from representation theory.
6
In Section 3, we study the dynamics of elements of A. We choose one particular
element X0 ∈ a with some nice properties, with the goal of eventually "modelling", in
some sense, generators of the group Γ on exp(X0 ).
In Section 4, we study the dynamics of elements g of the affine group G ⋉ V that are of
type X0 (see Definition 4.14). This section culminates in the definition of the Margulis
invariant of g, which measures the translation part of g along its "axis".
In Section 5, we study some quantitative properties of such elements g. In particular we
define a quantitative measure of being "in general position", and a quantitative measure
of being "strongly contracting"; both of these notions are tailored to the choice of ρ
and of X0 . We also define analogous notions for proximal maps, and prove a theorem
(Proposition 5.12) about products of proximal maps.
Section 6 is where most of the new ideas of this paper are exploited. Here we apply the
theory of products of proximal maps to a selection of "fundamental representations" ρi
(defined in Proposition 2.12). The goal is to show that the product of two strongly
contracting maps of type X0 in general position is still of type X0 .
In Section 7, we now apply the theory of products of proximal maps to suitable exterior
powers of the maps g, in order to study the quantitative properties of products of elements
of type X0 . This section follows Section 3.2 of [Smi16] very closely.
Section 8 contains the key part of the proof. We prove that if we take two strongly
contracting maps of type X0 in general position, the Margulis invariant of their product
is close to the sum of their Margulis invariants. This section follows Section 4 of [Smi16]
very closely.
The very short Section 9 uses induction to extend the results of the two previous
sections to products of an arbitrary number of elements. We omit the proof, as it is a
straightforward generalization of Section 5 in [Smi16].
Section 10 contains the proof of the Main Theorem. It follows Section 6 of [Smi16]
quite closely, but there are a couple of additions.
1.6 Acknowledgements
I am very grateful to my Ph.D. advisor, Yves Benoist, whose help while I was working
on this paper has been invaluable to me.
I would also like to thank Bruno Le Floch for some interesting discussions, which in
particular helped me gain some insight about weights of representations.
2 Algebraic preliminaries
In this section, we give some background about real finite-dimensional representations of
semisimple real Lie groups.
In Subsection 2.1, for any element g ∈ G, we relate the eigenvalues and singular values
of ρ(g) (where ρ is some representation) to some "absolute" properties of g.
In Subsection 2.2, we enumerate some properties of restricted weights of a real finitedimensional representation of a real semisimple Lie group.
7
2.1 Eigenvalues in different representations
The goal of this subsection is to prove Proposition 2.6, which expresses the eigenvalues
and singular values of a given element g ∈ G acting in a given representation ρ, exclusively in terms of the structure of g in the abstract group G (respectively its Jordan
decomposition and its Cartan decomposition).
Proposition 2.1 (Jordan decomposition). Let g ∈ G. There exists a unique decomposition of g as a product g = gh ge gu , where:
• gh is conjugate in G to an element of A ( hyperbolic);
• ge is conjugate in G to an element of K ( elliptic);
• gu is conjugate in G to an element of N + ( unipotent);
• these three maps commute with each other.
Proof. This is well-known, and given for example in [Ebe96], Theorem 2.19.24. Note
however that the latter theorem uses representation-dependent definitions of a hyperbolic,
elliptic or unipotent element (applied to the case of the adjoint representation). To state
the theorem with the representation-agnostic definitions that we used, we need to apply
Theorems 2.19.18 and 2.19.16 from the same book.
Proposition 2.2 (Cartan decomposition). Let g ∈ G. Then there exists a decomposition
of g as a product g = k1 ak2 , with k1 , k2 ∈ K and a = exp(X) with X ∈ a+ . Moreover,
the element X is uniquely determined by g.
Proof. This is a classical result; see e.g. Theorem 7.39 in [Kna96].
Definition 2.3. For every element g ∈ G, we define:
• the Jordan projection of g, written Jd(g), to be the unique element of the closed
dominant Weyl chamber a+ such that the hyperbolic part gh (from the Jordan
decomposition g = gh ge gu given above) is conjugate to exp(Jd(g));
• the Cartan projection of g, written Ct(g), to be the element X from the Cartan
decomposition given above.
To talk about singular values, we need to introduce a Euclidean structure. We are
going to use a special one.
Lemma 2.4. Let ρ∗ be some real representation of G on some space V∗ . There exists a
K-invariant positive-definite quadratic form B∗ on V∗ such that all the restricted weight
spaces are pairwise B∗ -orthogonal.
We want to reserve the plain notation ρ for the "default" representation, to be fixed
once and for all at the beginning of Section 3. We use the notation ρ∗ so as to encompass
both this representation ρ and the representations ρi defined in Proposition 2.12.
Such quadratic forms have been considered previously: see for example Lemma 5.33.a)
in [BQ].
8
Example 2.5. If ρ∗ = Ad is the adjoint representation, then B∗ is the form Bθ given by
∀X, Y ∈ g,
Bθ (X, Y ) = −B(X, θY ),
where B is the Killing form and θ is the Cartan involution (see (6.13) in [Kna96]).
Proof. First note that there is a unique representation ρC
∗ (the complexified representation) such that the diagram
g
ρ∗
gl(V∗ )
(2.1)
gC
ρC
∗
gl(V∗C )
commutes (where the vertical arrows represent the canonical inclusions). Let u = k ⊕ iq
be the compact real form of gC . Then the subgroup of GL(V∗C ) with Lie algebra ρC
∗ (u)
C
is compact; by averaging over this subgroup, we may find on V∗ a u-invariant positivedefinite Hermitian form B∗C . By construction, we then have:
(
C
∀X ∈ k, ρC
∗ (X) is B∗ -anti-self-adjoint;
C
∀X ∈ q, ρC
∗ (iX) is B∗ -anti-self-adjoint.
Now we define on V∗ a bilinear form B∗ by
∀x, y ∈ V∗ ,
B∗ (x, y) := Re B∗C (x ⊗ 1, y ⊗ 1) .
(2.2)
It is straightforward to check that B∗ is still positive definite, and that we have:
(
∀X ∈ k, ρ∗ (X) is B∗ -anti-self-adjoint;
∀X ∈ q, ρ∗ (X) is B∗ -self-adjoint.
The first property implies that B∗ is k-invariant, hence (since K is connected) Kinvariant. The second property implies that any element of q has pairwise B∗ -orthogonal
eigenspaces in V∗ . Applying this to any X ∈ a ⊂ q such that all restricted weights of ρ∗
take pairwise distinct values on X, we conclude that all the restricted weight spaces of ρ∗
are pairwise B∗ -orthogonal.
Recall that the singular values of a map g in a Euclidean space are defined as the square
roots of the eigenvalues of g∗ g, where g∗ is the adjoint map. The largest and smallest
singular values of g then give respectively the operator norm of g and the reciprocal of
the operator norm of g−1 .
Proposition 2.6. Let ρ∗ : G → GL(V∗ ) be any representation of G on some vector
space V∗ ; let λ1∗ , . . . , λd∗∗ be the list of all the restricted weights of ρ∗ , repeated according
to their multiplicities. Let g ∈ G; then:
9
(i) The list of the moduli of the eigenvalues of ρ∗ (g) is given by
i
eλ∗ (Jd(g))
.
1≤i≤d∗
(ii) The list of the singular values of ρ∗ (g), with respect to a K-invariant Euclidean
norm B∗ on V∗ that makes the restricted weight spaces of V∗ pairwise orthogonal
(such a norm exists by Lemma 2.4), is given by
i
eλ∗ (Ct(g))
.
1≤i≤d∗
Proof.
(i) Let g = gh ge gu be the Jordan decomposition of g.
We begin with the remark that:
• ρ∗ (ge ) is conjugate to a map that is orthogonal with respect to the norm B∗
constructed in Lemma 2.4;
• for any X ∈ n+ , the linear map ρ∗ (X) is nilpotent (this essentially follows from
the identity ρ∗ (gα ) · V∗λ ⊂ V∗α+λ for any restricted root α and any restricted
root λ of ρ∗ ); hence by exponentiating, for any n ∈ N + , the linear map ρ∗ (n)
is unipotent; hence also ρ∗ (gu ) is unipotent.
If follows that ρ∗ (ge ) and ρ∗ (gu ) both have eigenvalues of modulus 1. Since gh , ge
and gu all commute with each other, we deduce that the eigenvalues of ρ∗ (g) are
equal, in modulus, to those of ρ∗ (gh ).
On the other hand, gh is by definition conjugate to exp(Jd(g)), so ρ∗ (gh ) has the
same eigenvalues as ρ∗ (exp(Jd(g)).
Finally, since exp(Jd(g)) is in A (the group corresponding to the Cartan subspace),
the list of the eigenvalues of ρ∗ (exp(Jd(g))) is by definition given by
i
eλ∗ (Jd(g))
.
1≤i≤d∗
(ii) Let g = k1 exp(Ct(g))k2 be the Cartan decomposition of g. Since ρ∗ (k1 ) and ρ∗ (k2 )
are B∗ -orthogonal maps, the B∗ -singular values of ρ∗ (g) coincide with those of the
map exp(Ct(g)); since exp(Ct(g)), being an element of A, is self-adjoint, its singular
values coincide with its eigenvalues. We conclude as in the previous point.
2.2 Properties of restricted weights
In this subsection, we introduce a few properties of restricted weights of real finitedimensional representations. (Proposition 2.7 is actually a general result about Coxeter
groups.) The corresponding theory for ordinary weights is well-known: see for example
Chapter V in [Kna96].
10
Let α1 , . . . , αr be an enumeration of the set Π of simple restricted roots generating Σ+ .
For every i, we set
(
2αi if 2αi is a restricted root
′
αi :=
(2.3)
αi
otherwise.
For every index i such that 1 ≤ i ≤ r, we define the i-th fundamental restricted weight ̟i
by the relationship
h̟i , α′j i
2
= δij
(2.4)
kα′j k2
for every j such that 1 ≤ j ≤ r.
By abuse of notation, we will often allow ourselves to write things such as "for all i in
some subset Π′ ⊂ Π, ̟i satisfies..." (tacitly identifying the set Π′ with the set of indices
of the simple restricted roots that are inside).
In the following proposition, for any subset Π′ ⊂ Π, we denote:
• by WΠ′ the Weyl subgroup of type Π′ :
WΠ′ := hsα iα∈Π′ ;
• by a+
Π′ the fundamental domain for the action of WΠ′ on a:
′
a+
Π′ := X ∈ a ∀α ∈ Π , α(X) ≥ 0 ,
(2.5)
(2.6)
which is a kind of prism whose base is the dominant Weyl chamber of WΠ′ .
Proposition 2.7. Take any Π′ ⊂ Π, and let us fix X ∈ a+
Π′ . Let Y ∈ a. Then the
following two conditions are equivalent:
(i) the vector Y is in a+
Π′ and satisfies the system of linear inequalities
(
∀i ∈ Π′ ,
̟i (Y ) ≤ ̟i (X),
′
∀i ∈ Π \ Π , ̟i (Y ) = ̟i (X);
(ii) the vector Y is in a+
Π′ and also in the convex hull of the orbit of X by WΠ′ .
Proof. For Π′ = Π, this is well known: see e.g. [Hal15], Proposition 8.44.
Now let Π′ be an arbitrary subset of Π. We may translate everything by the vector
X
̟i (X)Hi
i∈Π\Π′
(where (Hi )i∈Π is a basis of a dual to the basis (̟i )i∈Π of a∗ ), which is obviously fixed
by WΠ′ . Thus we reduce the problem to the case where
∀i ∈ Π \ Π′ ,
̟i (X) = 0.
(2.7)
Now let Σ′ be the intersection of Σ with the vector space aΠ′ determined by this system
of equations, which is also the linear span of (αi )i∈Π′ . Then Σ′ is a root system that has:
11
• Π′ as a simple root system;
• WΠ′ as the Weyl group;
• a+
Π′ ∩ aΠ′ as the dominant Weyl chamber.
This reduces the problem to the case Π′ = Π.
Proposition 2.8. Every restricted weight of every representation of g is a linear combination of fundamental restricted weights with integer coefficients.
Proof. This is a particular case of Proposition 5.8 in [BT65]. For a correction concerning
the proof, see also Remark 5.2 in [BT72].
Proposition 2.9. If ρ∗ is an irreducible representation of g, there is a unique restricted
weight λ∗ of ρ∗ , called its highest restricted weight, such that no element of the form
λ∗ + αi with 1 ≤ i ≤ r is a restricted weight of ρ∗ .
Remark 2.10. In contrast to the situation with non-restricted weights, the highest restricted weight is not always of multiplicity 1; nor is a representation uniquely determined
by its highest restricted weight.
Proof. Let h be a Cartan subalgebra of g containing a (i.e. h is just a maximal abelian
subalgebra of g, without the requirement of being contained in q). Let ∆ be the set
of (non restricted) roots in h∗ ; then restricted roots are just restrictions of roots to a.
(Similarly, restricted weights are just restrictions of weights.) Let ∆+ be the subset of ∆
formed by roots whose restrictions are positive restricted roots (i.e. elements of Σ+ ); then
∆+ forms a system of positive roots.
It is well known (see e.g. [Kna96], Theorem 5.5 (d)) that ρ∗ has a unique (non restricted) weight with respect to h that is the highest with respect to ∆+ . Let λ∗ be
its restriction to a; then the required property for the highest restricted weight follows
from the corresponding property for the highest non restricted weight, and the fact that
restrictions of positive roots are positive restricted roots.
Proposition 2.11. Let ρ∗ be an irreducible representation of g; let λ∗ be its highest
restricted weight. Let Λλ∗ be the restricted root lattice shifted by λ∗ :
Λλ∗ := {λ∗ + c1 α1 + · · · + cr αr | c1 , . . . , cr ∈ Z} .
Then the set of restricted weights of ρ∗ is exactly the intersection of the lattice Λλ∗ with
the convex hull of the orbit {w(λ∗ ) | w ∈ W } of λ∗ by the restricted Weyl group.
Proof. Once again, this follows from the corresponding result for non restricted weights
(see e.g. [Hal15], Theorem 10.1) by passing to the restriction. In the case of restricted
weights, one of the inclusions is stated in [Hel08], Proposition 4.22.
Proposition 2.12. For every index i such that 1 ≤ i ≤ r, there exists an irreducible
representation ρi of G on a space Vi whose highest restricted weight is equal to ni ̟i (for
some positive integer ni ) and has multiplicity 1.
12
Proof. This follows from the general Theorem 7.2 in [Tit71].
Lemma 2.13. Fix an index i such that 1 ≤ i ≤ r. Then all restricted weights of ρi other
than ni ̟i have the form
r
X
cj αj ,
ni ̟i − αi −
j=1
with cj ≥ 0 for every j.
Proof. Let λ be some restricted weight of ρi . By Proposition 2.11 taken together with
Proposition 2.7, we already know that it can be written as
λ = ni ̟i −
r
X
c′j αj ,
j=1
where all coefficients c′j are nonnegative integers. It remains to show that if λ 6= ni ̟i ,
then necessarily c′i > 0.
Assume that c′i = 0. By Proposition 8.42 in [Hal15], we lose no generality in assuming
that λ is dominant. Let Πi := Π \ {i}; by Proposition 2.7, it follows that λ is then in the
convex hull of the orbit of ni ̟i by WΠi . But clearly WΠi fixes ̟i , hence also ni ̟i . The
conclusion follows.
3 Choice of a reference Jordan projection
For the remainder of the paper, we fix ρ an irreducible representation of G on a finitedimensional real vector space V . For the moment, ρ may be any representation; but in
the course of the paper, we shall gradually introduce several assumptions on ρ (namely
Assumptions 3.2, 3.8, 4.23 and 10.1) that will ensure that ρ satisfies the hypotheses of
the Main Theorem.
<
We call Ω the set of restricted weights of ρ. For any X ∈ a, we call Ω>
X (resp. ΩX ,
≥
≤
=
ΩX , ΩX , ΩX ) the set of all restricted weights of ρ that take a positive (resp. negative,
zero, nonnegative, nonpositive) value on X:
Ω≥
X := {λ ∈ Ω | λ(X) ≥ 0}
Ω>
X := {λ ∈ Ω | λ(X) > 0}
Ω≤
X := {λ ∈ Ω | λ(X) ≤ 0}
Ω<
X := {λ ∈ Ω | λ(X) < 0}
Ω=
X := {λ ∈ Ω | λ(X) = 0} .
The goal of this section is to study these sets, and to choose a vector X0 ∈ a+ for
which the corresponding sets have some nice properties. The motivation for their study
is that they parametrize the dynamical spaces (defined in Subsection 4.3) of exp(X0 )
(obviously), and actually of any element g ∈ G whose Jordan projection "has the same
type" as X0 (see Proposition 4.16).
In Subsection 3.1, we introduce the notion of a generic vector X ∈ a, and impose a
first constraint on ρ: that 0 be a restricted weight.
13
In Subsection 3.2, we introduce an equivalence relation on the set of generic vectors
that identifies elements with the same dynamics, and give several examples.
In Subsection 3.3, we introduce the notion of a symmetric vector X ∈ a, and ensure
that ρ does not exclude generic vectors from being symmetric.
In Subsection 3.4, we define parabolic subgroups and subalgebras of type X; we also
associate to every X ∈ a+ a set ΠX of simple restricted roots and a subgroup WX of the
restricted Weyl group.
In Subsection 3.5, we prove Proposition 3.17, which shows that every equivalence
class of generic vectors has a representative that has "as much symmetry" as the whole
equivalence class, called an "extreme" representative.
At the end of this section, we shall fix once and for all an extreme, symmetric, generic
vector X0 ∈ a+ , which will serve as a reference Jordan projection (see the definition at
the beginning of Subsection 4.4).
3.1 Generic elements
We say that an element X ∈ a is generic if
Ω=
X ⊂ {0}.
Remark 3.1. This is indeed the generic case: it happens as soon as X avoids a finite
collection of hyperplanes, namely the kernels of all nonzero restricted weights of ρ.
Assumption 3.2. From now on, we assume that 0 is a restricted weight of ρ:
0 ∈ Ω,
or equivalently
dim V 0 > 0.
Remark 3.3. By Proposition 2.11, this is the case if and only if the highest restricted
weight of ρ is a Z-linear combination of restricted roots.
We lose no generality in assuming this, because this assumption is necessary for condition (i)(a) of the Main Theorem (which is also Assumption 4.23, see below) to hold.
Indeed, any nonzero vector fixed by L is in particular fixed by A ⊂ L, which means that
it belongs to the zero restricted weight space.
In that case, for generic X we actually have
Ω=
X = {0}.
3.2 Types of elements of a
For two vectors X, Y ∈ a, we say that Y has the same type as X if
(
>
Ω>
Y = ΩX ;
<
Ω<
Y = ΩX ,
(3.1)
i.e. if every restricted weight takes the same sign on both of them. For generic X and Y ,
this implies that all five spaces Ω≥ , Ω≤ , Ω= , Ω> and Ω< coincide for X and Y , hence
that exp(X) and exp(Y ) have the same dynamical spaces (see Subsection 4.3).
14
This is an equivalence relation, which partitions the set of generic vectors into finitely
many equivalence classes. Some generic X ∈ a being fixed, we call
(
)
(
∀λ ∈ Ω>
,
λ(Y
)
>
0;
X
(3.2)
aρ,X := Y ∈ a
∀λ ∈ Ω<
X , λ(Y ) < 0
its equivalence class in a. If X is dominant, we additionally call
+
a+
ρ,X := aρ,X ∩ a
(3.3)
its equivalence class in the closed dominant Weyl chamber a+ .
Remark 3.4. Every such equivalence class is obviously a convex cone. Also, these equivalence classes actually coincide with connected components of the set of generic vectors.
Example 3.5.
1. If G is any noncompact semisimple real Lie group and ρ = Ad is its adjoint representation (so that V = g):
• A vector X ∈ a is generic if and only if it lies in one of the open Weyl chambers.
In particular a vector X in a+ is generic if and only if it lies in a++ .
• All elements of a++ have the same type; so there is only one generic equivalence
++ .
class in a+ . For any X ∈ a++ , we have aAd,X = a+
Ad,X = a
2. Take G = SO+ (3, 2). The root system is then B2 :
−e1 + e2
e2
e1 + e2
e1
−e1
−e1 − e2
−e2
e1 − e2
As this group is split, the roots are also the restricted roots. Let ρ be the representation with highest weight 2e1 + e2 (in the notations of [Kna96], Appendix C).
This is a representation of dimension 35, whose weights (also restricted weights)
are as follows:
15
(a) ρ of highest weight 2e1 + e2
(b) ρ of highest weight e1
Figure 1: Equivalence classes and Weyl chambers for two different representations of G =
SO+ (3, 2). Dashed lines represent walls of Weyl chambers. Thick gray lines
represent kernels of nonzero weights, which separate the different equivalence
classes. The dominant Weyl chamber a+ is hatched. All equivalence classes
in a that intersect a+ are shaded, with different shades if there are more than
one.
(the number of dots at each node represents multiplicity.) Then every equivalence
class in a is contained in some Weyl chamber: see Figure 1a. The dominant Weyl
chamber a+ is split into two equivalence classes by the line of slope 12 , kernel of the
weight −e1 + 2e2 .
3. Take G = SO+ (3, 2) and ρ the standard representation on V = R5 . Using once
again the notations of [Kna96], Appendix C, its highest weight is e1 and its weights
are ±e1 , ±e2 and 0 (of course all with multiplicity 1). Then (see Figure 1b):
• A vector X ∈ a+ is generic if and only if it avoids the "horizontal" wall of the
dominant Weyl chamber (the one normal to e2 ).
• All such vectors have the same type. So for a generic X ∈ a+ , the equivalence
class a+
ρ,X is the half-open dominant Weyl chamber, with the diagonal wall
included and the horizontal wall excluded.
• The whole equivalence class aρ,X is then an open quadrant of the plane a,
consisting of two half-open Weyl chambers glued back-to-back along their
shared diagonal wall.
4. Suppose that G and ρ are such that the set of its restricted weights of ρ neither
contains nor is contained in the set of multiples of the restricted roots of G. Then
16
both phenomena occur at the same time: equivalence classes in a neither contain
nor are contained in the Weyl chambers.
Examples are not immediate to come up with: the author even mistakenly believed
for some time that no such representations existed. However, here is one such
example:
• Take G = PSp4 (R) (which is a split form). In the notations of [Kna96],
Appendix C, its roots are all the possible expressions of the form ±ej ± ei
or ±2ei .
• Take ρ to be the representation with highest weight e1 + e2 + e3 + e4 . It has:
– the 16 weights of the form ±e1 ± e2 ± e3 ± e4 , with multiplicity 1;
– the 24 weights of the form ±ei ± ej , with multiplicity 1;
– the zero weight with multiplicity 2,
for a total dimension of 42.
The reader may check that there are then three different "types" of generic vectors
in the dominant Weyl chamber a+ , and two of the equivalence classes contain a
slice of the wall normal to 2e4 .
3.3 Swinging
We start this subsection with the following observation: if the Jordan projection of g
is X, then the Jordan projection of g−1 is −w0 (X), where w0 is the "longest element" of
the Weyl group that interchanges positive and negative restricted roots (see Section 1.2).
We would like to ensure that for every element g of the group Γ we are trying to
construct, the element g itself and its inverse g−1 have similar dynamics. To do that, we
would like X and −w0 (X) to be of the same type. Replacing if necessary X and −w0 (X)
by their midpoint, we lose no generality in assuming they are actually equal.
Definition 3.6. We say that an element X ∈ a is symmetric if it is invariant by −w0 :
−w0 (X) = X.
Unfortunately, it is not always possible to find a vector X that is both symmetric and
generic, as shown by the following example:
Example 3.7. Take G = SL3 (R). It is a split form, so its restricted root system is the
17
same as its root system, namely A2 :
e1 − e3
e2 − e3
e1 − e2
e2 − e1
e3 − e2
e3 − e1
For this group, −w0 is the map that exchanges the two simple positive roots e1 − e2
and e2 − e3 (we use the notations of [Kna96], Appendix C); in the picture above, it
corresponds to the reflection about the vertical axis. So a vector X ∈ a+ is symmetric if
and only if it lies on that vertical axis (which bisects the dominant Weyl chamber a+ ).
Now consider the representation ρ of G with highest weight 2e1 − e2 − e3 . Note that
this is three times the first fundamental weight, so ρ is actually the third symmetric
product S 3 R3 of the standard representation. Here are its weights:
e1 − e3
e2 − e3
2e1 − e2 − e3
e1 − e2
−e1 + 2e2 − e3
0
−e2 + e3
−e1 + e2
−e1 + e3
−e1 − e2 + 2e3
We see that any symmetric vector necessarily annihilates the weight −e1 +2e2 −e3 , hence
it cannot be generic.
We call this phenomenon "swinging". Here is the picture to have in mind: when we
apply the involution −w0 to some generic X, the annihilator of X (i.e. the hyperplane
of a∗ consisting of linear forms that vanish on X) "swings" past the weight −e1 +2e2 −e3 ,
thus switching it from the set Ω> to the set Ω< .
From now on, we assume that this issue does not arise:
Assumption 3.8 ("No swinging"). From now on, we assume that ρ is such that there
exists a symmetric generic element of a.
This is precisely condition (ii) from the Main Theorem.
18
Remark 3.9.
• It is well-known that when the restricted root system of G has any type other than
An (with n ≥ 2), D2n+1 or E6 , we actually have w0 = − Id. For those groups, every
vector X ∈ a is symmetric, and so every representation satisfies this condition.
• For the remaining groups, a straightforward linear algebra manipulation shows that
this condition is equivalent to the following: no nonzero restricted weight of ρ must
fall into the linear subspace
{λ ∈ a∗ | w0 λ = λ}
(3.4)
(the "axis of symmetry" of w0 in a∗ ). For example, this is always true for the
adjoint representation (any restricted root fixed by w0 would need to be positive
and negative at the same time). Heuristically, this seems to hold when the highest
restricted weight is "small", but to quickly fail when it gets "large enough".
3.4 Parabolic subgroups and subalgebras
A parabolic subgroup (or subalgebra) is usually defined in terms of a subset Π′ of the set Π
of simple restricted roots. We find it more convenient however to use a slightly different
language. To every to every such subset corresponds a facet of the Weyl chamber, given
by intersecting the walls corresponding to elements of Π′ . We may exemplify this facet
by picking some element X in it that does not belong to any subfacet. Conversely, for
every X ∈ a+ , we define the corresponding subset
ΠX := {α ∈ Π | α(X) = 0} .
(3.5)
The parabolic subalgebras and subgroups of type ΠX can then be very conveniently
rewritten in terms of X, as follows.
Remark 3.10. The set ΠX actually encodes the "type" of X with the respect to the
adjoint representation.
Definition 3.11. For every X ∈ a+ , we define:
−
• p+
X and pX the parabolic subalgebras of type X, and lX their intersection:
p+
X := l ⊕
M
gα ;
M
gα ;
α(X)≥0
p−
X := l ⊕
α(X)≤0
lX := l ⊕
M
α(X)=0
19
gα .
• PX+ and PX− the corresponding parabolic subgroups, and LX their intersection:
PX+ := NG (p+
X );
PX− := NG (p−
X );
LX := PX+ ∩ PX− .
An object closely related to these parabolic subgroups (see formula (4.4), the Bruhat
decomposition for parabolic subgroups) is the stabilizer of X in the Weyl group:
Definition 3.12. For any X ∈ a+ , we set
WX := {w ∈ W | wX = X} .
Remark 3.13. The group WX is also closely related to the set ΠX . Indeed, it follows
immediately that a simple restricted root α belongs to ΠX if and only if the corresponding
reflection sα belongs to WX . Conversely, it is well-known (Chevalley’s lemma, see e.g.
[Kna96], Proposition 2.72) that these reflections actually generate the group WX .
Thus WX is actually the same thing as WΠ′ (as defined before Proposition 2.7) where
we substitute Π′ = ΠX .
Example 3.14. To help understand the conventions we are taking, here are the extreme
cases:
1. If X lies in the open Weyl chamber a++ , then:
• PX+ = P + is the minimal parabolic subgroup; PX− = P − ; LX = L;
• ΠX = ∅;
• WX = {Id}.
2. If X = 0, then:
• PX+ = PX− = LX = G;
• ΠX = Π;
• WX = W .
3.5 Extreme vectors
Besides WX , we are also interested in the group
Wρ,X := {w ∈ W | wX has the same type as X} ,
(3.6)
which is the stabilizer of X "up to type". It obviously contains WX . The goal of this
subsection is to show that in every equivalence class, we can actually choose X in such
a way that both groups coincide.
20
Example 3.15. In Example 3.5.3 (G = SO+ (3, 2) acting on V = R5 ), the group Wρ,X
corresponding to any generic X is a two-element group. If we take X to be generic not
only with respect to ρ but also with respect to the adjoint representation (in other terms
if X is in an open Weyl chamber), then the group WX is trivial. If however we take as X
any element of the diagonal wall of the Weyl chamber, we have indeed WX = Wρ,X .
Definition 3.16. We call an element X ∈ a+ extreme if WX = Wρ,X , i.e. if it satisfies
the following property:
∀w ∈ W,
wX has the same type as X ⇐⇒ wX = X.
Proposition 3.17. For every generic X ∈ a+ , there exists a generic X ′ ∈ a+ that has
the same type as X and that is extreme.
If moreover X is symmetric, then X ′ is still symmetric.
Remark 3.18. The following statement will never be used in the paper (so we leave it
without proof), but might help to understand what is going on: for every generic X, we
have
+
aρ,X = Wρ,X a+
ρ,X = WX ′ aρ,X .
Also, it can be shown that a representative X ′ of a given equivalence class a+
ρ,X is extreme
+
if and only if it lies in every wall of the Weyl chamber that "touches" aρ,X (or, equivalently,
passes through aρ,X ), hence the term "extreme".
Proof. To construct an element that has the same type as X but has the whole group Wρ,X
as stabilizer, we simply average over the action of this group: we set
X
X′ =
wX.
(3.7)
w∈Wρ,X
(As multiplication by positive scalars does not change anything, we have written it as a
sum rather than an average for ease of manipulation.) Then obviously:
• By definition every wX for w ∈ Wρ,X has the same type as X; since the equivalence
class aρ,X is a convex cone, their sum X ′ also has the same type as X.
• In particular X ′ is generic.
• By construction whenever wX has the same type as X, we have wX ′ = X ′ ; conversely if w fixes X ′ , then wX has the same type as wX ′ = X ′ which has the same
type as X. So X ′ is extreme.
Let us now show that X ′ ∈ a+ , i.e. that for every α ∈ Π, we have α(X ′ ) ≥ 0:
• If sα X ′ = X ′ , then obviously α(X ′ ) = 0.
21
• Otherwise, since X ′ is extreme, it follows that sα X ′ does not even have the same
type as X ′ . Since X ′ is generic, this means that there exists a restricted weight λ
of ρ such that
(
λ(X ′ ) > 0,
(3.8)
sα (λ)(X ′ ) < 0.
By definition, the same inequalities then hold for any Y with the same type as X ′
(or as X):
(
λ(Y ) > 0,
sα (λ)(Y ) < 0.
In particular the form λ − sα (λ), which is a multiple of α, takes a positive value on
every such Y ; hence α never vanishes on the equivalence class aρ,X . By hypothesis X ∈ a+ , so α(X) ≥ 0. Since aρ,X is connected, we conclude that α(X ′ ) > 0.
Finally, assume that X0 is symmetric, i.e. −w0 (X) = X. Then since w0 belongs to the
Weyl group, it induces a permutation on Ω, hence we have:
>
>
<
w0 Ω>
X = Ωw0 X = Ω−X = ΩX ,
(3.9)
<
so that w0 swaps the sets Ω>
X and ΩX . Now by definition we have
<
Wρ,X = StabW (Ω>
X ) ∩ StabW (ΩX ),
(3.10)
hence w0 normalizes Wρ,X . Obviously the map X 7→ −X commutes with everything, so
−w0 also normalizes Wρ,X . We conclude that
X
−w0 (X ′ ) =
−w0 (w(X))
w∈Wρ,X
=
X
w′ (−w0 (X))
w ′ ∈Wρ,X
= X ′,
(3.11)
so that X ′ is still symmetric.
Remark 3.19. In practice, it can be shown that if G is simple, the set ΠX for extreme,
symmetric, generic X can actually only be one of the following:
(a) empty;
(b) the set of long simple restricted roots;
(c) the whole set Π.
Case (a) accounts for the vast majority of representations. Case (b) obviously only
occurs when the restricted root system has a non-simply-laced Dynkin diagram (G2 , F4 ,
Bn , Cn or BCn ), and then only occurs in finitely many representations of each group.
Case (c) only occurs in trivial situations, namely when either dim a = 0 (i.e. the group G
is compact) or the representation is trivial.
The proof of this fact mostly relies on the following two observations:
22
• As soon as Ω is large enough to include some simple restricted root α, no set ΠX ′
may contain α. Indeed in that case, α(X ′ ) never vanishes for generic X ′ .
• The Weyl group acts transitively on the set of restricted roots of the same length;
so as soon as Ω contains one restricted root of a given length, it contains all of
them.
For the remainder of the paper, we fix some symmetric generic vector X0 in the closed
dominant Weyl chamber a+ that is extreme.
4 Dynamics of elements of type X0
Now we take an element g in the affine group ρ(G) ⋉ V such that the Jordan projection
of its linear part has the same type as X0 . The goal of this section is to understand the
dynamics of g acting on the affine space corresponding to V , in particular its "dynamical
spaces" defined in Subsection 4.3. There is a lot of parallelism between this section and
Section 2 in [Smi16].
In Subsection 4.1, we introduce the dynamical subspaces of X0 . We also show that the
stabilizers in G of those subspaces (except for the neutral one) are precisely the parabolic
subgroups introduced in Subsection 3.4.
In Subsection 4.2, we introduce some formalism that reduces the study of the affine
space VAff corresponding to V to the study of a vector space called A. We also introduce
affine equivalents of linear notions defined previously.
In Subsection 4.3, we define the linear and affine dynamical subspaces associated to
an element of the affine group ρ(G) ⋉ V . This is very similar to Section 2.1 in [Smi16].
In Subsection 4.4, we describe the dynamical subspaces of an element g ∈ ρ(G) ⋉ V
whose Jordan projection has the same type as X0 .
In Subsection 4.5, we show that the action of any such element on its affine neutral
space is a "quasi-translation", and explain what that means. This is a generalization of
Section 2.4 in [Smi16].
In Subsection 4.6, we introduce a family of canonical identifications between different affine neutral spaces, and use them to define the "Margulis invariant" for any such
element, which is a vector measuring its translation part along a subspace of its affine
neutral space. This is a generalization of Section 2.5 in [Smi16].
4.1 Reference dynamical spaces
Recall that X0 is some generic, symmetric, extreme vector in the closed dominant Weyl
chamber a+ , chosen once and for all.
Definition 4.1. We define the following subspaces of V :
L
• V0> := λ(X0 )>0 V λ , the reference expanding space;
• V0< :=
L
λ(X0 )<0 V
λ,
the reference contracting space;
23
• V0= :=
• V0≥ :=
• V0≤ :=
L
λ(X0 )=0 V
L
λ(X0 )≥0 V
L
λ(X0 )≤0 V
λ,
the reference neutral space;
λ,
the reference noncontracting space;
λ,
the reference nonexpanding space.
In other terms, V0≥ is the direct sum of all restricted weight spaces corresponding to
weights in Ω≥
X0 , and similarly for the other spaces.
Clearly these are precisely the dynamical spaces (see Subsection 4.3) associated to the
map exp(X0 ) (acting on V by ρ).
Remark 4.2. Note that since X0 is generic, V0= is actually just the zero restricted weight
space:
V0= = V 0 ;
moreover by Assumption 3.2, zero is a restricted weight, so this space is nontrivial.
Example 4.3.
1. For G = SO+ (p, q) acting on V = Rp+q (where p ≥ q), there is only one generic
type. The spaces V0> and V0< are some maximal totally isotropic subspaces (transverse to each other), V0≥ and V0≤ are their respective orthogonal supplements,
and V0= is the (p − q)-dimensional space orthogonal to both V0> and V0< .
2. If G is any semisimple real Lie group acting on V = g (its Lie algebra) by the
adjoint representation, then the reference noncontracting space g≥
0 is obviously
equal to p+
.
There
is
once
again
only
one
generic
type,
given
by
any
X0 ∈ a++ ;
X0
+ is actually the (reference) minimal
we then have ΠX0 = ∅, so that p+
X0 = p
parabolic subalgebra. We have similar identities for the other dynamical spaces,
namely:
+
g≥
0 =p ;
+
g>
0 =n ;
−
g≤
0 =p ;
−
g<
0 =n ;
g=
0 = l.
Let us now understand what happens when we apply an element of G to one of those
subspaces. The motivation for this, as well as the explanation of the term "reference
subspace", comes from Corollary 4.17.
Proposition 4.4. We have:
(i) StabW (V0≥ ) = StabW (V0> ) = StabW (V0≤ ) = StabW (V0< ) = WX0 .
(ii) StabG (V0≥ ) = StabG (V0> ) = PX+0 .
(iii) StabG (V0≤ ) = StabG (V0< ) = PX−0 .
24
Remark 4.5. Note that every restricted weight space is invariant by ZG (A) = L, and that
NG (A) permutes these spaces; so if we have a direct sum of several restricted weights, it
makes sense to talk about its image by an element of W . Moreover, we have the obvious
identity
∀w ∈ W, ∀λ ∈ a∗ , wV λ = V wλ .
(4.1)
Proof of Proposition 4.4.
(i) First note that since X0 is generic, the only restricted weight that vanishes on X0
is the zero weight, so we have indeed
StabW (V0≥ ) = StabW (V0> ) = StabW (V0≤ ) = StabW (V0< ).
Moreover, this group is obviously included in the group Wρ,X0 = StabW (aρ,X0 ),
which is equal to WX0 since X0 is extreme. Conversely, let w ∈ WX0 ; then X0 is
fixed by w, and so is (say) the set Ω≥
X0 of restricted weights nonnegative on X0 . It
follows that StabW (V0≥ ) contains WX0 .
(ii) We first show that both StabG V0> and StabG V0≥ contain the group P + . Indeed:
• The group L stabilizes every restricted weight space V λ . Indeed, take some
λ ∈ a∗ , v ∈ V λ , l ∈ L, X ∈ a; then we have:
X · l(v) = l(Ad(l−1 )(X) · v) = l(X · v) = λ(X)l(v).
(4.2)
• Let α be a positive restricted root and λ a restricted weight such that the
value λ(X0 ) is positive (resp. nonnegative). Then clearly we have
gα · V λ ⊂ V λ+α ,
and (λ + α)(X0 ) is still positive (resp. nonnegative). Hence n+ stabilizes V0>
and V0≥ .
• The statement follows as P + = L exp(n+ ).
Now take any element g ∈ G. Let us apply the Bruhat decomposition: we may
write
g = p1 wp2 ,
with p1 , p2 some elements of the minimal parabolic subgroup P + and w some element of the restricted Weyl group W (see e.g. [Kna96], Theorem 7.40). (Technically
we need to replace w ∈ W = NG (A)/ZG (A) by some representative w̃ ∈ NG (A);
but by the remark preceding this proof, we may ignore this distinction.) From the
statement that we just proved it immediately follows that
StabG (V0≥ ) = StabG (V0> ) = P + StabW (V0≥ )P + = P + WX0 P + .
(4.3)
On the other hand, we have the Bruhat decomposition for parabolic subgroups:
+
+
PX+0 := StabG (p+
X0 ) = P W X0 P .
25
(4.4)
This can be shown by applying a similar reasoning to the adjoint representation:
indeed in that case the space V0≥ corresponding to the same X0 is just p+
X0 . (There
is just a small difficulty due to the fact that X0 is not, in general, generic with
respect to the adjoint representation.)
The conclusion follows.
(iii) Replacing P + and PX+0 respectively by P − and PX−0 , the same reasoning applies.
4.2 Extended affine space
Let VAff be an affine space whose underlying vector space is V .
Definition 4.6 (Extended affine space). We choose once and for all a point of VAff which
we take as an origin; we call R0 the one-dimensional vector space formally generated by
this point, and we set A := V ⊕ R0 the extended affine space corresponding to V . (We
hope that A, the extended affine space, and A, the group corresponding to the Cartan
space, occur in sufficiently different contexts that the reader will not confuse them.)
Then VAff is the affine hyperplane "at height 1" of this space, and V is the corresponding
vector hyperplane:
V = V × {0} ⊂ V × R0 ;
VAff = V × {1} ⊂ V × R0 .
Definition 4.7 (Linear and affine group). Any affine map g with linear part ℓ(g) and
translation vector v, defined on VAff by
g : x 7→ ℓ(g)(x) + v,
can be extended in a unique way to a linear map defined on A, given by the matrix
ℓ(g) v
.
0
1
From now on, we identify the abstract group G with the group ρ(G) ⊂ GL(V ), and
the corresponding affine group G ⋉ V with a subgroup of GL(A).
Definition 4.8 (Affine subspaces). We define an extended affine subspace of A to be a
vector subspace of A not contained in V . There is a one-to-one correspondence between
extended affine subspaces of A and affine subspaces of VAff of dimension one less. For
any extended affine subspace of A denoted by A1 (or A2 , A′ and so on), we denote by V1
(or V2 , V ′ and so on) the space A ∩ V (which is the linear part of the corresponding affine
space A ∩ VAff ).
Definition 4.9 (Translations). By abuse of terminology, elements of the normal subgroup V ⊳ G ⋉ V will still be called translations, even though we shall see them mostly
as endomorphisms of A (so that they are formally transvections). For any vector v ∈ V ,
we denote by τv the corresponding translation.
26
Definition 4.10 (Reference affine dynamical spaces). We now give a name for (the
vector extensions of) the affine subspaces of VAff parallel respectively to V0≥ , V0≤ and V0=
and passing through the origin: we set
≥
A≥
0 := V0 ⊕ R0 ,
the reference affine noncontracting space;
A0 := V0 ⊕ R0 ,
the reference affine nonexpanding space;
A0 := V0 ⊕ R0 ,
the reference affine neutral space.
≤
=
≤
=
These are obviously the affine dynamical spaces (see next subsection) corresponding to
the map exp(X0 ), seen as an element of G ⋉ V by identifying G with the stabilizer of R0
in G ⋉ V .
Definition 4.11 (Affine Jordan projection). Finally, we extend the notion of Jordan
projection to the whole group G ⋉ V , by setting
∀g ∈ G ⋉ V,
Jd(g) := Jd(ℓ(g)).
Remark 4.12.
1. It is tempting to try to define an "affine Jordan decomposition": to wit, any affine
map g ∈ G ⋉ V may be written as g = τv gh ge gu with gh (resp. ge , gu ) conjugate
in G ⋉ V to an element of A (resp. of K, of N + ) and v some element of V .
Unfortunately, we can neither require that τv commute with the other three factors,
nor (as erroneously claimed in the author’s previous paper [Smi16]) determine v in
a unique fashion. The trouble comes from unipotent elements; to understand the
problem, examine the affine transformation
x
1 1
x
0
g:
7→
+
.
y
0 1
y
1
So we must be a little more careful; see the proof of Proposition 4.16 for a more
detailed study of conjugacy classes in G ⋉ V .
2. We do not extend similarly the Cartan projection to G⋉V , for the following reason.
While eigenvalues of an element of G ⋉ V depend only on the eigenvalues of its
linear part, the same statement does not hold for its singular values.
4.3 Definition of dynamical spaces
For every g ∈ G ⋉ V , we define its linear dynamical spaces as follows:
• Vg> , the expanding space associated to g:
the largest vector subspace of V stable by g such that all eigenvalues λ of the
restriction of g to that subspace satisfy |λ| > 1;
• Vg< , the contracting space associated to g:
the same thing with |λ| < 1;
27
• Vg= , the neutral space associated to g:
the same thing with |λ| = 1;
• Vg≥ , the noncontracting space associated to g:
the same thing with |λ| ≥ 1;
• Vg≤ , the nonexpanding space associated to g:
the same thing with |λ| ≤ 1.
Equivalently, Vg> is the direct sum of all the generalized eigenspaces E λ of g associated to
eigenvalues λ of modulus larger than 1 (defined as E λ = ker(g −λ Id)n where n = dim V ),
and similarly for the four other spaces. We then obviously have
Vg≥
z }| {
V = Vg> ⊕ Vg= ⊕ Vg< .
| {z }
Vg≤
(4.5)
Also note that the restriction of g from A to V is just its linear part, so that the linear
dynamic subspaces of g only depend on ℓ(g).
For every g ∈ G ⋉ V , we define its affine dynamical subspaces:
• A≥
g , the affine noncontracting space associated to g,
• A≤
g , the affine nonexpanding space associated to g,
• and A=
g , the affine neutral space associated to g,
in the same way as the linear dynamical subspaces, but with V replaced everywhere
by A.
Remark 4.13.
<
• Note that if we defined in the same way A>
g (resp. Ag ), it would actually be
contained in V and so just be equal to Vg> (resp. Vg< ). Indeed an element of G ⋉ V
can never act on a vector in A\V (i.e. an element of A with a nonzero R0 component)
with an eigenvalue other than 1.
• Hence the decomposition (4.5) now becomes:
A≥
g
z }| {
>
<
A = Vg ⊕ A=
g ⊕ Vg
| {z }
A≤
g
(pay attention to the distribution of A’s and V ’s).
28
(4.6)
≥
≤
• From this identity, it immediately follows that neither A=
g , Ag nor Ag are contained
in V .
• Finally, it is obvious that the intersections of these three spaces with V are respectively Vg= , Vg≥ and Vg≤ . Thus this notation is consistent with the convention
outlined above.
In purely affine terms, these spaces may be understood as follows:
=
• A=
g ∩ VAff is the unique g-invariant affine space parallel to Vg (the "axis" of g);
=
≥
• A≥
g ∩ VAff is the unique affine space parallel to Vg and containing Ag ∩ VAff , and
≤
similarly for Ag ∩ VAff .
4.4 Description of dynamical spaces
We shall now characterize the dynamical subspaces of those elements of G ⋉ V that
satisfy the following property.
Definition 4.14. We say that an element g ∈ G ⋉ V is of type X0 if Jd(g) has the same
type as X0 , i.e. if
Jd(g) ∈ aρ,X0 .
Example 4.15.
1. For G = SO+ (p, q) acting on V = Rp+q (where p ≥ q), there is only one generic
type. For every g ∈ G, we have
dim Vg> = dim Vg< ≤ q.
(4.7)
An element g ∈ G is of generic type if and only if equality is attained. Such elements
have been called pseudohyperbolic in the previous literature ([AMS02, Smi14]).
2. If G is any semisimple real Lie group acting on V = g (its Lie algebra) by the
adjoint representation, there is only one generic type and an element g ∈ G is
of that type if and only if Jd(g) ∈ a++ . Such elements are called R-regular or
(particularly in [BQ]) loxodromic.
Here is a partial description of the dynamical spaces of an element of type X0 .
Proposition 4.16. Let g ∈ G ⋉ V be a map of type X0 . In that case:
(i) There exists a map φ ∈ G ⋉ V , called a canonizing map for g, such that
(
≥
φ(A≥
g ) = A0 ;
≤
φ(A≤
g ) = A0 .
<
(ii) The space Vg> is uniquely determined by A≥
g . The space Vg is uniquely determined
≤
by Ag .
29
(Compare this with Claim 2.5 in [Smi16].)
Proof.
(i)
• We start with the obvious decomposition
g = τv ℓ(g),
(4.8)
where ℓ(g) ∈ G is the linear part of g (seen as an element of G ⋉ V by
identifying G with the stabilizer of the origin R0 ) and v ∈ V is its translation
part. We then observe that we may rewrite this as
g = τv′ τw−1 ℓ(g)τw
(4.9)
for some w ∈ V , where v ′ is now actually an element of Vg= . Indeed, for any
translation vector v ∈ V and linear map f ∈ G, we have
f τv = τf (v) f.
(4.10)
The statement then follows from the fact that the map induced by ℓ(g) − Id
on Vg> ⊕ Vg< does not have 0 as an eigenvalue, hence is surjective. (In fact, this
argument shows that we could even require v ′ to lie in the actual characteristic
space corresponding to the eigenvalue 1.)
• Now let ℓ(g) =: gh ge gu be the Jordan decomposition of ℓ(g), so that
τw gτw−1 = τv′ gh ge gu ;
(4.11)
let φℓ ∈ G be any map that realizes the conjugacy φℓ gh φ−1
ℓ = exp(Jd(g)); and
let φ := φℓ τw .
Calling g′ := φgφ−1 and τv′′ , ge′ , gu′ the respective conjugates of τv′ , ge , gu
by φℓ (so that v ′′ = φℓ (v ′ )), we then have
g′ = τv′′ exp(Jd(g))ge′ gu′ ,
(4.12)
where ge′ ∈ G is elliptic, gu′ ∈ G is unipotent, both commute with exp(Jd(g)),
and v ′′ ∈ Vg=′ .
As already seen in the proof of Proposition 2.6, ge′ and gu′ have all eigenvalues
of modulus 1 and commute with exp(Jd(g)). Hence the linear dynamical
spaces of g′ coincide with those of exp(Jd(g)).
=
Now since exp(Jd(g)) ∈ G fixes R0 , the space A=
exp(Jd(g)) is just Vexp(Jd(g)) ⊕R0 ;
and since v ′′ ∈ Vg=′ , that space is still invariant by g ′ . It follows that we have
=
=
A=
g ′ = Aexp(Jd(g)) = Vexp(Jd(g)) ⊕ R0 .
(4.13)
By taking the direct sum with V > and with V < , we deduce that all the affine
dynamical spaces of g′ coincide with those of exp(Jd(g)).
30
Now since g is of type X0 , by definition, Jd(g) is a vector in a that has the
same type as X0 . Hence the affine dynamical subspaces of exp(Jd(g)) coincide
with those of exp(X0 ), which are the reference subspaces. We conclude that
(
≥
A≥
g ′ = A0 ;
≤
A≤
g ′ = A0 .
≥
≤
Since obviously A≥
g ′ = φ(Ag ) and similarly for A , the conclusion follows.
≥
>
>
(ii) Suppose that g is a map of type X0 such that A≥
g = A0 ; let us show that Vg = V0 .
≥
′
−1
Define g = φgφ as in the previous point; then φ stabilizes A0 . On the other
hand, we have Vg> = φ−1 (Vg>′ ); but from the previous point, it follows also that
Vg>′ = V0> .
Clearly the linear part of φ stabilizes V0≥ . It follows from Proposition 4.4 that it
also stabilizes V0> . Since the latter space is contained in V , the translation part
of φ acts trivially on it. We conclude that Vg> = V0> as required.
≥
Now if we have two maps g and g ′ of type X0 such that A≥
g = Ag ′ , by conjugating,
it follows that we also have Vg> = Vg>′ . The same proof works for A≤ and V < .
This immediately allows us to describe the remaining dynamical spaces of g:
Corollary 4.17. Let g ∈ G ⋉ V be a map of type X0 . Then if φ ∈ G ⋉ V is any
canonizing map of g, we have:
≥
φ(A≥
g ) = A0
φ(Vg≥ ) = V0≥
φ(Vg> ) = V0>
≤
φ(A≤
g ) = A0
φ(Vg≤ ) = V0≤
φ(Vg< ) = V0<
=
φ(A=
g ) = A0
φ(Vg= ) = V0= .
In other terms, if φ is a canonizing map of g then all eight dynamical spaces of the
conjugate φgφ−1 coincide with the reference dynamical spaces. This explains why we
called them "reference" spaces.
Proof. The equalities for A≥ and A≤ hold by definition of a canonizing map. The equality
for A= follows by taking the intersection. The equalities for V ≥ , V ≤ and V = follow by
taking the linear part. The equalities for V > and V < follow from Proposition 4.16 (ii).
4.5 Quasi-translations
Let us now investigate the action of a map g ∈ G ⋉ V of type X0 on its affine neutral
space A=
g . The goal of this subsection is to prove that it is "almost" a translation
(Proposition 4.20).
We fix on V a Euclidean form B satisfying the conditions of Lemma 2.4 for the representation ρ.
Definition 4.18. We call quasi-translation any affine automorphism of A=
0 induced by
an element of L ⋉ V0= .
31
Let us explain and justify this terminology. First note that the action of L on V0=
preserves B: indeed, the action of M does so because M ⊂ K, and the action of A on
this space is just trivial. The following statement is then immediate:
Proposition 4.19. Let V0t be the set of fixed points of L in V0= :
V0t := {v ∈ V0= | ∀l ∈ L, lv = v} .
(Note that this is also the set of fixed points of M ). Let V0r be the B-orthogonal complement of V0t in V0= , and let O(V0r ) denote the set of B-preserving automorphisms of V0r .
Then any quasi-translation is an element of
O(V0r ) ⋉ V0r × V0t .
In other words, quasi-translations are affine isometries of V0= that preserve the directions of V0r and V0t and act only by translation on the V0t component. You may think of a
quasi-translation as a kind of "screw displacement"; the superscripts t and r respectively
stand for "translation" and "rotation".
We now claim that any map of type X0 acts on its affine neutral space by quasitranslations:
Proposition 4.20. Let g ∈ G ⋉ V be a map of type X0 , and let φ ∈ G ⋉ V be any
canonizing map for g. Then the restriction of the conjugate φgφ−1 to A=
0 is a quasitranslation.
Let us actually formulate an even more general result, which will have another application in the next subsection:
≤
=
Lemma 4.21. Any map f ∈ G ⋉ V stabilizing both A≥
0 and A0 acts on A0 by quasitranslation.
Proof.
−
=
• We begin by showing that any element of lX0 = p+
X0 ∩ pX0 acts on V0 in the same
way as some element of l. Recall that by definition
M
lX0 = l ⊕
gα ;
α(X0 )=0
hence it is sufficient to show that for every restricted root α such that α(X0 ) = 0,
we have gα · V0= = 0. Indeed, since V0= = V 0 (because X0 is generic), we have
gα · V0= ⊂ V α .
On the other hand, we know by Proposition 4.4 that for such α, the action of gα
stabilizes both V0≥ and V0≤ ; it follows that the image gα · V0= lies in both of these
spaces, hence in their intersection V0= , which is also V 0 . Since α is nonzero, we
have V 0 ∩ V α = 0, which yields the desired equality.
32
• Let PX+0 ,e and PX−0 ,e denote the identity components of PX+0 and PX−0 respectively;
by integrating the previous statement, it follows that any element of PX+0 ,e ∩ PX−0 ,e
acts on V0= in the same way as some element of L.
• Now it follows from [Kna96] Proposition 7.82 (d) (using 7.83 (e)) that
LX0 = PX+0 ∩ PX−0 ⊂ M (PX+0 ,e ∩ PX−0 ,e ).
(4.14)
(Here we are using the assumption that G is connected.) We deduce that any
element of LX0 acts on V0= in the same way as some element of L.
≤
• Finally, any f ∈ G ⋉ V stabilizing both A≥
0 and A0 has linear part stabilizing both
≥
≤
V0 and V0 (hence lying in LX0 , by Proposition 4.4), and translation part contained
both in V0≥ and in V0≤ (in other words, in V0= ). The conclusion follows.
Proof of Proposition 4.20. The Proposition follows immediately by taking f = φgφ−1 .
≤
Indeed, by definition the "canonized" map φgφ−1 has A≥
0 and A0 as dynamical spaces;
in particular it stabilizes them.
Example 4.22.
1. For G = SO+ (p, q) acting on V = Rp+q (with p ≥ q), we have:
• M ≃ SOp−q (R),
• V0= = V 0 ≃ Rp−q ,
and the action of M on V0= is the obvious one. We may then distinguish two cases:
a. If p − q ≥ 2, then the action of M is transitive. The space V0t is trivial
and V0r = V0= . Any affine isometry of V0= may be a quasi-translation.
b. If p − q = 1, then the group M is trivial. We have on the contrary V0t = V0=
and V0r is trivial. A quasi-translation is just a translation.
(We exclude the case p = q because in that case V 0 = 0, which violates Assumption 3.2.)
2. More generally if G is split, then we have m = 0. The group M is in general a
nontrivial finite group; however, it can be shown that we still always have V0t = V0= ,
and a quasi-translation is still just a translation.
3. If G is any semisimple real Lie group acting on V = g (its Lie algebra) by the
adjoint representation, then:
0
• g=
0 = g = l;
• gt0 is the direct sum of a and of the center of m;
• gr0 is the semisimple part of m (in other terms, its derived subalgebra).
The example of G = SO+ (4, 1) (acting on so(4, 1), not on R5 ) shows that V0t and V0r
can both be nontrivial at the same time.
33
We would like to treat quasi-translations a bit like translations; for this, we need to
have at least a nontrivial space V0t . So from now on, we exclude cases like 1.a. above:
Assumption 4.23. The representation ρ is such that
dim V0t > 0.
This is precisely condition (i)(a) from the Main Theorem.
4.6 Canonical identifications and the Margulis invariant
The main goal of this subsection is to associate to every map g ∈ G ⋉ V of type X0 a
vector in V0t , called its "Margulis invariant" (see Definition 4.30). The two Propositions
and the Lemma that lead up to this definition are important as well, and will be often
used subsequently.
Corollary 4.17 has shown us that the "geometry" of any map g of type X0 (namely
the position of its dynamical spaces) is entirely determined by the pair of spaces
≥
≤
≤
(A≥
g , Ag ) = φ(A0 , A0 ).
In fact, such pairs of spaces play a crucial role. Let us begin with a definition; its
connection with the observation we just made will become clear after Proposition 4.26.
Definition 4.24.
• We define a parabolic space to be any subspace of V that is the image of either V0≥
or V0≤ (no matter which one, since X is symmetric) by some element of G.
• We define an affine parabolic space to be any subspace of A that is the image of A≥
0
by some element of G ⋉ V . Equivalently, a subspace A≥ ⊂ A is an affine parabolic
space iff it is not contained in V and its linear part V ≥ = A≥ ∩ V is a parabolic
space.
• We say that two parabolic spaces (or two affine parabolic spaces) are transverse if
their intersection has the lowest possible dimension.
Example 4.25.
1. For G = SO+ (p, q) acting on V = Rp+q (where p ≥ q), a subspace F ⊂ Rp+q
is a parabolic space if and only if F ⊥ is a maximal totally isotropic subspace.
Equivalently, F is a parabolic space if and only if F contains F ⊥ and is minimal
for that property (namely p-dimensional). Pairs of transverse parabolic spaces were
called frames in [Smi14].
2. If G is any semisimple real Lie group acting on V = g (its Lie algebra) by the
adjoint representation, a parabolic space is just an arbitrary minimal parabolic
subalgebra of g (hence the name "parabolic space").
34
Proposition 4.26. A pair of parabolic spaces (resp. of affine parabolic spaces) is trans≤
verse if and only if it may be sent to (V0≥ , V0≤ ) (resp. to (A≥
0 , A0 )) by some element of G
(resp. of G ⋉ V ).
In particular, it follows from Proposition 4.16 that for any map g ∈ G ⋉ V of type X0 ,
≤
the pair (A≥
0 , A0 ) is a transverse pair of affine parabolic spaces.
This Proposition, as well as its proof, is very similar to Claim 2.8 in [Smi16].
Proof. Let us prove the linear version; the affine version follows immediately. Let (V1 , V2 )
be any pair of parabolic spaces. By definition, for i = 1, 2, we may write Vi = φi (V0≥ ) for
some φi ∈ G. Let us apply the Bruhat decomposition to the map φ−1
1 φ2 : we may write
φ−1
1 φ2 = p1 wp2 ,
(4.15)
where p1 , p2 belong to the minimal parabolic subgroup P + , and w is an element of the
restricted Weyl group W (or, technically, some representative thereof). Let φ := φ1 p1 =
−1
+ stabilizes V ≥ , we have
φ2 p−1
2 w ; since P
0
V1 = φ(V0≥ ) and V2 = φ(wV0≥ ).
(4.16)
Thus V1 and V2 are transverse if and only if wV0≥ is transverse to V0≥ , which happens if
and only the sum of the multiplicities of restricted weights contained in the intersection
≥
Ω≥
X0 ∩ wΩX0
is the smallest possible.
Clearly, this intersection always contains {0}. Since by assumption X0 is generic and
≤
symmetric, we deduce that if we have wΩ≥
X0 = ΩX0 , and only in that case, it is equal
to {0}. Since X0 is symmetric, that case is realized when w = w0 . Then we have
V1 = φ(V0≥ ) and V2 = φ(V0≤ ) as required.
Remark 4.27.
• It follows from Proposition 4.4 that the set of all parabolic spaces can be identified
with the flag variety G/PX+0 , by identifying every parabolic space φ(V0≥ ) with the
coset φPX+0 .
• Using the Bruhat decomposition of PX0 (see (4.4)), we may show that two parabolic
spaces V1 = φ1 (V0≥ ) and V2 = φ2 (V0≥ ) = φ2 ◦ w0 (V0≤ ) are transverse if and only if
the corresponding pair of cosets
(φ1 PX+0 , φ2 w0 PX−0 )
is in the open G-orbit of G/PX+0 × G/PX−0 .
Consider a transverse pair of affine parabolic spaces. Their intersection may be seen as
a sort of "abstract affine neutral space". We now introduce a family of "canonical identifications" between those spaces. Unfortunately, these identifications have an inherent
ambiguity: they are only defined up to quasi-translation.
35
Proposition 4.28. Let (A1 , A2 ) be a pair of transverse affine parabolic spaces. Then any
≤
map φ ∈ G ⋉ V such that φ(A1 , A2 ) = (A≥
0 , A0 ) gives, by restriction, an identification of
the intersection A1 ∩ A2 with A=
0 , which is unique up to quasi-translation.
Here by φ(A1 , A2 ) we mean the pair (φ(A1 ), φ(A2 )). Note that if A1 ∩A2 is obtained in
another way as an intersection of two affine parabolic spaces, the identification with A=
0
will, in general, no longer be the same, not even up to quasi-translation: there could also
be an element of the Weyl group involved.
Compare this with Corollary 2.14 in [Smi16].
Proof. The existence of such a map φ follows from Proposition 4.26. Now let φ and φ′
be two such maps, and let f be the map such that
φ′ = f ◦ φ
(4.17)
≤
(i.e. f := φ′ ◦ φ−1 ). Then by construction f stabilizes both A≥
0 and A0 . It follows from
=
Lemma 4.21 that the restriction of f to A0 is a quasi-translation.
Let us now explain why we call these identifications "canonical". The following lemma,
while seemingly technical, is actually crucial: it tells us that the identifications defined in
Proposition 4.28 commute (up to quasi-translation) with the projections that naturally
arise if we change one of the parabolic subspaces in the pair while fixing the other.
Lemma 4.29. Take any affine parabolic space A1 .
Let A2 and A′2 be any two affine parabolic spaces both transverse to A1 .
Let φ (resp. φ′ ) be an element of G ⋉ V that sends the pair (A1 , A2 ) (resp. (A1 , A′2 ))
≤
to (A≥
0 , A0 ); these two maps exist by Proposition 4.26.
Let W1 be image of V0> by any map φ such that A1 = φ(A≥
0 ) (which is unique by
Proposition 4.4).
Let
ψ : A1 −−−−→ A1 ∩ A′2
be the projection parallel to W1 .
Then the map ψ defined by the commutative diagram
A=
0
ψ
φ′
φ
A1 ∩ A2
A=
0
ψ
A1 ∩ A′2
is a quasi-translation.
The space W1 is, in a sense, the "abstract linear expanding space" corresponding to
the "abstract affine noncontracting space" A1 : more precisely, for any map g ∈ G ⋉ V
>
of type X0 such that A≥
g = A1 , we have Vg = W1 (by Proposition 4.16 (ii)).
36
≥
≤
>
>
=
The projection ψ is well-defined because A≥
0 = V0 ⊕ V0 = V0 ⊕ (A0 ∩ A0 ), and so
≥
A1 = φ′−1 (A0 ) = W1 ⊕ (A1 ∩ A′2 ).
This statement generalizes Lemma 2.18 in [Smi16]. The proof is similar, but care must
be taken to replace minimal parabolics by parabolics of type X0 .
Proof. Without loss of generality, we may assume that φ = Id (otherwise we simply
replace the three affine parabolic spaces by their images under φ−1 .) Then we have
≤
′−1 (A≤ ), where φ′ can be any map stabilizing the
′
A1 = A≥
0
0 , A2 = A0 and A2 = φ
≥
space A0 . We want to show that the map φ′ ◦ ψ is a quasi-translation.
We know that φ′ lies in the stabilizer StabG⋉V (A≥
0 ); by Proposition 4.4, the latter is
equal to PX+0 ⋉ V0≥ . We now introduce the algebra
n+
X0 :=
M
gα
(4.18)
α(X0 )>0
+
:= exp n+
and the group NX
X0 . We then have the Langlands decomposition
0
+
PX+0 = LX0 NX
0
(4.19)
(see e.g. [Kna96], Proposition 7.83). Since LX0 stabilizes V0> , this generalizes to the
"affine Langlands decomposition"
+
⋉ V0> ).
PX+0 ⋉ V0≥ = (LX0 ⋉ V0= )(NX
0
(4.20)
+
⋉ V0> .
Thus we may write φ′ = l ◦ n with l ∈ LX0 ⋉ V0= and n ∈ NX
0
+
⋉ V0> stabilizes the
We shall use the following fact: every element n of the group NX
0
≥
>
>
space V0 and induces the identity map on the quotient space A0 /V0 . Indeed, when the
+
+
element n lies in the "linear" group NX
, since NX
is connected, this follows from the
0
0
≥
+
>
+
fact that nX0 · V0 ⊂ V0 (which, in turn, is true because n+
X0 ⊂ n ). When n is a pure
translation by a vector of V0> , this is obvious.
>
By definition, ψ also stabilizes V0> and induces the identity on A≥
0 /V0 ; hence so does
the map n ◦ ψ. But we also know that n ◦ ψ is defined on A1 ∩ A2 = A=
0 , and sends it
onto
=
n ◦ ψ(A1 ∩ A2 ) = n(A1 ∩ A′2 ) = l−1 (A=
0 ) = A0 .
′
Hence the map n ◦ ψ is the identity on A=
0 . It follows that ψ = φ ◦ ψ = l ◦ n ◦ ψ = l (in
restriction to A=
0 ); by Lemma 4.21, ψ is a quasi-translation as required.
Now let g be a map of type X0 . We already know that it acts on its neutral affine space
by quasi-translation; now the canonical identifications we have just introduced allow us
to compare the actions of different elements on their respective neutral affine spaces, as
if they were both acting on the same space A=
0 . However there is a catch: since the
identifications are only canonical up to quasi-translation, we lose information about the
rotation part; only the translation part along V0t remains.
Formally, we make the following definition. Let πt denote the projection from V0=
onto V0t parallel to V0r .
37
Definition 4.30. Let g ∈ G ⋉ V be a map of type X0 . Take any point x in the affine
≥
≤
≥
≤
space A=
g ∩ VAff and any map φ ∈ G such that φ(Vg , Vg ) = (V0 , V0 ). Then we define
the Margulis invariant of g to be the vector
M (g) := πt (φ(g(x) − x)) ∈ V0t .
We call it the Margulis invariant of g.
This vector does not depend on the choice of x or φ: indeed, composing φ with a
quasi-translation does not change the V0t -component of the image. See Proposition 2.16
in [Smi16] for a detailed proof of this claim (for V = g).
5 Quantitative properties
In this section, we define and study two important quantitative properties of maps of
type X0 :
• C-non-degeneracy, which means that the geometry of the map is not too close to
a degenerate case;
• and contraction strength, which measures the extent to which the map g is "much
more contracting" on its contracting space than on its affine nonexpanding space.
In Subsection 5.1, we define these and several other quantitative properties. Several
definitions coincide with those from Section 2.6 in [Smi16] or generalize them.
In the very short Subsection 5.2 (which is a straightforward generalization of Section 2.7
from [Smi16]), we compare these properties for an affine map and its linear part.
In Subsection 5.3, we define analogous quantitative properties for proximal maps, and
relate properties of a product of a (sufficiently contracting and nondegenerate) pair of
proximal maps to the properties of the factors. This is almost the same thing as Section 3.1 in [Smi16], but with one additional result.
5.1 Definitions
We endow the extended affine space A with a Euclidean norm (written simply k • k)
whose restriction to V coincides with the norm B defined in Lemma 2.4 and that makes
R0 orthogonal to V . Then the subspaces V0> , V0< , V0r , V0t and R0 are pairwise orthogonal,
and the restriction of this norm to V0r is invariant by quasi-translations. For any linear
map g acting on A, we write kgk := supx6=0 kg(x)k
kxk its operator norm.
Consider a Euclidean space E (for the moment, the reader may suppose that E = A;
later we will also need the case E = Λp A for some integer p). We introduce on the
projective space P(E) a metric by setting, for every x, y ∈ P(E),
α(x, y) := arccos
|hx, yi|
∈ [0, π2 ],
kxkkyk
38
(5.1)
where x and y are any vectors representing respectively x and y (obviously, the value
does not depend on the choice of x and y). This measures the angle between the lines x
and y. For shortness’ sake, we will usually simply write α(x, y) with x and y some actual
vectors in E \ {0}.
For any vector subspace F ⊂ E and any radius ε > 0, we shall denote the εneighborhood of F in P(E) by:
BP (F, ε) := {x ∈ P(E) | α(x, P(F )) < ε} .
(5.2)
(You may think of it as a kind of "conical neighborhood".)
Consider a metric space (M, δ); let X and Y be two subsets of M. We shall denote
the ordinary, minimum distance between X and Y by
(5.3)
δ(X, Y ) := inf inf δ(x, y),
x∈X y∈Y
as opposed to the Hausdorff distance, which we shall denote by
δHaus (X, Y ) := max
sup δ {x}, Y , sup δ {y}, X
x∈X
y∈Y
!
.
(5.4)
Finally, we introduce the following notation. Let X and Y be two positive quantities,
and p1 , . . . , pk some parameters. Whenever we write
X .p1 ,...,pk Y,
we mean that there is a constant K, depending on nothing but p1 , . . . , pk , such that
X ≤ KY . (If we do not write any subscripts, this means of course that K is an "absolute"
constant — or at least, that it does not depend on any "local" parameters; we consider
the "global" parameters such as the choice of G and of the Euclidean norms to be fixed
once and for all.) Whenever we write
X ≍p1 ,...,pk Y,
we mean that X .p1 ,...,pk Y and Y .p1 ,...,pk X at the same time.
Definition 5.1. Take a pair of affine parabolic spaces (A1 , A2 ). An optimal canonizing
map for this pair is a map φ ∈ G ⋉ V satisfying
≤
φ(A1 , A2 ) = (A≥
0 , A0 )
and minimizing the quantity max kφk, kφ−1 k . By Proposition 4.26 and a compactness
argument, such a map exists iff A1 and A2 are transverse.
We define an optimal canonizing map for a map g ∈ G ⋉ V of type X0 to be an optimal
≤
canonizing map for the pair (A≥
g , Ag ).
Let C ≥ 1. We say that a pair of affine parabolic spaces (A1 , A2 ) (resp. a map g of
type X0 ) is C-non-degenerate if it has an optimal canonizing map φ such that
kφk ≤ C
and
φ−1 ≤ C.
Now take g1 , g2 two maps of type X0 in G ⋉ V . We say that the pair (g1 , g2 ) is
≤
C-non-degenerate if every one of the four possible pairs (A≥
gi , Agj ) is C-non-degenerate.
39
The point of this definition is that there are a lot of calculations in which, when we
treat a C-non-degenerate pair of spaces as if they were perpendicular, we err by no more
than a (multiplicative) constant depending on C. The following result will often be
useful:
Lemma 5.2. Let C ≥ 1. Then any map φ ∈ GL(E) such that kφ±1 k ≤ C induces a
C 2 -Lipschitz continuous map on P(E).
This is exactly Lemma 2.20 from [Smi16].
Remark 5.3. The set of transverse pairs of extended affine spaces is characterized by two
open conditions: there is of course transversality of the spaces, but also the requirement
that each space not be contained in V . What we mean here by "degeneracy" is failure of
one of these two conditions. Thus the property of a pair (A1 , A2 ) being C-non-degenerate
actually encompasses two properties.
First, it implies that the spaces A1 and A2 are transversal in a quantitative way. More
precisely, this means that some continuous function that would vanish if the spaces were
not transversal is bounded below. An example of such a function is the smallest non
identically vanishing of the "principal angles" defined in the proof of Lemma 7.2 (iv).
Second, it implies that both A1 and A2 are "not too close" to the space V (in the same
sense). In purely affine terms, this means that the affine spaces A1 ∩ VAff and A2 ∩ VAff
contain points that are not too far from the origin.
Both conditions are necessary, and appeared in the previous literature (such as [Mar87]
and [AMS02]). However, they were initially treated separately. The idea of encompassing
both in the same concept of "C-non-degeneracy" seems to have been first introduced in
the author’s previous paper [Smi16].
Definition 5.4. Let g ∈ GL(E), let n = dim E, and let p be an integer such that
1 ≤ p < n. Let λ1 , . . . , λn be the eigenvalues of g ordered by nondecreasing modulus.
Then we define the p-th spectral gap of g to be the quotient
κp (g) :=
|λp+1 |
.
|λp |
(5.5)
Note that we chose the convention where the gap is a number smaller or equal than 1.
When E = A, we will most often use the p-th spectral gap for p = dim A≥
0 . In this
case we will omit the index:
κ(g) := κdimA≥ (g).
(5.6)
0
Also, we denote the spectral radius of g, i.e. the largest modulus of any eigenvalue, by:
r(g) := |λ1 |.
(5.7)
(The usual notation, ρ(g), is already taken to mean "g in the representation ρ".)
Definition 5.5. Let s > 0. For a map g ∈ G⋉V of type X0 , we say that g is s-contracting
if we have:
kg(y)k
kg(x)k
≤s
.
(5.8)
∀(x, y) ∈ Vg< × A≥
g,
kxk
kyk
40
(Note that by Corollary 4.17 the spaces Vg< and A≥
g always have the same dimensions as
≥
<
V0 and A0 respectively, hence they are nonzero.)
We define the strength of contraction of g to be the smallest number s(g) such that g
is s(g)-contracting. In other words, we have
s(g) = g|Vg<
g−1
A≥
g
.
(5.9)
Remark 5.6. This strength of contraction s(g) is defined as a kind of "mixed gap": it
measures the gap between singular values of the restrictions of g to some sums of its
eigenspaces. It turns out that this definition is the most convenient for our purposes.
However, if the map g from the above definition is C-non-degenerate, then we may
pretend that s(g) is a "purely singular" gap, as long as we do not care about multiplicative
constants. Indeed, let g′ = φgφ−1 , where φ is an optimal canonizing map for g; then it
is easy to see that we have
s(g) ≍C s(g′ ).
(5.10)
≥
On the other hand, since Vg<′ = V0< and A≥
g ′ = A0 are orthogonal (by convention), every
′
singular value of g is either a singular value of g′ |V < or of g′ |A≥ . It follows that s(g′ ) is
0
0
the quotient between two actual singular values of g ′ , and two consecutive singular values
if s(g) is small enough. See the proof of Lemma 5.1 (iii) for a more detailed discussion.
Remark 5.7. The spectral gap and contraction strength are somewhat related. Take
some affine map g ∈ G ⋉ V of type X0 ; then since the norm of any linear map is at least
equal to its spectral radius, we obviously have
s(g) ≥ κ(g).
(5.11)
On the other hand, for any map g ∈ G ⋉ V , we have
log s(gN ) = N log κ(g) + O (log N ).
N →∞
(5.12)
If g is of type X0 , then κ(g) < 1, so that
s(gN ) → 0.
N →∞
(5.13)
5.2 Affine and linear case
For any map f ∈ G ⋉ V , we denote by ℓ(f ) the linear part of f , seen as an element
of G ⋉ V by identifying G with the stabilizer of the "origin" R0 . In other words, for
every vector (x, t) ∈ V ⊕ R0 = A, we set
ℓ(f )(x, t) = f (x, 0) + (0, t).
(5.14)
(Seeing G as a subgroup of G ⋉ V allows us to avoid introducing new definitions of
C-non-degeneracy and contraction strength for elements of G.)
41
Lemma 5.8. Let C ≥ 1, and take any C-non-degenerate map g (or pair of maps (g, h))
of type X0 in G ⋉ V . Then:
(i) The map ℓ(g) (resp. the pair (ℓ(g), ℓ(h))) is still C-non-degenerate;
(ii) We have s(ℓ(g)) ≤ s(g);
.
(iii) Suppose that s(g−1 ) ≤ 1. Then we actually have s(g) ≍C s(ℓ(g)) g|A=
g
Proof. The proof is exactly the same as the proof of Lemma 2.25 in [Smi16], mutatis
mutandis.
5.3 Proximal maps
Let E be a Euclidean space. The goal of this section is to show Proposition 5.12. We
begin with a few definitions.
Definition 5.9. Let γ ∈ GL(E); let λ1 , . . . , λn be its eigenvalues repeated according to
multiplicity and ordered by nonincreasing modulus. We define the proximal spectral gap
of γ as its first spectral gap:
|λ2 |
κ̃(γ) := κ1 (γ) =
.
|λ1 |
We say that γ is proximal if κ̃(γ) < 1. We may then decompose E into a direct sum of a
line Eγs , called its attracting space, and a hyperplane Eγu , called its repelling space, both
stable by γ and such that:
(
γ|Eγs = λ1 Id;
for every eigenvalue λ of γ|Eγu , |λ| < |λ1 |.
Definition 5.10. Consider a line E s and a hyperplane E u of E, transverse to each other.
An optimal canonizing map for the pair (E s , E u ) is a map φ ∈ GL(E) satisfying
φ(E s ) ⊥ φ(E u )
and minimizing the quantity max kφk, kφ−1 k .
We define an optimal canonizing map for a proximal map γ ∈ GL(E) to be an optimal
canonizing map for the pair (Eγs , Eγu ).
Let C ≥ 1. We say that the pair formed by a line and a hyperplane (E s , E u ) (resp.
that a proximal map γ) is C-non-degenerate if it has an optimal canonizing map φ such
that φ±1 ≤ C.
Now take γ1 , γ2 two proximal maps in GL(E). We say that the pair (γ1 , γ2 ) is C-nondegenerate if every one of the four possible pairs (Eγsi , Eγuj ) is C-non-degenerate.
Definition 5.11. Let γ ∈ GL(E) be a proximal map. We define the proximal strength
of contraction of γ by
γ|Eγu
γ|Eγu
=
s̃(γ) :=
r(γ)
γ|Eγs
42
(where r(γ) is the spectral radius of γ, equal to |λ1 | in the notations of the previous
definition). We say that γ is s̃-contracting if s̃(γ) ≤ s̃.
Note that these definitions are different from the ones we used in the context of maps
of type X0 (hence the new notations s̃ and κ̃).
Proposition 5.12. For every C ≥ 1, there is a positive constant s̃5.12 (C) with the
following property. Take a C-non-degenerate pair of proximal maps γ1 , γ2 in GL(E), and
suppose that both γ1 and γ2 are s̃5.12 (C)-contracting. Then γ1 γ2 is proximal, and we
have:
(i) α Eγs1 γ2 , Eγs1 .C s̃(γ1 );
(ii) s̃(γ1 γ2 ) .C s̃(γ1 )s̃(γ2 ).
(iii) r(γ1 γ2 ) ≍C kγ1 kkγ2 k.
(The constant s̃5.12 (C) is indexed by the number of the proposition, a scheme that
we will stick to throughout the paper.)
Similar results have appeared in the literature for a long time, see e.g. Lemma 5.7
in [AMS02] or Proposition 6.4 in [Ben96].
Proof. The first two points have already been proved in the author’s previous paper:
see Proposition 3.4 in [Smi16]. To prove (iii), we start with the following observation.
Let η = 2Cπ 2 ; then by Lemma 5.2 we have:
α(Eγs1 , Eγu2 ) ≥ η.
On the other hand, we have already seen in the proof of Proposition 3.4 in [Smi16] that
we have
Eγs1 γ2 ∈ B(Eγs1 , η3 ).
The triangular inequality immediately gives us
2η
α Eγs1 γ2 , Eγu2 ≥
.
3
(5.15)
Take any nonzero x ∈ Eγs1 γ2 . We are going to show the estimates
kγ2 (x)k
≍C kγ2 k;
kxk
(5.16a)
kγ1 (γ2 (x))k
≍C kγ1 k.
kγ2 (x)k
(5.16b)
Since by definition, we have γ1 (γ2 (x)) = λx for some λ ∈ C having modulus r(γ1 γ2 ), the
estimate (iii) follows by multiplying (5.16a) and (5.16b) together.
43
Let us first show (5.16a). Let φ be an optimal canonizing map for γ2 ; since γ2 is Cnon-degenerate, we lose no generality by replacing γ2 and x respectively by γ2′ := φγ2 φ−1
and x′ := φ(x). Obviously we have:
kγ2′ (x′ )k ≤ kγ2′ kkx′ k.
(5.17)
To show the other inequality, let us decompose
x′ =: x′s + x′u .
|{z} |{z}
(5.18)
kγ2′ (x′ )k ≥ kγ2′ (x′s )k − kγ2′ (x′u )k.
(5.19)
∈E s′
γ2
Then we have
∈E u′
γ2
For the first term, we have:
kγ2′ (x′s )k = r(γ2′ ) · kx′s k
= kγ2′ k · sin α φ(Eγs1 γ2 ), Eγu′ · kx′ k
2
s
u
α
E
,
E
γ1 γ2
γ2
≥ kγ2′ k · sin
· kx′ k
C2
1 2η
≥ kγ2′ k · sin 2
· kx′ k
C 3
by Lemma 5.2
by (5.15).
(5.20)
For the second term, we have:
kγ2′ (x′u )k ≤ γ2′
E u′
≤ γ2′
E u′
γ
2
kx′u k
kx′ k
γ
2
= kγ2′ k s̃(γ2′ ) kx′ k
≤ kγ2′ k C 2 s̃(γ2 ) kx′ k.
Plugging those two estimates into (5.19), we obtain
2η
2
′
′
′
− C s̃(γ2 ) kx′ k.
kγ2 (x )k ≥ kγ2 k sin
3C 2
We may assume that s̃(γ2 ) ≤
we conclude that
1 1
2 C2
(5.21)
(5.22)
2η
sin 3C
2 . Since by construction η depends only on C,
kγ2′ (x′ )k &C kγ2′ kkx′ k.
(5.23)
Putting together (5.17) and (5.23), we get (5.16a) as required.
Now to show (5.16b), simply notice that
γ2 (Eγs1 γ2 ) = Eγs2 γ1
(5.24)
(since γ2 γ1 is the conjugate of γ1 γ2 by γ2 ), so that γ2 (x) ∈ Eγs2 γ1 . Hence we may follow
the same reasoning as for (5.16a), simply exchanging the roles of γ1 and γ2 .
44
6 Additivity of Jordan projections
The goal of this section is to prove Proposition 6.14, which says that the product of two
sufficiently contracting maps of type X0 in general position is still of type X0 . As it is
a purely linear property, we forget about translation parts and work exclusively in the
linear group G for the duration of this section. We proceed in four stages.
We start with Proposition 6.1, which shows that if an element of G is of type X0
and strongly contracting in the default representation ρ, it is proximal and strongly
contracting in some of the fundamental representations ρi defined in Proposition 2.12.
We continue with Proposition 6.6, which relates C-non-degeneracy in V and C ′ -nondegeneracy in the spaces Vi .
We then prove Proposition 6.10 (and a reformulated version, Corollary 6.12), which
constrains the Jordan projection of gh in terms of the Cartan projections of g and h.
Finally, we use Corollary 6.12 to prove Proposition 6.14.
Proposition 6.1. For every C ≥ 1, there is a constant s6.1 (C) with the following
property. Let g ∈ G be a C-non-degenerate map of type X0 such that s(g) ≤ s6.1 (C).
Then for every i ∈ Π \ ΠX0 , the map ρi (g) is proximal and we have
s̃(ρi (g)) .C s(g).
Remark 6.2.
• Note that since all Euclidean norms on a finite-dimensional vector space are equivalent, this estimate makes sense even though we did not specify any norm on Vi .
In the course of the proof, we shall choose one that is convenient for us.
• Recall that "i ∈ Π \ ΠX0 " is a notation shortcut for "i such that αi ∈ Π \ ΠX0 ".
Remark 6.3. Note that we have excluded the indices i that lie in ΠX0 . The latter should
be thought of as a kind of "exceptional set"; indeed, recall (Remark 3.19) that it is often
empty.
To pave the way for proving the Proposition, let us prove a couple of lemmas that
relate the contraction strength of an element of G to its Cartan projection.
Lemma 6.4. Let g ∈ LX0 . Then the Cartan decomposition
g = k1 exp(Ct(g))k2
may be done in such a way that both k1 and k2 are in K ∩ LX0 .
Proof. By Proposition 7.82 (a) in [Kna96], LX0 is the centralizer of the intersection of
the kernels of simple roots in ΠX0 :
(6.1)
LX0 = ZG {X ∈ a | ∀α ∈ ΠX0 , α(X) = 0} .
By Proposition 7.25 in [Kna96], it follows:
45
• that LX0 is reductive;
• that K ∩ LX0 is a maximal compact subgroup in LX0 .
Obviously a ⊂ lX0 is a Cartan subspace of lX0 . Thus if we do a Cartan decomposition in
the group LX0 (see Theorem 7.39 in [Kna96]), it will also be a valid Cartan decomposition
in G.
Lemma 6.5. For every C ≥ 1, there is a constant ε6.5 (C) with the following property.
Let g ∈ G be a C-non-degenerate map of type X0 . Then we have
min λ(Ct(g)) − max
λ(Ct(g)) ≥ − log s(g) − ε6.5 (C).
<
λ∈Ω≥
X
λ∈ΩX
0
0
(Recall that Ω≥
X0 is the set of restricted weights that take nonnegative values on X0 ,
and Ω<
is
its
complement in Ω.)
X0
Note that the first term on the left-hand side is certainly nonpositive, as 0 ∈ Ω≥
X0 .
Proof. First of all let φ be an optimal canonizing map for g, and let g′ = φgφ−1 . Then
it is easy to see that we have
s(g′ ) ≍C s(g)
(we already mentioned this in Remark 5.6), and the difference Ct(g′ ) − Ct(g) is bounded
by a constant that depends only on C. Hence it is enough to show the corresponding
inequality for g′ instead of g and without the ε6.5 (C) error term. In fact, let us prove
that equality holds:
min λ(Ct(g′ )) − max
λ(Ct(g′ )) = − log s(g′ ).
<
λ∈Ω≥
X
λ∈ΩX
0
(6.2)
0
By construction we have Vg≥′ = V0≥ and Vg<′ = V0< . Obviously g ′ stabilizes these two
vector spaces, hence (using Proposition 4.4) we have
g ′ ∈ PX+0 ∩ PX−0 = LX0 .
(6.3)
By Lemma 6.4, we may then write
g′ = k1 exp(Ct(g′ ))k2
(6.4)
with k1 , k2 ∈ K ∩ LX0 . In particular both k1 and k2 stabilize both V0≥ and V0< . Hence
so does the Cartan projection Ct(g′ ), and we have
g′ |V ≥ = exp(Ct(g′ ))|V ≥ ;
0
0
(6.5)
(g′ )−1 < = exp(Ct(g′ ))−1 < .
V
V
0
0
46
Now we know that exp(Ct(g′ )) (seen in the default representation ρ) is self-adjoint (by
choice of the Euclidean structure B), hence its singular values coincide with its eigenvalues. (Moreover V0≥ and V0< are orthogonal.) As exp(Ct(g′ )) ∈ A, obviously it acts on
every restricted weight space V λ with the eigenvalue
exp(λ(Ct(g′ ))).
This implies (6.2) as desired.
Proof of Proposition 6.1. Let s6.1 (C) be a constant small enough to satisfy all the
constraints that will appear in the course of the proof. Let us fix i ∈ Π \ ΠX0 , and let
g ∈ G be a map satisfying the hypotheses. Let us prove the two estimates
κ̃(ρi (g)) = exp(αi (Jd(g)))−1 ≤ κ(g),
(6.6)
which will show that ρi (g) is proximal; and then the two estimates
s̃(ρi (g)) ≍C exp(αi (Ct(g)))−1 .C s(g),
(6.7)
whose combination completes the proof.
• Let us start with the right part of (6.6). Since i ∈ Π \ ΠX0 and since X0 is extreme,
sαi (X0 ) does not have the same type as X0 . Since X0 is generic, we may then find
a restricted weight λ of ρ such that
λ(X0 ) > 0
and
sαi (λ)(X0 ) < 0
(6.8)
(we already made this observation in (3.8)). Since λ is a restricted weight, by
Proposition 2.8, the number
hλ, αi i
(6.9)
nλ :=
2hαi , αi i
is an integer. We have, on the one hand:
nλ αi (X0 ) = (λ − sαi (λ)) (X0 ) > 0;
on the other hand, αi (X0 ) ≥ 0 (because X0 ∈ a+ ); hence nλ is positive.
By Proposition 2.11, every element of the sequence
λ, λ − αi , . . . , λ − nλ αi
is a restricted weight of ρ. There must then exist an integer iλ such that:
(
(λ − iλ αi ) (X0 ) ≥ 0;
(λ − (iλ + 1)αi ) (X0 ) < 0.
47
(6.10)
Now since g is of type X0 , by definition, any restricted weight of ρ has the same
sign when evaluated at Jd(g) or at X0 . Thus we also have
(
(λ − iλ αi ) (Jd(g)) ≥ 0;
(λ − (iλ + 1)αi ) (Jd(g)) < 0.
From Proposition 2.6, it then follows that
κ(g) ≥ exp(αi (Jd(g)))−1
as desired.
• Similarly we may establish the right part of (6.7). By considering once again a
restricted weight λ and an integer iλ satisfying (6.10), it follows from Lemma 6.5
that
αi (Ct(g)) ≥ − log s(g) − ε6.5 (C).
By taking the opposite on both sides and exponentiating, the desired estimate
follows immediately.
• Let us now prove the left part of (6.6). By Proposition 2.6 (i), the list of the moduli
of the eigenvalues of ρi (g) is precisely
j
eλi (Jd(g))
,
1≤j≤di
where di is the dimension of Vi and (λji )1≤j≤di is the list of restricted weights of ρi
listed with multiplicity.
Up to reordering that list, we may suppose that λ1i = ni ̟i is the highest restricted
weight of ρi . We may also suppose that λ2i = ni ̟i − αi . Indeed we have
sαi (ni ̟i ) = sα′i (ni ̟i )
= ni ̟i − 2ni
h̟i , α′i i ′
α
hα′i , α′i i i
= ni ̟i − 2ni α′i
(6.11)
(recall that α′i is equal to 2αi if 2αi is a restricted root and to αi otherwise). But
by Proposition 2.11, sαi (ni ̟i ) is a restricted weight of ρi (because it is the image
of a restricted weight of ρi by an element of the Weyl group) and then ni ̟i − αi
is also a restricted weight of ρi (as a convex combination of two restricted weights
of ρi , that belongs to the restricted root lattice shifted by ni ̟i ).
Take any j > 2. Since by hypothesis, the restricted weight ni ̟i has multiplicity 1,
we have λji 6= λ1i . By Lemma 2.13, it follows that this restricted weight has the
form
48
λji = ni ̟i − αi −
r
X
ci′ αi′ ,
i′ =1
with ci′ ≥ 0 for every index i′ .
Finally, since by definition Jd(g) ∈ a+ , for every index i′ we have αi′ (Jd(g)) ≥ 0.
It follows that for every j > 2, we have
λ1i (Jd(g)) ≥ λ2i (Jd(g)) ≥ λji (Jd(g)).
(6.12)
In other words, among the moduli of the eigenvalues of ρi (g), the largest is
exp(λ1i (Jd(g))) = exp(ni ̟i (Jd(g))),
and the second largest is
exp(λ2i (Jd(g))) = exp(ni ̟i (Jd(g)) − αi (Jd(g))).
It follows that
κ̃(ρi (g)) = exp(αi (Jd(g)))−1
as desired.
• Let us finish with the left part of (6.7). We start with the following observation:
for every C ≥ 1, the set
φ ∈ G kφk ≤ C, kφ−1 k ≤ C
(6.13)
is compact. It follows that the continuous map
φ 7→ max ρi (φ) , ρi (φ−1 )
(6.14)
is bounded on that set, by some constant Ci′ that depends only on C (and on the
choice of a norm on Vi , to be made soon). Let φ be the optimal canonizing map
of g, and let g ′ = φgφ−1 ; then we get
s̃(ρi (g)) ≍C s̃(ρi (g′ )).
(6.15)
Now let us choose, on the space Vi where the representation ρi acts, a K-invariant
Euclidean form Bi such that all the restricted weight spaces for ρi are pairwise
Bi -orthogonal (this is possible by Lemma 2.4 applied to ρi ). Then s̃(ρi (g′ )) is simply
the quotient of the two largest singular values of ρi (g ′ ). By Proposition 2.6 (ii)
(giving the singular values of an element of G in a given representation) and by a
calculation analogous to the previous point, we have
s̃(ρi (g′ )) = exp(αi (Ct(g)))−1 .
The desired estimate follows by combining (6.15) with (6.16).
49
(6.16)
Proposition 6.6. Let (g1 , g2 ) be a C-non-degenerate pair of elements of G of type X0 .
Then for every i ∈ Π \ ΠX0 , the pair (ρi (g1 ), ρi (g2 )) is a Ci′ -non-degenerate pair of
proximal maps in GL(Vi ), where Ci′ is some constant that depends only on C and i.
For the duration of the proof of this Proposition, let us fix some i ∈ Π \ ΠX0 . Before
going on, we need a couple of lemmas.
Lemma 6.7.
(i) The restricted weight space Vini ̟i is stable by ρi (PX+0 ).
(ii) The direct sum of all restricted weight spaces Viλ with λ 6= ni ̟i is stable by ρi (PX−0 ).
Proof.
(i) Let us first prove that this space is stable by p+
X0 . By definition, we have:
p+
X0 = l ⊕
M
gβ ;
β(X0 )≥0
Since l centralizes a, it preserves the restricted weight space decomposition; so
clearly l stabilizes Vini ̟i .
Now let β be a root such that β(X0 ) > 0; let us write
X
β=
cα α.
α∈Π
By definition of the set ΠX0 , we then have
cα ≥ 0 for α ∈ Π \ ΠX0 .
Now we know that
(6.17)
gβ · Vini ̟i ⊂ Vini ̟i +β .
The latter space is actually zero. Indeed, otherwise, ni ̟i + β would have to be a
restricted root. But from Lemma 2.13, we know that this would imply
cαi ≤ −1,
which contradicts the inequality above, since i (or, technically, αi ) is in Π \ ΠX0 .
It follows that for every β such that β(X0 ) ≥ 0, the space Vini ̟i is stable by gβ ;
we conclude that it is stable by p+
X0 .
By integration, we deduce that this space is also stable by PX+0 ,e . Now we know (it
follows from [Kna96], Proposition 7.82 (d)) that PX+0 = M PX+0 ,e . Since M centralizes a, it preserves the restricted weight space decomposition, so it stabilizes Vini ̟i .
We conclude that PX+0 stabilizes Vini ̟i .
(ii) The proof is completely analogous.
50
In the following lemma, we denote by PS the set of all parabolic spaces of V ; we also
identify the projective space P(Vi ) with the set of vector lines in Vi and the projective
space P(Vi∗ ) with the set of vector hyperplanes of Vi .
Remark 6.8. Recall (Remark 4.27) that by Proposition 4.4, the manifold PS is diffeomorphic to G/PX+0 .
Lemma 6.9. There exist two continuous maps
Φsi : PS → P(Vi )
and
Φui : PS → P(Vi∗ )
with the following properties:
(i) for every map g ∈ G of type X0 , we have
(
Eρsi (g) = Φsi (Vg≥ );
Eρui (g) = Φui (Vg≤ ).
(ii) if V1 , V2 ∈ PS are transverse, then Φsi (V1 ) 6∈ Φui (V2 ).
Proof. We define the maps Φsi and Φui in the following way. For every φ ∈ G, we set
( s
ni ̟ i
Φi (φ(V0≥ )) = ρi (φ) (V
);
i
L
≤
u
λ .
Φi (φ(V0 )) = ρi (φ)
V
λ6=ni ̟i i
(6.18)
Let us first check that these maps are well-defined. Clearly it is enough to check that
whenever some φ ∈L
G stabilizes the space V0≥ (resp. V0≤ ), it also stabilizes the line Vini ̟i
λ
(resp. hyperplane
λ6=ni ̟i Vi ). Since i ∈ Π \ ΠX0 , this follows from Lemma 6.7 and
Proposition 4.4.
The fact that these maps are continuous is then obvious. To show property (i), we
essentially use:
• the identities
(
Eρsi (exp(Jd(g))) = Vini ̟i ;
L
Eρui (exp(Jd(g))) = λ6=ni ̟i Viλ ,
(6.19)
which follow from the inequality (6.12) ranking the values of different restricted
weights of ρi evaluated at Jd(g);
• and the simple observation that any eigenspace of φgφ−1 is the image by φ of the
eigenspace of g with the same eigenvalue.
Property (ii) essentially follows from Proposition 4.26, which says that G acts transitively
on the set of transverse pairs of parabolic spaces.
51
Proof of Proposition 6.6. Let C ≥ 1. Then the set of C-non-degenerate pairs of parabolic
spaces is compact. On the other hand, the function
(V1 , V2 ) 7→ α(Φsi (V1 ), Φui (V2 ))
is continuous, and (by Lemma 6.9 (ii)) takes positive values on that set. Hence it is
bounded below. So there is a constant Ci′ ≥ 1, depending only on C, such that whenever
a pair (V1 , V2 ) of parabolic spaces is C-non-degenerate, the pair (Φsi (V1 ), Φui (V2 )) is Ci′ non-degenerate.
The conclusion then follows by Lemma 6.9 (i).
Proposition 6.10. For every C ≥ 1, there are positive constants s6.10 (C) and ε6.10 (C)
with the following property. Take any C-non-degenerate pair (g, h) of elements of G of
type X0 such that s(g) ≤ s6.10 (C) and s(h) ≤ s6.10 (C). Then we have:
(i) ∀i ∈ Π,
̟i (Jd(gh) − Ct(g) − Ct(h)) ≤ 0;
(ii) ∀i ∈ Π \ ΠX0 ,
̟i (Jd(gh) − Ct(g) − Ct(h)) ≥ −ε6.10 (C).
See Figure 2 for a picture explaining both this proposition and the corollary below.
Remark 6.11. Though we shall not use it, a very important particular case is g = h.
We then obviously have Jd(gh) = 2 Jd(g) and Ct(g) + Ct(h) = 2 Ct(g), so that the
inequalities (i) and (ii) give a relationship between the Cartan and Jordan projections of
a C-non-degenerate, sufficiently contracting map of type X0 .
Before proving the Proposition, let us give a more palatable (though slightly weaker)
reformulation.
Corollary 6.12. For every C ≥ 1, there exists a positive constant ε6.12 (C) with the
following property. For any pair (g, h) satisfying the hypotheses of the Proposition, we
have
Jd(gh) ∈ Conv WX0 · Ct′ (g, h) ,
(6.20)
where Conv denotes the convex hull and Ct′ (g, h) is some vector in a satisfying
k Ct′ (gh) − Ct(g) − Ct(h)k ≤ ε6.12 (C).
(6.21)
Note that the vector Ct′ (g, h) might not lie in the closed dominant Weyl chamber a+
(even though it is very close to the vector Ct(g) + Ct(h) which does).
Proof. We define Ct′ (g, h) by the linear system
(
∀i ∈ Π \ ΠX0 , ̟i (Ct′ (g, h)) = ̟i (Jd(gh));
∀i ∈ ΠX0 ,
̟i (Ct′ (g, h)) = ̟i (Ct(g) + Ct(h)),
(6.22)
which is possible because (̟i )i∈Π is a basis of a. The estimate (6.21) then immediately
follows from the inequalities of Proposition 6.10. But we may now rewrite Proposition 6.10 without the epsilons: we now have
(
∀i ∈ Π,
̟i (Jd(gh) − Ct′ (g, h)) ≤ 0;
(6.23)
∀i ∈ Π \ ΠX0 , ̟i (Jd(gh) − Ct′ (g, h)) = 0.
52
WX0 = {Id, sα1 }
ε6.10 (C)
X0
sα1 · Ct′ (g, h)
α2
̟2
Jd(gh)
Ct(g) + Ct(h)
̟1
Ct′ (g, h)
α1
Figure 2: This picture represents the situation of Example 3.5.3, namely G = SO+ (3, 2)
acting on R5 . We have chosen a generic, symmetric, extreme vector X0 . The
set ΠX0 is then {α1 } (or {1} with the usual abuse of notations), and the
group WX0 is generated by the single reflection sα1 . Proposition 6.10 states
that Jd(gh) lies in the shaded trapezoid. Corollary 6.12 states that it lies on
the thick line segment. In any case it lies by definition in the dominant open
Weyl chamber (the shaded sector).
53
Let us now show the inequalities
∀i ∈ ΠX0 ,
αi (Ct′ (g, h)) ≥ 0.
(6.24)
Let (Hi )i∈Π be the basis of a dual to (̟i )i∈Π , i.e. the unique basis such that the
identity
X
̟i
cj Hj = ci
(6.25)
j∈Π
holds for any i ∈ Π and any tuple (ci ) ∈ RΠ . By definition of the fundamental restricted
weights ̟i , it then follows that we also have the identity
∀i ∈ Π, ∀λ ∈ a∗ ,
λ(Hi ) =
2hλ, α′i i
.
kα′i k2
(6.26)
By decomposing the vector Ct(g)+Ct(h)−Ct′ (g, h) in the basis (Hi )i∈Π and by plugging
the formula (6.25) into the second line of the defining system (6.22), we find that we may
write
X
Ct′ (g, h) = Ct(g) + Ct(h) −
cj Hj ;
(6.27)
j∈Π\ΠX0
by combining the first line of the defining system (6.22) with Proposition 6.10 (i), we
also obtain that cj ≥ 0 for every j ∈ Π \ ΠX0 .
Finally, take any index i ∈ ΠX0 . Then we have
X
X
αi
cj Hj =
cj αi (Hj )
j∈Π\ΠX0
j∈Π\ΠX0
=
X
j∈Π\ΠX0
cj
2hαi , α′j i
kα′j k2
≤0:
(6.28)
indeed since j varies in Π \ ΠX0 and i ∈ ΠX0 , we have i 6= j hence hαi , αj i ≤ 0; and α′j is
by construction a positive multiple of αj . We conclude that
αi (Ct′ (g, h)) ≥ αi (Ct(g)) + αi (Ct(h)) ≥ 0
(since Ct(g), Ct(h) ∈ a+ ), which gives us (6.24).
Now the system of inequalities (6.24) is equivalent to saying that
Ct′ (g, h) ∈ a+
X0 ,
(6.29)
where a+
X0 is a fundamental domain for the action of Weyl subgroup WX0 on a, more
specifically the one that contains the dominant Weyl chamber a+ . The statement (6.20)
then follows from this and from (6.23), by applying Proposition 2.7 which characterizes
convex hulls of orbits of WX0 .
54
Proof of Proposition 6.10. Let i ∈ Π. We know (see (6.12) above) that for any X ∈ a+ ,
the number ni ̟i (X) is the largest eigenvalue of ρi (X). From Proposition 2.6, it then
follows that:
(
ni ̟i (Ct(g)) = log kρi (g)k;
(6.30)
ni ̟i (Jd(g)) = log r(ρi (g))
(recall that r denotes the spectral radius).
(i) is straightforward from here: indeed,
ni ̟i Jd(gh) = log r(ρi (gh))
≤ log kρi (gh)k
≤ log kρi (g)kkρi (h)k
= ni ̟i Ct(g) + Ct(h) .
(6.31)
(ii) Assume that i ∈ Π \ ΠX0 . By Proposition 6.1, we know that the maps ρi (g)
and ρi (h) are proximal. By Proposition 6.6, they form a C ′ -non-degenerate pair,
for some Ci′ that depends only on C. By Proposition 6.1, if we take s6.10 (C) small
enough, we may then assume that both ρi (g) and ρi (h) are s̃5.12 (Ci′ )-contracting.
We may then apply Proposition 5.12 (iii) to these two maps: we get
r(ρi (g)ρi (h)) ≍C kρi (g)kkρi (h)k.
Now from Proposition 2.6, it follows that we have:
r(ρi (gh)) = exp(ni ̟i (Jd(gh)));
kρi (g)k = exp(ni ̟i (Ct(g)));
kρi (h)k = exp(ni ̟i (Ct(h))).
Taking the logarithm, we deduce that there exists a constant εi (C) such that for
sufficiently contracting g and h, we have
(6.32)
ni ̟i Jd(gh) − Ct(g) − Ct(h) ∈ [−εi (C), εi (C)].
Taking
ε6.10 (C) := max
i∈Π\Π
X0
1
εi (C),
ni
(6.33)
the conclusion follows.
Remark 6.13. Corollary 6.12 generalizes a result given by Benoist in [Ben97]. More
specifically, by taking together Lemma 4.1 and Lemma 4.5.1 from that paper, we obtain
that under suitable conditions, the vector
Jd(gh) − Ct(g) − Ct(h)
(which is λ(gh) − µ(g) − µ(h) in Benoist’s notations) is bounded. This seems to be
stronger than our result; but in fact, it also relies on stronger assumptions. There are
two possible ways to interpret it:
55
• Either we may take his set θ to be our Π \ ΠX0 . In that case, [Ben97] uses the
additional assumption that g and h are "of type θ", which means that their Jordan
projections must be WX0 -invariant. If we make this assumption on g and h, then
the estimate can also be deduced from our Corollary 6.12.
• Or we may take θ to be the whole set Π. But in that case, [Ben97] needs the
assumption that g and h are proximal (and in general position) in all representations ρi , which is stronger than the hypotheses we have made.
Proposition 6.14. For every C ≥ 1, there is a positive constant s6.14 (C) ≤ 1 with the
following property. Take any C-non-degenerate pair (g, h) of maps of type X0 in G such
that s(g±1 ) ≤ s6.14 (C) and s(h±1 ) ≤ s6.14 (C). Then gh is still of type X0 .
Proof. Let C ≥ 1, and let (g, h) be a C-non-degenerate pair of maps in G⋉ V of type X0 ,
such that
s(g±1 ) ≤ s6.14 (C) and s(h±1 ) ≤ s6.14 (C)
for some constant s6.14 (C) to be specified later.
Lemma 6.5 then gives us
max λ(Ct(g)) ≤ log s(g) + ε6.5 (C) + min λ(Ct(g))
λ∈Ω<
X
λ∈Ω≥
X
0
0
≤ log s(g) + ε6.5 (C).
(6.34)
(Indeed by Assumption 3.2, λ = 0 is a restricted weight that is certainly contained
in Ω≥
X0 , so the minimum above is nonpositive.)
Taking s6.14 (C) small enough, we may assume that
1
<
max kλk ε6.12 (C).
(6.35)
∀λ ∈ ΩX0 , λ(Ct(g)) < −
2 λ∈Ω
Of course a similar estimate holds for h:
<
∀λ ∈ ΩX0 ,
1
λ(Ct(h)) < −
2
max kλk ε6.12 (C).
λ∈Ω
(6.36)
Now let λ be any restricted weight that does not vanish on X0 . We distinguish two
cases:
• Suppose that λ(X0 ) < 0. Let Ct′ (g, h) be the vector defined in Corollary 6.12.
Then on the one hand, we deduce from (6.21) that:
|λ(Ct′ (g, h)) − λ(Ct(g)) − λ(Ct(h))| ≤ kλkk Ct′ (g, h) − Ct(g) − Ct(h)k
≤ max kλk ε6.12 (C).
(6.37)
λ∈Ω
Adding together the three estimates (6.35), (6.36) and (6.37), we get
56
λ(Ct′ (g, h)) < 0;
(6.38)
and this is true for any λ ∈ Ω<
X0 .
On the other hand, we have (6.20) which says that
Jd(gh) ∈ Conv(WX0 · Ct′ (g, h)).
Now take any w ∈ WX0 . Since X0 is extreme, λ is still negative on wX0 , so that
the restricted weight w−1 (λ) still satisfies
w−1 (λ)(X0 ) < 0.
Since (6.38) holds for any such restricted weight, we also have
λ(w(Ct′ (g, h))) = w−1 (λ)(Ct′ (g, h)) < 0.
Thus λ takes negative values on every point of the orbit WX0 · Ct′ (g, h); hence it
also takes negative values on every point of its convex hull. In particular, we have
λ(Jd(gh)) < 0.
(6.39)
• Suppose that λ(X0 ) > 0. Since the set of restricted weights Ω is invariant by W ,
the form w0 (λ) is still a restricted weight; since by hypothesis X0 is symmetric (i.e.
−w0 (X0 ) = X0 ), we then have
w0 (λ)(X0 ) < 0.
We may thus apply the previous point to the weight w0 (λ) and to the map (gh)−1 =
h−1 g−1 (since g−1 and h−1 verify the same hypotheses as g and h); this gives us
w0 (λ)(Jd((gh)−1 )) < 0.
Since Jd((gh)−1 ) = −w0 (Jd(gh)), we conclude that
λ(Jd(gh)) > 0.
(6.40)
We conclude that gh is indeed of type X0 .
Remark 6.15. If we assume that both g and g−1 are sufficiently contracting, then clearly
Lemma 6.5 implies that Ct′ (g, g) and then Ct(g) also has the same type as X0 . Conversely, we may show (by a version of Lemma 6.5 with the inequality going both ways)
that if Ct(g) has the same type as X0 and is "far enough" from the borders of aρ,X0 ,
then g and g −1 are strongly contracting.
57
7 Products of maps of type X0
The goal of this section is to prove Proposition 7.4, which not only says that a product of
a C-non-degenerate, sufficiently contracting pair of maps of type X0 is itself of type X0 ,
but allows us to control the geometry and contraction strength of the product. To do
this, we proceed almost exactly as in Section 3.2 in [Smi16]: we reduce the problem to
Proposition 5.12, by considering the action of G ⋉ V on a suitable exterior power Λp A
(rather than on the spaces Vi as in the previous section).
There is however one crucial difference from [Smi16]: while it is still true that when
g is of type X0 , its exterior power Λp g is proximal, the converse no longer holds. Filling
that gap is what the whole previous section was about.
Remark 7.1. The reader might wonder why we did not (developing upon the final remark
from the previous section) prove an additivity theorem for Cartan projections similar to
Proposition 6.10, and use it to estimate s(gh) in terms of s(g±1 ) and s(h±1 ). Since we
need to study the action on the spaces Vi anyway, this would seemingly allow us to forgo
the additional introduction of Λp A.
The reason is that this approach only works for linear maps g and h: for g ∈ G⋉V , the
Cartan projection is only defined for ℓ(g) and only gives information about the singular
values of ℓ(g), not those of g. So while possible, this approach would force us, on the
other hand, to abandon the unified treatment of quantitative properties of affine maps
(as outlined in Remark 5.3).
We introduce the integers:
≥
p := dim A≥
0 = dim V0 + 1;
q := dim V0< ;
(7.1)
d := dim A = dim V + 1 = q + p.
For every g ∈ G ⋉ V , we may define its exterior power Λp g : Λp A → Λp A. The Euclidean
structure of A induces in a canonical way a Euclidean structure on Λp A.
Lemma 7.2.
(i) Let g ∈ G ⋉ V be a map of type X0 . Then Λp g is proximal, and the attracting (resp.
<
repelling) space of Λp g depends on nothing but A≥
g (resp. Vg ):
(
EΛs p g = Λp A≥
g p
u
EΛp g = x ∈ Λ A x ∧ Λq Vg< = 0 .
(ii) For every C ≥ 1, whenever (g1 , g2 ) is a C-non-degenerate pair of maps of type X0 ,
(Λp g1 , Λp g2 ) is a C p -non-degenerate pair of proximal maps.
(iii) For every C ≥ 1, for every C-non-degenerate map g ∈ G ⋉ V of type X0 , we have
s(g) .C s̃(Λp g).
58
(7.2)
If in addition s(g) ≤ 1, we have
s(g) ≍C s̃(Λp g).
(7.3)
(Recall the Definitions 5.5 and 5.11 of the "contraction strengths" s(g) and s̃(γ),
respectively.)
(iv) For any two p-dimensional subspaces A1 and A2 of A, we have
αHaus (A1 , A2 ) ≍ α (Λp A1 , Λp A2 ) .
This is similar to Lemma 3.8 in [Smi16], except for point (i) which here is weaker than
there.
Proof. For (i), let g ∈ G ⋉ V be a map of type X0 . Let λ1 , . . . , λd be the eigenvalues of
g (acting on A) counted with multiplicity and ordered by nondecreasing modulus; then
|λq+1 | = 1 and |λq | < 1. On the other hand, we know that the eigenvalues of Λp g counted
with multiplicity are exactly the products of the form λi1 · · · λip , where 1 ≤ i1 < · · · <
ip ≤ d. As the two largest of them (by modulus) are λq+1 · · · λd and λq λq+2 · · · λd , it
follows that Λp g is proximal.
As for the expression of E s and E u , it follows immediately by considering a basis that
trigonalizes g.
For (ii), (iii) and (iv), the proof is exactly the same as for the corresponding points in
Lemma 3.8 in [Smi16], mutatis mutandis.
We also need the following technical lemma, which generalizes Lemma 3.9 in [Smi16]:
Lemma 7.3. There is a constant ε > 0 with the following property. Let A1 , A2 be any
two affine parabolic spaces such that
(
αHaus (A1 , A≥
0) ≤ ε
≤
Haus
α
(A2 , A0 ) ≤ ε.
Then they form a 2-non-degenerate pair.
(Of course the constant 2 is arbitrary; we could replace it by any number larger than 1.)
Proof. The proof is exactly the same as the proof of Lemma 3.9 in [Smi16], mutatis
mutandis.
Proposition 7.4. For every C ≥ 1, there is a positive constant s7.4 (C) ≤ 1 with the
following property. Take any C-non-degenerate pair (g, h) of maps of type X0 in G ⋉ V ;
suppose that we have s(g±1 ) ≤ s7.4 (C) and s(h±1 ) ≤ s7.4 (C). Then gh is of type X0 ,
2C-non-degenerate, and we have:
59
≥
αHaus A≥
.C s(g)
,
A
g
gh
(i)
αHaus A≤ , A≤ . s(h−1 )
C
gh
h
;
(ii) s(gh) .C s(g)s(h).
(This generalizes Proposition 3.6 in [Smi16].)
Before giving the proof, let us first formulate a particular case:
Corollary 7.5. Under the same hypotheses, we have
≥
αHaus Vgh
, Vg≥ .C s(ℓ(g))
αHaus V ≤ , V ≤ . s(ℓ(h)−1 ).
C
gh
h
Proof. This follows from Lemma 5.8. The proof is the same as for Corollary 3.7 in [Smi16].
Proof of Proposition 7.4. Let us fix some constant s7.4 (C), small enough to satisfy all
the constraints that will appear in the course of the proof. Let (g, h) be a pair of maps
satisfying the hypotheses.
First note that by Lemma 5.8, we have
s(ℓ(g)±1 ) ≤ s(g±1 ) ≤ s7.4 (C)
(7.4)
and similarly for h. If we take s7.4 (C) ≤ s6.14 (C), then Proposition 6.14 tells us
that ℓ(gh), hence gh (indeed the Jordan projection depends only on the linear part), is
of type X0 .
The remainder of the proof works exactly like the proof of Proposition 3.6 in [Smi16],
namely by applying Proposition 5.12 to the maps γ1 = Λp g and γ2 = Λp h. Taking into
account the central position occupied in the paper by the Proposition we are currently
proving, let us reproduce these details nevertheless. Let us check that γ1 and γ2 satisfy
the required hypotheses:
• By Lemma 7.2 (i), γ1 and γ2 are proximal.
• By Lemma 7.2 (ii), the pair (γ1 , γ2 ) is C p -non-degenerate.
• Since we have supposed s7.4 (C) ≤ 1, it follows by Lemma 7.2 (iii) that s̃(γ1 ) .C
s(g) and s̃(γ2 ) .C s(h). If we choose s7.4 (C) sufficiently small, then γ1 and γ2 are
s̃5.12 (C p )-contracting, i.e. sufficiently contracting to apply Proposition 5.12.
Thus we may apply Proposition 5.12. It remains to deduce from its conclusions the
conclusions of Proposition 7.4.
• We already know that gh is of type X0 .
60
• From Proposition 5.12 (i), using Lemma 7.2 (i), (iii) and (iv), we get
≥
,
A
αHaus A≥
g .C s(g),
gh
which shows the first line of Proposition 7.4 (i).
• By applying Proposition 5.12 to γ2−1 γ1−1 instead of γ1 γ2 , we get in the same way
the second line of Proposition 7.4 (i).
≤
• Let φ be an optimal canonizing map for the pair (A≥
g , Ah ). By hypothesis, we have
φ±1 ≤ C. But if we take s7.4 (C) sufficiently small, the two inequalities that we
have just shown, together with Lemma 7.3, allow us to find a map φ′ with kφ′ k ≤ 2,
kφ′ −1 k ≤ 2 and
≤
≥
≤
φ′ ◦ φ(A≥
gh , Agh ) = (A0 , A0 ).
It follows that the composition map gh is 2C-non-degenerate.
• The last inequality, namely Proposition 7.4 (ii), now is deduced from Proposition 5.12 (ii) by using Lemma 7.2 (iii).
8 Additivity of Margulis invariant
Proposition 8.1 below is the key ingredient of the proof. It explains how the Margulis
invariant behaves under group operations (inverse and composition).
The first point is easy to prove, but still important. It is a generalization of Proposition 4.1 (i) in [Smi16]; as the general case is slightly harder, we have now given more
details.
The proof of the second point occupies the remainder of this section. We prove it by
reducing it successively to Lemma 8.6 (which is proved using the technical lemma 8.7),
then to Lemma 8.9. The proof follows very closely that of Proposition 4.2 (ii) in [Smi16],
and we have actually omitted the proofs of Lemmas 8.7 and 8.9. We did repeat the proof
of the Proposition itself (to help the reader figure out precisely what is to be changed),
as well as the proof of Lemma 8.6 (to clear up a small confusion in the original proof:
see Remark 8.8).
Proposition 8.1.
(i) For every map g ∈ G ⋉ V of type X0 , we have
M (g−1 ) = −w0 (M (g)).
(ii) For every C ≥ 1, there are positive constants s8.1 (C) ≤ 1 and ε8.1 (C) with the
following property. Let g, h ∈ G⋉V be a C-non-degenerate pair of maps of type X0 ,
with g ±1 and h±1 all s8.1 (C)-contracting. Then gh is of type X0 , and we have:
kM (gh) − M (g) − M (h)k ≤ ε8.1 (C).
61
Remark 8.2. Note that M (g) is (by definition) an element of the space V0t , which (again
by definition) is the set of fixed points of L = ZG (A). From this, it is straightforward
to deduce that V0t is invariant by NG (A). Hence w0 induces a linear involution on V0t ,
which does not depend on the choice of a representative of w0 in G.
Let C ≥ 1. We choose some constant s8.1 (C) ≤ 1, small enough to satisfy all the
constraints that will appear in the course of the proof. For the remainder of this section,
we fix g, h ∈ G ⋉ V a C-non-degenerate pair of maps of type X0 such that g±1 and h±1
are s8.1 (C)-contracting.
The following remark will be used throughout this section.
≤
≥
≤
≤
≥
Remark 8.3. We may suppose that the pairs (A≥
gh , Agh ), (Ahg , Ahg ), (Ag , Agh ) and
≤
(A≥
hg , Ag ) are all 2C-non-degenerate. Indeed, recall that (by Proposition 7.4), we have
≥
αHaus A≥
,
A
g .C s(g)
gh
αHaus A≤ , A≤ . s(h−1 )
C
gh
h
and similar inequalities with g and h interchanged. On the other hand, by hypothe≤
sis, (A≥
g , Ah ) is C-non-degenerate. If we choose s8.1 (C) sufficiently small, these four
statements then follow from Lemma 7.3.
Proof of Proposition 8.1.
(i) Let φ be a canonizing map for g. Since Vg≥−1 = Vg≤ and vice-versa (obviously) and
since V0≤ = w0 V0≥ and vice-versa (because X0 is symmetric), it follows that w0 φ is
a canonizing map for g −1 .
It remains to show that w0 commutes with πt . Indeed, it is well-known that the
group W , that we defined as the quotient NG (A)/ZG (A), is also equal to the
quotient NK (A)/ZK (A) (see [Kna96], formulas (7.84a) and (7.84b)); hence
NG (A) = W ZG (A) = W ZK (A)A = NK (A)A ⊂ KA.
(8.1)
Let w̃0 be any representative of w0 in NG (A). We already know (Remark 8.2)
that both V0= = V 0 and V0t are invariant by w̃0 . Now by definition the group A
acts trivially on V 0 , and by construction K acts on V 0 by orthogonal transformations (indeed the Euclidean structure was chosen in accordance with Lemma 2.4);
hence V0r , which is the orthogonal complement of V0t in V 0 , is also invariant by w̃0 .
The desired formula now immediately follows from the definition of the Margulis
invariant.
(ii) The proof of this point is a straightforward generalization of the proof of Proposition 4.1 (ii) in [Smi16].
If we take s8.1 (C) ≤ s7.4 (C), then Proposition 7.4 ensures that gh is of type X0 .
=
To estimate M (gh), we decompose gh : A=
gh → Agh into a product of several maps.
62
• We begin by decomposing the product gh into its factors. We have the commutative diagram
gh
A=
gh
A=
hg
g
A=
gh
h
(8.2)
Indeed, since hg is the conjugate of gh by h and vice-versa, we have h(A=
gh ) =
=
=
A=
and
g(A
)
=
A
.
hg
hg
gh
=
=
=
• Next we factor the map g : A=
hg → Agh through the map g : Ag → Ag , which
is better known to us. We have the commutative diagram
A=
gh
A=
hg
g
πg
πg
A=
g
g
(8.3)
A=
g
<
>
where πg is the projection onto A=
g parallel to Vg ⊕ Vg . (It commutes with g
=
>
<
because Ag , Vg and Vg are all invariant by g.)
• Finally, we decompose again every diagonal arrow from the last diagram into
two factors. For any two maps u and v of type X0 , we introduce the notation
≥
≤
A=
u,v := Au ∩ Av .
=
We call P1 (resp. P2 ) the projection onto A=
g,gh (resp. Ahg,g ), still parallel
<
>
to Vg ⊕ Vg . To justify this definition, we must check that A=
g,gh (and sim=
>
<
ilarly Ahg,g ) is supplementary to Vg ⊕ Vg . Indeed, by Remark 8.3, A≤
gh is
>
;
,
hence
(by
Proposition
4.16
(ii))
supplementary
to
V
transverse to A≥
g thus
g
=
>
<
≥
<
=
>
.
Then
we
have
the
⊕
A
⊕
V
=
V
⊕
A
and
A
=
V
⊕
A
=
V
A≥
g
g
g
g
g
g
g,gh
g,gh
commutative diagrams
A=
gh
P1
A=
g,gh
πg
A=
g
(8.4a)
A=
g
(8.4b)
πg
and
A=
hg
P2
A=
hg,g
πg
πg
The second and third step can be repeated with h instead of g. The way to adapt
=
the second step is straightforward; for the third step, we factor πh : A=
hg → Ah
=
=
=
=
through Ah,hg and πh : Agh → Ah through Agh,h .
63
A=
0
A=
0
ggh
P1
A=
0
P2
A=
0
A=
0
gg,gh
ψ1
ψ2
A=
0
φgh
hgh
φg,gh
φg
A=
gh
A=
0
g=
φhg,g
φg
φhg
A=
hg
g
P1
φgh
h
A=
gh
P2
A=
g,gh
A=
hg,g
πg
πg
A=
g
A=
g
g
Diagram 3
Combining these three decompositions, we get the lower half of Diagram 3. (We
left out the expansion of h; we leave drawing the full diagram for especially brave
readers.) Let us now interpret all these maps as endomorphisms of A=
0 . To do this,
we choose some optimal canonizing maps
φg , φgh , φhg , φg,gh , φhg,g
≤
≥
≤
respectively of g, of gh, of hg, of the pair (A≥
g , Agh ) and of the pair (Ahg , Ag ). This
allows us to define ggh , hgh , gg,gh , g= , P1 , P2 , ψ1 , ψ2 to be the maps that make the
whole Diagram 3 commutative.
Now let us define
(
Mgh (g) := πt (ggh (x) − x)
Mgh (h) := πt (hgh (x) − x)
(8.5)
=
=
=
:= A=
, where VAff,0
for any x ∈ VAff,0
0 ∩ VAff is the affine space parallel to V0 and
64
passing through the origin. Since gh is the conjugate of hg by g and vice-versa,
the elements of G ⋉ V (defined in an obvious way) whose restrictions to A=
0 are
≤
ggh and hgh stabilize the spaces A≥
g
and
h
and
A
.
By
Lemma
4.21,
gh
gh are
0
0
thus quasi-translations. It follows that these values Mgh (g) and Mgh (h) do not
depend on the choice of x. Compare this to the definition of a Margulis invariant
=
(Definition 4.30): we have M (gh) = πt (ggh ◦ hgh (x) − x) for any x ∈ VAff,0
. It
immediately follows that
M (gh) = Mgh (g) + Mgh (h).
(8.6)
Thus it is enough to show that kMgh (g)−M (g)k .C 1 and kMgh (h)−M (h)k .C 1.
This is an immediate consequence of Lemma 8.6 below. (Note that while the vectors
Mgh (g) and Mgh (h) are elements of V0t , the maps ggh and hgh are extended affine
isometries acting on the whole subspace A=
0 .)
Remark 8.4. In contrast to actual Margulis invariants, the values Mgh (g) and Mgh (h) do
depend on our choice of canonizing maps. Choosing other canonizing maps would force
us to subtract some constant from the former and add it to the latter.
Definition 8.5. We shall say that a linear bijection f between two subspaces of the
extended affine space A is K(C)-bounded if it is bounded by a constant depending only
on C, that is, kf k .C 1 and kf −1 k .C 1. We say that two automorphisms f1 , f2 of A=
0
(depending somehow on g and h) are K(C)-almost equivalent, and we write f1 ≈C f2 , if
they satisfy the condition
kf1 − ξ ◦ f2 ◦ ξ ′ k .C 1
for some K(C)-bounded quasi-translations ξ, ξ ′ . This is indeed an equivalence relation.
Lemma 8.6. The maps ggh and hgh are K(C)-almost equivalent to g= and h= , respectively.
To show this, we use the following property:
Lemma 8.7. All the non-horizontal arrows in Diagram 3 represent K(C)-bounded, bijective maps.
Note that Lemma 8.7 alone does not imply Lemma 8.6: indeed, while the maps ψ 1
and ψ 2 are quasi-translations by Lemma 4.29, the maps P 1 and P 2 need not be. This
issue will be addressed in Lemma 8.9.
Proof of Lemma 8.7. The proof is exactly the same as the proof of Lemma 4.6 in [Smi16],
mutatis mutandis.
Proof of Lemma 8.6. We shall concentrate on the estimate ggh ≈C g= ; the proof of the
estimate hgh ≈C h= is analogous.
We now use Lemma 4.29 which shows that canonical identifications commute up to
quasi-translation with suitable projections; it implies that the maps ψ1 and ψ2 are quasitranslations. Hence gg,gh is also a quasi-translation.
65
A=
0
A=
0
′
ggh
ℓ(ggh )
A=
0
P2′
P1
P2
A=
0
A=
0
′
gg,gh
ψ2′
ψ1
A=
0
φgh
ℓ(gg,gh )
φg,gh
ψ2
A=
0
g=
φg
A=
gh
A=
0
φ′hg,g φhg,g φ′hg
φg
φhg
A=
hg
g
P1
P2
A=
hg,g
A=
g,gh
πg
πg
A=
g
A=
g
g
Diagram 4
We would like to pretend that ggh and gg,gh are actually translations. To do that, we
modify slightly the upper right-hand corner of Diagram 3. We set
(
φ′hg := ℓ(ggh ) ◦ φhg
(8.7)
φ′hg,g := ℓ(gg,gh ) ◦ φhg,g ,
where ℓ stands for the linear part as defined in Section 5.2, and we define P2′ , ψ2′ ,
′ , g′
ggh
g,gh so as to make the new diagram commutative (see Diagram 4). The factors
ℓ(ggh ) and ℓ(gg,gh ) we introduced (the short horizontal arrows in Diagram 4) have norm 1:
indeed, being quasi-translations of A=
0 fixing R0 , they are orthogonal linear transformations (by Lemma 4.21). Thus Lemma 8.7 still holds for Diagram 4; but now, the modified
′ and g ′
maps ggh
g,gh are translations by construction.
We may write:
−1
′ = (P −1 ◦ g ′
ggh
◦ P2′ ).
(8.8)
1
g,gh ◦ P1 ) ◦ (P1
′ and g ′
Then, since ggh
g,gh are translations, P1
−1
66
◦ P2′ is also a translation. By Lemma 8.7
(applied to Diagram 4), it is the composition of two K(C)-bounded maps, hence K(C)bounded. Thus we have
′ ≈ P −1 ◦ g ′
ggh
(8.9)
C 1
g,gh ◦ P1 .
Since ℓ(ggh ), ℓ(gg,gh ), ψ1 and ψ2 are K(C)-bounded quasi-translations, ggh is K(C)′ and g is K(C)-almost equivalent to g ′
almost equivalent to ggh
=
g,gh . It remains to check
′
that the map gg,gh
is K(C)-almost equivalent to its conjugate P1
−1
′
◦ gg,gh
◦ P1 .
This follows from Lemma 8.9 below. Indeed, let P1′′ be the quasi-translation con′
structed in Lemma 8.9. Let v ∈ V0= be the translation vector of gg,gh
, so that
′
gg,gh
=: τv .
(8.10)
Keep in mind that while we call the map
τv a "translation", it is formally a transvection:
v . Then we have
its matrix in a suitable basis is Id
0 1
P1
−1
′
◦ gg,gh
◦ P1 − P1′′
−1
′
◦ P1′′ = τP
◦ gg,gh
−1
1
= P1
−1
≤ (P1
(v)
− τP ′′ −1 (v)
(v) −
−1
1
−1
P1′′ (v)
− P1′′
−1
)
V0=
kvk
(8.11)
(as v ∈ V0= ).
Remark 8.8. While the corresponding calculation in [Smi16] does not technically contain
any explicit falsehoods (the inequality just happens to be slightly weaker than what it
should be), it implicitly relies on the false "identity" τu − τv = τu−v . Here we have
corrected this confusion.
Now by Lemma 4.21, we know that the quasi-translation P1′′ restricted to V0= is a
linear map preserving the Euclidean norm. We also know that the map ρ 7→ ρ−1 (defined
on GL(V0= )) is Lipschitz-continuous on a neighborhood of the orthogonal group (which
is compact). Finally, by Lemma 5.8, s(ℓ(g)) does not exceed s(g) which is by hypothesis
smaller than or equal to s8.1 (C). Taking s8.1 (C) small enough, we may deduce from
Lemma 8.9 that
(P1
−1
− P1′′
−1
)
V0=
.C s(ℓ(g)).
(8.12)
′
′
, since gg,gh
On the other hand, we have kvk ≤ kτv k = gg,gh
.C g|A=
is the compog
with several K(C)-bounded maps. It follows that
sition of g|A=
g
P1
−1
′
◦ gg,gh
◦ P1 − P1′′
−1
′
.
◦ P1′′ .C s(ℓ(g)) g|A=
◦ gg,gh
g
(8.13)
.C s(g); and we know that s(g) ≤ 1. Finally
By Lemma 5.8 (iii), we have s(ℓ(g)) g|A=
g
we get
−1
−1
′
′
P1 ◦ gg,gh
◦ P1 − P1′′ ◦ gg,gh
◦ P1′′ .C 1.
(8.14)
67
To complete the proof of Lemma 8.6, and hence also the proof of Proposition 8.1, it
remains only to prove Lemma 8.9.
Lemma 8.9. The linear part of the map P1 is "almost" a quasi-translation. More precisely, there is a quasi-translation P1′′ such that
(P1 − P1′′ )
V0=
.C s(ℓ(g)).
Recall that ℓ(g) is the map with the same linear part as g, but with no translation
part: see subsection 5.2. We use the double prime because the relationship between P1′′
and P1 is not the same as the relationship between P2′ and P2 .
Proof. The proof is exactly the same as the proof of Lemma 4.7 in [Smi16], mutatis
mutandis.
9 Margulis invariants of words
We have already studied how contraction strengths (Proposition 7.4) and Margulis invariants (Proposition 8.1) behave when we take the product of two C-non-degenerate,
sufficiently contracting maps of type X0 . The goal of this section is to generalize these
results to words of arbitrary length on a given set of generators. It is a straightforward
generalization of Section 5 in [Smi16] (we slightly changed the notations).
Definition 9.1. Take k generators g1 , . . . , gk . Consider a word g = giσ11 · · · giσll with
length l ≥ 1 on these generators and their inverses (for every m we have 1 ≤ im ≤ k
and σm = ±1). We say that g is reduced if for every m such that 1 ≤ m ≤ l − 1, we have
(im+1 , σm+1 ) 6= (im , −σm ). We say that g is cyclically reduced if it is reduced and also
satisfies (i1 , σ1 ) 6= (il , −σl ).
Proposition 9.2. For every C ≥ 1, there is a positive constant s9.2 (C) ≤ 1 with the
following property. Take any family of maps g1 , . . . , gk ∈ G ⋉ V satisfying the following
hypotheses:
(H1) Every gi is of type X0 .
(H2) Any pair taken among the maps {g1 , . . . , gk , g1−1 , . . . , gk−1 } is C-non-degenerate, except of course if it has the form (gi , gi−1 ) for some i.
(H3) For every i, we have s(gi ) ≤ s9.2 (C) and s(gi−1 ) ≤ s9.2 (C).
Take any nonempty cyclically reduced word g = giσ11 · · · giσll (with 1 ≤ im ≤ k, σm = ±1
for every m). Then g is of type X0 , 2C-non-degenerate, and we have
M (g) −
l
X
M (giσmm ) ≤ lε8.1 (2C)
m=1
(where ε8.1 (2C) is the constant introduced in Proposition 8.1).
68
The proof proceeds by induction, with Proposition 7.4 and Proposition 8.1 providing
the induction step.
Proof. The proof is exactly the same as proof of Proposition 5.2 in [Smi16], mutatis
mutandis.
10 Construction of the group
Here we prove the Main Theorem. We closely follow Section 6 from [Smi16], with only
two substantial differences:
• While in the case of the adjoint representation, existence of a −w0 -invariant vector
in V0t was automatic, here we must postulate it explicitly (Assumption 10.1).
• Where we originally relied on Lemma 7.2 in [Ben96], we now need the more general
Lemma 4.3.a in [Ben97].
In the next-to-last paragraph of the proof, we have also made more explicit the relationship between sMain (C) and s9.2 (C).
Let us recall the outline of the proof. We begin by showing (Lemma 10.3) that if we
take a group generated by a family of C-non-degenerate, sufficiently contracting maps of
type X0 with suitable Margulis invariants, it satisfies all of the conclusions of the Main
Theorem, except Zariski-density. We then exhibit such a group that is also Zariski-dense
(and thus prove the Main Theorem).
The idea is to ensure that the Margulis invariants of all elements of the group lie
almost on the same half-line. Obviously if −w0 maps every element of V0t to its opposite,
Proposition 8.1 (i) makes this impossible. So we now exclude this case:
Assumption 10.1. The representation ρ is such that the action of w0 on V0t is not
trivial.
This is precisely condition (i) from the Main Theorem. More precisely, V0t is the set of
all vectors that satisfy (i)(a), and what we say here is that some of them also satisfy (i)(b).
Example 10.2.
1. Consider G = SO+ (p, q) acting on Rp+q (with p ≥ q), we have already seen that
the only case when V0t 6= 0 is when p − q = 1 (see Example 4.22.1) So let p = n + 1,
q = n; then we may show that
w0 |V t = (−1)n Id
0
(10.1)
(this is essentially the content of Lemma 3.1 in [AMS02] or of Proposition 2.7
in [Smi14]). So G = SO+ (n + 1, n) satisfies this assumption if and only if n is odd.
2. If G is any semisimple real Lie group acting on V = g (its Lie algebra) by the adjoint
representation, then gt0 contains the Cartan subspace a (see Example 4.22.2), on
which w0 obviously acts nontrivially unless a is itself trivial. So this assumption is
satisfied whenever G is noncompact.
69
Thanks to Assumption 10.1, we choose once and for all some nonzero vector v ∈ V0t
that is a fixed point of −w0 (which is possible since w0 is an involution). We also choose
a vector M0 collinear to v and such that kM0 k = 2ε8.1 (2C).
Lemma 10.3. Take any family g1 , . . . , gk ∈ G ⋉ V satisfying the hypotheses (H1), (H2)
and (H3) from Proposition 9.2, and also the additional condition
(H4) For every i, M (gi ) = M0 .
Then these maps generate a free group acting properly discontinuously on the affine
space VAff .
Proof. The proof is exactly the same as the proof of Lemma 6.1 in [Smi16], mutatis
mutandis. The (orthogonal) projection
π̂z : ĝ → z ⊕ R0
parallel to d ⊕ n+ ⊕ n−
now becomes the (orthogonal) projection
π̂t : A → V0t ⊕ R0
parallel to V0r ⊕ V0> ⊕ V0< .
Proof of Main Theorem. First note that the assumptions we have made on ρ in the course
of the paper ensure that it satisfies the hypotheses of the Main Theorem: Assumptions 10.1 and 3.8 give the conditions (i) and (ii) respectively.
(The two other assumptions were "for free": Assumption 4.23 is just the weaker condition (i)(a), and Assumption 3.2 is an even weaker condition that follows from (i)(a).)
Once again, we use the same strategy as in the proof of the Main Theorem of [Smi16].
We find a positive constant C ≥ 1 and a family of maps g1 , . . . , gk ∈ G ⋉ V (with
k ≥ 2) that satisfy the conditions (H1) through (H4) and whose linear parts generate a
Zariski-dense subgroup of G, then we apply Lemma 10.3. We proceed in several stages.
• We begin by using a result of Benoist: we apply Lemma 4.3.a in [Ben97] to
– Γ = G;
– t = k + 1;
– Ωi = aρ,X0 ∩ a++ for every i.
This gives us, for any k ≥ 2, a family of maps γ1 , . . . , γk ∈ G (which we shall see
as elements of G ⋉ V , by identifying G with the stabiliser of R0 ), such that:
(i) Every γi is of type X0 (this is (H1)).
(ii) For any two indices i, i′ and signs σ, σ ′ such that (i′ , σ ′ ) 6= (i, −σ), the spaces
Vγ≥σ and V ≤σ′ are transverse.
i
γi′
(iii) Any single γi generates a Zariski-connected group.
(iv) All of the γi generate together a Zariski-dense subgroup of G.
70
Since in our case t is finite, Benoist’s item (v) is not relevant to us.
A comment about item (i): we actually get not only that every γi is of type X0 ,
but also that every γi is R-regular.
A comment about item (ii): since we have taken Benoist’s Γ to be the whole
group G, we have θ = Π, so that YΓ is the complete flag variety G/P + . Benoist’s
conclusion can then be restated by saying that the pair of cosets
(φg P + , φh P − )
(where φg and φh are respective canonizing maps of g and h as defined by us in
Proposition 4.16) is in the open G-orbit of G/P + ×G/P − . Once again it is actually
stronger than our conclusion, which is equivalent to saying that the pair of cosets
(φg PX+0 , φh PX−0 )
is in the open G-orbit of G/PX+0 × G/PX−0 .
• Clearly, every pair of transverse spaces is C-non-degenerate for some finite C; and
here we have a finite number of such pairs. Hence if we choose some suitable value
of C (which we fix for the rest of this proof), the hypothesis (H2) becomes a direct
consequence of the condition (ii) above.
• From condition (iii) (Zariski-connectedness), it follows that any algebraic group
containing some power γiN of some generator must actually contain the generator γi
itself. This allows us to replace every γi by some power γiN without sacrificing condition (iv) (Zariski-density). Clearly, conditions (i), (ii) and (iii) are then preserved
as well. If we choose N large enough, we may suppose that the numbers s(γi±1 ) are
as small as we wish: this gives us (H3). In fact, we shall suppose that for every i,
we have s(γi±1 ) ≤ sMain (C) for an even smaller constant sMain (C), to be specified
soon.
• To satisfy (H4), we replace the maps γi by the maps
gi := τφ−1 (M0 ) ◦ γi
i
(10.2)
(for 1 ≤ i ≤ k), where φi is a canonizing map for γi .
We need to check that this does not break the first three conditions. Indeed, for
every i, we have γi = ℓ(gi ); even better, since the translation vector φ−1
i (M0 )
=
lies in the subspace Vγi stable by γi , obviously the translation commutes with γi ,
≥
≥
hence gi has the same geometry as γi (meaning that A≥
gi = Aγi = Vγi ⊕ R0 and
≤
≤
≤
Agi = Aγi = Vγi ⊕ R0 ). Hence the gi still satisfy the hypotheses (H1) and (H2),
but now we have M (gi ) = M0 (this is (H4)). As for contraction strength, we have,
by Lemma 5.8:
s(gi ) .C s(γi )kτM0 k
≤ sMain (C)kτM0 k,
71
(10.3)
and similarly for gi−1 . Recall that kM0 k = 2ε8.1 (2C), hence kτM0 k depends only
on C: in fact it is equal to the norm of the 2-by-2 matrix 01 kM10 k . It follows that
if we choose
−1
1 2ε8.1 (2C)
,
(10.4)
sMain (C) ≤ s9.2 (C)
0
1
then the hypothesis (H3) is satisfied.
We conclude that the group generated by the elements g1 , . . . , gk acts properly
discontinuously (by Lemma 10.3), is free (by the same result), nonabelian (since
k ≥ 2), and has linear part Zariski-dense in G.
References
[Abe01] H. Abels. Properly discontinuous groups of affine transformations, a survey.
Geom. Dedicata, 87:309–333, 2001.
[AMS02] H. Abels, G.A. Margulis, and G.A. Soifer. On the Zariski closure of the linear
part of a properly discontinuous group of affine transformations. J. Differential
Geometry, 60:315–344, 2002.
[AMS13] H. Abels, G.A. Margulis, and G.A. Soifer. The Auslander conjecture for dimension less than 7. Preprint, arXiv:1211.2525, 2013.
[Aus64]
L. Auslander. The structure of compact locally affine manifolds. Topology,
3:131–139, 1964.
[Ben96]
Y. Benoist. Actions propres sur les espaces homogènes réductifs. Annals of
Mathematics, 144:315–347, 1996.
[Ben97]
Y. Benoist. Propriétés asymptotiques des groupes linéaires. Geom. and funct.
anal., 7:1–47, 1997.
[BQ]
Y.
Benoist
and
J.-F.
Quint.
Random
walks
on
ductive groups.
Ergebnisse.
To appear;
available
http://www.math.u-psud.fr/~benoist/prepubli/15walk.pdf.
[BT65]
A. Borel and J. Tits. Groupes réductifs. Publications Mathématiques de l’IHÉS,
27:55–151, 1965.
[BT72]
A. Borel and J. Tits. Compléments à l’article “ Groupes réductifs ”. Publications
Mathématiques de l’IHÉS, 41:253–276, 1972.
[DGK]
J. Danciger, F. Guéritaud, and F. Kassel. Proper affine action of right-angled
Coxeter groups. In preparation.
[Ebe96]
P.B. Eberlein. Geometry of Nonpositively Curved Manifolds. University of
Chicago Press, 1996.
72
reat
[FG83]
D. Fried and W.M. Goldman. Three-dimensional affine crystallographic groups.
Adv. in Math., 47:1–49, 1983.
[Hal15]
B.C. Hall. Lie Groups, Lie Algebras and Representations: An Elementary
Introduction. Springer International Publishing, second edition, 2015.
[Hel08]
S. Helgason. Geometric Analysis on Symmetric Spaces. Amer. Math. Soc.,
second edition, 2008.
[Kna96] A.W. Knapp. Lie Groups Beyond an Introduction. Birkhaüser, 1996.
[Mar83] G.A. Margulis. Free properly discontinuous groups of affine transformations.
Dokl. Akad. Nauk SSSR, 272:937–940, 1983.
[Mar87] G.A. Margulis. Complete affine locally flat manifolds with a free fundamental
group. J. Soviet Math., 134:129–134, 1987.
[Mil77]
J. Milnor. On fundamental groups of complete affinely flat manifolds. Adv. in
Math., 25:178–187, 1977.
[Smi14]
I. Smilga. Fundamental domains for properly discontinuous affine groups.
Geom. Dedicata, 171:203–229, 2014.
[Smi16]
I. Smilga. Proper affine actions on semisimple Lie algebras. Annales de l’Institut
Fourier, 66(2):785–831, 2016.
[Tit71]
J. Tits. Représentations linéaires irréductibles d’un groupe réductif sur un
corps quelconque. Journ. Reine Angw. Math., 247:196–220, 1971.
73
| 4math.GR
|
Approximately Counting Triangles in Sublinear Time
arXiv:1504.00954v3 [cs.DS] 22 Sep 2015
Talya Eden∗
Amit Levi†
Dana Ron‡
C. Seshadhri
§
Abstract
We consider the problem of estimating the number of triangles in a graph. This problem
has been extensively studied in both theory and practice, but all existing algorithms read the
entire graph. In this work we design a sublinear-time algorithm for approximating the number
of triangles in a graph, where the algorithm is given query access to the graph. The allowed
queries are degree queries, vertex-pair queries and neighbor queries.
We show that for any given approximation parameter 0 < ǫ < 1, the algorithm provides
an estimate b
t such that with high constant probability, (1 − ǫ) · t < b
t < (1 + ǫ) · t, where t is
the
number
of
triangles
in
the
graph
G.
The
expected
query
complexity
of the algorithm is
n
o
m3/2
n
+ min m, t
· poly(log n, 1/ǫ), where n is the number of vertices in the graph and
t1/3
3/2
n
· poly(log n, 1/ǫ). We
m is the number of edges, and the expected running time is t1/3
+ mt
n
o
n
m3/2
also prove that Ω t1/3 + min m, t
queries are necessary, thus establishing that the query
complexity of this algorithm is optimal up to polylogarithmic factors in n (and the dependence
on 1/ǫ).
∗
School of Computer Science, Tel Aviv University, [email protected]
School of Electrical Engineering, Tel Aviv University, [email protected]
‡
School of Electrical Engineering, Tel Aviv University, [email protected]. This research was partially supported
by the Israel Science Foundation grant No. 671/13 and by a grant from the Blavatnik fund.
§
University of California, Santa Cruz, [email protected]
†
1
Contents
1 Introduction
1.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Overview of the algorithm . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 A simple oracle-based procedure for a 1/3-estimate . . . . .
1.2.2 Assigning weights to triangles so as to improve the estimate
1.2.3 Deciding whether
a vertex is heavy . . . . . . . . . . . . . .
P
1.2.4 Estimating v∈S wt(v) . . . . . . . . . . . . . . . . . . . .
1.3 A high level discussion of the lower bound . . . . . . . . . . . . . .
1.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Approximating graph parameters in sublinear time . . . . .
1.4.2 Triangle counting . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Preliminaries
3 The
3.1
3.2
3.3
3.4
6
Algorithm
Heavy and light vertices . . . . . . . . . . . . . . .
A procedure for deciding whether a vertex is heavy
Estimating the number of triangles given m and t .
The final algorithm . . . . . . . . . . . . . . . . . .
4 A Lower Bound
4.1 A lower bound for t = m . . . . . . . . . . .
4.1.1 The lower-bound construction . . . .
4.1.2 Definition of the processes P1 and P2
4.1.3 The auxiliary graph for t = m . . . .
4.1.4 Statistical distance . . . . . . . . . .
4.2 A lower bound for m < t < m3/2 . . . . . .
4.2.1 The lower-bound construction . . . .
4.2.2 The processes P1 and P2 . . . . . . .
4.2.3 The auxiliary graph . . . . . . . . .
4.2.4 Statistical distance . . . . . . . . . .
√
4.3 A lower bound for m ≤ t ≤ 14 m . . . . . .
4.3.1 The lower-bound construction . . . .
4.3.2 The processes P1 and P2 . . . . . . .
4.3.3 The auxiliary graph . . . . . . . . .
4.3.4 Statistical distance . . . . . . . . . .
√
4.4 Lower Bound for t < 14 m. . . . . . . . . .
4.4.1 The construction . . . . . . . . . . .
4.4.2 The processes P1 and P2 . . . . . . .
4.4.3 The auxiliary graph . . . . . . . . .
4.4.4 Statistical distance . . . . . . . . . .
4.5 Wrapping things up . . . . . . . . . . . . .
2
2
3
3
3
3
4
5
5
5
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
7
9
11
15
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
17
20
20
20
23
24
28
28
28
28
29
30
30
30
31
32
35
35
37
37
38
38
1
Introduction
Counting the number of triangles in a graph is a fundamental algorithmic problem. In the study
of complex networks and massive real-world graphs, triangle counting is a key operation in graph
analysis for bioinformatics, social networks, community analysis, and graph modeling [HL70, Col88,
Por00, EM02, MSOI+ 02, Bur04, BBCG08, FWVDC10, BHLP11, SKP12]. In the theoretical
computer science community, the primary tool for counting the number of triangles is fast matrix multiplication [IR78, AYZ97, BPWZ14]. On the more applied side, there is a plethora of
provable and practical algorithms that employ clever sampling methods for approximate triangle counting [CN85, SW05b, SW05a, Tso08, TKMF09, Avr10, KMPT12, CC11, SV11, TKM11,
AKM13, SPK13, TPT13]. Triangle counting has also been a popular problem in the streaming
setting [BYKS02, JG05, BFL+ 06, AGM12, KMSS12, JSP13, PTTW13, TPT13, ADNK14].
All these algorithms read the entire graph, which may be time consuming when the graph is
very large. In this work, we focus on sublinear algorithms for triangle counting. We assume the
following query access to the graph, which is standard for sublinear algorithms that approximate
graph parameters. The algorithm can make: (1) Degree queries, in which the algorithm can query
the degree dv of any vertex v. (2) Neighbor queries, in which the algorithm can query what vertex
is the ith neighbor of a vertex v, for any i ≤ dv . (3) Vertex-pair queries, in which the algorithm
can query for any pair of vertices v and u whether (u, v) is an edge.
Gonen et al. [GRS11], who studied the problem of approximating the number of stars in a
graph in sublinear time, also considered the problem of approximating the number of triangles in
sublinear time. They proved that there is no sublinear approximation algorithm for the number
of triangles when the algorithm is allowed to perform degree and neighbor queries (but not pair
queries). 1
They asked whether a sublinear algorithm exists when allowed vertex-pair queries in addition
to degree and neighbor queries. We show that this is indeed the case.
1.1
Results
Let G be a graph with n vertices, m edges, and t triangles. We describe an algorithm that, given
an approximation parameter 0 < ǫ < 1 and query access to G, outputs an estimate b
t, such that
with high constant probability (over the randomness of the algorithm), (1 − ǫ) · t ≤ b
t ≤ (1 + ǫ) · t.
The expected query complexity of the algorithm is
(
)!
n
m3/2
+ min m,
· poly(log n, 1/ǫ) ,
t
t1/3
3/2
n
+ mt
and its expected running time is t1/3
· poly(log n, 1/ǫ). We show that this result is almost
optimal by proving that the number of queries performed by any multiplicative-approximation
algorithm for the number of triangles in a graph is
(
)!
m3/2
n
Ω 1/3 + min m,
.
t
t
1
To be precise, they showed that there exist two families of graphs over m = Θ(n) edges, such that all graphs in
one family have Θ(n) triangles, all graphs in the other family have no triangles, but in order to distinguish between
a random graph in the first family and random graph in the second family, it is necessary to perform Ω(n) degree
and neighbor queries.
2
1.2
Overview of the algorithm
For the sake of clarity, we suppress any dependencies on the approximation parameter ǫ and on
log n using the notation O∗ (·).
1.2.1
A simple oracle-based procedure for a 1/3-estimate
First, let us assume access to an P
oracle that, given a vertex v, returns tv , the number of triangles that
v is incident to. Note that t = v tv /3. An unbiased estimate is obtained
by sampling, uniformly
P
at random, a (multi-)set S of s vertices, and outputting tS = 31 · ns · v∈S tv . Yet this estimate can
have extremely large variance (consider the “wheel” graph where there is one vertex with tv = Θ(n)
and all other tv ’s are constant). Inspired by work on estimating the average degree [Fei06, GR08],
we can reduce the variance by simply “cutting off” the contribution of vertices v for which tv is
above a certain threshold. Call such vertices heavy, and denote the remaining light. If the threshold
is set to Θ(t2/3 /ǫ1/3 ), then the number of heavy vertices is O((ǫt)1/3 ). This implies that the total
number of triangles in which all three endpoints are heavy is O(ǫt).
Hence, suppose we define e
tv to be tv if tv ≤ t2/3 /ǫ1/3 and 0 otherwise, and consider e
tS =
1 n P
e
e
t
.
We
can
argue
that
E[
t
]
∈
[(1/3
−
ǫ)t,
t],
since
(roughly
speaking)
every
triangle
·
·
v
S
v∈S
3 s
that contains at least one light vertex is counted at least once. Since e
tv ranges
0 and
between
n
2/3
1/3
∗
t /ǫ , by applying the multiplicative Chernoff bound, a sample of size s = O t1/3 is sufficient
to ensure that with high constant probability e
tS is in the range 1 − 2ǫ · t, (1 + ǫ) · t .
3
1.2.2
Assigning weights to triangles so as to improve the estimate
To improve the approximation, we assign weights to triangles inversely proportional to the number
of their light endpoints (rather than assigning a uniform weight of 31 as is done when defining
P
e
tS = ns · v∈S 31 · e
tv ). If for each light vertex v we let wt(v) be the sum over the weights of all
triangles that
v participates in and for each heavy vertex v we let wt(v) = e
tv = 0, then the expected
n P
value of s · v∈S wt(v) is in [(1 − O(ǫ)) · t, (1 + O(ǫ)) · t].
To get rid of the fictitious oracle, we must resolve two issues. The first P
issue is efficiently deciding
n
whether a vertex is heavy or light, and the second is approximating s · v∈S wt(v), assuming we
have a procedure for deciding whether a vertex is heavy or light. We next discuss each of these two
issues. For convenience, we will assume that the algorithm already has constant factor estimates
for m and t. This can be removed by approximating m and performing a geometric search on t.
1.2.3
Deciding whether a vertex is heavy
Let v be a fixed vertex with degree dv . Consider an edge e incident to v, and let u be the other
endpoint of this edge. Let te denote the number of triangles that e belongs to. Consider the random
variable Y defined by selecting, uniformly at random, a neighbor w of u, and setting Y = du if (v, w)
is an edge (so that (v, u, w) is a triangle) and Y = 0 otherwise. Since the number of neighbors
of u that form a triangle with v is te , the expected value of Y is dteu · du = te . Now consider
selecting (uniformly at random) several edges incident to v, denoted e1 , . . . , er , and for each
edge
1 Pr
ej selected, defining the corresponding random variable Yj . Then the expected value of r j=1 Yj
P
is d1v · e=(v,u) te = d2v · tv . If we multiply by dv /2, then we get an unbiased estimator for tv , which
in particular can indicate whether v is heavy or light.
However, once again the difficulty is with the variance of this estimator and the implication
on the complexity of the resulting decision procedure. To address these difficulties we modify
3
the procedure described above as follows. First, if dv is above a certain
threshold, then v is also
1/3
considered heavy (where this threshold is of order O m/(ǫt)
, so that the total number of
triangles in which all three endpoints are heavy remains O(ǫt). Second, observe that when trying
to estimate the number of triangles that an edge ej = (v, xj ) participates in, we can either select
a random neighbor w of v and check whether (xj , w) ∈ E, or we can select a random neighbor
w of xj and check whether (v, w) ∈ E. Since it is advantageous for the sake of the complexity
to consider the endpoint that has a smaller degree, we do the following. Each time we select an
edge ej = (v, xj ) incident to v, we let uj be the endpoint of ej that has smaller degree. If duj is
√
√
relatively large (larger than m), then we select k = ⌈duj / m⌉ neighbors of uj and let Yj equal
duj times the fraction among these neighbors that close a triangle with ej . The setting of k implies
√
a bound on the variance of Yj (conditioned on the choice of ej ), which is m times its expected
value, tej . Third, in order to bound the variance due to the random choice of edges ej incident to
v, we do the following. We assign each triangle that v participates in to a unique edge incident to
v and modify the definition of te to be the number of such triangles that are assigned to e. The
√
assignment is such that te is always upper bounded by O( m). Finally, we perform a standard
median selection over O(log n) repetitions of the procedure.
Our analysis shows
it suffices to set r (the number of random edges incident to v that are
3/2that
m
so as to ensure the correctness of the procedure (with high probability).
selected) to be O∗ t
In the analysis of the expected query complexity and running time of the procedure we have to
√
take into account the number of iterations k = ⌈duj / m⌉ for each selected (lower degree endpoint)
uj and argue that for every vertex v, the expected number of these iterations is a constant.
1.2.4
Estimating
P
v∈S
wt(v)
n
s
4
·
P
wt(v) is indeed in
[(1− O(ǫ)) ·
n
∗
t, (1 + O(ǫ)) · t] (which we know occurs with high probability if we select s = O t1/3
vertices
uniformly at random). Consider the set of edges incident to vertices in S, where we view edges as
directed, so that if there is an edge between v and v ′ that both belong to S, then (v, v ′ ) and (v ′ , v)
are considered
P as two different edges. We denote this set of edges by ES , and their number by dS ,
where dS = v∈S dv . Suppose that for each edge e = (v, x) we assign a weight wt(e), which is the
sum of the weights of all triangles that v participates in and are assigned to e (where the weight
of
as defined previously based on the number of light endpoints that it has). Then
P a triangle is P
wt(e)
=
v∈S wt(v).
e∈ES
The next idea is to sample edges in ES uniformly at random, and for each selected edge e = (v, u)
to estimate wt(e). An important observation is that since we can query the degrees of all vertices
in S, we can efficiently select uniform random edges in ES (as opposed to the more difficult task
of selecting random edges from the entire graph). Similarly to what was described in the decision
procedure for heavy vertices, given an edge e ∈ ES we let u be its endpoint that has smaller degree.
√
We then select ⌈ m/du ⌉ random neighbors of u and for each check whether it closes a triangle with
e. For each triangle found that is assigned to e, we check how many heavy endpoints it has (using
the aforementioned procedure for detecting heavy vertices) so as to compute the weight
of the
1 P
triangle. In this manner we can obtain random variables whose expected value is dS v∈S wt(v),
√
and whose variance is not too large (upper bounded
3/2by
m times this expected value). We can
now take an average over sufficiently many (O ∗ m t ) such random variables and multiply by
dS · n. By upper bounding the probability that dS is much larger than its expected value we can
prove that the output of the algorithm is as desired. The expected query complexity and running
Suppose we have a (multi-)set S of vertices such that
v∈S
time of the algorithm are shown to be O ∗
n
t1/3
+
m3/2
t
3/2
.
3/2
Finally we note that if t < m1/2 so that m t > m, then we can replace m t with m in the
upper bound on the query complexity since we can store all queried edges so that no edge needs to
be queried more than twice (once from each endpoint).
1.3
A high level discussion of the lower bound
n
queries is fairly
Proving that every multiplicative-approximation algorithm must perform Ω t1/3
n
o
3/2
straightforward, and our main focus is on proving that Ω min m, m t
queries are necessary
as well. In order to prove this claim we define, for every n, every 1 ≤ m ≤ n2 and every
1 ≤ t ≤ min{ n3 , m3/2 }, a graph G1 and a family of graphs G2 for which the following holds:
(1) The graph G1 and all the graphs in G2 have n vertices and m edges. (2) In G1 there are no
triangles, while the number of triangles
in each graph G ∈ G2 is Θ(t). We prove that for values of t
√
m3/2
queries are required in order to distinguish with high constant
such that t ≥ m, at least Ω t
probability between G1 and a random graph in G2 . We then prove that for values of t such that
√
t < m, at least Ω(m) queries are required for this task. We give three different constructions for
G1 and G2 depending on the value of t as a function of m (where two of the constructions are for
√
subcases of the case that t ≥ m). For further discussion of the lower bound, see Section 4.
1.4
1.4.1
Related Work
Approximating graph parameters in sublinear time
We build on previous work on approximating the average degree of a graph and the number of
stars [Fei06, GR08, GRS11]. Feige [Fei06] investigated the problem of estimating the average degree of a graph, denoted d, when given query access
q to thedegrees of the vertices. By performing
n/d/ǫ queries are sufficient in order to oba careful variance analysis, Feige proved that O
tain a ( 21 − ǫ)-approximation of d. He also proved that a better approximation ratio cannot be
achieved in sublinear time using only degree queries. The same problem was considered by Goldreich and
[GR08]. Goldreich and Ron proved that a (1 + ǫ)-approximation can be achieved
q Ron
√
n/ d · poly(log n, 1/ǫ) queries, if neighbor queries are also allowed.
with O
Building on these ideas, Gonen et al. [GRS11] considered the problem of approximating the
number of s-stars in a graph. Their algorithm only used neighbor and degree queries. A major
difference between stars and triangles is that the former are non-induced subgraphs, while the
latter are. Additional work on sublinear algorithms for estimating other graph parameters include
those for approximating the size of the minimum weight spanning tree [CRT05, CS09, CEF+ 05],
maximum matching [NO08, YYI09] and of the minimum vertex cover [PR07, NO08, MR09, YYI09,
HKNO09, ORRR12].
1.4.2
Triangle counting
Triangle counting has a rich history. A classic result of Itai and Rodeh showed that triangles
can be enumerated in O(m3/2 ) time, and a more elegant algorithm was given by Chiba and
Nishizeki [CN85]. The connections to matrix multiplication have been exploited for faster theoretical algorithms [IR78, AYZ97, BPWZ14]. In practice, there is a diverse body on work on
5
counting triangles using different techniques, for different models. There are serial algorithms based
on eigenvalue methods [Tso08, Avr10], graph sparsification [TDM+ 09, KMPT12, TKM11, PT12],
and sampling paths [SW05b, SPK13]. Triangle counters have been given for MapReduce [Coh09,
SV11, KPP+ 13]; external memory models [CC11]; distributed settings [AKM13]; semi-streaming
models [BBCG08, KMPT12]; one-pass streaming [BYKS02, JG05, BFL+ 06, AGM12, KMSS12,
JSP13, PTTW13, TPT13, ADNK14]. It is worth noting that across the board, all these algorithms
required reading the entire graph.
Most relevant to our work are various sampling algorithms, that set up a random variable whose
expectation is directly related to the triangle count [SW05b, KMPT12, JG05, BFL+ 06, SPK13,
JSP13, PTTW13, TPT13, ADNK14]. Typically, this involves sampling some set of vertices or edges
to get a set of three vertices. The algorithm checks whether the sampled set induces a triangle, and
uses the probability of success to estimate the triangle count. We follow the basic same philosophy.
But it is significantly more challenging to set up the “right” random experiment, since we cannot
read the entire graph.
2
Preliminaries
Let G = (V, E) be a simple graph with |V | = n vertices and |E| = m edges. For a vertex v ∈ V ,
we denote by dv the degree of the vertex, by Γv the set of v’s neighbors, and by Ev the set of edges
incident to v. We denote by Tv the set of triangles incident to the vertex v, and let tv = |Tv |.
Similarly, the set of triangles in the graph G is denoted by T , and the number of triangles in the
graph in denote by t. We use c, c1 , . . . to denote sufficiently large constants.
We consider algorithms that can sample uniformly in V and perform three types of queries:
1. Degree queries, in which the algorithm may query for the degree dv of any vertex v of its
choice.
2. Neighbor queries, in which the algorithm may query for the ith neighbor of any vertex v of
its choice. If i > dv , then a special symbol (e.g. †) is returned. No assumption is made on
the order of the neighbors of any vertex.
3. Pair queries, in which the algorithm may ask if there is an edge (u, v) ∈ E between any pair
of vertices u and v.
We sometimes use set notations for operations on multisets. We use the notation O∗ (·) to suppress
dependencies on the approximation parameter ǫ or on log n.
We use the following variant of the multiplicative Chernoff bound. Let χ1 , . . . , χr be r independent random variables, such that χi ∈ [0, B] for some B > 0 and E[χi ] = b for every 1 ≤ i ≤ r. For
every γ ∈ (0, 1] the following holds:
#
" r
2
γ ·b·r
1X
χi > (1 + γ)b < exp −
,
(1)
Pr
r
3B
i=1
and
#
2
r
γ ·b·r
1X
χi < (1 − γ)b < exp −
.
Pr
r
2B
"
(2)
i=1
We will also make an extensive use of Chebyshev’s inequality: For a random variable X and for
γ > 0,
Var[X]
.
Pr [|X − E[X]| ≥ γ] ≤
γ2
6
We fix a total order on vertices denoted by ≺ as follows: u ≺ v if du < dv or du = dv and u < v
(in terms of id number). Given u and v, two degree queries suffice to decide their ordering.
√
Claim 1. Fix any vertex v. The number of neighbors w of v such that v ≺ w is at most 2m.
Proof. P
Let S = {w|w ∈ Γv , v ≺ w}.
√ Naturally, dv ≥ |S|. By definition of ≺, ∀w ∈ S, dw ≥ dv ≥ |S|.
Thus, w∈S dw ≥ |S|2 and |S| ≤ 2m.
3
The Algorithm
We start by introducing the notions of heavy and light vertices and how they can be utilized
in the context of estimating the number of triangles. We then give a procedure for deciding
(approximately) whether a vertex is heavy or light. Using this procedure we give an algorithm for
estimating the number of triangles based on the following assumption (which is later removed).
Assumption 1. Our initial algorithm takes as input estimates t and m on the number of edges
and triangles in the graph respectively, such that
1. t/4 ≤ t ≤ t.
2. m/6 ≤ m.
Assumption 1 can be easily removed by performing a geometric search on t and using the
algorithm from [Fei06] to approximate m, as explained precisely in the proof of Theorem 13.
For every vertex v, we view the set of edges Ev as directed edges originating from v. We then
associate each triangle (v, x, w) ∈ Tv with a unique edge e ∈ Ev , as defined next.
−−−→
Definition 1. We say that a triangle (v, x, w) ∈ Tv is associated with the directed edge (v, x) if
−−−→
−−−→
→
→
x ≺ w, and to (v, w) otherwise. For a directed edge −
e = (v, x) we let T−
e denote the set of triangles
that it is associated with, that is, the set of triangles (v, x, w) such that x ≺ w.
Since it will always be clear from the context from which vertex an edge we consider originates,
for the sake of succinctness, we drop the
Pdirected notation and use the notation Te . We let te = |Te |,
te .
and for a fixed vertex v, we get tv =
e∈Ev
In all the follows we assume that ǫ < 1/2, and otherwise we run the algorithm with ǫ = 1/2.
3.1
Heavy and light vertices
Definition 2. We say that a vertex v is heavy if dv >
dv ≤
2m
(ǫt)1/3
and tv ≤
2/3
t
,
2ǫ1/3
2m
(ǫt)1/3
or if tv >
2/3
2t
ǫ1/3
. If v is such that
then we say that v is light.
We shall say that a partition (H, L) of V is appropriate (with respect to m and t) if every heavy
vertex belongs to H and every light vertex belongs to L.
Note that for an appropriate partition (H, L) both H and L may contain vertices that are
neither heavy nor light (but no light vertex belongs to H and no heavy vertex belongs to L).
For a fixed partition (H, L) we associate with each triangle ∆ a weight depending on the number
of its endpoints that belong to L.
7
Definition 3. For a triangle ∆ we define its weight wtL (∆) to be
(
0
if no endpoints of ∆ belong to L
wtL (∆) =
1/ℓ if ∆ has ℓ > 0 endpoints that belong to L .
Whenever it is clear for the context, we drop the subscript L and use the notation wt(·) instead
of wtL (·).
Claim 2. If (H, L) is appropriate and Assumption 1 holds, then the number of triangles with weight
0 is at most cH · ǫt for some constant cH .
Proof. By Assumption 1, the number of vertices v such that dv is greater than (2m/(ǫt)1/3 , is at
2/3
most 2m/(2m/(ǫt)1/3 ) ≤ 6(ǫt)1/3 , and the number of vertices v such that tv > 2t /ǫ1/3 is at most
1/3
2/3
< 2000ǫt triangles with all three
3t/(2t /ǫ1/3 ) ≤ 6(ǫt)1/3 . Therefore, there are at most 12(ǫt)
3
endpoints in H. Setting cH = 2000 completes the proof.
P
wt(∆). For a vertex v ∈ L we
Definition 4. For any set T of triangles we define wt(T ) =
∆∈T
P
wt(∆), and wt(v) = 0 for v ∈ H.
define wt(v) =
∆∈Tv
P
wt(v) ≤ t. If (H, L) is appropriate and Assumption 1
Lemma 3. For any partition (H, L),
v∈L
P
wt(v) ∈ [t(1 − cH · ǫ), t].
holds, then
v∈L
Proof. Let χ(v, ∆) be an indicator variable such that χ(v, ∆) = 1 if ∆ contains the vertex v, and
χ(v, ∆) = 0 otherwise. Consider a triangle ∆ that contains ℓ > 0 light vertices. Then
X
χ(v, ∆) = ℓ = 1/wt(∆) .
v∈L
If ℓ = wt(∆) = 0, then the above expression equals 0. By interchanging summations,
X
X
X
X X
wt(∆) =
wt(∆)
χ(v, ∆) = t − |{∆ | wt(∆) = 0}|.
wt(v) =
v∈L
∆∈T
v∈L ∆∈Tv
v∈L
Clearly for any partition (H, L) the above expression is at most t. On the other hand, If (H, L) is
appropriate and Assumption 1 holds, then by Claim 2 we have that |{∆ | wt(∆) = 0}| ≤ cH · ǫt,
and the lemma follows.
1/3
Theorem 4. Let s = (c log(n/ǫ)/ǫ3 )n/t
where c is a constant, and let S be a sample of s vertices
v1 , v2 , . . . , vs that are selected uniformly, independently at random. Then
" s
#
1X
t
E
wt(vi ) ≤ .
s
n
i=1
Furthermore, if (H, L) is appropriate and Assumption 1 holds, then
" s
#
1X
E
wt(vi ) ∈ [t(1 − cH · ǫ)/n, t/n]
s
i=1
and for a sufficiently large constant c,
#
" s
1X
wt(vi ) < t(1 − 2cH · ǫ)/n < ǫ2 /n .
Pr
s
i=1
8
s
P
Proof. Let Y denote the random variable Y = 1s
wt(vi ). By the first part of Lemma 3,
i=1
s
P
E 1s
wt(vi ) ≤ t/n. Now assume that (H, L) is appropriate and Assumption 1 holds. The
i=1
claim regarding the expected value of Y follows from the second part of Lemma 3, so it remains to
prove the claim regarding the deviation from the expected value. Note that wt(v) ≤ tv for every
vertex v, which for v ∈ L is at most
Assumption 1,
2/3
2t
ǫ1/3
. By the multiplicative Chernoff bound and by Item 1 in
ǫ2 E[Y ]s
Pr [Y < (1 − ǫ)E[Y ]] < exp − 2/3 1/3
4t /ǫ
< exp −
ǫ2 · c log(n/ǫ)(n/ǫt
4t
2/3
1/3
/ǫ1/3
) · t/(2n)
!
<
ǫ2
,
n
where the last inequality holds for a sufficiently large constant c.
3.2
A procedure for deciding whether a vertex is heavy
In this subsection we provide a procedure for deciding (approximately) whether a given vertex v is
heavy or light. Recall that a high-level description of the procedure appears in Subsection 1.2.3 of
the introduction.
Heavy(v)
1. If dv > 2m/(ǫt)1/3 , output heavy.
2. For i = 1, 2, . . . , 10 log n:
(a) For j = 1, 2, . . . , s = 20m3/2 /ǫ2 t:
i. Select an edge e ∈ Ev uniformly, independently and at random, and let u
be the smaller endpoint according
to the order ≺.
√
ii. For k = 1, 2, . . . , r = ⌈du / m⌉:
A. Pick a neighbor w of u uniformly at random. Let x denote the endpoint
of e that is not v.
B. If e andPw form a triangle and x ≺ w, set Zk = du , else Zk = 0.
iii. Set Yj = 1r
Zk .
k
P
Yj .
(b) Set Xi = dsv
j
3. If the median of the Xi variables is greater than t
light.
2/3
/ǫ1/3 , output heavy, else output
We have three nested loops, with loop variables i, j, k respectively. We refer to these as “iteration
i”, “iteration j”, and “iteration k”.
Lemma 5. For any iteration i, Pr[|Xi − tv | > ǫ · max(tv , tdv /m)] < 1/4.
Proof. Recall that we associate each
P triangle (v, x, w) ∈ Tv with (v, x) if x ≺ w and with (v, w)
otherwise, so that we have tv = e∈Ev te . For an edge e = (v, x), te is upper bounded by the
√
number of neighbors w of x such that x ≺ w. By Claim 1, te ≤ 2m.
Fix an iteration j and let ej denote the edge chosen in the j th iteration and uj denote its smaller
degree endpoint. We use Ej to denote the event of ej being chosen. Conditioned on the event Ej ,
the probability that Zk is non-zero is is tej /duj . Hence,
E[Zk | Ej ] =
te j
· duj = tej ,
duj
9
and
Var[Zk | Ej ] ≤ E[Zk2 | Ej ] ≤ duj · E[Zk | Ej ].
By linearity of expectation,
#
r
r
1X
1X
Zk | Ej =
E [Zk | Ej ] = tej .
E[Yj | Ej ] = E
r
r
"
k=1
(3)
k=1
By the independence of the Zk variables,
" r
#
r
r
1X
1 X
1 X
Var[Yj | Ej ] = Var
Var [Zk | Ej ] ≤ 2
duj · E [Zk | Ej ]
Zk | Ej = 2
r
r
r
k=1
k=1
k=1
√
du
= 2j · r · tej ≤ m · tej .
r
(4)
The conditioning can be removed to yield
E[Yj ] =
X 1
tv
1 X
te = .
· E[Yj | Ej ] =
·
dv
dv
dv
e∈Ev
By the law of total variance, the law of total expectation, the bounds tej ≤
by Equations (3) and (4):
Let Y =
1
s
P
(5)
e∈Ev
√
2m and m ≤ 6m, and
Var[Yj ] = Eej [Var [Yj | Ej ]] + Varej [E [Yj | Ej ]]
h√
i
≤ Eej
m · E [Yj | Ej ] + Varej [tej ]
√
= m · E [Yj ] + Eej [t2ej ]
√
1 X 2
= m · E [Yj ] +
te j
·
dv
ej ∈Ev
√
√
√
≤ m · E [Yj ] + 2m · E[Yj ] < 5 mE[Yj ] .
(6)
Yj . By Equation (5), E[Y ] = tv /dv . By Equation (6),
j
√
s
s
s
X
X
X
√
5 m
1
1
1
Var[Y ] = Var
Var[Yj ] < 2
5 m · E [Yj ] =
Yj = 2
·E
Yj
s
s
s
s
s
j=1
j=1
j=1
j=1
√
√
2
5 m · (tv /dv )
5 m
ǫ tv t
=
=
E Y =
·
·
.
(7)
3/2
2
s
4 dv m
20 · m /ǫ t
s
1X
By Chebyshev’s inequality and Equation (7),
tv t
Var[Y ]
1
tv
Pr Y −
> ǫ · max
,
≤ 2
< .
2
dv
dv m
4
ǫ max(tv /dv , t/m)
Since Xi = dv · Y , we have that Pr[|Xi − tv | > ǫ · max(tv , tdv /m)] < 1/4.
Lemma 6. For every vertex v, if v is heavy, then a call to Heavy(v) returns heavy with probability
at least 1−1/n2 . If v is light, then a call to Heavy(v) returns light with probability at least 1−1/n2 .
10
Proof. First consider a heavy vertex v. Clearly, if dv > 2m/(ǫt)1/3 , then v is declared heavy.
2/3
2/3
Therefore, assume that tv > 2t /ǫ1/3 and dv ≤ 2m/(ǫt)1/3 , so that tdv /m ≤ 2t /ǫ1/3 , and hence
max(tdv /m, tv ) = tv . By Lemma 5, and since ǫ <≤ 1/2, for any iteration i, Pr [|Xi − tv | > ǫtv ] <
2/3
1/4. Therefore, Pr[Xi < t /ǫ1/3 ] < 1/4, and by Chernoff, the probability that the median of the
2/3
Xi variables (where i = 1, . . . , 10 log n) will be greater than t /ǫ1/3 is at least 1 − 1/n2 . Hence
Heavy(v) outputs heavy with probability at least 1 − 1/n2 .
2/3
Now consider a light vertex v. Since dv ≤ 2m/(ǫt)1/3 and tv ≤ t /2ǫ1/3 , it holds that tdv /m ≤
2/3
2/3
2t /ǫ1/3 , implying that max(tdv /m, tv ) = tv ≤ 2t /ǫ1/3 . Therefore, by Lemma 5, Pr[|Xi − tv | >
2/3
2/3
ǫ(2t /ǫ1/3 )] < 1/4, and the probability that the median will be less than t /ǫ1/3 is at least
1 − 1/n2 . Hence v is declared light with probability at least 1 − 1/n2 .
The following is a corollary of Lemma 6.
Corollary 7. Consider running Heavy for all the vertices in the graph. Let H denote the set of
vertices that are declared heavy and let L denote the set of vertices that are declared light. Then,
with probability at least 1 − 1/n, the partition (H, L) is appropriate (as defined in Definition 2).
We now turn to analyze the running time of Heavy. The proof will be similar to the complexity
analysis of the exact triangle counter of Chiba and Nishizeki [CN85].
Lemma 8. If Item 2 in Assumption 1 holds, then for every vertex v the expected running time of
Heavy(v) is O ∗ (m3/2 /t).
Proof. We first argue that the expected time to generate a single sample of Yj is O(1). Our query
√
model allows for selecting an edge in Ev uniformly at random by√a single query. If dv ≤ m,
then the degree of the smaller endpoint for √
any e ∈ Ev is at most m. Hence a sample is clearly
generated in O(1) time.
√ Suppose that dv > m. If an edge e = (v, u) is sampled, then the runtime
is O(1 + min(dv , du )/ m). Hence, the expected runtime to generate Yj is, up to constant factors,
at most:
X
X
1
2m
1
min {dv , du }
1 X
√
du ≤ 1 + √
du ≤ 1 + √
≤ 5,
≤1+ √
1+
dv
m
m
·
d
m
·
d
m
·
d
v
v
v
u∈Γ
u∈Γ
u∈V
v
v
where the last inequality follows from Item 2 in Assumption 1
By the above, each iteration of the ‘for’ loop in Step 2a takes O(1) time in expectation. Therefore, together, all iterations of Step 2a take O(m3/2 /(ǫt)) time in expectation, and since it is repeated
O(log n) times, the expected running time of the procedure is (m3/2 /t) · poly(log n, 1/ǫ).
3.3
Estimating the number of triangles given m and t
We are now ready to present an algorithm Estimate-with-advice that takes m, t as input (“advice”), and outputs an estimate of t. Later, we employ the the average degree approximation
algorithm of Feige [Fei06] and a geometric search to get the bonafide algorithm that estimates t
without any initial estimates m and t. Recall that a high-level description of the procedure appears
in Subsection 1.2.4 of the introduction. In what follows we rely on the following assumption.
Assumption 2. We will assume that the random coins used by Heavy are fixed in advance, and
that the partition (H, L) as defined in Corollary 7 is indeed appropriate.
By Corollary 7 this assumption only adds 1/n to the error probability in all subsequent probability bounds. Recall that we use c, c1 , . . . to denote sufficiently large constants.
Recall that cH is the constant defined in Claim 2.
11
Estimate-with-advice(m, t, ǫ)
1/3
1. Sample s1 = c1 ǫ−3 log(n/ǫ)(n/t ) vertices, uniformly, independently and at random. Denote the chosen multiset S.
2. Set up a data structure to enable sampling vertices in S proportional to their degree.
3. For i = 1, 2, . . . , s2 = c2 ǫ−4 (log2 n)(m3/2 /t):
(a) Sample v ∈ S proportional to dv and sample e ∈ Ev uniformly at random. Let
u be the smaller endpoint according to the order ≺. Let x be the endpoint of
e that is√not v.
√
(b) If du √
≤ m, set r = 1√with probability du / m and set r = 0 otherwise. If
du > m, set r = ⌈du / m⌉.
(c) Repeat for j = 1, 2, . . . , r:
i. Pick a neighbor w of u uniformly at random.
ii. If e and w do not form a triangle, set Zj = 0.
iii. If e and w form a triangle and w ≺ x, set Zj = 0.
iv. If e and w form a triangle ∆ and x ≺ w: call Heavy for all vertices in ∆,
and let (
0
if Heavy(v) returned heavy
√
.
Zj =
max(du , m) · wt(∆) otherwise
r
P
(d) Set Yi = 1r
Zj . (If r = 0, set Yi = 0.)
j=1
s
2
P
P
n
dv ·
Yi .
4. Output X = s1 s2 ·
v∈S
i=1
Theorem 9. For X as defined in Step 4 of Estimate-with-advice, E[X] ≤ t. Moreover, if (H, L)
is appropriate and Assumption 1 holds, then E[X] ∈ [t(1 − cH · ǫ), t] and Pr[X < t(1 − 3cH · ǫ)] <
3ǫ/ log n.
There are three “levels” of randomness. First is the choice of S, second is the choice of e
(Step 3a), and finally the Zj ’s. When analyzing the randomness in any level, we condition on the
previous levels. Before proving the theorem, we present the following definition and claim.
P
wt(v)/s1 ≥ t(1 −
Definition 5. Let S be a multiset of s1 vertices. We say that S is good if
v∈S
P
dv ≤ s1 (2m/n)(log n/ǫ).
2cH · ǫ)/n. We say that S is great if, in addition to being good, dS =
v∈S
P
P
wt(v)
dv . For every i, E[Yi | S] = d−1
Claim 10. Fix the choice of the set S, and let dS =
S
v∈S
v∈S
√
and Var[Yi | S] < 5 m · E[Yi | S].
Proof. This is similar to the argument in Lemma 5. Let vi be the chosen vertex in the ith iteration of
the algorithm, and let ei be the chosen edge. We refer to this event by Ei , and condition over the set
S being chosen and the event Ei . Denote by ui the lower degree endpoint of ei . If Heavy(vi )=heavy,
then E[Y√
i | S, Ei ] = 0 and Var[Yi | S, Ei ] = 0. If Heavy(vi )=light, then there are two possibilities.
If dui ≤ m then,
du X 1 √
· m · wt(∆) = wt(Tei ) .
E[Yi | S, Ei ] = √ i
m ∆∈T dui
ei
12
Since the maximum value of Yi in this case is at most
√
m,
√
(8)
Var[Yi | S, Ei ] ≤ E[Yi2 | S, Ei ] ≤ m · E[Yi | S, Ei ] .
√
Now consider the case that dui > m. In order to bound the variance of the Yi variables we first
analyze the expectation and variance of the Zj variables. Note that Zj is non-zero when a triangle
∆ ∈ Tei is found. It holds that
E[Zj | S, Ei ] =
X
∆∈Tei
1
· dui · wt(∆) = wt(Tei ),
dui
and
Var[Zj | S, Ei ] ≤ dui · E[Zj | S, Ei ].
(9)
By linearity of expectation,
E[Yi | S, Ei ] = wt(Tei ).
By independence of the (Zj | S, Ei ) variables, linearity of expectation and Equation (9),
r
r
r
X
1 X
1 X
1
Var [Zj | S, Ei ] ≤ 2
dui · E [Zj | S, Ei ]
Zj | S, Ei = 2
Var[Yi | S, Ei ] = Var
r
r
r
j=1
j=1
j=1
r
X
√
du
1
= i · E
Zj | S, Ei ≤ m · E[Yi | S, Ei ].
(10)
r
r
j=1
We remove the conditioning on Ei :
E[Yi | S] =
X
X X
X dv 1 X
−1
wt(v) .
wt(T
)
=
d
wt(Te ) = d−1
·
e
S
S
dS dv
v∈S∩L
v∈S
v∈S∩L e∈Ev
e∈Ev
√
Recall that by Claim 1, wt(e) ≤ 2m. Therefore, by the law of total variance, the law of total
expectation, the bound m ≤ 6m, and Equations (8) and (10),
Var[Yi | S] = Eei [Var [Yi | S, Ei ]] + Varei [E [Yi | S, Ei ]]
i
h√
m · E [Yi | S, Ei ] + Eei [wt(Tei )2 ]
≤ Eei
√
√
√
m · E[Yi | S] + 2mEei [wt(Tei )] < 5 m · E[Yi | S] .
This completes the proof of Claim 10.
Proof of Theorem 9: For a fixed set S, let XS denote the sum XS =
n
s1 s2
P
v∈S
s
2
P
dv ·
Yi
i=1
(as defined in Step 4 of Estimate-with-advice), given that the set S in chosen in Step 1. By the
definition of XS and by Claim 10,
E[XS ] =
ndS
n X
E[Yi | S] =
wt(v).
s1
s1
v∈S
By Theorem 4, ES
1
s1
P
v∈S
wt(v) ∈ [t(1 − cH · ǫ)/n, t/n], implying that
E[XS ] ∈ [t(1 − cH · ǫ), t].
13
(11)
By Theorem 4, Definition 5 and Assumption 2, S
expected value, over S, of dS is Es [dS ] = s1 · 2m
n .
2m
PrS dS > s1 ·
n
is good with probability at least 1 − ǫ2 /n. The
By Markov’s inequality,
log n
ǫ
·
.
<
ǫ
log n
By taking a union bound, the probability that S is great is at least 1 − 2ǫ/ log n. For a fixed choice
s2
P
Yi . By the independence of the Yi variables and by Claim 10,
of S, let YS = s12
i=1
√
s2
s2
√
1 X
5 m
1 X
Var [Yi | S] < 2
5 m · E[Yi | S]=
Var [YS ] = 2
· E [YS ] .
s2
s2 i=1
s2 i=1
(12)
By Chebyshev’s inequality, the setting of s2 and Equation (12), we get that
√
5 m · E[YS ]
Var[YS ]
≤ 2
Pr |YS − E[YS ]| > ǫE[YS ] < 2
ǫ · E[YS ]2
ǫ (c2 ǫ−4 log2 n)(m3/2 /t) · E[YS ]2
ǫ2
.
=
c2 (log2 n)(m/t) · E[YS ]
P
wt(v), which for a great S is at least
By Claim 10, E[YS ] = d−1
S
v∈S
t(1 − 2cH · ǫ)/n)/s1
t
ǫ
≥
·
.
s1 (2m/n)(ǫ/ log n)
4m log n
Therefore, by Assumption 1, for a sufficiently large constant c2 ,
ǫ
.
log n
Pr [|YS − E[YS ]| > ǫE[YS ]] ≤
By the definition of XS in Step 4 of the algorithm, XS is just a scaling of YS . Therefore,
Pr [|XS − E[XS ]| > ǫE[XS ]] ≤
By Equation (11), E[XS ] =
great S,
n
s1
P
v∈S
ǫ
.
log n
wt(v), which for a great S is at least t(1 − 2cH · ǫ). Hence, for a
Pr [XS < (1 − 3cH · ǫ) · t] ≤
ǫ
.
log n
The probability of S not being great is at most 2ǫ/ log n. We apply the union bound to remove the
conditioning, so we get
3ǫ
,
Pr [X < (1 − 3cH · ǫ) · t] ≤
log n
which completes the proof.
Theorem 11. If Item 2 in Assumption 1 holds then the expected running time of Estimate-with-advice
1/3
is O∗ (n/t + m3/2 /t).
14
1/3
Proof. The sampling of S is done in O∗ (n/t ) time. Generating the Zj variables, without the
calls to Heavy, takes time O∗ (m3/2 /t) in expectation, by an argument identical to that in the proof
of Lemma 8. Therefore, it remains to bound the running time resulting from calls to Heavy.
Let us compute the expected number of triangles found during the run of the algorithm. In each
iteration
√ choosing an edge e, the expected number of triangles found is at most
√ i, conditioned on
2(du / m)(te /du ) = 2te / m. Averaging
over the edges, the expected number of triangles found in
√
a single iteration is at most 6t/(m · m), which by Item 2 in Assumption 1 is O(t/m3/2 ). There are
O(m3/2 /t) · poly(log n, 1/ǫ) iterations, leading to a total of O∗ (1) expected triangles. Thus, there
are O ∗ (1) expected calls to Heavy, each taking O ∗ (m3/2 /t) time by Lemma 8. Together with the
1/3
above, we get an expected running time of O(n/t + m3/2 /t) · poly(log n, 1/ǫ).
3.4
The final algorithm
We are now ready to present an algorithm that requires no prior knowledge regarding m and t.
Estimate(ǫ)
1. Let ǫ′ = ǫ/3cH , where cH is the constant defined in Claim 2.
2. Invoke Feige’s algorithm [Fei06] for approximating the average degree of a graph
10 log n times. Let d be the median value of all invocations.
3. Let m = nd/2.
4. Let e
t = n3 .
5. While e
t≥1
t:
(a) For t = n3 , n3 /2, n3 /4, . . . , e
i. For i = 1, . . . , cǫ−1 log log n:
A. Let Xi =Estimate-with-advice(ǫ′, m, t).
ii. Let X = mini {Xi }.
iii. If X ≥ t return X.
(b) Let e
t=e
t/2.
Before analyzing the correctness and running time of the algorithm, we present the following
simple proposition, whose proof we give for the sake of completeness.
Proposition 12. For every graph G, t ≤ 43 m3/2 .
Proof.
1X1
1
t=
tv ≤
3
2
6
v∈V
X
√
v: dv > m
tv +
X
√
v: dv ≤ m
√
√
1
2d2v ≤ 2 m · 2m + 2 m
6
X
√
v: dv ≤ m
4
dv ≤ m3/2 .
3
Theorem 13. Algorithm Estimate(ǫ) returns a value X, such that (1 − ǫ)t ≤ X ≤
(1 + ǫ)t, with probability
at least 5/6. The expected query complexity of the algorithm is
O∗ n/t1/3 + max m, m3/2 /t
and the expected running time of the algorithm is O ∗ (n/t1/3 +
m3/2 /t).
Proof. We first prove that the value of X is as stated in the theorem. Let davg denote the average
degree of vertices in G. The algorithm from [Fei06] returns a value d such that, with probability
at least 2/3, d ∈ [davg /(2 + γ), davg ] for a constant γ. Since we take the median value of 10 log n
15
invocations, it follows from Chernoff’s inequality that m is as stated in Item 2 of Assumption 1
with probability at least 1 − 1/poly(n). Assume that this is indeed the case.
Before analyzing the algorithm Estimate as described above, first consider executing Step 5a
with e
t = 1. That is, rather than running both an outer loop over decreasing values of e
t and an
inner loop over decreasing values of t, we only run a single loop over decreasing value of t, starting
with t = n3 . By the first part of Theorem 9 and by Markov’s inequality, for each value of t and for
each i, Pr[Xi ≤ (1 + ǫ)t] > ǫ/2, where Xi as defined in Step 5(a)iA. Therefore, for each value of t,
the minimum estimate X (as defined in Step 5(a)iii) is at most (1 + ǫ)t, with probability at least
1 − 1/ log 3 n. It follows that for each t such that t > 2t, we have that X < t with probability at
least 1 − 1/ log 3 n, and the algorithm will continue with t = t/2. Once we reach a value of t for
which t/4 ≤ t ≤ t/2, Item 1 in Assumption 1, regarding t, holds. By the second part of Theorem 9,
Xi ∈ [(1 − ǫ)t, (1 + ǫ)t] for every i with probability at least 1 − c/ log n. Hence, we have that
t≤
1
t ≤ (1 − ǫ)t ≤ X ≤ (1 + ǫ)t,
2
with probability at least 1 − c/ log n. Therefore, we halt and return correct X.
If however we do reach a value t such that t ≤ t/4, since Assumption 1 does not hold, we cannot
lower bound X, implying that we can no longer bound the probability that X < t. Therefore we
might continue running with decreasing values of t, causing the running time to exceed the desired
bound of O∗ (n/t1/3 + m3/2 /t). In order to avoid this scenario, we run both an outer loop over e
t and
3
e
e
an inner loop over t. Specifically, starting with t = n , whenever we halve t, we run over all values
of t = n3 , n3 /2, . . ., until we reach e
t. This implies that for every value of e
t > 2t the probability of
returning an incorrect estimate, that is, outside the range of (1 − ǫ)t ≤ X ≤ (1 + ǫ)t, is at most
1 − 1/ log 2 n. On the other hand, for values of e
t such that e
t ≤ t/2 the probability of returning a
correct estimate (within (1 − ǫ)t ≤ X ≤ (1 + ǫ)t) is at least 1 − c/ log n. A union bound over all
failure probabilities gives a success probability of at least 5/6.
We now turn to analyze the query complexity and running time of the algorithm. By [Fei06],
√
the expected running time of the average degree approximation algorithm is O ∗ (n/ m). By
Theorem 11, conditioned on m satisfying Item 2 in Assumption 1, the expected running time
1/3
of Estimate-with-advice(ǫ, m, t) is O∗ (n/t
+ m3/2 /t). It follows from Proposition 12 that
√
n/ m = O(n/t1/3 ), implying that the running time is determined by the value of m and by the
smallest value of t that Estimate-with-advice(ǫ, m, t) is invoked with.
Recall that whenever we halve the value of t̃, we run with all values t = n3 , n3 /2, . . .. This,
together with the fact that when running with t/4 ≤ t ≤ t/2 we halt with probability at least
1−c/ log n, implies that the probability of reaching a value t̃ = t/2k is at most (c/ log n)k . Therefore,
the expected running time, conditioned on m satisfying Item 2 in Assumption 1, is bounded by
! log n
!
!
X
m3/2
m3/2
m3/2
n
n
n
2
∗
∗
k
k
∗
log n · O
+
+
+
+
=O
.
(c/ log n) · 2 · O
t
t
t
t1/3
t1/3
t1/3
k=1
Now consider the value of m computed in Step 3 of Estimate(ǫ). As stated previously, with
probability at least 1 − 1/poly(n) (e.g., 1 − 1/n4 ), the estimate m is within a constant factor from
m. Therefore the expected running time of the algorithm (without the conditioning on the value
of m) is bounded by
!
!
3/2
3/2
m
m
1
n
n
1
+ 4 · O(n3 ) = O ∗ 1/3 +
.
1 − 4 · O ∗ 1/3 +
n
t
n
t
t
t
16
Observe that we can always assume that the algorithm does not perform queries it can answer by
itself. That is, we can allow the algorithm to save all the information it obtained from past queries,
and assume it does not query for information it can deduce from its past queries. Further observe
that any pair query is preceded by a neighbor query. Therefore, if at any point the algorithm
performs more than 2m queries, it can abort. It follows that the expected query complexity is
O∗ (n/t1/3 + min{m, m3/2 /t}).
4
A Lower Bound
In this section we present a lower bound on the number of queries necessary for estimating the
number of triangles in a graph. Since we sometimes refer to the number of triangles in different
graphs, we use the notation t(G) for the number of triangles in a graph G. Our lower bound matches
our upper bound in terms of the dependence on n, m and t(G), up to polylogarithmic factors in
n and the dependence in 1/ǫ. In what follows, when we refer to approximation algorithms for the
number of triangles in a graph, we mean multiplicative-approximation algorithms that output with
high constant probability an estimation b
t such that t(G)/C ≤ b
t ≤ C · t(G) for some predetermined
approximation factor C.
We consider multiplicative-approximation algorithms that are allowed the following three types
of queries: Degree queries, pair queries and random new-neighbor queries. Degree queries and pair
queries are as defined in Section 2. A random new-neighbor query qi is a single vertex u and the
corresponding answer is a vertex v such that (u, v) ∈ E and the edge (u, v) is selected uniformly
at random among the edges incident to u that have not yet been observed by the algorithm. In
Corollary 34 we show that this implies a lower bound when the algorithm may perform (standard)
neighbor queries instead of random new-neighbor queries.
We first give a simple lower bound that depends on n and t(G).
Theorem 14. Any
algorithm for the number of triangles in a graph
multiplicative-approximation
n
must perform Ω t(G)1/3 queries, where the allowed queries are degree queries, pair queries and
random new-neighbor queries.
Proof. For every n and every 1 ≤ t ≤ n3 we next define a graph G1 and a family of graphs G2 for
which the following holds. The graph G1 is the empty graph over n vertices.
In G2 , each graph
1/3
1/3
consists of a clique of size t
and an independent set of size n − t
. See Figure 1 for an
illustration. Within G2 the graphs differ only in the labeling of the vertices. By construction, G1
contains no triangles and each graph in G2 contains Θ(t) triangles. Clearly, unless the algorithm
“hits” a vertex in the clique it cannot distinguish between the two cases.
The probability of hitting
1/3 /n. Thus, in order for this
such a vertex in a graph selected uniformly at random
from
G
is
t
2
event to occur with high constant probability, Ω
n
t1/3
queries are necessary.
We next state our main theorem.
Theorem 15. Any multiplicative-approximation
algorithm for the number of triangles in a graph
n 3/2 o
m
must perform at least Ω min t(G) , m
queries, where the allowed queries are degree queries,
pair queries and random new-neighbor queries.
For every n, every 1 ≤ m ≤ n2 and every 1 ≤ t ≤ min n3 , m3/2 we define a graph G1 and
a family of graphs G2 for which the following holds. The graph G1 and all the graphs in G2 have
n vertices and m edges. For the graph G1 , t(G1 ) = 0, and for every graph G ∈ G2 , t(G) = Θ(t).
17
n
∆1/3
n − ∆1/3
Figure 1: An illustration of the two families.
n 3/2 o
We prove it is necessary to perform Ω min m t , m
queries in order to distinguish with high
constant probability between G1 and a random graph in G2 . For the sake of simplicity, in everything
√
that follows we assume that m is even.
1√
We prove that for values of t such that
t
<
3/2 4 m, at least Ω(m) queries are required, and for
√
values of t such that t ≥ m at least Ω m t
queries are required.
We delay the discussion on the former case to Subsection 4.4, and start with the case that
√
t ≥ m. Our construction of G2 depends on the value of t as a function of m where we deal
separately with the following two ranges of t:
1. t ∈ [Ω(m), O(m3/2 )].
√
2. t ∈ [Ω( m, O(m)].
We prove that for every t as above, Ω(m3/2 /t) queries are needed in order to distinguish between
the graph G1 and a random
graph in G2 . Observe that by Proposition 12, for every graph G, it
holds that t(G) = O m3/2 . Hence, the above ranges indeed cover all the possible values of t as a
function of m.
A high level discussion of the lower bound. The constructions for the different ranges of
√
t ≥ m are all based on the same basic idea, and have the following in common. In all construction
√
for t as above, G1 consists of a complete bipartite graph (L ∪ R, E) with |L| = |R| = m and an
√
independent set of n − 2 m vertices. The basic structure of the graphs in the family G2 is the same
as that of G1 with the following modifications:
√
• For every value of t, we add t/ m edges between vertices in L (and similarly in R). Since
√
each edge contributes (roughly) m triangles, this gives the desired total number of triangles
in the graph. In the case that t = m this is done by adding a perfect matching within L and
a perfect matching within R. In the case that t > m we add several such perfect matchings,
√
√
and in the case that m ≤ t ≤ m/4 we add a (non-perfect) matching of size t/ m.
• In order to maintain the degrees of all the vertices in the bipartite component, we remove
edges between vertices in L and R.
For an illustration of the case t = m, see Figure 2. In what follows we assume that the algorithm
knows in advance which vertices are in L and which are in R, and consider only the bipartite
component of the graphs. In order to give the intuition for the m3/2 /t lower bound we consider
each type of query separately, starting with degree queries.
18
Since both in the graph G1 and in all the graphs in G2 , all the vertices in L ∪ R have the
√
same degree (of m), degree queries do not reveal any information that is useful for distinguishing
between the two.
As for pair queries, unless the algorithm queries a pair in L × L (or R × R) and receives a
positive answer, or queries a pair in L × R and receives a negative answer, the algorithm cannot
distinguish between the bipartite component of the graph G1 and those of the graphs in G2 . We
√
refer to these pairs as witness pairs. Roughly speaking, since there are Θ(t/ m) such pairs, and
m pairs in total, it takes Ω(m3/2 /t) queries in order to “catch a witness pair”.
We are left to deal with neighbor queries. Here too, distinguishing between the graph G1 and
the graphs in G2 can be done by “catching a witness”. That is, if the algorithm queries for a
neighbor of a vertex in L and the answer is another vertex in L (analogously for a vertex in R).
As before, the probability for hitting such a witness pair is small. However, there is another source
of difference resulting from neighbor queries. When the algorithm queries a vertex v ∈ L there is
a difference in the conditional distribution on answers v ∈ R when the answer is according to the
graph G1 or according to a graph in the family G2 . The reason for the difference, is that in the
√
graph G1 every vertex has exactly m neighbors in the opposite side, while for graphs in G2 , each
√
√
vertex has Θ( m − t/m) neighbors in the opposite side (for the range Ω( m) ≤ t ≤ O(m) this is
true on average). We prove that this difference in sufficiently small so as to ensure the Ω(m3/2 /t)
lower bound.
Our formal analysis is based on defining two processes that interact with an algorithm for
approximating the number of triangles, denoted ALG. The first process answer queries according
to G1 , and the second process answers queries while constructing a uniformly selected graph in
G2 . An interaction between ALG and each of these processes induces a distribution over sequences
of queries and answers. We prove that if the number of queries performed by ALG is smaller
than m3/2 /(ct) for a sufficiently large constant c, then the statistical distance between the two
distributions is a small constant.
We start by addressing the case that t = m in Subsection 4.1, and deal with the case that
√
3/2
m < t ≤ m8 in Subsection 4.2, and with the case that m ≤ t ≤ m
4 in Subsection 4.3.
Before embarking on the proof for t = m, we introduce the notion of a knowledge graph (as
defined previously in e.g., [GR02]), which will be used in all lower bound proofs. Let ALG be an
algorithm for approximating he number of triangles, which performs Q queries. Let qt denote its tth
query and let at denote the corresponding answer. Then ALG is a (possibly probabilistic) mapping
from query-answer histories π , h(q1 , a1 ), . . . , (qt , at )i to qt+1 , for every t < Q, and to N for t = Q.
We assume that the mapping determined by the algorithm is determined only on histories that
are consistent with the graph G1 or one of the graphs in G2 . Any query-answer history π of length
kn
t can be used to define a knowledge graph Gkn
π at time t. Namely, the vertex set of Gπ consists
of n vertices. For every new-neighbor query ui answered by vi for i ≤ t, the knowledge graph
contains the edge (ui , vi ), and similarly for every pair query (uj , vj ) that was answered by 1. In
addition, for every pair query (ui , vi ) that is answered by 0, the knowledge graph maintains the
information that (ui , vi ) is a non-edge. The above definition of the knowledge graph is a slight
abuse of the notation of a graph since Gkn
π is a subgraph of the graph tested by the algorithm, but
it also contains additional information regarding queried pairs that are not edges. For a vertex u,
kn
kn
we denote its set of neighbors in the knowledge graph by Γkn
π (u), and let dπ (u) = Γπ (u) . We
denote by Nπkn (u) the set of vertices v such that (u, v) is either an edge or a non-edge in Gkn
π .
19
4.1
A lower bound for t = m
4.1.1
The lower-bound construction
√
The graph G1 has two components. The first component is a complete bipartite graph with m
√
vertices on each side, i.e, K√m,√m , and the second component is an independent set of size n−2 m.
We denote by L the set of vertices ℓ1 , . . . , ℓ√m on the left-hand side of the bipartite component and
by R the set of vertices r1 , . . . , r√m on its right-hand side. The graphs in the family G2 have the
same basic structure with a few modifications. We first choose for each graph a perfect matching
M C between the two sides R and L and remove the edges in M C from the graph. We refer to the
removed matching as the “red matching” and its pairs as “crossing non-edges” or “red pairs”. Now,
we add two perfect matching from L to L and from R to R, denoted M L and M R respectively.
We refer to these matchings as the blue matchings and their edges as “non-crossing edges” or “blue
pairs”. Thus for each choice of three perfect matchings M C , M L and M R as defined above, we
have a corresponding graph in G2 .
√
Consider a graph G ∈ G2 . Clearly, every blue edge participate in m − 2 triangles. Since, every
√
√
triangle in the graph contains exactly one blue edge, there are 2 m · ( m − 2) = Θ(m) triangles
in G.
√
√
n−2 m
m
Figure 2: An illustration of the family G2 for t = m.
4.1.2
Definition of the processes P1 and P2
In what follows we describe two random processes, P1 and P2 , which interact with an arbitrary
algorithm ALG. The process P1 answers ALG’s queries consistently with G1 . The process P2
answers ALG’s queries while constructing a uniformly selected random graph from G2 . We assume
without loss of generality that ALG does not ask queries whose answers can be derived from its
knowledge graph, since such queries give it no new information. For example, ALG does not ask
a pair query about a pair of vertices that are already known to be connected by an edge due to a
neighbor query. Also, we assume ALG knows in advance which vertices belong to L and which to
to R, so that ALG need not query vertices in the independent set. Since the graphs in G2 differ
from G1 only in the edges of the subgraph induced by L ∪ R, we think of G1 and graphs in G2 as
consisting only of this subgraph. Finally, since in our constructions all the vertices in L ∪ R have
√
the same degree of m, we assume that no degree queries are performed.
For every, Q, every t ≤ Q and every query-answer history π of length t − 1 the process P1
answers the tth query of the algorithm consistently with G1 . Namely:
• For a pair query qt = (u, v) if the pair (u, v) is a crossing pair in G1 , then the process replies
1, and otherwise it replies 0.
20
• For a random new-neighbor query qt = u the process answers with a random neighbor of u that
has yet been observed by the algorithm. That is, for every vertex v such that v ∈ Γ(u)\Γkn
π (u)
√
the process replies at = v with probability 1/( m − dkn
(u)).
π
The process P2 is defined as follows:
• For a query-answer history π we denote by G2 (π) ⊂ G2 the subset of graphs in G2 that are
consistent with π.
• For every t ≤ Q and every query-answer history π of length t − 1, the process P2 selects a
graph in G2 uniformly at random and answers the tth query as follows.
1. If the tth query is a pair query qt = (u, v), then P2 answers the query qt according to the
selected graph.
2. If the tth query is a random new-neighbor query qt = ut , then P2 ’s answer is a uniform
new neighbor of ut in the selected graph.
• After all queries are answered (i.e., after Q queries), uniformly choose a random graph G
from G2 (π).
For a query-answer history π of length Q we denote by π ≤t the length t prefix of π and by π ≥t
the Q − t + 1 suffix of π.
We note that the selected graph is only used to answer the tth query and is then “discarded
back to” the remaining graphs that are consistent with that answer (and all previous answers in
π).
Claim 16. Let π be a query-answer history of length t − 1. We use ◦ to denote concatenation.
• If the tth query is a pair query, then at = 1 with probability
|G2 (π ◦ (qt , 1))|
,
|G2 (π)|
and at = 0 with probability
|G2 (π ◦ (qt , 0))|
.
|G2 (π)|
• If the tth query is a random new-neighbor query qt = ut , then for every v ∈ V \ Γkn
π (u) the
probability that the process P2 answers at = v is
1
|G2 (π ◦ (qt , v))|
·√
.
|G2 (π)|
m − dkn
π (ut )
If v ∈ Γkn
π (u) then the probability that P2 answers at = v is 0.
Proof. First consider a pair query qt = (ut , vt ). The probability that (ut , vt ) is an edge in the graph
chosen by the process P2 is the fraction of graphs in G2 (π) in which (ut , vt ) is an edge. This is
t ,1))|
. Similarly, the probability of choosing a graph in which (ut , vt ) is not an edge
exactly |G2 (π◦(q
|G2 (π)|
is
|G2 (π◦(qt ,0))|
.
|G2 (π)|
Now consider a random new-neighbor query qt = ut . We start with the case that v ∈ V \ Γkn
π .
The probability that v is chosen by P2 is the probability that a graph G in which v is a neighbor
of ut is chosen in the first step, and that v is the chosen new neighbor among all of u’s neighbors
21
in the second step. Since there are |G2 (π ◦ (u, v))| graphs in which v is a neighbor of ut , and ut has
√
m − dkn
π (ut ) neighbors, this happens with probability
|G2 (π ◦ (u, v))|
1
·√
.
|G2 |
m − dkn
π (ut )
For a vertex v such that v ∈
/ V \ Γkn
π , in every graph G ∈ G2 , v is not a neighbor of ut , implying
that the probability that the process replies at = v is 0.
Lemma 17. For every algorithm ALG, the process P2 , when interacting with ALG, answers ALG’s
queries according to a uniformly generated graph G in G2 .
Proof. Consider a specific graph G ∈ G2 . Let π be the query-answer history generated by the
interaction between ALG and P2 . Let Q be the number of queries performed during the interaction.
The probability that G is the resulting graph from that interaction is
1
Pr[G ∈ G2 (π ≤1 )] · Pr[G ∈ G2 (π ≤2 ) | G ∈ G2 (π ≤1 )] · . . . · Pr[G ∈ G2 (π ≤Q )|G ∈ G2 (π ≤Q−1 )] ·
|G(π ≤Q )|
|G2 (π ≤1 )| |G2 (π ≤2 )|
|G2 (π ≤Q )|
1
1
=
·
·
.
.
.
·
·
=
,
|G2 |
|G2 (π ≤1 )|
|G2 (π ≤Q−1 )| |G2 (π Q )|
|G2 |
and the lemma follows.
b
For a fixed algorithm ALG that performs Q queries, and for b ∈ {1, 2}, let DALG
denote the
distribution on query-answers histories of length Q induced by the interaction between ALG and
3/2
Pb . We shall show that for every algorithm ALG that performs at most
Q = m
100t queries, the
statistical distance between D1ALG and D2ALG , denoted d D1ALG , D2ALG , is at most 31 . This will
imply that the lower bound stated in Theorem 15 holds for the case that t(G) = m. In order to
obtain this bound we introduce the notion of a query-answer witness pair, defined next.
Definition 6. We say that ALG has detected a query-answer witness pair in three cases:
1. If qt is a pair query for a crossing pair (ut , vt ) ∈ L × R and at = 0.
2. If qt is a pair query for a non-crossing pair (ut , vt ) ∈ (L × L) ∪ (R × R) and at = 1.
3. If qt = ut is a random new-neighbor query and at = v for some v such that (ut , v) is a
non-crossing pair.
We note that the source of the difference between D1ALG and D2ALG is not only due to the
probability that the query-answer history contains a witness pair (which is 0 under D1ALG and
non-0 under D2ALG ). There is also a difference in the distribution over answers to random new
neighbor queries when the answers do not result in witness pairs (in particular when we condition
on the query-answer history prior to the tth query). However, the analysis of witness pairs serves
us also in bounding the contribution to the distance due to random new neighbor queries that do
not result in a witness pairs.
Let w be a “witness function”, such that for a pair query qt on a crossing pair, w(qt ) = 0, and
for a non-crossing pair, w(qt ) = 1. The probability that ALG detects a witness pair when qt is a
pair query (ut , vt ) and π is a query-answer history of length t − 1, is
PrP2 [w(qt ) | π] =
|G2 (π ◦ (qt , w(qt )))|
|G2 (π ◦ (qt , w(qt )))|
≤
.
|G2 (π)|
|G2 (π ◦ (qt , w(qt )))|
Therefore, to bound the probability that the algorithm observes a witness pair it is sufficient to
bound the ratio between the number of graphs in G2 (π ◦ (q, w(qt ))) and the number of graphs in
G2 (π ◦ (q, w(qt ) )). We do this by introducing an auxiliary graph, which is defined next.
22
4.1.3
The auxiliary graph for t = m
For every t ≤ Q, every query-answer history π of length t − 1 for which π is consistent with G1 (that
is, no witness pair has yet been detected), and every pair (u, v), we consider a bipartite auxiliary
graph Aπ,(u,v) . On one side of Aπ,(u,v) we have a node for every graph in G2 (π) for which the pair
(u, v) is a witness pair. We refer to these nodes as witness graphs. On the other side of the auxiliary
graph, we place a node for every graph in G2 (π) for which the pair is not a witness. We refer
to these nodes as non-witness graphs. We put an edge in the auxiliary graph between a witness
graph W and a non-witness graph W if the pair (u, v) is a crossing (non-crossing) pair and the two
graphs are identical except that their red (blue) matchings differ on exactly two pairs – (u, v) and
one additional pair. In other words, W can be obtained from W by performing a switch operation,
as defined next.
Definition 7. We define a switch between pairs in a matching in the following manner. Let (u, v)
and (u′ , v ′ ) be two matched pairs in a matching M . A switch between (u, v) and (u′ , v ′ ) means
removing the edges (u, v) and (u′ , v ′ ) from M and adding to it the edges (u, v ′ ) and (u′ , v).
Note that the switch process maintains the cardinality of the matching. We denote by dw (Aπ,(u,v) )
the minimal degree of any witness graph in Aπ,(u,v) , and by dnw (Aπ,(u,v) ) the maximal degree of the
non-witness graphs. See Figure 3 for an illustration.
edge in
Aπ,(u,v)
dw
witness
graphs
non-witness ′
u
graphs
v′
u′
v′
u
v
u
v
dw
A witness graph W
(a) The auxiliary graph with witness nodes on
the left and non-witness nodes on the right.
W – A neighbor of W
(b) An illustration of two neighbors in the auxiliary graph
for t = m.
Figure 3
3/2
Lemma 18. Let t = m and Q = m
100t . For every t ≤ Q, every query-answer history π of length
t − 1 such that π is consistent with G1 and every pair (u, v),
dnw (Aπ,(u,v) )
2
2t
≤ √ = 3/2 .
dw (Aπ,(u,v) )
m
m
Proof. Recall that the graphs in G2 are as defined in Subsection 4.1.1 and illustrated in Figure 2.
In the following we consider crossing pairs, as the proof for non-crossing pairs is almost identical.
Recall that a crossing pair is a pair (u, v) such that u ∈ L and v ∈ R or vise versa. A witness graph
W with respect to the pair (u, v) is a graph in which (u, v) is a red pair, i.e., (u, v) ∈ M C . There
is an edge from W to every non-witness graph W ∈ G2 (π) such that M C (W ) and M C (W ) differ
exactly on (u, v) and one additional edge.
23
Every red pair (u′ , v ′ ) ∈ M C (W ) creates a potential non-witness graph W (u′ ,v′ ) when switched
with (u, v) (as defined in Definition 7). However, not all of the these non-witness graphs are in
′
kn
G2 (π). If u′ is a neighbor of v in the knowledge graph Gkn
π , i.e., u ∈ Γπ (v), then W (u′ ,v′ ) is not
/ G2 (π). This is also the case for a
consistent with the knowledge graph, and therefore W (u′ ,v′ ) ∈
′ , v ′ ) ∈ M C such that u′ ∈
(u).
Therefore,
only
pairs
(u
/ Γkn
pair (u′ , v ′ ) such that v ′ ∈ Γkn
π (v) and
π
′
kn
v ∈
/ Γπ (u) produce a non-witness graph W (u′ ,v′ ) ∈ G2 (π) when switched with (u, v). We refer to
√
m
m
, both u and v each have at most 100
neighbors in the
these pairs as consistent pairs. Since t ≤ 100
√
knowledge graph, implying that out of the m − 1 potential pairs, the number of consistent pairs
is at least
√
√
√
m
1√
kn
kn
m − 1 − dπ (u) − dπ (v) ≥ m − 1 − 2 ·
m.
≥
100
2
√
Therefore, the degree of every witness graph W ∈ Aπ,(u,v) is at least 12 m, implying that
√
dw (Aπ,(u,v) ) ≥ 12 m.
In order to prove that dnw (Aπ,(u,v) ) = 1, consider a non-witness graph W . Since W is a nonwitness graph, the pair (u, v) is not a red pair. This implies that u is matched to some vertex
v ′ ∈ R, and v is matched to some vertex u′ ∈ L. That is, (u, v ′ ), (v, u′ ) ∈ M C . By the construction
of the edges in the auxiliary graph, every neighbor W of W can be obtained by a single switch
between two red pairs in the red matching. The only possibility to switch two pairs in M C (W )
and obtain a matching in which (u, v) is a red pair is to switch the pairs (u, v ′ ) and (v, u′ ). Hence,
every non-witness graph W has at most one neighbor.
√
We showed that dw (Aπ,(u,v) ) ≥ 21 m and that dnw (Aπ,(u,v) ) ≤ 1, implying
dnw (Aπ,(u,v) )
2
2t
≤ √ = 3/2 ,
dw (Aπ,(u,v) )
m
m
and the proof is complete.
4.1.4
Statistical distance
For a query-answer history π of length t − 1 and a query qt , let Ans(π, qt ) denote the set of possible
answers to the query qt that are consistent with π. Namely, if qt is a pair query (for a pair that
does not belong to the knowledge graph Gkn
π ), then Ans(π, qt ) = {0, 1}, and if qt is a random
new-neighbor query, then Ans(π, qt ) consists of all vertices except those in Nπkn .
3/2
Lemma 19. Let t = m and Q = m
100t . For every t ≤ Q, every query-answer history π of length
t − 1 such that π is consistent with G1 and for every query qt :
X
12t
12
PrP1 [a | π, qt ] − PrP2 [a | π, qt ] ≤ √ = 3/2 .
m
m
a∈Ans(π,qt )
Proof. We prove the lemma separately for each type of query.
• We start with a crossing pair query (ut , vt ). In this case the witnesses are red pairs. Namely,
our witness graphs for this case are all the graphs in G2 (π ◦(qt , 0)), and the non-witness graphs
are all the graphs in G2 (π ◦ (qt , 1)). By the construction of the auxiliary graph
|G2 (π ◦ (qt , 0))| · dw (Aπ,(u,v) ) ≤ |G2 (π ◦ (qt , 1))| · dnw (Aπ,(u,v) ).
This, together with Lemma 18, implies
dnw (Aπ,(u,v) )
|G2 (π ◦ (qt , 0))|
|G2 (π ◦ (qt , 0))|
2
2t
≤
≤
= √ = 3/2 .
|G2 (π)|
|G2 (π ◦ (qt , 1))|
dw (Aπ,(u,v) )
m
m
24
For a pair query qt , the set of possible answers Ans(π, qt ) is {0, 1}. Therefore,
X
PrP1 [a | π, qt ] − PrP2 [a | π, qt ]
a∈{0,1}
= PrP1 [0 | π, qt ] − PrP2 [0 | π, qt ] + PrP1 [1 | π, qt ] − PrP2 [1 | π, qt ]
2t
4t
4
2t
= 3/2 + 1 − 1 − 3/2 = 3/2 = √ .
m
m
m
m
(13)
• For a non-crossing pair query qt = (u, v) our witness graphs are graphs that contain qt as a
blue pair, i.e., graphs from G2 (π, (qt , 1)), and our non-witness graphs are graphs in which no
blue pair had been queried, i.e., graphs from G2 (π, (qt , 0)). From Lemma 18 we get that for
a non-crossing pair query qt :
dnw (Aπ,(u,v) )
2
|G2 (π ◦ (qt , 1))|
2t
|G2 (π ◦ (qt , 1))|
≤
≤
= 3/2 = √ .
|G2 (π)|
|G2 (π ◦ (qt , 0))|
dw (Aπ,(u,v) )
m
m
Therefore,
X
a∈{0,1}
PrP1 [a | π, qt ] − PrP2 [a | π, qt ]
=
PrP1 [0 | π, qt ] − PrP2 [0 | π, qt ] + PrP1 [1 | π, qt ] − PrP2 [1 | π, qt ]
2t
2t
4t
4
= 1 − 1 − 3/2 + 3/2 = 3/2 = √ .
m
m
m
m
(14)
• For a new-neighbor query qt = ut , the set of possible answers Ans(π, qt ) is the set of all the
vertices in the graph. Therefore,
X
PrP1 [a | π, qt ] − PrP2 [a | π, qt ]
a∈Ans(π,qt )
=
X
v∈R
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] +
X
v∈L
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] .
Recall that for a vertex v ∈ Γkn
π (u), PrP1 [v | π, qt ] = PrP2 [v | π, qt ] = 0. Therefore, it suffices to
consider only vertices v such that v ∈
/ Γkn
π (u). Assume without loss of generality that u ∈ L,
kn
and consider a vertex v ∈ R, v ∈
/ Γπ (u). Since for every v ∈ R we have that (ut , v) ∈ E(G1 ),
by the definition of P1 ,
PrP1 [v | π, qt ] = √
1
.
m − dkn
π (ut )
Now consider the process P2 . By its definition,
G2 (π ◦ (qt , v))
1
·√
G2 (π)
m − dkn
π (u)
1
G2 (π ◦ ((u, v), 1))
·√
=
G2 (π)
m − dkn
(u)
π
1
G2 (π ◦ ((u, v), 0))
.
·√
= 1−
G2 (π)
m − dkn
π (u)
PrP2 [v | π, qt ] =
25
(15)
By the first item in the proof, for any crossing pair qt = (u, v),
4t
G2 (π ◦ (qt , 0))
4
= 3/2 = √ ,
G2 (π)
m
m
and it follows that
PrP2 [v | π, qt ] =
4t
1 − 3/2
m
·√
1
.
m − dkn
π (u)
(16)
/ Γkn
By Equations (15) and (16), we get that for every v ∈ R such that v ∈
π (u),
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] = √
Therefore,
X
v∈R
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] =
=
√
X
v∈R,v∈Γ
/ kn
π (u)
4t/m3/2
.
m − dkn
π (u)
(17)
PrP1 [v | π, qt ] − PrP2 [v | π, qt ]
4t/m3/2
4
4t
√
m − dkn
(u)
·
= 3/2 = √ .
π
kn
m − dπ (u)
m
m
(18)
Now consider a vertex v ∈ L. Observe that for every v ∈ L, it holds that v ∈
/ Γkn
π (u) since
otherwise π is not consistent with G1 . For the same reason,
PrP1 [v | π, qt ] = 0 .
(19)
As for P2 , as before,
PrP2 [v | π, qt ] =
1
G2 (π, (ut , v))
·√
.
G2 (π)
m − dkn
π (ut )
By the second item of the claim, since for every v ∈ L, (ut , v) is a non-crossing pair, we have
that
4
4t
|G2 (π, (ut , v))|
= 3/2 = √ .
(20)
|G2 (π)|
m
m
Combining Equations (19) and (20) we get that for every v ∈ L
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] = √
√
3/2
m
kn
Since Q = m
100t = 100 , for every t ≤ Q, dπ (u) <
bounded by 2. Hence,
X
v∈L
4t/m3/2
.
m − dkn
π (u)
1√
2 m,
and it follows that
√
√
m−1
m−dkn (u)
is
√
4t/m3/2
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] = ( m − 1) · √
m − dkn
π (u)
=
8
8t
=√ .
3/2
m
m
(21)
By Equations (18) and (21) we get
X
X
PrP1 [v | π, qt ] − PrP2 [v | π, qt ]
PrP1 [v | π, qt ] − PrP2 [v | π, qt ] +
v∈L
v∈R
12t
12
= 3/2 = √ .
m
m
(22)
26
This completes the proof.
Recall that DbALG , b ∈ {1, 2}, denotes the distribution on query-answer histories of length Q,
induced by the interaction of ALG and Pb . We show that the two distributions are indistinguishable
for Q that is sufficiently small.
Lemma 20. Let t = m. For every algorithm ALG that asks at most Q =
statistical distance between D1ALG and D2ALG is at most 31 .
m3/2
100t
queries, the
ALG be the distribution over query-answer
Proof. Consider the following hybrid distribution. Let D1,t
histories of length Q, where in the length t prefix ALG is answered by the process P1 and in the
ALG = D ALG and that
length Q − t suffix ALG is answered by the process P2 . Observe that D1,Q
1
ALG = D ALG . Let π = (π , π , . . . , π ) denote a query-answer history of length ℓ. By the triangle
D1,0
1
2
ℓ
2
Q−1
P
ALG , D ALG ) .
inequality d(D1ALG , D2ALG ) ≤
d(D1,t+1
1,t
t=0
d(D1ALG , D2ALG ) ≤
Q−1
X
t=0
1
2
ALG , D ALG ) =
It thus remains to bound d(D1,t+1
1,t
ALG
ALG
d(D1,t+1
, D1,t
).
P
π
PrDALG [π] − PrDALG [π] for every t such that
1,t+1
1,t
0 ≤ t ≤ Q − 1. Let Q denote the set of all possible queries.
X
π
PrDALG [π] − PrDALG [π] =
1,t+1
1,t
·
·
X
π1 ,...,πt−1
PrP1 ,ALG [π1 , . . . , πt−1 ] ·
X
a∈Ans((π1 ,...,πt−1 ),q)
X
πt+1 ,...,πQ
X
q∈Q
PrALG [q | π1 , . . . , πt−1 ]
PrP1 [a | π1 , . . . , πt−1 , q] − PrP2 [a | π1 , . . . , πt−1 , q]
PrP2 ,ALG [πt+1 , . . . , πQ | π1 , . . . , πt−1 , (q, a)] .
By Lemma 19, for every 1 ≤ t ≤ Q − 1, and every π1 , . . . , πt−1 and q,
X
12t
PrP1 [a | π1 , . . . , πt−1 , q] − PrP2 [a | π1 , . . . , πt−1 , q] ≤ 3/2 .
m
a∈Ans((π1 ,...,πt−1 ),q)
We also have that for every pair (q, a),
X
PrP2 ,ALG [πt+1 , . . . , πQ | π1 , . . . , πt−1 , (q, a)] = 1 .
πt+1 ,...,πQ
Therefore,
X
PrDALG [π] − PrDALG [π] ≤
π
1,t+1
1,t
Hence, for Q =
√
m
100 ,
X
PrP1 ,ALG [π1 , . . . , πt−1 ]
π1 ,...,πt−1
q∈Q
Q−1
d(D1ALG , D2ALG ) =
and the proof is complete.
X
PrALG [q | π1 , . . . , πt−1 ] ·
1
12t
1XX
1
PrDALG [π] − PrDALG [π] ≤ · Q · 3/2 ≤ ,
1,t+1
1,t
2 π
2
3
m
t=1
27
12t
12t
= 3/2 .
3/2
m
m
3/2
In the next subsection we turn to prove the theorem for the cases where m < t ≤ m8 , and for
√
the case where m ≤ t ≤ m
4 . We start with the former case. The proof will follow the building
blocks of the proof for t = m, where the only difference is in the description of the auxiliary graph
dnw (A
)
√2r .
Aπ,(u,v) and in the proof that dw (A π,(u,v)) ≤ m2t
3/2 =
m
π,(u,v)
4.2
A lower bound for m < t < m3/2
√
Let t = r · m for an integer r such that 1 < r ≤ 18 m. It is sufficient for our needs to consider only
values of t for which r is an integer. The proof of the lower bound for this case is a fairly simple
extension of the proof for the case of t = m, that is, r = 1. We next describe the modifications we
make in the construction of G2 .
4.2.1
The lower-bound construction
Let G1 be as defined in Subsection 4.1.1. The construction of G2 for t = r · m can be thought
of as repeating the construction of G2 for t = m (as described in Subsection 4.1.1) r times. We
√
again start with a complete bipartite graph K√m,√m and an independent set of size n − 2 m. For
each graph G ∈ G2 we select r perfect matchings between the two sides R and L and remove these
edges from the graph. We denote the r perfect matchings by M1C , . . . , MrC and refer to them as the
red matchings. We require that each two perfect matchings MiC and MjC do not have any shared
edges. That is, for every i and for every j, for every (u, v) ∈ MiC it holds that (u, v) ∈
/ MjC . In
order to maintain the degrees of the vertices, we next select r perfect matchings for each side of the
bipartite graph (L to L and R to R). We denote these matchings by M1R , ..., MrR and M1L , ..., MrL
respectively. Again we require that no two matchings share an edge. We refer to these matchings
as the blue matchings and their edges as blue pairs. Each such choice of 3r matchings defines a
graph in G2 .
Let G be a graph in G2 . We say that a triangle is blue if all its edges are blue. Otherwise
√
we say the triangle is mixed. Observe that every blue edge in G participates in at least m − 2r
mixed triangles, and at most r blue triangles. Also note that every two mixed triangles are disjoint.
√
√
√
√
√
Therefore, there are at least 12 r m · (2 m − 2r) = Ω(r · m) and at most 21 r m · (2 m − 2r) + r 2 m
√
triangles in G. Since r < 18 m, we get that every graph in G has Θ(r · m) triangles.
4.2.2
The processes P1 and P2
The definition of the processes P1 and P2 is the same as in Subsection 4.1.2 (using the modified
definition of G2 ), and Lemma 17 holds here as well.
4.2.3
The auxiliary graph
As before, for every t ≤ Q, every query-answer history π of length t−1 such that π is consistent with
G1 and every pair (u, v), we define a bipartite auxiliary graph Aπ,(u,v) , such that on one side there
is a node for every witness graph W ∈ G2 (π), and on the other side a node for every non-witness
graph W ∈ G2 (π). The witness graphs for this case are graphs in which (u, v) is a red (blue) edge
in one of the red (blue) matchings. If (u, v) is a crossing pair, then for every witness graph W ,
(u, v) ∈ MiC (W ) for some 1 ≤ i ≤ r. If (u, v) is a non-crossing pair, then for every witness graph
W , (u, v) ∈ MiL (W ) or (u, v) ∈ MiL (W ). There is an edge from W to every graph W such that
the matching that contains (u, v) in W and the corresponding matching in W differ on exactly two
pairs – (u, v) and one additional pair. For example, if (u, v) ∈ MiC (W ), there is an edge from W
to every graph W such that MiC (W ) and MiC (W ) differ on exactly (u, v) and one additional pair.
28
√
3/2
Lemma 21. Let t = r · m for an integer r such that 1 < r ≤ 8m and let Q = m
100t . For every
t ≤ Q, every query-answer history π of length t − 1 such that π is consistent with G1 and every
pair (u, v),
dnw (Aπ,(u,v) )
2r
2t
≤ 3/2 = √ .
dw (Aπ,(u,v) )
m
m
Proof. We again analyze the case in which the pair is a crossing pair (u, v), as the proof for a
non-crossing pair is almost identical. We first consider the minimal degree of the witness graphs in
Aπ,(u,v) . Let MiC be the matching to which (u, v) belongs. As before, only pairs (u′ , v ′ ) ∈ MiC such
′ / Γkn (v) result in a non-witness graph W ∈ G (π) when switched with (u, v).
that u′ ∈
/ Γkn
2
π
π (u), v ∈
However, we have an additional constraint. Since by our construction no two red matchings share
an edge, it must be that u′ is not matched to v in any of the other r red matching, and similarly
√
m3/2
)
that u is not matched to v ′ in any of the other matchings. It follows that of the ( m−1−2· 100·r·m
√
potential pairs (as in the proof of Lemma 18), we discard 2r additional pairs. Since 1 ≤ r ≤ 8m
√
√
√
√
√
we remain with ( m − 1 − 50m − 14 m) ≥ 12 m potential pairs. Thus, dw (Aπ,(u,v) ) ≥ 12 m.
We now turn to consider the degree of the non-witness graphs and prove that dnw (Aπ,(u,v) ) ≤ r.
Consider a non-witness graph W . To prove that W has at most r neighbors it is easier to consider
all the possible options to “turn” W from a non-witness graph into a witness graph. It holds that
for every j ∈ [r], (u, v) ∈
/ MjC (W ). Therefore for every matching MjC , u is matched to some vertex,
′
denoted vj and v is matched to some vertex, denoted u′j . If we switch between the pairs (u, vj′ )
and (v, u′j ), this results in a matching in which (u, v) is a witness pair. We again refer the reader
to Figure 3b, where the illustrated matching can be thought of as the j th matching. Denote the
resulting graph by W(u′j ,vj′ ) . If the pair (u′j , vj′ ) has not been observed yet by the algorithm then
W(u′j ,vj′ ) is a witness graph in Aπ,(u,v) . Therefore there are at most r options to turn W into a
√
witness graph, and dnw (Aπ,(u,v) ) ≤ r. We showed that dw (Aπ,(u,v) ) ≥ 21 m and dnw (Aπ,(u,v) ) ≤ r,
implying
dnw (Aπ,(u,v) )
2r
2t
≤ √ = 3/2 ,
dw (Aπ,(u,v) )
m
m
as required.
4.2.4
Statistical distance
The proof of the next lemma is exactly the same as the proof of Lemma 19, except that occurrences
√
√
of the term (t/m3/2 ) are replaced by (r/ m) instead of (1/ m), and we apply Lemma 21 instead
of Lemma 18.
√
3/2
Lemma 22. Let t = r · m for an integer r such that 1 < r ≤ 8m and let Q = m
100t . For every
t ≤ Q, every query-answer history π of length t − 1 such that π is consistent with G1 and for every
query qt ,
X
12t
12r
PrP1 [a | π, qt ] − PrP2 [a | π, qt ] = 3/2 = √ .
m
m
a∈Ans(π,qt )
The proof of the next lemma is same as the proof of Lemma 20 except that we replace the
application of Lemma 19, by an application of Lemma 22.
Lemma 23. Let t = r · m for an integer r such that 1 < r ≤
performs at most Q =
m3/2
100t
√
m
8 .
For every algorithm ALG that
queries, the statistical distance between D1ALG and D2ALG is at most 13 .
29
4.3
A lower bound for
√
m ≤ t ≤ 14 m
√
Similarly√to the previous section, we let t = k m and assume that k is an integer such that
1 ≤ k ≤ 4m .
4.3.1
The lower-bound construction
The construction of the graph G1 is as defined in Subsection 4.1.1, and we modify the construction of
the graphs in G2 . As before, the basic structure of every graph is a complete bipartite graph K√m,√m
√
and an independent set of size n−2 m vertices. In this case, for each graph in G2 , we do not remove
a perfect matching from the bipartite graph, but rather a matching M C of size k. In order to keep
√
the degrees of all vertices to be m, we modify the way we construct the blue matchings. Let M C =
{(ℓi1 , ri1 ), (ℓi2 , ri2 ), . . . , (ℓik , rik )} be the crossing matching. The blue matchings will be M L =
{(ℓi1 , ℓi2 ), (ℓi3 , ℓi4 ), . . . , (ℓik −1 , ℓik )} and M R = {(ri1 , ri2 ), (ri3 , ri4 ), . . . , (rik −1 , rik )}. Note that every
matched pair belongs to a four-tuple hℓij , ℓij+1 , rij+1 , rij i such that (ℓij , rij ) and (ℓij+1 , rij+1 ) are red
pairs and (ℓij , ℓij+1 ) and (rij , rij+1 ) are blue pairs. We refer to these structures as matched squares
and to four-tuples (ℓx , ℓy , rz , rw ) such that no pair in the tuple is matched as unmatched squares.
See Figure 4 for an illustration. Every graph in G2 is defined by its set of k four-tuples.
Similarly to previous constructions, in every graph G ∈ G2 , every blue edge participates in
√
m − 2 triangles. Since every triangle in the G contains exactly one blue edge, we have that G has
√
√
k · ( m − 2) = Θ(k m) triangles.
√
The ith
matched square
m
Figure 4: An illustration of the bipartite component in the family G2 for
4.3.2
√
m ≤ t ≤ 14 m.
The processes P1 and P2
We introduce a small modification to the definition of the processes P1 and P2 . Namely, we leave
the answering process for pair queries as described in Subsection 4.1.2 and modify the answering
process for random new-neighbor queries as follows. Let t ≤ Q, and π be a query-answer history of
length t − 1 such that π is consistent with G1 . If the tth query is a new-neighbor query qt = u and
1√
in Subsection 4.1.2. However, if
dkn
π (u) < 2 m, then the processes P1 and P2 answer as described √
1
the tth query is a new-neighbor query qt = u such that dkn
(u)
≥
π
2 m, then the processes answers
as follows.
• The process P1 answers with the set of all neighbors of u in G1 . That is, if u is in L, then the
process replies with a = R = {r1 , . . . , r√m }, and if u is in R, then the process replies with
a = L = {ℓ1 , . . . , ℓ√m }.
30
The process P2 answers with a = {v1 , . . . , v√m }, where {v1 , . . . , v√m } is the set of neighbors
of u in a subset of the graphs in G2 . By the definition of G2 , if u is in L, then this set is either
R, or it is R \ {ri } ∪ {ℓj } for some ri ∈ R and ℓj ∈ L, and if u is in R, then this set is either
L, or it is L \ {ℓi } ∪ {rj } for some ℓi ∈ L and rj ∈ R. For every such set a ∈ Ans(π, qt ), the
process returns a as an answer with probability
|G2 (π ◦ (qt , a))|
.
|G2 (π)|
We call this query an all-neighbors query.
First note that the above modification makes the algorithm “more powerful”. That is, every
algorithm that is not allowed all-neighbors query can be emulated by an algorithm that is allowed
this type of query. Therefore this only strengthen our lower bound results.
Also note that this modification does not affect the correctness of Lemma 17. We can redefine
the function αt (π) to be
if qt (π) is a pair query
1 √
kn
αt (π) = 1/ m − dπ≤t−1 (u) if qt (π) = u is a random new-neighbor query ,
1
if qt (π) is an all-neighbors query
and the rest of the proof follows as before.
4.3.3
The auxiliary graph
For every t ≤ Q, every query-answer history π of length t − 1 such that π is consistent with G1
and every pair (u, v), the witness graphs in Aπ,(u,v) are graphs in which (u, v) is either a red pair
or a blue pair. There is an edge between a witness graph W and a non-witness graph W if the two
graphs have the same set of four-tuples except for two matched squares – one that contains the
pair (u, v), hu, v, u′ , v ′ i and another one.
Definition 8. We define a switch between a matched square and an unmatched square in the following manner. Let hu, v, u′ , v ′ i be a matched square and hx, y, x′ , y ′ i be an un matched squares.
Informally, a switch between the squares is “unmatching” the matched square and instead “matching” the unmatched square.
Formally, a switch consists of two steps. The first step is removing the edges (u, v) and (u′ , v ′ )
from the red matching M C and the edges (u, u′ ) and (v, v ′ ) from the blue matchings M L and M R
respectively. The second step is adding the edges (x, y) and (x′ , y ′ ) from the red matching M C and
the edges (x, x′ ) and (y, y ′ ) from the blue matchings M L and M R respectively. See Figure 5 for an
illustration.
√
√
3/2
Lemma 24. Let t = k · m for an integer k such that 1 < k ≤ 4m and let Q = m
600t . For every
t ≤ Q, every query-answer history π of length t − 1 such that π is consistent with G1 and every
pair (u, v),
dnw (Aπ,(u,v) )
16k
16t
=
= 3/2 .
dw (Aπ,(u,v) )
m
m
Proof. We start with proving that dw (Aπ,(u,v) ) ≥ 12 m. A witness graph in Aπ,(u,v) with respect
to a pair (u, v) is a graph in which (u, v) is part of a matched square hu, v, u′ , v ′ i. Potentially,
31
x′
y′
x′
y′
x
y
x
y
A switch
u′
v′
u′
v′
u
v
u
v
Figure 5: An illustration of a switch between the squares hu, v, u′ , v ′ i and hx, y, x′ , y ′ i.
hu, v, u′ , v ′ i could be switched with every unmatched square to √
get a non-witness
pair. There are
√
√
m−k
m−k
m − k unmatched vertices on each side, so that there are
·
≥ 18 m2 potential
2
2
′
′
squares. To get a graph that is in G2 (π), the unmatched square hx, y, x , y i must be such that
none of the induced pairs between the vertices x, x′ , y, y ′ have been observed yet by the algorithm.
When all-neighbor queries are allowed, if at most Q queries has been performed, then at most 4Q
m
≤ 14 m of the potential
pairs have been observed by the algorithm. Therefore, for at most 4 100k
squares, an induced pair was queried. Hence, every witness square can be switched with at least
1 2
1
1
1
2
2
8 m − 4 m ≥ 16 m consistent unmatched squares, implying that dw (Aπ,(u,v) ) ≥ 16 m .
To complete the proof it remains to show that dnw (Aπ,(u,v) ) ≤ mk. To this end we would like
to analyze the number of witness graphs that every non-witness W can be “turned” into. In every
non-witness graph W the pair (u, v) is unmatched, and in order to turn W into a witness graph,
one of the k matched squares should be removed and the pair (u, v) with an additional pair (u′ , v ′ )
should be “matched”. There are k options to remove an existing square, and at most m options to
choose a pair u′ , v ′ to match (u, v) with. Therefore, the number of potential neighbors of W is at
most mk. It follows that
dnw (Aπ,(u,v) )
16mk
16t
16k
=
= 3/2 ,
=
dw (Aπ,(u,v) )
m2
m
m
and the proof is complete.
4.3.4
Statistical distance
For an all-neighbors query q = u we say that the corresponding answer is a witness answer if u ∈ L
and a 6= R, or symmetrically if u ∈ R and a 6= L. Let E Q be the set of all query-answer histories π
of length Q such that there exists a query-answer pair (q, a) in π in which q is an all-neighbors pair
Q
Q
and a is a witness answer with respect to that query, and let E = ΠQ \ E Q . That is, E is the set
of all query-answer histories of length Q such that no all-neighbors query is answered with a witness
answer. Let Pe1 and Pe2 by the induced distributions of the processes P1 and P2 conditioned on the
event that the process do not reply with a witness answer. Observe that for every query-answer
history π of length t − 1, for every query qt that is either a pair query or a random new-neighbor
query and for every a ∈ Ans(π, qt ),
PrPeb [a | π, qt ] = PrPb [a | π, qt ].
for b ∈ {1, 2}. Therefore, the proof of the next lemma is exactly the same as the proof of Lemma 19,
√
except that occurrences of the term (t/m3/2 ) are replaced by (k/m) instead of (1/ m) and we apply
Lemma 24 instead of Lemma 18.
32
√
√
3/2
Lemma 25. Let t = k · m for an integer k such that 1 < k ≤ 4m and let Q = m
600t . For every
t ≤ Q, every query-answer history π of length t − 1 such that π is consistent with G1 and for every
pair or random new-neighbors query qt ,
X
a∈Ans(π,qt )
PrPe1 [a | π, qt ] − PrPe2 [a | π, qt ] =
96k
96t
= 3/2 .
m
m
Note that Lemma 25 does not cover all-neighbors queries, and hence we establish the next
lemma.
√
√
3/2
Lemma 26. Let t = k · m for an integer k such that 1 < k ≤ 4m and let Q = m
600t . For every
t ≤ Q, every query-answer history π of length t − 1 such that π is consistent with G1 and for every
all-neighbors query qt ,
16k
PrP2 [at is a witness answer | π, qt ] ≤ √ .
m
Proof. Assume without loss of generality that u ∈ L. By the definition of the process P2 , it
answers the query consistently with a uniformly selected random graph G2 ∈ G2 (π) by returning
the complete set of u’s neighbors in G2 . In G2 , there are two types of graphs. First, there are graphs
in which u is not matched, that is (u, u′ ) ∈
/ M L for every vertex u′ ∈ L. In these graphs the set
√
of u’s neighbors is R ={r1 , . . . , r m }. We refer to these graphs as non-witness graphs. The second
type of graphs are those in which (u, u′ ) ∈ M L for some u′ ∈ L and (u, v) ∈ M C for some v ∈ R.
In these graphs the set of u’s neighbors is (R \ {v}) ∪ {u′ }. We refer to these graphs as witness
graphs. As before, let Ans(π, qt ) be the set of all possible answers for an all-neighbors query qt . It
holds that
X
PrP2 [at is a witness answer | π, qt ] =
PrP2 [a | π, qt ]
a∈Ans(π,qt )
a6=R
=
X
u′ ∈L,v∈R
=
|G2 (π ◦ ((u, u′ ), 1) ◦ ((u, v), 0))|
|G2 (π)|
X |G2 (π ◦ ((u, u′ ), 1))| X |G2 (π ◦ ((u, u′ ), 1) ◦ ((u, v), 0))|
·
|G2 (π)|
|G2 (π)|
′
v∈R
u ∈L
X |G2 (π ◦ ((u, u′ ), 1))|
.
=
|G2 (π)|
′
u ∈L
Similarly to the proof of Lemma 19, for every u and u′ in L,
PrP2 [at is a witness answer | π, qt ] =
|G2 (π◦((u,u′ ),1))|
|G2 (π)|
≤
16k
m .
Therefore,
X |G2 (π ◦ ((u, u′ ), 1))| √
16k
16k
≤ m·
=√ ,
|G2 (π)|
m
m
′
u ∈L
and the lemma follows.
√
It remains to prove that a similar lemma to Lemma 20 holds for m ≤ t ≤ 14 m (and the
distributions D1ALG and D2ALG as defined in this subsection).
√
√
Lemma 27. Let t = k · m for an integer k such that 1 < k ≤ 4m . For every algorithm ALG that
performs at most Q =
m3/2
600t
queries, the statistical distance between D1ALG and D2ALG is at most 13 .
33
Q
Proof. Let the sets E Q and E be as defined in the beginning of this subsection. By the definition
of the statistical distance, and since PrP1 ,ALG [E Q ] = 0,
X
X
1
PrP1 ,ALG [π] − PrP2 ,ALG [π] +
PrP1 ,ALG [π] − PrP2 ,ALG [π]
d(D1ALG , D2ALG ) =
2
Q
Q
π∈E
π∈E
X
1
= PrP2 ,ALG [E Q ] +
(23)
PrP1 ,ALG [π] − PrP2 ,ALG [π] .
2
Q
π∈E
By Lemma 26, the probability of detecting a witness as a result of an all-neighbors query is at most
√
16k
√
. Since in Q queries, there can be at most 4Q/ m all-neighbors queries, we have that
m
PrDALG [E Q ] ≤
2
1
.
6
(24)
We now turn to upper bound the second term. Let α = PrP2 ,ALG [E Q ].
X
π∈E
Q
PrP1 ,ALG [π] − PrP2 ,ALG [π]
=
X
π∈E
=
π∈E
≤
Q
X
π∈E
≤
Q
X
Q
X
π∈E
Q
Q
Q
PrPe1 ,ALG [π] · PrP1 ,ALG [E ] − PrPe2 ,ALG [π] · PrP2 ,ALG [E ]
(25)
PrPe1 ,ALG [π] − (1 − α) · PrPe2 ,ALG [π]
Q
PrPe1 ,ALG [π] − PrPe2 ,ALG [π] + α · PrPe2 ,ALG [E ]
PrPe1 ,ALG [π] − PrPe2 ,ALG [π] +
1
,
6
(26)
Q
where in Equation (25) we used the fact that PrP1 ,ALG [E ] = 1, and in Equation (26) we used the
Q
fact that PrPe2 ,ALG [E ] = 1 and that α ≤ 1/6.
Therefore, it remains to bound
X
PrPe1 ,ALG [π] − PrPe2 ,ALG [π] .
π∈E
Q
ALG for t ∈ [Q − 1] be as defined in Lemma 20 (based on the
Let the hybrid distributions D1,t
ALG
ALG
that are induced by the processes P1 and P2 that were defined in
and D2
distributions D1
ALG
ALG conditioned on the event that no
e
this subsection). Also, let D1,t be the hybrid distribution D1,t
e ALG is the distribution over query-answer
all-neighbors query is answered with a witness. That is, D
1,t
histories π of length Q, where in the length t prefix ALG is answered by the process P1 , in the
length Q − t suffix ALG is answered by the process P2 , and each all-neighbors query is answered
consistently with G1 (so that no witness is observed). By the above definitions and the triangle
inequality,
X
π∈E
Q
PrPe1 ,ALG [π] − PrPe2 ,ALG [π] ≤
Q−1
X
34
t
X
π∈E
Q
PrDe ALG [π] − PrDe ALG [π] .
1,t+1
1,t
(27)
As in the proof of Lemma 20 we have that for every t ∈ [Q − 1],
X
PrDe ALG [π] − PrDe ALG [π]
π∈E
1,t+1
Q
1,t
X
=
π ′ =π1 ,...,πt−1 ,qt :
t−1
π ′ ∈E
PrPe1 ,ALG [π ′ , qt ] ·
X
PrPe1 [a | π ′ , qt ] − PrPe2 [a | π ′ , qt ] .
a∈Ans(π ′ ,qt ):
t
π ′ ◦(qt ,a)∈E
(28)
By Lemma 25 (and since for an all-neighbor query qt we have that the (unique) answer according
to Pe2 is the same as according to Pe1 ),
X
PrPe1 [a | π ′ , qt ] − PrPe2 [a | π ′ , qt ] ≤
a∈Ans(π ′ ,qt ):
t
π ′ ◦(qt ,a)∈E
and it follows that
X
π∈E
Hence, for Q =
PrDe ALG [π] − PrDe ALG [π] ≤
1,t+1
Q
1,t
96k
96t
= 3/2 ,
m
m
96t
96k
= 3/2 .
m
m
m3/2
600t ,
Q−1
X
t
X
π∈E
Q
PrDe ALG [π] − PrDe ALG [π] ≤ Q ·
1,t
1,t+1
1
48t
≤ .
3/2
6
m
Combining Equations (23), (24), (26), (27) and (29), we get
1 1 1 1
1
ALG
ALG
+ +
d(D1 , D2 ) ≤
≤ ,
2 6 6 6
3
(29)
(30)
and the proof is complete.
4.4
4.4.1
Lower Bound for t <
1√
m.
4
The construction
In this case the basic structure of G1 and G2 is a bit different. Also, for the sake of simplicity, we
present graphs with 2m edges, and either 0 or 4t triangles. The graph G1 has three components –
√
√
two complete bipartite graphs, each over 2 m vertices, and an independent set of size n − 4 m.
Let A and B be the left-hand side and the right-hand side sets, respectively, of the first bipartite
component, and C and D of the second one. We refer to the edges between
A and B and the edges
√
m
between C and D as black edges. We divide each of these
sets into t subsets of size t, denoted
√
{Λ1 , . . . , Λ √m } for Λ ∈ {A, B, C, D}. For every 1 ≤ i ≤
t
m
t ,
we first remove a complete bipartite
graph between Ai and Bi and between Ci and Di , and refer to the removed edges as red edges.
We then add a complete bipartite graph between Bi and Ci and between Di and Ai , and refer to
√
added edges as blue edges. Note that this maintains the degrees of all the vertices to be m.
In G2 the basic structure of all the graphs is the same as of G1 with the following modifications.
Each graph is defined by the choice of four “special” vertices a∗ , b∗ , c∗ , d∗ such that a∗ ∈ Aia∗ , b∗ ∈
Bib∗ , c∗ ∈ Cic∗ and d∗ ∈ Did∗ for some indices ia∗ , ib∗ , ic∗ and id∗ such that no two indices are equal.
We then add edges (a∗ , c∗ ) and (b∗ , d∗ ), referred to as green edges, and remove edges (a∗ , b∗ ) and
35
A
B
C
D
c∗
a∗
d∗
b∗
Figure 6: An illustration of a graph in G2 . The broken thin (red) edges describe edges that were
removed and the thin (blue) edges describe edges that were added. The broken thick (purple) edges
describe the special non-edges (a∗ , b∗ ) and (c∗ , d∗ ). The curly (green) edges describe the special
edges (a∗ , c∗ ) and (b∗ , d∗ ).
(c∗ , d∗ ), referred to as purple edges. We also refer to the green and purple edges as special edges.
Note that we add one edge and remove one edge from each special vertex, thus maintaining their
initial degrees. See Figure 6.
We first prove that t(G1 ) = 0 and then that for every graph G in G2 , t(G) = 4t.
Claim 28. The graph G1 has no triangles.
Proof. Consider an edge (u, v) in G1 . First assume u and v are connected by a black edge, that
is, they are on different sides of the same bipartite component. Hence we can assume without loss
of generality that u ∈ A and that v ∈ B. Since u is in A it is only connected to vertices in B or
vertices in D. Since v is in B it is only connected to vertices in A or vertices in C. Thus u and v
cannot have a common neighbor. A similar analysis can be done for a pair (u, v) that is connected
by a blue edge. Therefore t(G) is indeed zero as claimed.
Claim 29. For every graph G ∈ G2 , t(G) = 4t.
Proof. Since the only differences between G1 and graphs in G2 are the two added green edges and
the two removed red edges, any triangle in G2 must include a green edge. Therefore we can count
all the triangles that the green edges form. Consider the green edge (a∗ , c∗ ) and recall that a∗ is
in Aia∗ and c∗ is in Cic∗ . The only common neighbors of (a∗ , c∗ ) are all the vertices in Bic∗ and all
the vertices in Dia∗ . A vertex v such that v ∈
/ Bic∗ and v ∈
/ Dia∗ is either (1) in A or in D \ Dia∗ , in
which case it is not a neighbor of a∗ , or it is (2) in C or in B \Bic∗ , in which case it is not a neighbor
36
of c∗ . Since both Bic∗ and Dia∗ are of size t, the edge (a∗ , c∗ ) participates in 2t triangles. Similarly
the edge (b∗ , d∗ ) participate in 2t triangles, and together we get that t(G) = 4t, as claimed.
4.4.2
The processes P1 and P2
The definition of the processes P1 and P2 is the same as in Subsection 4.3.2 (using the modified
definitions of G1 and G2 ).
4.4.3
The auxiliary graph
We define a switch for this case as well. Informally, a switch between a matched pair (u∗ , v ∗ ) and
an unmatched pair (u, v) is “unmatching” (u∗ , v ∗ ) and “matching” (u, v) instead. Formally stating
we define a switch as follows.
Definition 9. A switch between a green pair (a∗ , c∗ ) and a pair (a, c) such that a ∈ Ai , c ∈ Cj
and none of the indices i, j, ib∗ , id∗ are equal, is the following two steps process. In the first step we
“unmatch” (a∗ , c∗ ) by removing the green edge (a∗ , c∗ ) and adding the edges (a∗ , b∗ ) and (c∗ , d∗ ).
In the second step we “match” (a, c) by adding the green edge (a, c) and removing the edges (a, b∗ )
and (c, d∗ ). A switch with the pair (b∗ , d∗ ) can be defined in a similar manner.
a
a
c
c∗
c
c∗
A switch
a∗
a∗
d∗
d∗
b∗
b∗
Figure 7: An illustration of a switch between the pairs (a∗ , c∗ ) and (a, c).
√
m
. For every t ≤ Q, every query-answer history π of length t − 1 and
Let t < m and let Q = 600
every pair (u, v) we define the following auxiliary graph. The witness nodes are graphs in which
(u, v) is one of the four special pairs. If the pair is a green matched pair then there is an edge in
the auxiliary graph between a witness graph W and a non-witness graph W , if W can be obtained
from W by a single switch between (u, v) and another unmatched pair.
√
m
Lemma 30. For t < 14 m let Q = 600
. For every t ≤ Q, every query-answer history π of length
t − 1 such that π is consistent with G1 and every pair (u, v),
dnw (Aπ,(u,v) )
8
= .
dw (Aπ,(u,v) )
m
Proof. We analyze the case where the pair (u, v) is such that u ∈ A and v ∈ C, as the proof for
the other cases is almost identical. We first prove that dw (Aπ,(u,v) ) ≥ 81 m. A witness graph W
is a graph in which (a, c) is a special pair. That is (u, v) = (a∗ , c∗ ). Potentially, for every pair
(a′ , c′ ) such that a′ ∈ Ai , c′ ∈ Cj and none of the indices i, j, ib∗ , id∗ are equal, the graph resulting
√
from a switch between (a∗ , c∗ ) and (a′ , c′ ) is a non-witness graph. There are m − 2t vertices a′ in
√
A \ (Aib∗ ∪ Aid∗ ) and for each such a′ there are m − 3t vertices c′ in C \ (Cib∗ ∪ Cid∗ ∪ Cia′ )). Since
37
√
√
√
t < 14 m, there are at least ( m − 2t) · ( m − 3t) = m − 6t2 ≥ 41 m potential pairs (a′ , c′ ) that
(a∗ , c∗ ) could be switched with. For the resulting graph to be consistent, that is, to be in G2 (π),
the pair (a′ , c′ ) must be such that the pairs (a′ , c′ ), (a∗ , b∗ ) and (c∗ , d∗ ) have not been observed yet
1
1
by the algorithm. Since the number of queries is at most 600
m, at least 14 m − 125
m ≥ 18 m of the
potential pairs (a′ , c′ ) can be switched with (a∗ , c∗ ) such that the resulting graph is consistent with
G2 (π). Therefore, dw (Aπ,(u,v) ) ≥ 81 m.
Now consider a non-witness graph W . There is only one possibility to turn W into a witness
graph, which is to switch the pair (u, v) with the green pair (a∗ , c∗ ). Therefore, the maximal degree
of every non-witness graph, dnw (Aπ,(u,v) ), is 1.
Together we get that
dnw (Aπ,(u,v) )
8
≤ ,
dw (Aπ,(u,v) )
m
and the proof is complete.
4.4.4
Statistical distance
A similar proof to the ones of Lemma 25 and Lemma 26 using Lemma 30 gives the following
√
lemmas for the case that 1 ≤ t < 14 m.
√
m
. For every t ≤ Q, every query-answer history π of
Lemma 31. Let 1 ≤ t < 41 m and Q = 600
length t − 1 such that π is consistent with G1 and for every all-neighbors query qt ,
PrP2 [at is a witness answer | π, qt ] ≤
16
.
m
√
m
Lemma 32. Let 1 ≤ t < 41 m and Q = 600
. For every t ≤ Q, every query-answer history π of
length t − 1 such that π is consistent with G1 and for every pair or random new-neighbors query qt ,
X
a∈Ans(π,qt )
PrPe1 [a | π, qt ] − PrPe2 [a | π, qt ] =
96
.
m
The next lemma is proven in a similar way to 1.3.4 based on the above two lemma.
√
m
, the statistical
Lemma 33. Let 1 ≤ t < 41 m. For every algorithm ALG that asks at most Q = 600
1
ALG
ALG
is at most 3 .
and D2
distance between D1
4.5
Wrapping things up
Theorem 15 follows from Lemmas 20, 23, 27 and 33, and the next corollary is proved using Theorems
15 and 14.
Corollary 34. Any
algorithm for the number of triangles in a graph
n
multiplicative-approximation
o
m3/2
n
must perform Ω t(G)1/3 + min m, t(G)
queries, where the allowed queries are degree queries,
pair queries and neighbor queries.
Proof. Assume towards a contradiction that there exists an algorithm ALG’ for which the following
holds:
1. ALG’ is allowed to ask neighbor queries as well as degree queries and pair queries.
2. ALG’ asks Q′ queries.
38
3. ALG’ outputs a (1 ± ǫ)-approximation to the number of triangles of any graph G with probability greater than 2/3.
Using ALG’ we can define an algorithm ALG that is allowed random new-neighbor queries, performs
at most Q = 3Q′ queries and answers correctly with the same probability as ALG’ does. ALG runs
ALG’ and whenever ALG’ performs a query qt′ , ALG does as follows:
• If qt′ is a degree query, ALG performs the same query and sets a′t = at .
• If qt′ is a pair query (u, v), then ALG performs the same query q = q ′ . Let at be the
corresponding answer.
– If at = 0, then ALG sets a′t = at .
– If at = 1, then ALG sets a′t = (at , i, j), such that i and j are randomly chosen labels
that have not been previously used for neighbors of u and v, and are within the ranges
[1..d(u)] and [1..dv ] respectively.
• If qt′ is a neighbor query (u, i), ALG performs a random new-neighbor query qt = u, and
returns the same answer a′t = at .
We note that the above requires the algorithm ALG to store for every vertex v, all the labels used for
its neighbors in the previous steps. Once ALG’ outputs an answer, ALG outputs the same answer.
It follows that ALG performs at most 3Q queries to the graph G. By the third assumption above,
ALG outputs a (1 ± ǫ)-approximation
to the n
number of
otriangles of anygraph G withnprobability
o
3/2
n
m3/2
′
then Q ∈
/ Ω t(G)n1/3 + min m, m
greater than 2/3. If Q ∈
/ Ω t(G)1/3 + min m, t(G)
t(G)
which is a contradiction to Theorem 14 and Theorem 15.
References
[ADNK14]
N. K. Ahmed, N. Duffield, J. Neville, and R. Kompella. Graph sample and hold: A
framework for big graph analytics. In Conference on Knowledge Discovery and Data
Mining (KDD), 2014. 2, 6
[AGM12]
K. J. Ahn, S. Guha, and A. McGregor. Graph sketches: sparsification, spanners, and
subgraphs. In Principles of Database Systems, pages 5–14, 2012. 2, 6
[AKM13]
Shaikh Arifuzzaman, Maleq Khan, and Madhav Marathe. Patric: A parallel algorithm for counting triangles in massive networks. In Proceedings of the 22nd ACM
international conference on Conference on information & knowledge management,
pages 529–538. ACM, 2013. 2, 6
[Avr10]
Haim Avron. Counting triangles in large graphs using randomized matrix trace estimation. In Workshop on Large-scale Data Mining: Theory and Applications, volume 1, 2010. 2, 6
[AYZ97]
Noga Alon, Raphael Yuster, and Uri Zwick. Finding and counting given length cycles.
Algorithmica, 17(3):209–223, 1997. 2, 5
[BBCG08]
Luca Becchetti, Paolo Boldi, Carlos Castillo, and Aristides Gionis. Efficient semistreaming algorithms for local triangle counting in massive graphs. In Proceedings of
the 14th ACM SIGKDD international conference on Knowledge discovery and data
mining, pages 16–24. ACM, 2008. 2, 6
39
[BFL+ 06]
Luciana S Buriol, Gereon Frahling, Stefano Leonardi, Alberto Marchetti-Spaccamela,
and Christian Sohler. Counting triangles in data streams. In Proceedings of the
twenty-fifth ACM SIGMOD-SIGACT-SIGART symposium on Principles of database
systems, pages 253–262. ACM, 2006. 2, 6
[BHLP11]
J. W. Berry, B. Hendrickson, R. A. LaViolette, and C. A. Phillips. Tolerating the
Community Detection Resolution Limit with Edge Weighting. Physical Review E,
83(5), May 2011. 2
[BPWZ14]
A. Björklund, R. Pagh, V. Vassilevska Williams, and U. Zwick. Listing triangles. In
Proceedings of International Colloquium on Automata, Languages, and Programming
(ICALP), pages 223–234, 2014. 2, 5
[Bur04]
R. S. Burt. Structural holes and good ideas.
110(2):349–399, 2004. 2
[BYKS02]
Ziv Bar-Yossef, Ravi Kumar, and D Sivakumar. Reductions in streaming algorithms,
with an application to counting triangles in graphs. In Proceedings of the Thirteenth
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 623–632. Society for Industrial and Applied Mathematics, 2002. 2, 6
[CC11]
Shumo Chu and James Cheng. Triangle listing in massive networks and its applications. In Proceedings of the 17th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 672–680. ACM, 2011. 2, 6
[CEF+ 05]
Artur Czumaj, Funda Ergün, Lance Fortnow, Avner Magen, Ilan Newman, Ronitt
Rubinfeld, and Christian Sohler. Approximating the weight of the euclidean minimum
spanning tree in sublinear time. SIAM Journal on Computing, 35(1):91–109, 2005. 5
[CN85]
Norishige Chiba and Takao Nishizeki. Arboricity and subgraph listing algorithms.
SIAM Journal on Computing, 14(1):210–223, 1985. 2, 5, 11
[Coh09]
J. Cohen. Graph twiddling in a MapReduce world. Computing in Science & Engineering, 11:29–41, 2009. 6
[Col88]
J. S. Coleman. Social capital in the creation of human capital. American Journal of
Sociology, 94:S95–S120, 1988. 2
[CRT05]
Bernard Chazelle, Ronitt Rubinfeld, and Luca Trevisan. Approximating the minimum
spanning tree weight in sublinear time. SIAM Journal on Computing, 34(6):1370–
1379, 2005. 5
[CS09]
Artur Czumaj and Christian Sohler. Estimating the weight of metric minimum spanning trees in sublinear time. SIAM Journal on Computing, 39(3):904–922, 2009. 5
[EM02]
J.-P. Eckmann and E. Moses. Curvature of co-links uncovers hidden thematic layers
in the World Wide Web. Proceedings of the National Academy of Sciences (PNAS),
99(9):5825–5829, 2002. 2
[Fei06]
Uriel Feige. On sums of independent random variables with unbounded variance and
estimating the average degree in a graph. SIAM Journal on Computing, 35(4):964–
984, 2006. 3, 5, 7, 11, 15, 16
40
American Journal of Sociology,
[FWVDC10] Brooke Foucault Welles, Anne Van Devender, and Noshir Contractor. Is a friend
a friend?: Investigating the structure of friendship networks in virtual worlds. In
CHI’10 Extended Abstracts on Human Factors in Computing Systems, pages 4027–
4032. ACM, 2010. 2
[GR02]
Oded Goldreich and Dana Ron. Property testing in bounded degree graphs. Algorithmica, pages 302–343, 2002. 19
[GR08]
Oded Goldreich and Dana Ron. Approximating average parameters of graphs. Random Structures and Algorithms, 32(4):473–493, 2008. 3, 5
[GRS11]
Mira Gonen, Dana Ron, and Yuval Shavitt. Counting stars and other small subgraphs
in sublinear-time. SIAM Journal on Discrete Math, 25(3):1365–1411, 2011. 2, 5
[HKNO09]
Avinatan Hassidim, Jonathan A Kelner, Huy N Nguyen, and Krzysztof Onak. Local
graph partitions for approximation and testing. In Proceedings of the Fiftieth Annual
Symposium on Foundations of Computer Science (FOCS), pages 22–31. IEEE, 2009.
5
[HL70]
P. W. Holland and S. Leinhardt. A method for detecting structure in sociometric
data. American Journal of Sociology, 76:492–513, 1970. 2
[IR78]
Alon Itai and Michael Rodeh. Finding a minimum circuit in a graph. SIAM Journal
on Computing, 7(4):413–423, 1978. 2, 5
[JG05]
Hossein Jowhari and Mohammad Ghodsi. New streaming algorithms for counting
triangles in graphs. In Computing and Combinatorics, pages 710–716. Springer, 2005.
2, 6
[JSP13]
M. Jha, C. Seshadhri, and A. Pinar. A space efficient streaming algorithm for triangle
counting using the birthday paradox. In Proceedings of the 19th ACM SIGKDD
international conference on Knowledge discovery and data mining, KDD ’13, pages
589–597, New York, NY, USA, 2013. ACM. 2, 6
[KMPT12]
Mihail N Kolountzakis, Gary L Miller, Richard Peng, and Charalampos E
Tsourakakis. Efficient triangle counting in large graphs via degree-based vertex partitioning. Internet Mathematics, 8(1-2):161–185, 2012. 2, 6
[KMSS12]
D. M. Kane, K. Mehlhorn, T. Sauerwald, and H. Sun. Counting arbitrary subgraphs
in data streams. In International Colloquium on Automata, Languages, and Programming (ICALP), pages 598–609, 2012. 2, 6
[KPP+ 13]
Tamara G. Kolda, Ali Pinar, Todd Plantenga, C. Seshadhri, and Christine Task.
Counting triangles in massive graphs with MapReduce. arXiv:1301.5887, January
2013. 6
[MR09]
Sharon Marko and Dana Ron. Approximating the distance to properties in boundeddegree and general sparse graphs. ACM Transactions on Algorithms, 5(2), 2009.
5
[MSOI+ 02]
Ron Milo, Shai Shen-Orr, Shalev Itzkovitz, Nadav Kashtan, Dmitri Chklovskii, and
Uri Alon. Network motifs: simple building blocks of complex networks. Science,
298(5594):824–827, 2002. 2
41
[NO08]
Huy N Nguyen and Krzysztof Onak. Constant-time approximation algorithms via
local improvements. In Proceedings of the Forty-Ninth Annual Symposium on Foundations of Computer Science (FOCS), pages 327–336. IEEE, 2008. 5
[ORRR12]
Krzysztof Onak, Dana Ron, Michal Rosen, and Ronitt Rubinfeld. A near-optimal
sublinear-time algorithm for approximating the minimum vertex cover size. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 1123–1131. SIAM, 2012. 5
[Por00]
Alejandro Portes. Social capital: Its origins and applications in modern sociology.
LESSER, Eric L. Knowledge and Social Capital. Boston: Butterworth-Heinemann,
pages 43–67, 2000. 2
[PR07]
Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and a connection to distributed algorithms. Theoretical Computer Science,
381(1-3):183–196, 2007. 5
[PT12]
R. Pagh and C. Tsourakakis. Colorful triangle counting and a mapreduce implementation. Information Processing Letters, 112:277–281, 2012. 6
[PTTW13]
A. Pavan, K. Tangwongsan, S. Tirthapura, and K.-L. Wu. Counting and sampling
triangles from a graph stream. In International Conference on Very Large Databases
(VLDB), 2013. 2, 6
[SKP12]
C. Seshadhri, T. G. Kolda, and A. Pinar. Community structure and scale-free collections of Erdös-Rényi graphs. Physical Review E, 85(5):056109, May 2012. 2
[SPK13]
C. Seshadhri, A. Pinar, and T. G. Kolda. Fast triangle counting through wedge
sampling. In Proceedings of the SIAM Conference on Data Mining, 2013. 2, 6
[SV11]
S. Suri and S. Vassilvitskii. Counting triangles and the curse of the last reducer. In
World Wide Web (WWW), pages 607–614, 2011. 2, 6
[SW05a]
T. Schank and D. Wagner. Approximating clustering coefficient and transitivity.
Journal of Graph Algorithms and Applications, 9:265–275, 2005. 2
[SW05b]
T. Schank and D. Wagner. Finding, counting and listing all triangles in large graphs,
an experimental study. In Experimental and Efficient Algorithms, pages 606–609.
Springer Berlin / Heidelberg, 2005. 2, 6
[TDM+ 09]
Charalampos E Tsourakakis, Petros Drineas, Eirinaios Michelakis, Ioannis Koutis,
and Christos Faloutsos. Spectral counting of triangles in power-law networks
via element-wise sparsification. In Social Network Analysis and Mining, 2009.
ASONAM’09. International Conference on Advances in, pages 66–71. IEEE, 2009.
6
[TKM11]
Charalampos E Tsourakakis, Mihail N Kolountzakis, and Gary L Miller. Triangle
sparsifiers. J. Graph Algorithms Appl., 15(6):703–726, 2011. 2, 6
[TKMF09]
Charalampos E Tsourakakis, U Kang, Gary L Miller, and Christos Faloutsos. Doulion:
counting triangles in massive graphs with a coin. In Proceedings of the 15th ACM
SIGKDD international conference on Knowledge discovery and data mining, pages
837–846. ACM, 2009. 2
42
[TPT13]
K. Tangwongsan, A. Pavan, and S. Tirthapura. Parallel triangle counting in massive
streaming graphs. In ACM Conference on Information & Knowledge Management
(CIKM), 2013. 2, 6
[Tso08]
Charalampos E Tsourakakis. Fast counting of triangles in large real networks without counting: Algorithms and laws. In Data Mining, 2008. ICDM’08. Eighth IEEE
International Conference on, pages 608–617. IEEE, 2008. 2, 6
[YYI09]
Yuichi Yoshida, Masaki Yamamoto, and Hiro Ito. An improved constant-time approximation algorithm for maximum. In Proceedings of the Fourty-First Annual ACM
Symposium on the Theory of Computing, pages 225–234. ACM, 2009. 5
43
| 8cs.DS
|
MULTIPLICATIVE STRUCTURES OF HYPERCYCLIC
FUNCTIONS FOR CONVOLUTION OPERATORS
LUIS BERNAL-GONZÁLEZ, J. ALBERTO CONEJERO, GEORGE COSTAKIS,
AND JUAN B. SEOANE-SEPÚLVEDA
arXiv:1710.10413v1 [math.FA] 28 Oct 2017
Dedicated to Professor Domingo Garcı́a on the occasion of his 60th birthday
Abstract. In this note, it is proved the existence of an infinitely generated
multiplicative group consisting of entire functions that are, except for the constant function 1, hypercyclic with respect to the convolution operator associated
to a given entire function of subexponential type. A certain stability under multiplication is also shown for compositional hypercyclicity on complex domains.
1. Introduction
Assume that X is a (Hausdorff) topological vector space over the real line R or
the complex plane C, and consider the vector space L(X) of all operators on X,
that is, the family of all continuous linear self-mappings T : X → X. An operator
T ∈ L(X) is said to be hypercyclic provided that it admits a dense orbit, that is,
provided that there is some vector x0 ∈ X (called hypercyclic for T ) such that
the orbit {T n x0 : n ∈ N} of x0 under T is dense in X (N := {1, 2, . . . }). The
set of hypercyclic vectors for T will be denoted by HC(T ). In this paper, we are
concerned with the size of HC(T ), mainly from an algebraic point of view and for
certain differential operators. For background on hypercyclic operators we refer
the reader to the excellent books [6, 20]. An account of concepts and results about
algebraic structures inside nonlinear sets can be found in [2, 5, 11, 15, 17, 29, 30].
It is well known that if X is an F-space (i.e., complete and metrizable) and T is
a hypercyclic operator then the set HC(T ) is residual, that is, it contains a dense
Gδ subset of the (Baire) space X; we can say that HC(T ) is topologically large.
Furthermore, for any topological vector space and any hypercyclic T ∈ L(X), the
family HC(T ) is algebraically large; specifically, it contains, except for 0, a dense
(even T -invariant) vector subspace of X (see [33]). Starting from [25], a number of
criteria have been established for an operator T ∈ L(X) to support a closed infinite
dimensional vector subspace all of whose nonzero vectors are T -hypercyclic on an
2010 Mathematics Subject Classification. 22A05, 30E10, 30K20, 46E10, 47A16.
Key words and phrases. hypercyclic operator, convolution operator, composition operator,
subexponential growth, group of non-vanishing entire functions, lineability, spaceability.
1
2
BERNAL, CONEJERO, COSTAKIS, AND SEOANE
F-space X; however, not all hypercyclic operators support such a subspace (see
[6, Chapter 8] and [20, Chapter 10]).
By a domain in C we mean a nonempty connected open subset G ⊂ C. It
is well known that the space H(G) of holomorphic functions G → C becomes
an F-space when endowed with the topology of uniform convergence on compact
subsets of G. We are mainly interested in operators defined on the space H(C)
of entire functions C → C. An operator T ∈ L(H(C)) is said to be a convolution
operator provided that it commutes with translations, that is,
T ◦ τa = τa ◦ T for all a ∈ C,
where τa f := f (· + a). Then T ∈ L(H(C)) happens to be a convolution operator
if and only if T is an infinite order linear differential operator with constant coefficients T = Φ(D), where Φ is an entire function with exponential type, that is,
there exist constants A,P
B ∈ (0, +∞) such that |Φ(z)| ≤ AeB|z| for all z ∈ C. For
such a function Φ(z) = n≥0 an z n and f ∈ H(C), we have
Φ(D)f =
∞
X
an f (n) .
n=0
Godefroy and Shapiro [18] proved that every non-scalar convolution operator is
hypercyclic, so covering the classical Birkhoff and MacLane results on hypercyclicity of the translation operator τa (take Φ(z) = eaz , a 6= 0) and of the derivative
operator D : f 7→ f ′ (take Φ(z) = z), respectively. It has been proved that,
for every non-scalar convolution operator T , the set HC(T ) contains, except for
0, a closed infinite dimensional vector subspace of H(C); see [24, 28, 32] and also
[2, Section 4.5] and [20, Section 10.1].
In view of the preceding paragraph, we can say that hypercyclic vectors of
convolution operators are rather stable under summation and scaling. However,
stability under multiplication does not seem to be so clear. For instance, for every
a 6= 0, every f ∈ HC(τa ) and every n ∈ N with n ≥ 2 we have that f n 6∈ HC(τa )
(see [3, Cor. 2.4]). As a positive result, it was proved in [3, Th. 2.3] the existence of
a function f ∈ H(C) –in fact, of a residual subset of them– such that f n ∈ HC(D)
for all n ∈ N, see also [4]. This result was extended by Bayart and Matheron
who proved that there is even a residual set of functions in H(C) generating a
hypercyclic algebra for the derivative operator, that is every non-null function in
one of these algebras is hypercyclic for the operator D [6, Th. 8.26]. Recall that
the algebra generated by a function f is nothing but the set {P ◦f : P polynomial,
P (0) = 0}. The existence of (one-generated) algebras of hypercyclic functions for
D was also independently proved by Shkarin in [32] and extended by Bès et al. in
[12] to operators of the form P (D), where P is a nonconstant polynomial with
P (0) = 0. The existence of hypercyclic algebras for many convolution operators
HYPERCYCLIC FUNCTIONS FOR CONVOLUTION OPERATOS
3
not induced by polynomials, such as cos(D), DeD , or eD − aI, where 0 < a ≤ 1
was recently shown in [13].
So far, the existence of multiply generated algebras of hypercyclic entire functions is unknown. The lack of knowledge about existence of families of hypercyclic
functions that are stable under multiplication seems to be one of the main obstacles. Our goal in this paper is to contribute to fill in this gap. Specifically, given
a nonconstant entire function Φ of subexponential type, we will construct in Section 3 an infinitely generated multiplicative group of entire functions all of whose
members, except for the constant function 1(z) = 1, are hypercyclic with respect
to the operator Φ(D). A certain multiplicative stability is also shown for hypercyclicity with respect to composition operators. Section 2 is devoted to provide the
necessary background and auxiliary results.
2. Non-vanishing hypercyclic entire functions
Recall that the exponential type of a Φ ∈ H(C) is defined as, see [14],
τ (Φ) = lim sup
r→∞
log max{|Φ(z)| : |z| = r}
.
r
Hence, f has subexponential type if and only if, given ε > 0, there exists A =
A(ε) ∈ (0, +∞) such that |f (z)| ≤ A eε|z| for all z ∈ C. In fact, if Φ has subexponential type then Φ(D) is a well defined operator on H(G) for any domain
G ⊂ C (see for instance [7] or [9, Theorem 4]). It should be underlined that if Φ
is of exponential type, then the operator Φ(D) does not act on proper domains
in general: take, for instance, G = D and Φ(z) = ez .
As we have mentioned in Section 1, there is no entire function f such that
f is hypercyclic for the translation operator τa = eaD (a 6= 0). If we observe
that Φ(z) := eaz has exponential type τ (Φ) = |a| > 0, then we may wonder
whether there are entire functions f such that f and f 2 are in HC(Φ(D)) if Φ has
subexponential type, i.e., if τ (Φ) = 0. This property happens to be true, in a strong
sense, in the special case of Φ = P , a nonconstant polynomial with P (0) = 0 [12].
The rate of growth of a hypercyclic function for D has been optimally estimated
in [19, 31].
2
It is well known (see, for instance, [1]) that a function f ∈ H(C) is never
zero if and only if there is g ∈ H(C) such that f = eg . Consistently, we can
denote the family of non-vanishing entire functions by eH(C) . This family forms,
trivially, a group under the pointwise multiplication. Moreover, eH(C) is –as it
is easily proven– a topological group (for background on topological groups, see
for instance [23]) under the topology of uniform convergence in compacta, that is,
both mappings
1
(f, g) ∈ eH(C) × eH(C) 7→ f g ∈ eH(C) and f ∈ eH(C) 7→ ∈ eH(C)
f
4
BERNAL, CONEJERO, COSTAKIS, AND SEOANE
are continuous, from which one can derive that each mapping f ∈ eH(C) 7→ f m ∈
eH(C) (m ∈ Z = the set of integers) is a continuous homomorphism (recall that a
mapping Φ : G1 → G2 between two groups G1 , G2 is an homomorphism whenever
Φ(ab) = Φ(a)Φ(b) for all a, b ∈ G1 ). Note that this mapping is not bijective
H(C)
(except for m = ±1), but if we consider the topological subgroup e+,0 := {f ∈
eH(C) : f (0) > 0} then each restriction
H(C)
H(C)
Φm : f ∈ e+,0 7→ f m ∈ e+,0
(m ∈ Z \ {0})
is not only continuous (and homomorphic) but also bijective, because for any
H(C)
g ∈ e+,0 there are exactly m entire functions f with f m = g (necessarily, such
f ’s are non-vanishing), but only one of them satisfies f (0) > 0. In fact, in the
H(C)
following two lemmas it is shown that the group e+,0 also enjoys a nice topological
structure and that every Φm is an automorphism of it.
H(C)
Lemma 2.1. The set e+,0 is a Gδ -subset of H(C). In particular, it is a separable
completely metrizable space, as well as a Baire space.
Proof. The second part is a consequence of the first one. Indeed, Alexandroff’s
theorem (see e.g. [27, pp. 47–48]) asserts that a Gδ -set is homeomorphic to a
complete metrizable space; separability is inherited by any subspace of a separable
metrizable space; finally, any completely metrizable space is a Baire space (see
T
H(C)
e.g. [26]). As for the first part, simply put e+,0 = ∞
k=1 Ak , where Ak is defined
as
1
Ak := f ∈ H(C) : inf |f (z)| > 0, |Imf (0)| < and Ref (0) > 0 .
|z|≤k
k
That each Ak is open follows from the facts that {z ∈ C : |z| ≤ k} is compact
and that both evaluation mappings f ∈ H(C) 7→ Ref (0) ∈ R, f ∈ H(C) 7→
Imf (0) ∈ R are continuous.
Lemma 2.2. For each m ∈ Z\ {0}, the mapping Φm is an onto homeomorphism.
Proof. The unique property to be proved is that every Φm has inverse continuous.
Since Φ−1 is an onto homeomorphism, it is enough to see that, given m ∈ N
H(C)
H(C)
with m ≥ 2, the mapping Ψ = (Φm )−1 : e+,0 → e+,0 is continuous. By
H(C)
Lemma 2.1, e+,0 is a completely metrizable topological group. Note that Ψ is a
H(C)
homomorphism from the group e+,0 into itself. According to the abstract closed
graph theorem (see, e.g., [16, Theorem 5.2]), it is enough to show that Ψ has closed
H(C)
graph. To this end, assume that (fk ) ⊂ e+,0 is a sequence such that there are
H(C)
f, g ∈ e+,0 with (fk , Ψ(fk )) → (f, g) as k → ∞ in the topology product. Then
fk → f and Ψ(fk ) → g. The continuity of Φm yields fk = Φm Ψ(fk ) → Φm (g),
and the uniqueness of the limit implies f = Φm (g) or, that is the same, g = Ψ(f ),
which proves that the graph of Ψ is closed, as required.
HYPERCYCLIC FUNCTIONS FOR CONVOLUTION OPERATOS
5
Finally, we will use in the next section the following theorem, that is contained
in [8, Theorem 5]. This theorem is, in turn, an extension of a result about inherited
D-hypercyclicity due to Herzog [21].
Theorem 2.3. If Φ is a nonconstant entire function of subexponential type, then
the set HC(Φ(D)) ∩ eH(C) is residual in eH(C) .
Additional statements on zero-free D-hypercyclic entire functions can be found
in [10].
3. Groups of hypercyclic functions
Firstly, we show that the family of entire functions f that are hypercyclic
with respect to certain convolution operators (including D) satisfying f (0) > 0 is
topologically large.
Proposition 3.1. If Φ is a nonconstant entire function of subexponential type,
H(C)
H(C)
then the set HC(Φ(D)) ∩ e+,0 is residual in e+,0 .
Proof. According to Birkhoff transitivity theorem (see e.g. [20]), if Tn : X → Y
(n ∈ N) is a sequence of continuous mappings between two Hausdorff topological
spaces, with X Baire and first-countable and Y second-countable, then the set
U = U((Tn )) of points x ∈ X whose orbit {Tn x : n ∈ N} is dense in Y is a Gδ
subset of X; consequently, U is residual as soon as it is dense. By Theorem 2.3,
the set HC(Φ(D)) ∩ eH(C) is residual in eH(C) . It is easy to see that the mapping
|f (0)|
H(C)
· f ∈ e+,0
f (0)
is continuous and onto. Then it takes dense sets into dense sets. In particular, the
H(C)
image Π(HC(Φ(D)) ∩ eH(C) ) is dense in e+,0 . Since any nonzero scalar multiple
of a hypercyclic vector is also hypercyclic, we obtain Π(HC(Φ(D)) ∩ eH(C) ) ⊂
H(C)
H(C)
HC(Φ(D)) ∩ e+,0 . But U := HC(Φ(D)) ∩ e+,0 equals the set of elements
H(C)
f ∈ X := e+,0 whose orbit under the sequence (Tn ) := ((Φ(D))n |eH(C) ) is dense
Π : f ∈ eH(C) 7→
+,0
in Y := H(C) (note that X is Baire and first-countable by Lemma 2.1). Therefore
H(C)
U is a Gδ dense subset of e+,0 , hence residual.
Our main result (Theorem 3.3) will be deduced as a consequence of the following,
more abstract, assertion.
Theorem 3.2. Let X be a separable infinite dimensional F-space. Assume that
X is also a unitary commutative linear algebra, and that the corresponding multiplication law (f, g) ∈ X × X 7→ f ∗ g ∈ X is continuous. Suppose also that Z is a
subset of X fulfilling the following conditions:
(i) Z is a Gδ -subset of X.
(ii) (Z, ∗) is a topological group.
6
BERNAL, CONEJERO, COSTAKIS, AND SEOANE
(iii) The mappings u ∈ Z 7→ uk ∈ Z (k ∈ N) are onto homeomorphisms.
(iv) No nonempty relatively open subset of Z is contained in a countable dimensional vector subspace of X.
Then for each residual subset R of Z there exists a subset G ⊂ Z satisfying the
following properties:
(a) G is a subgroup of Z.
(b) G is dense in Z.
(c) R ⊃ G \ {e}, where e denotes the unit element of X.
(d) G is infinitely generated in a strong sense, namely, the algebra generated
by G is infinitely generated.
Proof. By Alexandroff’s theorem and by the fact that X is a separable F-space,
we have (thanks to (i)) that Z is a completely metrizable space that, in addition,
is second-countable. Then Z is a Baire space and there is a countable open basis
{Gn }n≥1 for the topology of Z. From (ii) and (iii) one takes out that every
mapping Ψk : u ∈ Z 7→ uk ∈ Z (k ∈ Z \ {0}) is an onto homeomorphism.
Moreover, it follows from (ii) that, for every v ∈ Z, the multiplication mapping
Mv : u ∈ Z 7→ v ∗ u ∈ Z is also an onto homeomorphism.
Fix a residual subset R of Z. Then, for each k ∈ Z \ {0}, the set Ψ−1
k (R) is
residual in Z. Then the countable intersection
\
R0 :=
Ψ−1
k (R)
k∈Z\{0}
is also residual –and, in particular, dense– in Z. Let us set L0 := span{e}, which
equals the algebra generated by e. Since L0 is a finite dimensional subspace, it is
closed in X. Then G1 \ L0 is a nonempty (otherwise, G1 ⊂ L0 , contradicting (iv))
open set in the topology of Z. Then there exists v1 ∈ G1 ∩ R0 .
Let us proceed by induction. Assume that, for some n ∈ N, the vectors v1 , . . . , vn
have been selected. Let Ln denote the linear algebra generated by v0 , v1 , . . . , vn ,
that is, Ln = span{v1jn ∗ · · · ∗ vnjn : (j1 , . . . , jn ) ∈ Nn0 }, where N0 := N ∪ {0}.
Observe that Ln is a countable dimensional vector subspace of X. Since X is a
Baire space, if follows that Ln is an Fσ set with empty interior. Now, we set
\
\
−1
−1
M k1
Rn :=
Ψk
kn (R) .
k∈Z\{0}
(k1 ,...,kn )∈Zn
v1 ∗···∗vn
Observe that the choice (k1 , . . . , kn ) = (0, . . . , 0) gives Rn ⊂ R0 . Since the Ψk ’s
as well as the M −1
k1
kn ’s are onto homeomorphisms and countable intersections
v1 ∗···∗vn
of residual sets are residual, we get that each Rn is residual in Z. Assume, by
way of contradiction, that Gn+1 \ Ln is of first category (in the sense of Baire) in
Z. Then Gn+1 = (Gn+1 \ Ln ) ∪ (Gn+1 ∩ Ln ) and, by the assumption (iv), Ln ∩ Z
is a (relatively Fσ ) set with empty interior in Z, so of first category in Z. This
HYPERCYCLIC FUNCTIONS FOR CONVOLUTION OPERATOS
7
implies that its subset Gn+1 ∩Ln is also of first category in Z, hence Gn+1 is, which
is absurd because Z is a Baire space. Thus, Gn+1 \ Ln is of second category (in
the sense of Baire) in Z. As Rn is residual in Z, we have Rn ∩ (Gn+1 \ Ln ) 6= ∅.
Therefore, we can select vn+1 ∈ Rn ∩ (Gn+1 \ Ln ) and the induction procedure is
finished.
Define G as the group generated by {vn }n∈N , that is
G = v1k1 ∗ · · · ∗ vnkn : (k1 , . . . , kn ) ∈ Zn , n ∈ N .
Since each vn belongs to Z and vn ∈ Gn (n ∈ N), we obtain conclusions (a) and
(b). Property (c) is derived from the fact that, by construction, each combination
v1k1 ∗ · · · ∗ vnkn ∈ R as soon as (k1 , . . . , kn ) ∈ Zn \ {(0, . . . , 0)}. Finally, S
property
(d) is true because the algebra generated by G contains the algebra A := n≥0 Ln ,
and A is infinitely generated. Indeed, assume that A is generated by some finite
subset F ⊂ A. Since Ln ⊂ Ln+1 (n ≥ 0), there is N ∈ N such that F ⊂ LN .
Then LN generates A. But LN is itself an algebra, so A = LN , which is absurd
because vN +1 ∈ A \ LN . Hence A cannot be finitely generated, which concludes
the proof.
Theorem 3.3. Assume that Φ is a nonconstant entire function of subexponential type. Consider the convolution operator T = Φ(D) associated to Φ. Then
there exists an infinitely generated multiplicative group G ⊂ H(C) such that every
member of G –except for 1– is Φ(D)-hypercyclic. Moreover, G is a dense subgroup
H(C)
of e+,0 and the algebra generated by G is infinitely generated.
Proof. It suffices to apply Theorem 3.2 with the following “characters”: Take X :=
H(C)
H(C), Z := e+,0 , ∗ := the usual multiplication, e := 1 and R := HC(Φ(D)) ∩
H(C)
e+,0 . Now, Z is a topological group for which the mappings f 7→ f k (k ∈ N) are
onto homeomorphisms by Lemma 2.2, Z is a Gδ subset of X by Lemma 2.1, and
R is residual in Z by Proposition 3.1.
H(C)
Now, assume that O is a nonempty open subset in e+,0 and that, by way of
contradiction, Y is a countable dimensional space of H(C) with O ⊂ Y . Then
span{hn }n≥1 =
SY for certain hn ∈ H(C) (n ≥ 1). Setting Yn := span{h1 , . . . , hn }
we get Y = n≥1 Yn . But each Yn is a finite dimensional subspace, hence σcompact, and so Y is. In other words, there are countably many compact sets
S
H(C)
H(C)
Kn ⊂ e+,0 (n ∈ N) such that Y = n≥1 Kn . Therefore each set Zn := e+,0 ∩
S
H(C)
H(C)
Kn is compact (hence closed) in e+,0 and O ⊂ n≥1 Zn . Since e+,0 is a
completely metrizable space, Baire’s category theorem tells us that at least one
H(C)
H(C)
Zm has nonempty interior in e+,0 . Now, the group structure of e+,0 entails that
H(C)
the function 1 possesses a compact neighborhood W in e+,0 .
Now, consider the vector space S := {g ∈ H(C) : Im g(0) = 0} endowed
with the topology inherited from H(C). Trivially, S is a topological group for the
8
BERNAL, CONEJERO, COSTAKIS, AND SEOANE
H(C)
operation “+” and the same topology. For every f ∈ e+,0 there is a unique g ∈ S
H(C)
such that f = eg . Hence the mapping T : f = eg ∈ (e+,0 , ·) 7→ g ∈ (S, +) is an
algebraic group isomorphism. But it is in fact a topological isomorphism. Indeed,
T −1 : g 7→ eg is trivially continuous (any superposition mapping g ∈ H(C) 7→
ϕ ◦ g ∈ H(C), with ϕ ∈ H(C), is continuous), and T is continuous because it is
continuous at the neutral element 1. This means that if a sequence (fn = egn ) ⊂
H(C)
e+,0 (with gn = un + ivn ∈ S, so that vn (0) = 0 for all n) satisfies fn → 1 then
gn = un + ivn = T (fn ) → T (1) = 0 uniformly on compacta. This, in turn, follows
by invoking the continuity of the principal branch logp w := ln |w| + i argp w of
the logarithm on C \ (−∞, 0]. Indeed, since 1 ∈ C \ (−∞, 0], for given R > 0
we have that fn (K) ⊂ C \ (−∞, 0] eventually, where K = {z : |z| ≤ R}. Then
logp (fn ) → logp 1 = 0 uniformly on K. But logp (fn (z)) = un (z) + i argp (eivn (z) ) =
un (z) + i(vn (z) + 2kn π) for certain kn ∈ Z not depending on z. Hence un → 0 and
vn + 2kn π → 0 uniformly on K. In particular, 2kn π = vn (0) + 2kn π → 0, so kn → 0
and, since kn ∈ Z, we get kn = 0 eventually. This implies vn → 0 uniformly on
K, and so does gn = un + ivn as n → ∞, as required. We have proved that T
is a homeomorphism. Thus T (W ) is a compact neighborhood of 0 in S. But
Riesz’s theorem (see, e.g., [22, page 147, Theorem 3]) implies that dim (S) < ∞,
which is absurd because S contains all functions z n (n ∈ N). This is the desired
contradiction.
Hence O is not contained in any countable dimensional subspace of H(C). Consequently, conditions (i)–(iv) in Theorem 3.2 are fulfilled and the conclusion follows.
Another important class of operators is the one formed by the composition
operators. If G ⊂ C is a domain and ϕ : G → G is a holomorphic self-map of G,
the mapping
Cϕ : f ∈ (G) 7→ f ◦ ϕ ∈ H(G)
is well defined, linear and continuous, and it is called the composition operator
with symbol ϕ. By Aut(G) we denote the group of automorphisms of G, i.e., the
family of all bijective holomorphic functions G → G. The translations τa (a ∈ C)
form a special instance when G = C and ϕ(z) := z + a. We have already seen
that, even for translations, the set of hypercyclic functions is not stable under
products (f 2 is never τa -hypercyclic, whatever f is). Nevertheless, we obtain a
reasonably degree of stability if the products do not allow repetition of factors.
This can be done for many composition operators, as the following theorem –with
which we conclude this paper– shows. Recall that a domain G ⊂ C is called simply
connected if it lacks holes, that is, if C∞ \ G is connected.
Theorem 3.4. Let G ⊂ C be a simply connected domain. Assume that ϕ ∈
Aut(G) and ϕ has no fixed points. Then for every f ∈ HC(Cϕ ) there is a family
F ⊂ H(G) satisfying the following properties:
HYPERCYCLIC FUNCTIONS FOR CONVOLUTION OPERATOS
(a)
(b)
(c)
(d)
9
f ∈ F ⊂ HC(Cϕ ).
F is dense in H(G).
Q
For each finite nonempty subfamily F0 ⊂ F , one has g∈F0 g ∈ HC(Cϕ ).
F is a linearly free system.
Proof. Since ϕ has no fixed points, the operator Cϕ is mixing (see [20, Theorem
4.37]). In the context of F-spaces (as H(G) is), this is equivalent to the fact
that, for every (strictly increasing) subsequence (nk ) ⊂ N, the set U((Cϕnk )) of
functions having dense (Cϕnk )-orbit is dense (even residual) in H(G). Of course,
mixing property implies hypercyclicity. Observe that (Cϕ )m = Cϕm for m ∈ N,
where ϕ1 := ϕ and ϕm+1 := ϕ ◦ ϕm .
Fix f ∈ HC(Cϕ ) as well as a countable topological basis (Gn ) for the topology
of H(G) (recall that H(G) is metrizable and separable, hence second countable).
We set f0 := f and G0 := H(G). Then there is a sequence (p(1, k)) ⊂ N such that
f0 ◦ ϕp(1,k) → 1 (k → ∞) uniformly on each compact subset of G. Since the set
p(1,k)
p(1,k)
U((Cϕ )) is dense in H(G), there exists f1 ∈ G1 ∩ U((Cϕ )). In particular, for
the constant function z 7→ 1, there is a subsequence (p(2, k)) ⊂ (p(1, k)) satisfying
f1 ◦ ϕp(2,k) → 1 (k → ∞) in H(G). Of course, f0 ◦ ϕp(2,k) → 1 in H(G).
By following this procedure, we can obtain recursively a countable family F :=
{fn }n≥0 ⊂ H(G) as a well as countable family {(p(n, k))k≥1 }n≥0 of strictly increasing sequences of natural numbers, satisfying, for each n ≥ 0, the following:
p(n,k)
• fn ∈ Gn ∩ U((Cϕ
)),
• (p(n + 1, k)) ⊂ (p(n, k)), and
• fj ◦ ϕp(n+1,k) → 1 (k → ∞) in H(G) for all j = 0, . . . , n.
Since (Gn ) is a topological basis and fn ∈ Gn , the set F is dense in H(G).
p(n,k)
By construction, f ∈ F and, since HC(Cϕ ) ⊃ U((Cϕ
)) for every n, we get
F ⊂ HC(Cϕ ). Now, if ∅ 6= F0 ⊂ F with F0 finite, we can write F0 =
{fm(1) , . . . , fm(N ) }, with 0 ≤ m(1) < m(2) < · · · < m(N). Let
h :=
Y
g∈F0
g=
N
Y
fm(j) .
j=1
Trivially, h ∈ HC(Cϕ ) if N = 1. If N > 1 then h = h0 · fm(N ) , where h0 :=
QN −1
p(m(N ),k)
)), there is a subsequence
j=1 fm(j) . Fix F ∈ H(G). As fm(N ) ∈ U((Cϕ
(νk ) ⊂ (p(m(N), k)) such that fm(N ) ◦ ϕνk → F (k → ∞) uniformly on compacta
in G. But fm(j) ◦ ϕνk → 1 (k → ∞) uniformly on compacta in G for each
j ∈ {1, . . . , N − 1}. This entails h0 ◦ ϕνk → 1 in H(G). Thus
h ◦ ϕνk = (h0 ◦ ϕνk ) · (fm(N ) ◦ ϕνk ) −→ 1 · F = F in H(G).
Consequently, h ∈ HC(Cϕ ). So far, (a), (b) and (c) have been proved.
10
BERNAL, CONEJERO, COSTAKIS, AND SEOANE
In order to prove (d), assume, by way of contradiction, that there is some Ntuple (c0 , c1 , . . . , cN ) ∈ CN +1 \ {(0, . . . , 0)} such that c0 f0 + · · · + cN fN = 0, where,
without loss of generality, we can suppose cN 6= 0. Then c0 (f0 ◦ ϕp(N,k) ) + · · · +
cN (fN ◦ ϕp(N,k) ) = 0 for all k ≥ 1. Letting k → ∞ we get c0 · 1 + · · · + cN −1 · 1 +
cN (fN ◦ ϕp(N,k) ) → 0, so
fN ◦ ϕ
p(N,k)
−→ −
N
−1
X
cj c−1
in H(G) as k → ∞,
N
j=0
which is absurd because {fN ◦ ϕp(N,k) }k≥1 is dense in H(G). The theorem is now
totally proved.
Acknowledgments.
The authors are indebted to the referee for helpful comments and suggestions,
specially for nice simplifications of the proofs of some results.
The first author has been supported by the Plan Andaluz de Investigación
de la Junta de Andalucı́a FQM-127 Grant P08-FQM-03543 and by MEC Grant
MTM2015-65242-C2-1-P. The second author has been supported by MEC Grant
MTM2016-75963-P. The second and third authors were also supported by Generalitat Valenciana, Project GV/2010/091, and by Universitat Politècnica de València,
Project PAID-06-09-2932. The fourth author has been supported by Grant MTM201565825-P.
References
[1] L. V. Ahlfors, Complex Analysis, 3rd ed., McGraw-Hill, London, 1979.
[2] R. M. Aron, L. Bernal-González, D. Pellegrino, and J. B. Seoane-Sepúlveda, Lineability:
The search for linearity in Mathematics, Monographs and Research Notes in Mathematics,
Chapman & Hall/CRC, Boca Raton, FL, 2016.
[3] R. M. Aron, J. A. Conejero, A. Peris, and J. B. Seoane-Sepúlveda, Powers of hypercyclic
functions for some classical hypercyclic operators, Integr. Equ. Oper. Theory 58 (2007),
no. 4, 591–596.
[4] R. M. Aron, J. A Conejero, A. Peris, and J. B. Seoane-Sepulveda, Sums and products of bad
functions, Contemporary Mathematics 435 (2007), 47–52.
[5] R. M. Aron, V. I. Gurariy, and J. B. Seoane-Sepúlveda, Lineability and spaceability of sets
of functions on R, Proc. Amer. Math. Soc. 133 (2005), no. 3, 795–803.
[6] F. Bayart and E. Matheron, Dynamics of Linear Operators, Cambridge Tracts in Mathematics, Cambridge University Press, 2009.
[7] C. A. Berenstein and R. Gay, Complex analysis and special topics in harmonic analysis,
Springer, New York, 1995.
[8] L. Bernal-González, On universal functions with zero-free derivatives, Arch. Math. 68
(1997), 145–150.
, Hypercyclic sequences of differential and antidifferential operators, J. Approx. The[9]
ory 96 (1999), 323–337.
[10] L. Bernal-González, A. Bonilla, and G. Costakis, On the growth of zero-free MacLaneuniversal entire functions, Indag. Math. 23 (2012), 311–317.
HYPERCYCLIC FUNCTIONS FOR CONVOLUTION OPERATOS
11
[11] L. Bernal-González, D. Pellegrino, and J. B. Seoane-Sepúlveda, Linear subsets of nonlinear
sets in topological vector spaces, Bull. Amer. Math. Soc. (N.S.) 51 (2014), no. 1, 71–130.
[12] J. Bès, J. A. Conejero, and D. Papathanasiou, Convolution operators supporting hypercyclic
algebras, J. Math. Anal. Appl. 445 (2017), no. 2, 1232–1238.
, Hypercyclic algebras for convolution and composition operators, Preprint (2017).
[13]
[14] R. P. Boas Jr., Entire functions, Academic Press, New York, 1954.
[15] D. Cariello and J. B. Seoane-Sepúlveda, Basic sequences and spaceability in ℓp spaces, J.
Funct. Anal. 266 (2014), no. 6, 3797–3814, DOI 10.1016/j.jfa.2013.12.011.
[16] J. P. R. Christensen., Topology and Borel structure, North Holland, Amsterdam, 1974.
[17] P. H. Enflo, V. I. Gurariy, and J. B. Seoane-Sepúlveda, Some results and open questions on
spaceability in function spaces, Trans. Amer. Math. Soc. 366 (2014), no. 2, 611–625.
[18] G. Godefroy and J. H. Shapiro, Operators with dense, invariant, cyclic vectors manifolds,
J. Funct. Anal. 98 (1991), no. 2, 229–269.
[19] K.-G. Grosse-Erdmann, On the universal functions of G.R. MacLane, Complex Variables
and Elliptic Equations 15 (1990), no. 3, 193–196.
[20] K.-G. Grosse-Erdmann and A. Peris, Linear Chaos, Springer, London, 2011.
[21] G. Herzog, On zero-free universal entire functions, Arch. Math. 63 (1994), 329–332.
[22] J. Horváth, Topological vector spaces and distributions, Vol. I, Addison-Wesley, Reading,
Massachusetts, 1966.
[23] N. G. Markley, Topological Groups: An Introduction, Wiley, NJ, 2010.
[24] Q. Menet, Hypercyclic subspaces and weighted shifts, Adv. Math. 255 (2014), 305–337.
[25] A. Montes-Rodrı́guez, Banach spaces of hypercyclic vectors, Michigan Math. J. 43 (1996),
no. 3, 419–436.
[26] J. R. Munkres, Topology: A First Course, 2nd ed., Prentice-Hall Inc., Upper Saddle River,
NJ, 2000.
[27] J. C. Oxtoby, Measure and Category, 2nd edition, Springer-Verlag, New York, 1980.
[28] H. Petersson, Hypercyclic subspaces for Fréchet spaces operators, J. Math. Anal. Appl. 319
(2006), no. 2, 764–782.
[29] J. B. Seoane-Sepúlveda, Chaos and lineability of pathological phenomena in analysis, ProQuest LLC, Ann Arbor, MI, 2006. Thesis (Ph.D.)–Kent State University.
, Explicit constructions of dense common hypercyclic subspaces, Publ. Res. Inst.
[30]
Math. Sci. 43 (2007), no. 2, 373–384.
[31] S. A. Shkarin, On the growth of D-universal functions, On the growth of D-universal functions 48 (1993), no. 6, 49–51.
[32] S. Shkarin, On the set of hypercyclic vectors for the differentiation operator, Israel J. Math.
180 (2010), no. 1, 271–283.
[33] J. Wengenroth, Hypercyclic operators on nonlocally convex spaces, Proc. Amer. Math. Soc.
131 (2003), no. 6, 1759–1761.
12
BERNAL, CONEJERO, COSTAKIS, AND SEOANE
Departamento de Análisis Matemático,
Facultad de Matemáticas, Universidad de Sevilla,
Avenida Reina Mercedes,
41080 Sevilla, Spain.
E-mail address: [email protected]
Instituto Universitario de Matemática Pura y Aplicada,
Universitat Politècnica de València,
46022 València, Spain.
E-mail address: [email protected]
Department of Mathematics and Applied Mathematics,
University of Crete, Voutes Campus,
70013 Heraklion, Crete, Greece.
E-mail address: [email protected]
IMI and Departamento de Análisis Matemático,
Facultad de Ciencias Matemáticas,
Plaza de Ciencias 3,
Universidad Complutense de Madrid,
28040 Madrid, Spain.
E-mail address: [email protected]
| 4math.GR
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.